admin/mcp-nextcloud
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request #750 from cbcoutinho/feat/verify-on-read-semantic-search

Browse Source 
feat(search): verify-on-read for semantic search (ADR-019)
This commit is contained in:
Chris Coutinho
2026-05-02 01:03:43 +02:00
committed by GitHub
parent 3226117101 104bbd390d
commit 055e786007
22 changed files with 3112 additions and 151 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.claude/skills/pre-push-review/SKILL.md
+380
View File
@@ -0,0 +1,380 @@
---
name: pre-push-review
description: |
Self-review the current branch's diff against nextcloud-mcp-server review patterns
before pushing a PR. Runs ruff/ty/tests on changed files and produces a punch list
of likely review-round findings, calibrated to issues that have repeatedly surfaced
in this repo's automated PR reviews. Use when the user is about to push, says "ready
to push", "review my work", "check before PR", or invokes /pre-push-review.
Report-only — does not modify code.
allowed-tools:
- Bash
- Read
- Grep
- Glob
---
# Pre-Push Review
## Purpose
Catch the issues that have repeatedly surfaced in PR reviews on this repo (#733–#750)
*before* the human reviewer or the automated `claude` PR-review bot sees them. Output is
a labelled punch list. The main loop decides what to fix.
This skill is calibrated to **this repo's recurring patterns** — not a generic code
review. The point is to short-circuit the review-round-N loop, not to replace human
judgment.
## When to Use
**Trigger when:**
- User says "review my work", "ready to push", "check before PR", "pre-push", or
invokes `/pre-push-review`.
- After a substantive change (new module, new tool, refactor, behavior change) and
before `git push`.
- Before opening or updating a PR.
**Skip when:**
- Tiny diffs (typo fix, README tweak, dependency bump only).
- User has explicitly said "just push it" / "skip the review".
- Branch is `master` or has zero commits ahead of base.
## Workflow
### Phase 1 — Establish scope (≤ 30s)
Determine the base branch and diff range. Default base is `master`.
```bash
git fetch origin master --quiet
BASE=$(git merge-base HEAD origin/master)
git rev-list --count $BASE..HEAD # commits ahead
git diff --stat $BASE..HEAD # files touched
git log --format="%h %s" $BASE..HEAD # commit list
```
If the user names a different base (e.g. `main`, a stacked branch), use that instead.
**Identify the change shape** — these classifications drive scope-aware checks (Phase 3):
| Touches | Treat as |
|---|---|
| `nextcloud_mcp_server/auth/`, `*scope*`, `*token*`, `*verifier*` | **security-sensitive** (escalate B, J, K to blocking) |
| `nextcloud_mcp_server/server/*.py` (new `@mcp.tool`) | **MCP surface** (D4, I1 mandatory) |
| `nextcloud_mcp_server/client/*.py` | **client layer** (B1–B4, H1, retry/backoff) |
| `nextcloud_mcp_server/models/*.py` | **schema** (D1–D3, validators for PHP quirks) |
| `nextcloud_mcp_server/search/`, `nextcloud_mcp_server/vector/` | **vector subsystem** (E*, F3, latency budget) |
| Webhooks, `nextcloud_mcp_server/auth/webhook_*` | **webhook handler** (J3, J4, K1) |
| Tests only | **test scope** (skip A–E, run F1–F4) |
| Docs / ADR only | **docs scope** (run G3, G4 only) |
**Report up front:** branch, base, commits ahead, files changed, change shape.
### Phase 2 — Automated checks (in parallel, ≤ 30s)
Run these from the project root with separate `Bash` calls in **one message** so they
run concurrently:
```bash
uv run ruff check # lint
uv run ruff format --check # formatting
uv run ty check -- nextcloud_mcp_server # types
uv run pytest tests/unit/ -x --no-header -q # unit tests
```
If the diff touches `nextcloud_mcp_server/server/` or `tests/server/`, also run:
```bash
uv run pytest -m smoke -x --no-header -q
```
**If any automated check fails: stop and report.** Don't run the checklist on top of a
broken build — fixing the failures may eliminate findings or change the diff.
### Phase 3 — Read the diff and run the project checklist (2–5min)
```bash
git diff $BASE..HEAD # full diff
git diff $BASE..HEAD -- '*.py' | head -2000 # python only, capped
```
Read the **whole diff** before composing findings. Cross-file patterns (test symmetry,
single-source-of-truth, deduplicated logic) are only visible in aggregate.
For each violation, capture:
- **Severity label** (🔴 / 🟡 / 🟢 — see Severity Guide below)
- **Category tag** (A1, B2, J3, etc.)
- **File:line** (new-side line number)
- **One-sentence note** — what's wrong, with a concrete fix direction
### Phase 4 — Output the punch list
Format follows the template in [§ Output Template](#output-template). Group by
severity. Include a brief **Strengths** section to balance the findings — this
matches the tone of the bot review on PRs #733–#747 and avoids the impression that
the review is purely negative.
**Hard rules:**
- Do **not** edit any files.
- Do **not** propose patches inline (one-line fix direction is OK; full code blocks are not).
- Do **not** repeat findings the linter or type checker already surfaces.
- If a finding is style-only and CLAUDE.md doesn't mandate it, demote to 🟢.
---
## Severity Guide
| Label | Meaning | Examples |
|---|---|---|
| 🔴 **blocking** | Must fix before push. Real bug, security issue, or violates an ADR/CLAUDE.md mandate. | Cache-hit bypass on auth gate; raw `List[Dict]` from MCP tool; missing `await` on async call; HMAC compare with strings instead of bytes. |
| 🟡 **important** | Should fix. Not a bug today but a footgun that has caused review-rounds before. | Field description mismatches value; unbounded fan-out without semaphore; `int()` cast inside catch-all `except`; missing `openWorldHint=True`. |
| 🟢 **nit** | Polish. Not blocking, would be nice. | `Optional[X]` when surrounding file uses `X | None`; redundant regex anchors; missing test docstring; pluralization in error messages. |
| 💡 **suggestion** | Alternative approach to consider. No fix expected. | "Could use `frozenset` here for O(1) lookup"; "Could centralise this rewrite in `_make_request`". |
| 🎉 **praise** | Notable good pattern. Worth calling out for visibility. | Cache-hit AND cache-miss paths both enforce the gate; CI-guard test for registry exhaustiveness; explicit `*` keyword-only separator. |
**Escalation rules** (apply per change shape from Phase 1):
- Security-sensitive scope → all J* and K* findings are at least 🟡; auth bypass risks are 🔴.
- MCP surface → D4 (annotations) and I1 (response wrapping) are 🔴.
- Client layer → H1 (duplicate round-trips) is 🟡 minimum.
---
## Project Checklist
Each item is tagged for use in the punch list (e.g. `[B1]`). Items are mined from
PRs #733, #734, #735, #736, #737, #741, #745, #746, #747, #750.
### A. Style & typing (CLAUDE.md)
- **A1** — PEP 604 unions: `X | None`, never `Optional[X]`. (#737, #745, #735)
- **A2** — Lowercase generics: `dict[str, Any]`, `list[T]`, never `Dict`/`List`. (#736, #745)
- **A3** — `anyio` not `asyncio`: `anyio.create_task_group`, `anyio.Lock`, `anyio.run`. Flag any new `asyncio.gather`/`asyncio.Lock`/`asyncio.run`.
- **A4** — Lazy `%`-style logging: `logger.warning("msg %s", var)`, not f-strings, in any file touched on this branch. (memory: feedback_lazy_logging)
- **A5** — Typed signatures: every new/modified function has parameter and return type hints. (#733)
- **A6** — Stay consistent with surrounding file: if a touched file uses legacy `Dict`/`Optional`, flag the introduced inconsistency, not pre-existing debt. (#736, #735, #745)
### B. Error handling specificity
- **B1** — Hoist parsing/casts before network calls: `int()`, `json.loads()`, etc. in their own `try/except (TypeError, ValueError)` *before* the network call. Catch-all `except Exception` over the parse + network produces "unexpected error" instead of a specific log line. (#750-r3, #750-r5)
- **B2** — Narrow exception types: `except RuntimeError`, not bare `except Exception`, when only one failure mode is documented. (#750-r6)
- **B3** — Specific error messages with context: include the problematic value and its surrounding context (e.g. `f"expected int id for {doc_type}, got {type(value).__name__}: {value!r}"`), not opaque re-raises. (#750-r6)
- **B4** — Fail-open vs fail-closed is a deliberate choice: any new fail-open branch (transient error → keep going) needs a log line and a brief comment on *why* fail-open is correct here. (#750-r2)
- **B5** — Except clause order: when catching a subclass and superclass in the same `try`, the subclass must come first. `except TimeoutError` before `except Exception` (TimeoutError is OSError → Exception subtype). (#747)
- **B6** — Defensive parsing envelope: when extracting fields from external payloads, wrap in `try/except (KeyError, TypeError, ValueError)` and return a sentinel. Bubbling `ValueError` from `int("not-a-number")` to the caller is a finding. (#747)
### C. Comment honesty & documentation
- **C1** — Comments must match behavior: don't say "background" if the code runs inline; don't say "evicts asynchronously" if it `await`s. (#750-r1)
- **C2** — Document all return-None / sentinel cases: if a function returns `None` for both 404 *and* malformed XML, the docstring must enumerate both. (#750-r6)
- **C3** — Cross-references point at canonical/earlier definition: "Mirrors `_verify_X`" should reference the function defined first. (#750-r1)
- **C4** — TODOs include WHY: `# TODO(perf): get_items(batch_size=-1) fetches all items at query time` ✓; `# TODO: optimize` ✗. (#750-r3)
- **C5** — Defensive code carries its reason: `.get(key, default)` that "shouldn't fire in practice" needs a comment saying so — otherwise readers treat it as load-bearing. (#750-r6)
- **C6** — Surprising-on-first-read code needs a one-liner: e.g. `password=token` for bearer auth is surprising; a `# caldav reuses the password slot for bearer tokens` line saves readers from suspecting copy-paste bugs. (#734)
- **C7** — Mark deliberate omissions: when a code path *intentionally* skips a guard (e.g. DELETE returns empty body, no shape guard), say so explicitly. Otherwise future maintainers will treat it as an oversight. (#736)
### D. API & type boundaries
- **D1** — Public response models narrow, internal types may widen: if `SearchResult.id: int | str` for forward-compat, `SemanticSearchResult.id: int` and convert at the boundary with an explicit cast that fails loudly. (#750-r3)
- **D2** — Field descriptions match the value: a field described as "unique documents" must hold a unique-document count, not a chunk count. Watch for `len(results)` assigned to a field whose docstring implies dedup. (#750-r1, #735)
- **D3** — Symmetry between related fields: `verified_count` and `dropped_count` should count at the same granularity. (#750-r6)
- **D4** — MCP tool annotations (ADR-017): tools that hit Nextcloud have `openWorldHint=True`. Read-only ops have `readOnlyHint=True`. Destructive ops have `destructiveHint=True` + `idempotentHint=True`. Create ops have `idempotentHint=False`. Update ops have `idempotentHint=False` (etag changes mean different inputs). (#741)
- **D5** — Keyword-only separators: use `*` in signatures with multiple optional params of the same/related types, e.g. `def __init__(self, *, password=None, token=None)` to prevent positional misuse. (#734)
- **D6** — Defaults match documented constraints: if the docstring says "max 1000 characters", add `Field(max_length=1000)` — don't rely on the server returning 400. (#737)
- **D7** — Create/update tools return `*Response` wrappers, not raw models: any new `@mcp.tool` returning a raw `BaseModel` (not `BaseResponse` subclass) is a finding. (#737, CLAUDE.md MCP Response Patterns)
- **D8** — Pagination metadata: a field called `total` should be the server-side total, not page count. Use `count` for "returned in this page" or add `has_more`. (#737)
### E. Concurrency, resource bounds & lifespan
- **E1** — Bound concurrency: any new `asyncio.gather` / task-group fan-out over user data needs an `anyio.Semaphore` cap (default 20 in this repo). (#750-r1)
- **E2** — Bound over-fetching: `limit * K` for filtering must have a fixed `K`. Unbounded N-types or unbounded `K` is a finding. (#750-r1)
- **E3** — Don't snapshot mutable singletons: `eviction_task_group` set during lifespan startup must be read via `@property`/accessor, not snapshotted into another object's `__init__` (order-sensitive race). (#750-r5)
- **E4** — Comment safety of dict/list mutation across tasks: when multiple concurrent tasks write distinct keys to a shared dict, add a one-line comment explaining cooperative-multitasking safety. (#750-r1)
- **E5** — Lifespan context symmetry: when adding fields to one of `AppContext` / `OAuthAppContext`, mirror them in the other to prevent silent regressions in OAuth mode. (#746)
- **E6** — Security gates apply on cache-hit AND cache-miss paths: an allowlist or scope check must run after both branches resolve. A test asserting cache-hit enforcement is mandatory. (#745)
### F. Tests
- **F1** — Symmetry across verifiers/handlers: if you added a 403 test for the notes verifier, add one for files/deck/news too. Asymmetric coverage in a registry-style module is a finding. (#750-r5, #741)
- **F2** — Fail-open paths have a test: any verifier or handler with a fail-open branch (non-numeric id, malformed payload) has a unit test that exercises it. (#750-r5, #750-r6)
- **F3** — CI guard tests for registries: registries (e.g. `INDEXED_DOC_TYPES → verifier`) should have a test that the registry is exhaustive. (#750-r2)
- **F4** — New public functions have tests: any new MCP tool, client method, or Pydantic model without a unit test is a 🟡 finding (or 🟢 if integration-only).
- **F5** — `caplog` is logger-scoped: prefer `caplog.set_level(logging.WARNING, logger="nextcloud_mcp_server.client.notes")` over plain `caplog.at_level("WARNING")` to avoid false positives from other loggers. (#736)
- **F6** — Wire-through tests at the right layer: pure-unit tests on the leaf class are great, but a test that the parent (`NextcloudClient.from_env`) actually threads the param through is a common gap. (#734)
- **F7** — Production-evidence regressions get a unit test: any bug found via CloudWatch / production logs (not in tests) needs a unit test added in the fix PR. (#746)
- **F8** — Test docstring consistency: if 4 of 5 new tests have docstrings, give the 5th one too. (#735)
### G. Configuration, docs & ADRs
- **G1** — New config knobs go through `Settings` + dynaconf validator: any new env var read directly via `os.getenv` in module code (rather than via `Settings`) is a finding. (#750-r5)
- **G2** — Resolve config lazily: `default = settings.X` at function-call time, not at decoration / module-import time, so test overrides work. (#750-r5)
- **G3** — User-facing docs cover non-obvious costs: new code paths with measurable latency (unbounded fetches, per-item round-trips) need a note in `docs/configuration.md` or the relevant ADR. (#750-r2, #750-r6)
- **G4** — ADRs match shipped interface: example code blocks in any modified ADR reflect the actual module's public API. If the ADR shows `Verifier` but the code ships `BatchVerifier`, the ADR is wrong. (#750-r3)
- **G5** — Tool docstrings document hidden round-trips: if an MCP tool now does an extra fetch per result (e.g. post-verification race guard), the docstring says so. (#750-r6)
- **G6** — Migration notes for breaking changes: a PR that silently changes failure modes (e.g. previously-allowed clients now get 401) needs a migration note in the PR description and ideally a startup log line. (#745)
- **G7** — Priority order in env var fallbacks: if a setting reads from `WEBHOOK_INTERNAL_URL` → `NEXTCLOUD_MCP_SERVER_URL` → `/.dockerenv` → localhost, document the order in code comments. (#747)
### H. Performance hygiene
- **H1** — No duplicate round-trips: if data needed by a verifier/handler is already in `SearchResult.metadata` or a passed-in object, don't re-scroll/re-fetch it. (#750-r1)
- **H2** — Single source of truth: hardcoded lists of doc types / scopes / app names in multiple files should reference one constant (e.g. `INDEXED_DOC_TYPES`). (#750-r2)
- **H3** — Limit clamping at the right layer: `min(max(1, limit), 200)` at the client layer gives a clear error rather than a confusing server-side mismatch. (#741)
- **H4** — Avoid empty JSON bodies in HTTP requests: `json=body or None` (instead of `json=body`) prevents sending `{}` and a spurious `Content-Type: application/json` header for bodyless calls. (#741)
### I. MCP response patterns (CLAUDE.md)
- **I1** — No raw `List[Dict]` from MCP tools: tools return Pydantic models inheriting from `BaseResponse`. FastMCP mangles raw lists into dicts with numeric string keys. **🔴 blocking when violated.**
- **I2** — Co-author trailer in commits: each AI-assisted commit ends with `Co-Authored-By: Claude ...`. Missing trailer on AI-touched commits is a 🟢 note.
### J. Security & input validation
- **J1** — Validate untrusted identifiers before HTTP layer: room tokens, paths, IDs that go into URLs must be validated against an allowlist regex (e.g. `r"[A-Za-z0-9]+"` with `fullmatch`) *before* the request is built — not after. Path traversal (`../etc/passwd`) is the canonical example. (#741)
- **J2** — `hmac.compare_digest` on bytes: both sides encoded to `bytes` before comparison. Comparing strings is incorrect. (#747)
- **J3** — HTML-escape user-controlled values in HTML responses: even when the input is gated by an upstream check (e.g. `get_preset(preset_id)` returning `None` for unknown), escape on the success path too. Defense-in-depth, not single-layer. (#747)
- **J4** — Static-config-but-still-rendered fields should be escaped consistently: if dynamic fields are escaped and static fields aren't, the inconsistency itself is the smell. (#747)
- **J5** — Privacy defaults: defaults for polling/read-receipt parameters should minimize side effects (e.g. `no_status_update=True` for chat polling so users don't appear "online" on every poll). (#741)
- **J6** — Server-layer hardening even when client supports it: if an MCP tool calls a client with hardcoded safe values (e.g. `look_into_future=False`), don't expose the unsafe option as a tool parameter. (#741)
- **J7** — Redundant regex anchors with `fullmatch()`: `re.compile(r"^[A-Za-z0-9]+$").fullmatch(...)` — the `^`/`$` are redundant and misleading. Use `re.compile(r"[A-Za-z0-9]+")` with `fullmatch`. (#741)
### K. Trust boundaries
- **K1** — Comment where input comes from at trust boundaries: webhook payloads, request body fields, query params — say "user-controlled" or "static config" so future readers know the threat model. (#747)
- **K2** — Differentiate HTTP status codes by retry semantics: 503 for "try again" (queue saturated, dependency down), 500 for "broken" (genuine bug). 401 without `WWW-Authenticate` is correct when the caller has no auth state machine (e.g. Nextcloud webhook delivery). (#747)
- **K3** — Trust-boundary comments at API entrypoints: when a `uri` field in a request body is taken at face value (e.g. Astrolabe provides its own webhook URI), document the trust assumption. (#747)
### L. Sentinel values & PHP API quirks
- **L1** — `0` as a valid value vs `None` as "absent": fields like `lastReadMessage` or `expirationTimestamp` may use `0` for "no messages read" or "never expires". Don't truthy-test them. Document semantics in field docstrings. (#741)
- **L2** — PHP empty-array-as-list quirks: spreed and other PHP backends serialize empty objects as `[]`. Pydantic validators must coerce `[] → {}` for object-shaped fields and `[] → None` for nullable single-object fields. (#741)
- **L3** — `isinstance(value, (date, datetime))` is redundant: `datetime` is a subclass of `date`. Either drop `datetime` from the tuple, or distinguish behavior between the two. (#735)
- **L4** — Field description matches serialized form: a field documented as "ISO date format" but accepting `datetime` will serialize as `"2000-01-01T12:30:00"`, not `"2000-01-01"`. Either coerce to date or update the description. (#735)
### M. Defensive code patterns & symmetry
- **M1** — `getattr(obj, "field", None)` masks future typos: when an attribute is *expected* to exist post-fix, prefer direct access so a typo (`document_recieve_stream`) fails loudly. The defensive `getattr` is the right call only when the attribute is genuinely optional. (#746)
- **M2** — Both-credentials / multiple-mode case has explicit precedence: when two mutually-exclusive params can both be set, either raise `ValueError` or `logger.warning` — silent precedence is a footgun. (#734)
- **M3** — Defensive dict construction over passing `None`: `kwargs = {"password": password, "auth_type": "basic"} if password else {"auth_type": "bearer"}` — don't pass `None` into constructors that don't accept it. (#734)
- **M4** — Speculative defensive code without observed cause: if you add an unwrap for `[{...}] → {...}` "in case the API returns it", cite the version/route where it's been observed. Otherwise it silently masks contract changes. (#736)
- **M5** — Idempotency of URL/path transforms: `_resolve_url` should be safe to call twice. If it adds `/index.php` only when missing, document and test the idempotency. (#733)
---
## What NOT to Flag
- **Pre-existing technical debt** outside the diff. Note as 💡 if relevant, but don't list as a blocking finding.
- **Style violations the linter already catches.** ruff and ty are authoritative for what they cover. Manual style checklist is for things they don't cover (PEP 604 unions, lazy logging, `Optional` vs `|`).
- **Code formatting** — `ruff format --check` is the source of truth.
- **Bikeshed-tier preferences** (variable naming taste, line breaks, comment phrasing).
- **Anything you'd phrase as "I would have done it differently"** without a concrete repo-pattern violation.
- **Findings that duplicate the bot's previous review** if there's an open PR with one already (use `gh pr view --comments` to check).
If a category has zero applicable findings, omit it from the punch list. Don't pad.
---
## Output Template
```
# Pre-push review for <branch>
**Base:** <base> · **Commits:** <N> · **Files changed:** <M>
**Change shape:** <classification from Phase 1>
## Automated checks
- ruff: ✅ / ❌
- format: ✅ / ❌
- ty: ✅ / ❌
- unit tests: ✅ / ❌ (X passed, Y failed)
- smoke (if run): ✅ / ❌
[If any check failed, stop here and report. Otherwise continue.]
## Findings
### 🔴 Blocking (<count>)
1. **[J1]** `nextcloud_mcp_server/client/talk.py:42` — Token interpolated into
URL before validation; path-traversal risk. Move `_validate_token(token)`
above the `httpx.URL(...)` call.
### 🟡 Important (<count>)
1. **[D2]** `nextcloud_mcp_server/models/semantic.py:87` — `verified_count`
description says "unique documents" but value is `len(verified_results)`
(chunk count). Either reword or dedup the value.
2. **[F1]** `tests/unit/search/test_verification.py` — 403 tests exist for
notes/deck verifiers but not for files/news. Add for symmetry.
### 🟢 Nits (<count>)
1. **[A1]** `nextcloud_mcp_server/server/deck.py:118` — `Optional[int]` →
`int | None` per CLAUDE.md.
### 💡 Suggestions
1. `nextcloud_mcp_server/auth/client_registry.py` — Pre-existing legacy
`Dict`/`Optional` typing throughout. If touching anyway, opportunity
to modernize.
## Strengths
- 🎉 **[E6]** Allowlist enforced on both cache-hit and cache-miss paths in
`verify_token_for_management_api`, with a dedicated test
(`test_cache_hit_also_enforces_allowlist`). Exemplary security gate.
- 🎉 **[F3]** New `test_supported_doc_types_covers_indexed_types` enforces
registry exhaustiveness — will catch any future doc_type added without a
verifier.
## Notes
- 5 commits on branch; all have `Co-Authored-By` trailer ✓
- ADR-019 implementation checklist all marked [x] (G4 ✓)
```
### Clean diff (no findings)
```
# Pre-push review for fix/notes-etag-handling
**Base:** master · **Commits:** 2 · **Files changed:** 3
**Change shape:** client layer
## Automated checks
- ruff: ✅
- format: ✅
- ty: ✅
- unit tests: ✅ (412 passed)
## Findings
No checklist violations found.
## Strengths
- 🎉 **[B1]** ETag parsing hoisted into its own try/except before the
network call — error logs will be specific to malformed ETag, not
generic "unexpected error".
Safe to push.
```
---
## Notes for the Running Model
- Read the **whole diff** before composing the report. Cross-file patterns
(F1 symmetry, H2 single-source-of-truth, E5 lifespan symmetry) are only
visible in aggregate.
- Keep findings **specific and short**. One sentence each plus a fix direction.
The point is to feed them back into the main loop for fixing, not to write
a treatise.
- If an automated check fails, **stop and report** — don't run the checklist
on top of a broken build.
- The checklist is calibrated to *this repo's* recurring review feedback. If
the diff is in an unrelated area (CI config, random docs), most items won't
apply — say so rather than padding.
- **Don't fix anything.** This skill reports; the main loop fixes. Even if a
finding is a one-character change, leave it for the user to action.
- Use `gh pr view <PR#> --comments` (not `gh api`) when checking for prior bot
reviews on an existing PR — see memory feedback_use_gh_pr.
CLAUDE.md
+11 -1
View File
@@ -23,11 +23,21 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
```
### Code Quality
- **Run ruff and ty before committing**:
- **Before committing or pushing, invoke the `pre-push-review` skill**:
- Runs `ruff check`, `ruff format --check`, `ty check`, and unit tests, then audits the
branch diff against this repo's recurring PR-review patterns (mined from PRs #733–#750).
- Output is a labelled punch list (🔴 blocking / 🟡 important / 🟢 nit). The main loop
fixes; the skill reports.
- Skill location: `.claude/skills/pre-push-review/SKILL.md`. Invoke via the `Skill` tool
with `skill="pre-push-review"`, or when the user types `/pre-push-review`.
- Skip only for tiny diffs (typo, README tweak, single-line dependency bump) or when the
user explicitly says "just push it".
- **Manual fallback** (if the skill is unavailable):
```bash
uv run ruff check
uv run ruff format
uv run ty check -- nextcloud_mcp_server
uv run pytest tests/unit/ -x -q
```
- **Ruff configuration** in pyproject.toml (extends select: ["I"] for import sorting)
docker-compose.yml
+1 -1
View File
@@ -36,7 +36,7 @@ services:
# Mount OIDC development directory outside /var/www/html to avoid rsync conflicts
# The post-installation hook will register /opt/apps as an additional app directory
#- ./third_party:/opt/apps:ro
- ./third_party/astrolabe:/opt/apps/astrolabe:ro
#- ./third_party/astrolabe:/opt/apps/astrolabe:ro
#- ./third_party/oidc:/opt/apps/oidc:ro
environment:
- NEXTCLOUD_TRUSTED_DOMAINS=app
docs/ADR-019-verify-on-read-for-semantic-search.md
+279
View File
@@ -0,0 +1,279 @@
# ADR-019: Verify-on-Read for Semantic Search Results
**Status**: Accepted
**Date**: 2026-05-01
**Depends On**: ADR-007 (Background Vector Sync), ADR-010 (Webhook-Based Vector Sync)
## Context
The vector index in Qdrant is a *recall layer*, not the source of truth. Authoritative state for every indexed document — whether a note exists, whether a file is still shared with the user, whether a deck card is on a board the user can read — lives in Nextcloud, not in our index. Whenever those two views drift, semantic search returns **ghost records**: results that point to documents the user can no longer access (or that no longer exist at all).
### How drift happens
Two mechanisms keep Qdrant in sync with Nextcloud, and both have non-zero latency:
1. **Webhook delivery (ADR-010)**. Nextcloud's `webhook_listeners` app dispatches change notifications via background jobs. The default `cron` job runs every 5 minutes, so even a healthy webhook pipeline opens a 0–5 minute window where deletions/unshares are not yet reflected in the index. Operators with dedicated webhook workers can shrink this, but most production deployments stay on the default cadence.
2. **Periodic scanner (ADR-007)**. The fallback reconciliation scan runs on `vector_sync_scan_interval`. The dev default is 60 seconds, but ADR-010 explicitly recommends raising this to 1 hour or more in production once webhooks are in place, since the scanner exists primarily to recover from missed events. Large deployments may run it once per day.
Beyond cadence, several failure modes cause webhooks to be missed entirely:
- The MCP server is down or unreachable when the webhook fires (Nextcloud does not durably retry).
- Sharing changes (revoking a share, leaving a group) do not always emit file events that match what we registered for.
- Application-level deletions in apps without rich event support (older Deck versions, custom Tables flows) bypass the file-event hooks.
In all of these cases, the document remains in Qdrant until the next periodic scan reconciles it — which may be hours away. Until then, `nc_semantic_search` happily returns the stale entry.
### Why this matters more for semantic search than keyword search
A keyword search via the Notes API is naturally bounded by what the API returns: deleted notes are not in the result set. The vector index is a separate store with its own lifecycle. The further we extend semantic search across apps (notes, files, deck cards, news items today; calendar, contacts, tables, cookbook tomorrow), the more divergent surfaces we expose to drift. Every new doc_type is another path where a webhook can be missed and another type of "ghost" can leak into results.
The risk is not just a confusing UX. For RAG flows like `nc_semantic_search_answer` (ADR-008), a stale result means the LLM is asked to synthesize an answer over content the user no longer has access to — a privacy boundary violation, not just a relevance bug.
### Current state of verification
Verification today is ad-hoc and inconsistent:
| Surface | Verifies? | Mechanism |
|---|---|---|
| `nc_semantic_search` | No | Returns raw Qdrant results. The docstring at `search/semantic.py:52` and `search/bm25_hybrid.py:75` references a `verify_search_results()` helper that was never implemented. |
| `nc_semantic_search_answer` | Partially | `server/semantic.py:431-448` fetches `notes.get_note(id)` and drops on exception — but only for `doc_type == "note"`. Files, news items, and deck cards fall through to the `else` branch (`server/semantic.py:449`) and are returned with their excerpt unverified. |
| `get_chunk_with_context` (when `include_context=True`) | Implicitly, all types | `search/context.py::_fetch_document_text` re-fetches the document; on failure the *context expansion* is skipped but the original (unverified) result is still returned (`server/semantic.py:246-251`). |
There is no single point where the system asks: "is this document still accessible to this user, right now?"
### The four indexed doc types
The vector pipeline (`vector/scanner.py`, `vector/processor.py`) currently indexes four types, each with its own access-check shape:
| doc_type | Cheapest authoritative check | Notes |
|---|---|---|
| `note` | `notes.get_note(id)` — single REST call, 404 on deletion | Per-user store; access is binary (yours or not). |
| `news_item` | `get_items(batch_size=-1)` once per search + intersect | No per-item REST endpoint (`get_item()` is itself a fetch-all + filter); batching once is cheaper than N fetch-all-and-filter calls. |
| `file` | WebDAV `PROPFIND` with `Depth: 0` on `file_path` (already stored in Qdrant payload, see `server/semantic.py:161`) | `read_file()` works but downloads the body — too heavy for a verification check. PROPFIND is the WebDAV equivalent of HEAD. Catches both deletes and unshares. |
| `deck_card` | `deck.get_card(board_id, stack_id, card_id)` using metadata cached in Qdrant (`search/context.py::_get_deck_metadata_from_qdrant`) | Fallback iteration through all boards/stacks (used by context expansion) is O(boards × stacks) and far too expensive to run on every query. |
All four are query-time-cheap **if** we (a) deduplicate per-document before checking and (b) run checks concurrently.
## Decision
Implement **verify-on-read** as the authoritative access gate for semantic search. The vector index decides *what might be relevant*; Nextcloud decides *what the user can see*. We will:
1. Introduce a single `nextcloud_mcp_server/search/verification.py` module exposing `verify_search_results(client, results) -> list[SearchResult]`.
2. Wire it into both `nc_semantic_search` and `nc_semantic_search_answer` as the final step before results leave the server, replacing the ad-hoc note-only verification in the answer tool.
3. Dispatch per `doc_type` to a registry of verifiers using the cheapest authoritative check for each type.
4. Lazily evict from Qdrant when verification reveals a definitively-gone document, so the next query for the same content does not re-pay the verification cost.
The vector index becomes a **hint**, not a contract. We never trust it for access decisions.
## Implementation
### Module shape
```python
# nextcloud_mcp_server/search/verification.py
from typing import Awaitable, Callable
import anyio
import httpx
from nextcloud_mcp_server.search.algorithms import SearchResult
# A batch verifier takes a list of results for a single doc_type and returns
# the set of doc_ids that are currently accessible to the user. The shared
# semaphore caps concurrent Nextcloud round-trips across all verifier types.
#
# - Definitive 403/404 → omit the id from the returned set (drop the result).
# - Transient error (5xx, network, parse) → include the id (fail-open keep).
# - Verifier crash → caught by the dispatcher and treated as transient
# (all results for that type are kept; logged distinctly).
BatchVerifier = Callable[
["NextcloudClientProtocol", list[SearchResult], anyio.Semaphore],
Awaitable[set[int | str]],
]
async def verify_search_results(
client: "NextcloudClientProtocol",
results: list[SearchResult],
*,
max_concurrent: int = 20,
evict_on_missing: bool = True,
eviction_task_group: anyio.abc.TaskGroup | None = None,
) -> list[SearchResult]:
"""Filter search results to those the user can currently access.
Deduplicates by (doc_id, doc_type) before verifying, so multiple chunks
from the same document cost a single check. Each verifier owns its own
concurrency under the shared semaphore. Drops results whose verifier
omitted them; keeps results whose verifier raised or whose doc_type has
no registered verifier (transient failure should not silently shrink
results).
When evict_on_missing=True, schedules async deletion of the Qdrant points
for the missing document(s) so subsequent queries don't re-pay the cost.
Pass ``eviction_task_group`` (typically the lifespan-owned background task
group) to make eviction fire-and-forget; without it we run a local task
group that blocks the response until evictions complete.
"""
```
**Why batch?** A per-id `Verifier` would force one task-group creation per id, multiply the number of small tasks, and prevent the news single-fetch optimization (the News API has no per-item endpoint, so per-id verification would be O(N × all_items)). The batch interface lets each verifier own its own concurrency strategy: notes/files/deck cards parallelize per id under the shared semaphore; news fetches once and intersects.
### Verifier registry
```python
_VERIFIERS: dict[str, BatchVerifier] = {
"note": _verify_notes,
"news_item": _verify_news_items,
"file": _verify_files,
"deck_card": _verify_deck_cards,
}
```
Each verifier follows the same shape — accept a list of results for its type, fan out per-id under the shared semaphore (or fetch once and intersect, for news), and return the set of accessible ids:
```python
async def _verify_notes(
client,
results: list[SearchResult],
semaphore: anyio.Semaphore,
) -> set[int | str]:
accessible: set[int | str] = set()
async def check(result: SearchResult) -> None:
async with semaphore:
try:
await client.notes.get_note(int(result.id))
accessible.add(result.id)
except httpx.HTTPStatusError as e:
if e.response.status_code in (403, 404):
return # definitive — drop
accessible.add(result.id) # transient — keep
async with anyio.create_task_group() as tg:
for r in results:
tg.start_soon(check, r)
return accessible
```
For `file`, use WebDAV PROPFIND (`Depth: 0`) on the `file_path` from the Qdrant payload, not `read_file()`. For `deck_card`, use the cached `(board_id, stack_id)` from `_get_deck_metadata_from_qdrant`; if metadata is absent, treat the result as accessible and log — we will not run the iteration fallback in the hot path. For `news_item`, batch-fetch the user's items once via `get_items(batch_size=-1)` and intersect, since the News API has no per-item endpoint.
### Deduplication
A 10-result page typically references 3–4 unique documents because of chunking. Dedupe by `(doc_id, doc_type)` *before* invoking the verifiers, so each batch verifier sees only unique ids. The dispatcher then propagates each id's verdict to every chunk of that document:
```python
unique: list[SearchResult] = []
seen: set[tuple[int | str, str]] = set()
for r in results:
key = (r.id, r.doc_type)
if key not in seen:
seen.add(key)
unique.append(r)
# Group unique results by doc_type and run their batch verifiers in parallel.
by_type: dict[str, list[SearchResult]] = group_by_doc_type(unique)
accessible_by_type: dict[str, set[int | str]] = {}
async with anyio.create_task_group() as tg:
for dtype, items in by_type.items():
verifier = _VERIFIERS.get(dtype)
if verifier is None:
# Soft failure: keep all results for unknown doc_types.
accessible_by_type[dtype] = {r.id for r in items}
continue
tg.start_soon(_run_verifier, verifier, dtype, items, accessible_by_type)
return [
r for r in results
if r.id in accessible_by_type.get(r.doc_type, set())
]
```
A verifier crash maps to "keep all" for that type — we do not want a flaky network blip to silently shrink results.
### Lazy eviction
When a verdict is `False`, queue a Qdrant delete for all points matching `(user_id, doc_id, doc_type)`. The plumbing already exists in `vector/placeholder.py::delete_placeholder_point` (which uses a filter-based delete); we need a sibling `delete_document_points` that omits the `is_placeholder` filter, so it removes real chunks too.
Eviction is fire-and-forget from the verification path — wrap it in a background task group on the lifespan context to avoid blocking the response. If eviction fails, the next query will simply re-verify and re-attempt; this is self-healing.
### Wiring
In `server/semantic.py::nc_semantic_search`, after the existing dedup and `[:limit]` slice, but **before** context expansion (which is expensive and pointless on inaccessible results):
```python
search_results = all_results[:limit * 2] # fetch extra to absorb evictions
search_results = await verify_search_results(client, search_results)
search_results = search_results[:limit]
```
Note the over-fetch: verification can shrink the page, so we ask for `limit * 2` candidates and trim *after* verification. This preserves the user's requested page size when ghosts are present without paying for full re-search.
In `server/semantic.py::nc_semantic_search_answer`, replace the per-type `if result.doc_type == "note"` branch (lines 428-453) with a call to `verify_search_results` followed by the existing full-content fetch (which can stay note-specific, since only notes use full content; the rest still use excerpts).
### What we deliberately do NOT do
- **No verification cache.** The whole point of verify-on-read is that the answer can change between calls. A short-TTL cache (say, 30s) is plausible if benchmarks show verification dominating latency, but it is not in the v1 scope.
- **No verifier for unsupported doc_types.** If a future doc_type lands in Qdrant without a registered verifier, log a warning and pass the result through. Verification is opt-in per type; missing a verifier is a soft failure.
- **No deck-card iteration fallback.** The fallback in `_fetch_document_text` exists for context expansion, where O(boards × stacks) is acceptable for a single result. In verification we may run the check on every chunk in every search; the fallback would amplify search latency unacceptably.
## Consequences
### Positive
- **Correctness**: Deletes/unshares are reflected in search results within one query, regardless of webhook delivery delays or scanner intervals. Operators can safely raise `vector_sync_scan_interval` to its production-recommended value without leaking ghost records.
- **Privacy**: RAG flows (`nc_semantic_search_answer`) can no longer synthesize answers over content the user has lost access to.
- **Self-healing index**: Lazy eviction means the index converges toward correctness as users query, without needing the scanner to find every drifted record.
- **Single source of truth**: Removes the docstring/code mismatch where `verify_search_results()` was promised but never delivered.
### Negative
- **Latency tax on every search**: Each unique `(doc_id, doc_type)` adds one Nextcloud round-trip. With 3–4 unique docs and 20-way concurrency, this is one parallel batch — likely under 100ms on a healthy connection, but it *is* on the critical path.
- **API load on Nextcloud**: A query that previously hit only Qdrant now hits Nextcloud once per unique result. For high-QPS deployments this is non-trivial and may need rate limiting (already present in `BaseNextcloudClient` retry logic).
- **More moving parts in the search path**: Errors in verification can mask errors in search. Verifier exceptions must be logged distinctly so debugging stays tractable.
- **Doc_type coverage is now a correctness contract**: When we add a new indexable doc_type, we must add a verifier in the same PR, or accept that ghost records are possible for that type. CI should fail if a doc_type is indexed without a registered verifier.
### Neutral
- The `verify_search_results()` function name in existing docstrings becomes accurate. No public API breakage.
- Webhooks remain valuable — they keep the index *recall* fresh (so semantically-relevant new docs appear in results quickly). Verification only handles the *precision* side (filtering inaccessible ones out).
## Alternatives Considered
**1. Tighten webhook delivery cadence.** Reduce Nextcloud's webhook cron interval from 5 minutes to 1 minute, or run a dedicated webhook worker. *Rejected as a complete solution*: addresses average-case latency but does nothing for missed webhooks, server-down windows, or app surfaces that lack rich events. We still recommend operators do this — it improves recall freshness — but it cannot replace verification.
**2. Synchronous webhook acknowledgement.** Have the MCP server delete from Qdrant inside the webhook handler before returning 2xx. *Rejected*: still doesn't help missed webhooks, and adds a hard dependency from the webhook critical path to Qdrant being reachable. Already partially implemented; verify-on-read complements it rather than replacing it.
**3. Bloom filter / negative cache of recently-deleted IDs.** Maintain an in-memory set of "known deleted" IDs populated by webhook handlers, consulted before returning search results. *Rejected*: cannot answer for unshares (which are user-relative, not global), grows unbounded, and is essentially a worse verifier — verifying against Nextcloud is authoritative and not much slower for the page sizes we deal with.
**4. Verify only in `nc_semantic_search_answer`, not `nc_semantic_search`.** Argue that raw search is "advisory" and verification only matters when the LLM consumes content. *Rejected*: ghost records in raw search results are still misleading to users and to other tools that compose on top of search. The bar for a search tool is "results are accessible," not "results are accessible if you happen to feed them into a sampling tool."
**5. Pre-verification at index time only (no query-time check).** *Already what we have*, and the problem statement.
## Related Decisions
- ADR-007: Background Vector Sync — establishes the polling architecture that produces drift.
- ADR-008: MCP Sampling for Semantic Search — defines the RAG flow that most acutely needs verified results.
- ADR-010: Webhook-Based Vector Sync — reduces but does not eliminate drift; verify-on-read closes the residual gap.
- ADR-013: RAG Evaluation — verification policy should be exercised in eval suites (with both fresh and stale fixtures).
## References
- `nextcloud_mcp_server/search/semantic.py:52` and `search/bm25_hybrid.py:75` — orphaned `verify_search_results()` references.
- `nextcloud_mcp_server/server/semantic.py:428-453` — current note-only verification in `nc_semantic_search_answer`.
- `nextcloud_mcp_server/search/context.py::_fetch_document_text` — per-doc-type fetch logic that informs the verifier registry.
- `nextcloud_mcp_server/vector/placeholder.py::delete_placeholder_point` — filter-based Qdrant delete pattern to extend for full-document eviction.
## Implementation Checklist
- [x] Create `nextcloud_mcp_server/search/verification.py` with `verify_search_results()` and the verifier registry.
- [x] Implement `_verify_notes`, `_verify_news_items`, `_verify_files` (PROPFIND), `_verify_deck_cards` (metadata fast-path only). Names plural to reflect the batch-verifier interface (see "Module shape" above).
- [x] Add `delete_document_points()` in `nextcloud_mcp_server/vector/eviction.py` for non-placeholder filter-based deletes.
- [x] Wire into `nc_semantic_search` with `limit * 2` over-fetch, trim to `limit` after verification.
- [x] Wire into `nc_semantic_search_answer`; verification runs upstream in `nc_semantic_search`, and the note-only re-fetch is retained as a sub-second race guard.
- [x] Update existing docstrings in `search/semantic.py` and `search/bm25_hybrid.py` to reflect the new verify-on-read path.
- [x] Unit tests: each verifier handles 200/403/404/transient distinctly; dedup collapses chunks; eviction is scheduled on missing.
- [x] Integration test: index a note, delete via API (no webhook), confirm the next semantic search does not return it. (See `tests/integration/test_verify_on_read.py`. Coverage gap for `file`, `deck_card`, `news_item` integration tests is tracked as a follow-up.)
- [x] CI guard: enumerate indexed doc_types in `vector/scanner.py` and assert each has a registered verifier. (`INDEXED_DOC_TYPES` in `vector/scanner.py`; `tests/unit/search/test_verification.py::test_supported_doc_types_covers_indexed_types`.)
- [x] Document the latency budget and rate-limit posture in `docs/configuration.md`. (See "Verify-on-Read Latency Budget" section.)
docs/configuration.md
+45
View File
@@ -474,6 +474,51 @@ DOCUMENT_CHUNK_OVERLAP=100
**Important**: Changing chunk size requires re-embedding all documents. The collection naming strategy (see "Qdrant Collection Naming" above) helps manage this by creating separate collections for different configurations.
### Verify-on-Read Latency Budget
Every semantic search request runs an access-control verification pass over its
results before returning them, to filter out documents the user can no longer
access (deleted, unshared, permissions changed). See
[ADR-019](ADR-019-verify-on-read-for-semantic-search.md) for the full design.
This adds Nextcloud round-trips to the search path that operators should be
aware of:
- **Per-search cost**: one Nextcloud round-trip per *unique* `(doc_id, doc_type)`
in the result set. Chunking means a 10-result page typically references 3-5
unique documents, so verification adds 3-5 round-trips. With the default
20-way concurrency this is one parallel batch — usually under 100 ms on a
healthy connection.
- **Concurrency**: all verifications fan out under a shared semaphore.
Tunable via the `VERIFICATION_CONCURRENCY` env var (settings field
`verification_concurrency`, default 20) — lower it if your Nextcloud
backend struggles with the parallel fan-out, or raise it on a healthy
connection to speed up large result pages.
- **News API caveat**: the News app has no per-item endpoint, so the news
verifier issues a single `news.get_items(batch_size=-1, get_read=True)` call
per search that contains any news result, then intersects locally. The
payload is **unbounded** — for users with very large feed backlogs this can
dominate verification latency. As a rough guide on a healthy LAN connection:
a typical purged backlog (1k–5k items) returns in ~200–500 ms; very large
backlogs (>20k items) can exceed 2 s and become the dominant cost of any
search that surfaces news results. Disabling News in the indexer or running
with a smaller backlog mitigates this; per-item paginated verification is
tracked as a future improvement.
- **Eviction**: when verification finds a definitive miss (404 / 403), the
corresponding Qdrant points are deleted in the background on a lifespan-owned
task group — fire-and-forget, does **not** block the search response.
Eviction failures are logged but never propagated; the next query will
re-verify and re-attempt (self-healing).
- **Failure modes**: transient errors (5xx, network) keep results visible
(fail open) so a flaky link does not silently shrink result pages; only
*definitive* 404 / 403 drops them.
If eviction ever needs to be disabled (debugging, benchmarking), the
`evict_on_missing=False` keyword argument on `verify_search_results()` skips
the Qdrant deletes without changing what is returned to the caller. **This
is a developer/test flag, not an operator knob — it has no env-var
equivalent.** Operators who need a runtime toggle should open an issue.
### Environment Variables Reference
| Variable | Required | Default | Description |
nextcloud_mcp_server/app.py
+56 -19
View File
@@ -9,12 +9,13 @@ import traceback
from collections.abc import AsyncIterator
from contextlib import AsyncExitStack, asynccontextmanager
from dataclasses import dataclass
from typing import Optional, cast
from typing import cast
from urllib.parse import urlparse
import anyio
import click
import httpx
from anyio.abc import TaskGroup
from anyio.streams.memory import MemoryObjectReceiveStream, MemoryObjectSendStream
from mcp.server.auth.settings import AuthSettings
from mcp.server.fastmcp import Context, FastMCP
@@ -315,10 +316,14 @@ class VectorSyncState:
and FastMCP session lifespans (where MCP tools need access to the streams).
"""
document_send_stream: Optional[MemoryObjectSendStream] = None
document_receive_stream: Optional[MemoryObjectReceiveStream] = None
shutdown_event: Optional[anyio.Event] = None
scanner_wake_event: Optional[anyio.Event] = None
document_send_stream: MemoryObjectSendStream | None = None
document_receive_stream: MemoryObjectReceiveStream | None = None
shutdown_event: anyio.Event | None = None
scanner_wake_event: anyio.Event | None = None
# Long-lived task group used for fire-and-forget background work spawned
# from the request path (e.g. ADR-019 verify-on-read eviction). Set by the
# starlette lifespan after entering its task group; cleared on shutdown.
eviction_task_group: TaskGroup | None = None
# Module-level singleton for vector sync state
@@ -330,11 +335,20 @@ class AppContext:
"""Application context for BasicAuth mode."""
client: NextcloudClient
storage: Optional["RefreshTokenStorage"] = None
document_send_stream: Optional[MemoryObjectSendStream] = None
document_receive_stream: Optional[MemoryObjectReceiveStream] = None
shutdown_event: Optional[anyio.Event] = None
scanner_wake_event: Optional[anyio.Event] = None
storage: "RefreshTokenStorage | None" = None
document_send_stream: MemoryObjectSendStream | None = None
document_receive_stream: MemoryObjectReceiveStream | None = None
shutdown_event: anyio.Event | None = None
scanner_wake_event: anyio.Event | None = None
@property
def eviction_task_group(self) -> TaskGroup | None:
# Read dynamically from the module-level singleton instead of
# snapshotting at lifespan-yield time. Snapshotting is order-sensitive:
# if the FastMCP server lifespan ever runs before the Starlette
# lifespan assigns the task group, every session for the life of the
# process would see ``None`` and fall back to inline eviction.
return _vector_sync_state.eviction_task_group
@dataclass
@@ -343,16 +357,19 @@ class OAuthAppContext:
nextcloud_host: str
token_verifier: object # UnifiedTokenVerifier (ADR-005 compliant)
refresh_token_storage: Optional["RefreshTokenStorage"] = None
oauth_client: Optional[object] = None
refresh_token_storage: "RefreshTokenStorage | None" = None
oauth_client: object | None = None
oauth_provider: str = "nextcloud" # "nextcloud" or "keycloak"
server_client_id: Optional[str] = (
None # MCP server's OAuth client ID (static or DCR)
)
document_send_stream: Optional[MemoryObjectSendStream] = None
document_receive_stream: Optional[MemoryObjectReceiveStream] = None
shutdown_event: Optional[anyio.Event] = None
scanner_wake_event: Optional[anyio.Event] = None
server_client_id: str | None = None # MCP server's OAuth client ID (static or DCR)
document_send_stream: MemoryObjectSendStream | None = None
document_receive_stream: MemoryObjectReceiveStream | None = None
shutdown_event: anyio.Event | None = None
scanner_wake_event: anyio.Event | None = None
@property
def eviction_task_group(self) -> TaskGroup | None:
# See AppContext.eviction_task_group for rationale.
return _vector_sync_state.eviction_task_group
class BasicAuthMiddleware:
@@ -569,6 +586,8 @@ async def app_lifespan_basic(server: FastMCP) -> AsyncIterator[AppContext]:
document_receive_stream=_vector_sync_state.document_receive_stream,
shutdown_event=_vector_sync_state.shutdown_event,
scanner_wake_event=_vector_sync_state.scanner_wake_event,
# eviction_task_group is exposed via @property (reads
# _vector_sync_state at access time, not snapshot).
)
finally:
logger.info("Shutting down BasicAuth session")
@@ -1189,6 +1208,8 @@ def get_app(transport: str = "streamable-http", enabled_apps: list[str] | None =
document_receive_stream=_vector_sync_state.document_receive_stream,
shutdown_event=_vector_sync_state.shutdown_event,
scanner_wake_event=_vector_sync_state.scanner_wake_event,
# eviction_task_group is exposed via @property (reads
# _vector_sync_state at access time, not snapshot).
)
finally:
logger.info("Shutting down MCP server")
@@ -1604,6 +1625,12 @@ def get_app(transport: str = "streamable-http", enabled_apps: list[str] | None =
username,
)
# Expose this long-lived task group to request-path code that
# wants to spawn background work (e.g. ADR-019 verify-on-read
# eviction). Eviction coroutines have their own try/except, so
# they cannot panic the parent group.
_vector_sync_state.eviction_task_group = tg
logger.info(
f"Background sync tasks started: 1 scanner + "
f"{settings.vector_sync_processor_workers} processors"
@@ -1617,6 +1644,8 @@ def get_app(transport: str = "streamable-http", enabled_apps: list[str] | None =
# Shutdown signal
logger.info("Shutting down background sync tasks")
shutdown_event.set()
# Request path must not spawn into a cancelling group.
_vector_sync_state.eviction_task_group = None
await client.close()
# TaskGroup automatically cancels all tasks on exit
@@ -1786,6 +1815,12 @@ def get_app(transport: str = "streamable-http", enabled_apps: list[str] | None =
use_basic_auth, # Pass as positional arg (before task_status)
)
# Expose this long-lived task group to request-path code
# that wants to spawn background work (e.g. ADR-019
# verify-on-read eviction). Eviction coroutines have their
# own try/except, so they cannot panic the parent group.
_vector_sync_state.eviction_task_group = tg
logger.info(
f"Background sync tasks started: 1 user manager + "
f"{settings.vector_sync_processor_workers} processors"
@@ -1799,6 +1834,8 @@ def get_app(transport: str = "streamable-http", enabled_apps: list[str] | None =
# Shutdown signal
logger.info("Shutting down background sync tasks")
shutdown_event.set()
# Request path must not spawn into a cancelling group.
_vector_sync_state.eviction_task_group = None
# Close token broker HTTP client
if token_broker._http_client:
await token_broker._http_client.aclose()
nextcloud_mcp_server/client/webdav.py
+29 -21
View File
@@ -483,13 +483,6 @@ class WebDAVClient(BaseNextcloudClient):
"status_code": 409,
"message": "Parent directory of destination doesn't exist",
}
logger.debug(
f"Parent directory of destination '{destination_path}' doesn't exist"
)
return {
"status_code": 409,
"message": "Parent directory of destination doesn't exist",
}
else:
logger.error(
f"HTTP error moving resource from '{source_path}' to '{destination_path}': {e}"
@@ -1301,12 +1294,33 @@ class WebDAVClient(BaseNextcloudClient):
async def get_file_info(self, path: str) -> dict[str, Any] | None:
"""Get file info including file ID via WebDAV PROPFIND.
.. note::
**Behavior change (ADR-019):** previously this method returned
``None`` for HTTP 404. It now raises ``HTTPStatusError`` for any
non-2xx status, including 404. ``None`` is reserved for the
ambiguous *malformed PROPFIND* case (server returned 2xx with a
response body missing required XML elements). External callers
updating from the old contract must catch ``HTTPStatusError``
and inspect ``e.response.status_code`` to handle 404 explicitly.
Args:
path: Path to the file (relative to user's files directory)
Returns:
File info dictionary with id, name, size, content_type, etc.
Returns None if file not found.
Returns ``None`` ONLY when the server returned a malformed
PROPFIND response (missing ``<d:response>`` /
``<d:propstat>`` / ``<d:prop>`` elements) — an ambiguous
state where we cannot tell whether the file exists.
Raises:
HTTPStatusError: For any non-2xx HTTP status, including 404
("not found"). Callers that want to treat 404 as
"absent" should catch ``HTTPStatusError`` and check
``e.response.status_code``. This matches the convention
of the rest of this client and lets verify-on-read
distinguish a definitive absence (HTTP 404) from a
brittle response (None).
"""
webdav_path = f"{self._get_webdav_base_path()}/{path.lstrip('/')}"
@@ -1323,19 +1337,13 @@ class WebDAVClient(BaseNextcloudClient):
</d:prop>
</d:propfind>"""
try:
response = await self._client.request(
"PROPFIND",
webdav_path,
headers={"Depth": "0"},
content=propfind_body,
)
response.raise_for_status()
except HTTPStatusError as e:
if e.response.status_code == 404:
logger.debug(f"File not found: {path}")
return None
raise
response = await self._client.request(
"PROPFIND",
webdav_path,
headers={"Depth": "0"},
content=propfind_body,
)
response.raise_for_status()
# Parse XML response
root = ET.fromstring(response.content)
nextcloud_mcp_server/config.py
+11
View File
@@ -65,6 +65,8 @@ _DEFAULTS: dict[str, Any] = {
"vector_sync_processor_workers": 3,
"vector_sync_queue_max_size": 10000,
"vector_sync_user_poll_interval": 60,
# Verify-on-read concurrency cap (ADR-019)
"verification_concurrency": 20,
# Qdrant
"qdrant_url": None,
"qdrant_location": None,
@@ -170,6 +172,7 @@ _dynaconf = Dynaconf(
Validator("VECTOR_SYNC_PROCESSOR_WORKERS", gte=1),
Validator("VECTOR_SYNC_QUEUE_MAX_SIZE", gte=1),
Validator("VECTOR_SYNC_USER_POLL_INTERVAL", gte=1),
Validator("VERIFICATION_CONCURRENCY", gte=1),
Validator("DOCUMENT_CHUNK_SIZE", gte=1),
# Non-negative
Validator("DOCUMENT_CHUNK_OVERLAP", gte=0),
@@ -455,6 +458,12 @@ class Settings:
vector_sync_queue_max_size: int = 10000
vector_sync_user_poll_interval: int = 60 # seconds - OAuth mode user discovery
# Verify-on-read concurrency (ADR-019). Cap on parallel Nextcloud
# round-trips during search-result verification fan-out. Lower this if the
# Nextcloud backend struggles with the parallel load; raise it on a
# healthy connection to speed up large result pages.
verification_concurrency: int = 20
# Qdrant settings (mutually exclusive modes)
qdrant_url: str | None = None # Network mode: http://qdrant:6333
qdrant_location: str | None = None # Local mode: :memory: or /path/to/data
@@ -793,6 +802,8 @@ def get_settings() -> Settings:
"vector_sync_processor_workers": "VECTOR_SYNC_PROCESSOR_WORKERS",
"vector_sync_queue_max_size": "VECTOR_SYNC_QUEUE_MAX_SIZE",
"vector_sync_user_poll_interval": "VECTOR_SYNC_USER_POLL_INTERVAL",
# Verify-on-read (ADR-019)
"verification_concurrency": "VERIFICATION_CONCURRENCY",
# Qdrant settings
"qdrant_url": "QDRANT_URL",
"qdrant_location": "QDRANT_LOCATION",
nextcloud_mcp_server/models/semantic.py
+46 -16
View File
@@ -1,7 +1,5 @@
"""Pydantic models for semantic search responses."""
from typing import List, Optional
from pydantic import BaseModel, Field
from .base import BaseResponse
@@ -10,7 +8,15 @@ from .base import BaseResponse
class SemanticSearchResult(BaseModel):
"""Model for semantic search results with additional metadata."""
id: int = Field(description="Document ID (int for all document types)")
id: int = Field(
description=(
"Document ID. Numeric for all currently indexed types (notes, files, "
"deck cards, news items). The internal SearchResult.id is typed as "
"int|str to leave room for future doc types with string identifiers; "
"the MCP response narrows to int and a future widening here would be "
"a deliberate, breaking-by-design API change."
)
)
doc_type: str = Field(
description="Document type (note, calendar_event, deck_card, etc.)"
)
@@ -29,36 +35,36 @@ class SemanticSearchResult(BaseModel):
)
chunk_index: int = Field(description="Index of matching chunk in document")
total_chunks: int = Field(description="Total number of chunks in document")
chunk_start_offset: Optional[int] = Field(
chunk_start_offset: int | None = Field(
default=None, description="Character position where chunk starts in document"
)
chunk_end_offset: Optional[int] = Field(
chunk_end_offset: int | None = Field(
default=None, description="Character position where chunk ends in document"
)
page_number: Optional[int] = Field(
page_number: int | None = Field(
default=None, description="Page number for PDF documents"
)
page_count: Optional[int] = Field(
page_count: int | None = Field(
default=None, description="Total number of pages in PDF document"
)
# Context expansion fields (optional, populated when include_context=True)
has_context_expansion: bool = Field(
default=False, description="Whether context expansion was performed"
)
marked_text: Optional[str] = Field(
marked_text: str | None = Field(
default=None,
description="Full text with position markers around matched chunk",
)
before_context: Optional[str] = Field(
before_context: str | None = Field(
default=None, description="Text before the matched chunk"
)
after_context: Optional[str] = Field(
after_context: str | None = Field(
default=None, description="Text after the matched chunk"
)
has_before_truncation: Optional[bool] = Field(
has_before_truncation: bool | None = Field(
default=None, description="Whether before_context was truncated"
)
has_after_truncation: Optional[bool] = Field(
has_after_truncation: bool | None = Field(
default=None, description="Whether after_context was truncated"
)
@@ -66,7 +72,7 @@ class SemanticSearchResult(BaseModel):
class SemanticSearchResponse(BaseResponse):
"""Response model for semantic search across all indexed Nextcloud apps."""
results: List[SemanticSearchResult] = Field(
results: list[SemanticSearchResult] = Field(
description="Semantic search results with similarity scores"
)
query: str = Field(description="The search query used")
@@ -74,6 +80,30 @@ class SemanticSearchResponse(BaseResponse):
search_method: str = Field(
default="semantic", description="Search method used (semantic or hybrid)"
)
verified_chunk_count: int = Field(
default=0,
description=(
"Number of search result chunks that passed verify-on-read "
"access checks (ADR-019). Equals len(verified_results) before "
"trimming to limit. Sized in chunks (result rows), NOT in "
"unique documents — see dropped_document_count for the "
"per-document counterpart."
),
)
dropped_document_count: int = Field(
default=0,
description=(
"Number of unique (doc_id, doc_type) pairs dropped as ghost "
"records during verify-on-read (ADR-019). A short result page "
"(len(results) < limit) combined with a non-zero "
"dropped_document_count indicates ghost density rather than "
"scarcity of relevant content. Note: this counter is sized in "
"unique documents while verified_chunk_count is sized in "
"chunks — a single document can contribute multiple chunks, "
"so subtracting dropped_document_count from "
"verified_chunk_count is NOT a meaningful operation."
),
)
class SamplingSearchResponse(BaseResponse):
@@ -98,7 +128,7 @@ class SamplingSearchResponse(BaseResponse):
generated_answer: str = Field(
..., description="LLM-generated answer based on retrieved documents"
)
sources: List[SemanticSearchResult] = Field(
sources: list[SemanticSearchResult] = Field(
default_factory=list,
description="Source documents with excerpts and relevance scores",
)
@@ -106,10 +136,10 @@ class SamplingSearchResponse(BaseResponse):
search_method: str = Field(
default="semantic_sampling", description="Search method used"
)
model_used: Optional[str] = Field(
model_used: str | None = Field(
default=None, description="Model that generated the answer"
)
stop_reason: Optional[str] = Field(
stop_reason: str | None = Field(
default=None, description="Reason generation stopped"
)
nextcloud_mcp_server/search/algorithms.py
+9 -2
View File
@@ -67,6 +67,11 @@ class NextcloudClientProtocol(Protocol):
"""Tables client for accessing table row documents."""
...
@property
def news(self) -> Any:
"""News client for accessing news item documents."""
...
async def get_indexed_doc_types(user_id: str) -> set[str]:
"""Query Qdrant to get actually-indexed document types for a user.
@@ -127,7 +132,9 @@ class SearchResult:
"""A single search result with metadata and score.
Attributes:
id: Document ID (int for all document types)
id: Document ID. Numeric for indexed types today (notes, files,
deck cards, news items), but typed as ``int | str`` to allow
future doc types that use string identifiers (e.g., file paths).
doc_type: Document type (note, file, calendar, contact, etc.)
title: Document title
excerpt: Content excerpt showing match context
@@ -144,7 +151,7 @@ class SearchResult:
point_id: Qdrant point ID for batch vector retrieval (None if not from Qdrant)
"""
id: int
id: int | str
doc_type: str
title: str
excerpt: str
nextcloud_mcp_server/search/bm25_hybrid.py
+10 -3
View File
@@ -71,8 +71,10 @@ class BM25HybridSearchAlgorithm(SearchAlgorithm):
"""
Execute hybrid search using dense + sparse vectors with native RRF fusion.
Returns unverified results from Qdrant. Access verification should be
performed separately at the final output stage using verify_search_results().
Returns unverified results from Qdrant. Access verification is
performed separately at the server tool layer via
``nextcloud_mcp_server.search.verification.verify_search_results``
(see ADR-019).
Deduplicates by (doc_id, doc_type, chunk_start_offset, chunk_end_offset)
to show multiple chunks from the same document while avoiding duplicate chunks.
@@ -206,7 +208,7 @@ class BM25HybridSearchAlgorithm(SearchAlgorithm):
for result in search_response.points:
if result.payload is None:
continue
# doc_id can be int (notes) or str (files - file paths)
# doc_id can be int (files) or str (notes/news_items/deck_cards) — see scanner.py
doc_id = result.payload["doc_id"]
doc_type = result.payload.get("doc_type", "note")
chunk_start = result.payload.get("chunk_start_offset")
@@ -231,9 +233,14 @@ class BM25HybridSearchAlgorithm(SearchAlgorithm):
metadata["path"] = path
# Add deck_card-specific metadata for frontend URL construction
# and verify-on-read (ADR-019) — both board_id and stack_id are
# required to call deck.get_card without an O(boards × stacks)
# iteration fallback.
if doc_type == "deck_card":
if board_id := result.payload.get("board_id"):
metadata["board_id"] = board_id
if stack_id := result.payload.get("stack_id"):
metadata["stack_id"] = stack_id
# Return unverified results (verification happens at output stage)
results.append(
nextcloud_mcp_server/search/semantic.py
+9 -2
View File
@@ -48,8 +48,10 @@ class SemanticSearchAlgorithm(SearchAlgorithm):
) -> list[SearchResult]:
"""Execute semantic search using vector similarity.
Returns unverified results from Qdrant. Access verification should be
performed separately at the final output stage using verify_search_results().
Returns unverified results from Qdrant. Access verification is
performed separately at the server tool layer via
``nextcloud_mcp_server.search.verification.verify_search_results``
(see ADR-019).
Deduplicates by (doc_id, doc_type, chunk_start_offset, chunk_end_offset)
to show multiple chunks from the same document while avoiding duplicate chunks.
@@ -162,9 +164,14 @@ class SemanticSearchAlgorithm(SearchAlgorithm):
metadata["path"] = path
# Add deck_card-specific metadata for frontend URL construction
# and verify-on-read (ADR-019) — both board_id and stack_id are
# required to call deck.get_card without an O(boards × stacks)
# iteration fallback.
if doc_type == "deck_card":
if board_id := result.payload.get("board_id"):
metadata["board_id"] = board_id
if stack_id := result.payload.get("stack_id"):
metadata["stack_id"] = stack_id
# Return unverified results (verification happens at output stage)
results.append(
nextcloud_mcp_server/search/verification.py
+592
View File
@@ -0,0 +1,592 @@
"""Verify-on-read access checks for semantic search results (ADR-019).
The vector index is a recall layer; Nextcloud is the source of truth for
access. This module filters search results by checking each unique document
against Nextcloud at query time, dropping any that the user can no longer
access (deleted, unshared, etc.) and lazily evicting them from the index.
Per-doc_type verifiers are registered in ``_VERIFIERS``. Each takes the
authenticated client, the (deduplicated) list of ``SearchResult``s for that
doc_type, and a shared concurrency semaphore. They return the subset of
``doc_id`` values that are currently accessible. Verifiers read whatever
metadata they need (file path, deck card board/stack ids) directly from the
SearchResult — these fields are populated at index-time and propagated by
the algorithm layer (see ``search/bm25_hybrid.py`` and ``search/semantic.py``)
so verification adds zero extra Qdrant round-trips.
Concurrency is bounded by a shared semaphore (default 20) so a large search
result page (or a multi-doc_type query) cannot exhaust the httpx connection
pool or trigger Nextcloud rate limiting. The 20-slot default matches the
context-expansion convention in ``server/semantic.py``.
Failure policy:
- Definitive 403/404 from Nextcloud → drop the result and schedule eviction.
- Transient errors (5xx, network blips, unexpected exceptions) → keep the
result and log a warning. We never silently shrink result sets due to
flakes; the next query will re-verify.
- Unsupported doc_type (no registered verifier) → keep the result and log a
warning. Verification is opt-in per type; a missing verifier is a soft
failure, not a search failure.
"""
import logging
from collections.abc import Awaitable, Callable
import anyio
from anyio.abc import TaskGroup
from httpx import HTTPStatusError
from nextcloud_mcp_server.config import get_settings
from nextcloud_mcp_server.search.algorithms import (
NextcloudClientProtocol,
SearchResult,
)
from nextcloud_mcp_server.vector.eviction import delete_document_points
logger = logging.getLogger(__name__)
BatchVerifier = Callable[
[NextcloudClientProtocol, list[SearchResult], anyio.Semaphore],
Awaitable[set[int | str]],
]
"""(client, results, semaphore) -> set of doc_ids accessible to the user."""
# ---------------------------------------------------------------------------
# Per-doc-type verifiers
# ---------------------------------------------------------------------------
def _is_definitive_404_or_403(exc: BaseException) -> bool:
"""Return True if exc indicates the document is definitively inaccessible.
401 is intentionally excluded — it usually signals expired credentials
rather than permanent denial, so it is treated as transient (keep the
result; the next query will re-verify after the client refreshes).
"""
if isinstance(exc, HTTPStatusError):
return exc.response.status_code in (403, 404)
return False
async def _verify_notes(
client: NextcloudClientProtocol,
results: list[SearchResult],
semaphore: anyio.Semaphore,
) -> set[int | str]:
# safe: cooperative concurrency, no lock needed (see verify_search_results)
accessible: set[int | str] = set()
async def check(result: SearchResult) -> None:
doc_id = result.id
# Parse defensively before the network call so a malformed payload
# produces a specific log line, not a generic "unexpected error"
# from the catch-all ``except Exception`` below. ``_verify_notes``
# is the canonical shape; ``_verify_deck_cards`` and
# ``_verify_news_items`` mirror this hoisted-cast pattern.
try:
note_id_int = int(doc_id)
except (TypeError, ValueError) as e:
logger.warning(
"Non-numeric note id %r: %s; keeping result",
doc_id,
e,
)
accessible.add(doc_id)
return
async with semaphore:
try:
await client.notes.get_note(note_id_int)
accessible.add(doc_id)
except HTTPStatusError as e:
if _is_definitive_404_or_403(e):
return
logger.warning(
"Transient error verifying note %s: %s %s; keeping result",
doc_id,
e.response.status_code,
e,
)
accessible.add(doc_id)
except Exception as e:
logger.warning(
"Unexpected error verifying note %s: %s; keeping result",
doc_id,
e,
)
accessible.add(doc_id)
async with anyio.create_task_group() as tg:
for r in results:
tg.start_soon(check, r)
return accessible
async def _verify_files(
client: NextcloudClientProtocol,
results: list[SearchResult],
semaphore: anyio.Semaphore,
) -> set[int | str]:
# safe: cooperative concurrency, no lock needed (see verify_search_results)
accessible: set[int | str] = set()
async def check(result: SearchResult) -> None:
doc_id = result.id
# file_path is propagated from the Qdrant payload by the algorithm
# layer (bm25_hybrid.py / semantic.py). No extra Qdrant round-trip.
file_path = (result.metadata or {}).get("path")
if not file_path:
# Cannot verify without a path; treat as accessible to avoid
# silently dropping legitimate results when payload is missing
# (legacy data, or a future doc_type that doesn't propagate path).
logger.warning(
"No file path in metadata for file_id %s; keeping result "
"(verification skipped)",
doc_id,
)
accessible.add(doc_id)
return
async with semaphore:
try:
info = await client.webdav.get_file_info(file_path)
if info is None:
# Contract (see WebDAVClient.get_file_info docstring):
# `None` means a malformed PROPFIND response — an
# ambiguous state, not a definitive 404. Treat as
# transient and KEEP the result rather than evicting.
# Real 404s raise HTTPStatusError and land in the
# _is_definitive_404_or_403 branch below.
logger.warning(
"Malformed PROPFIND response verifying file %s (%s); "
"keeping result (ambiguous state, not a definitive 404)",
doc_id,
file_path,
)
accessible.add(doc_id)
return
accessible.add(doc_id)
except HTTPStatusError as e:
if _is_definitive_404_or_403(e):
return
logger.warning(
"Transient error verifying file %s (%s): %s %s; keeping result",
doc_id,
file_path,
e.response.status_code,
e,
)
accessible.add(doc_id)
except Exception as e:
logger.warning(
"Unexpected error verifying file %s (%s): %s; keeping result",
doc_id,
file_path,
e,
)
accessible.add(doc_id)
async with anyio.create_task_group() as tg:
for r in results:
tg.start_soon(check, r)
return accessible
async def _verify_deck_cards(
client: NextcloudClientProtocol,
results: list[SearchResult],
semaphore: anyio.Semaphore,
) -> set[int | str]:
# safe: cooperative concurrency, no lock needed (see verify_search_results)
accessible: set[int | str] = set()
async def check(result: SearchResult) -> None:
doc_id = result.id
# board_id and stack_id are propagated from the Qdrant payload by the
# algorithm layer. No extra Qdrant round-trip.
meta = result.metadata or {}
board_id = meta.get("board_id")
stack_id = meta.get("stack_id")
if board_id is None or stack_id is None:
# Without metadata we cannot run the cheap fast-path. Per ADR-019
# we deliberately do NOT fall back to O(boards × stacks) iteration
# in the search hot path; treat as accessible.
logger.warning(
"Incomplete deck metadata for card %s (board_id=%s, stack_id=%s); "
"keeping result (verification skipped, legacy data)",
doc_id,
board_id,
stack_id,
)
accessible.add(doc_id)
return
# Parse defensively before the network call so a malformed payload
# produces a specific log line, not a generic "unexpected error"
# from the catch-all ``except Exception`` below. Mirrors the
# canonical hoisted-cast pattern in ``_verify_notes``.
try:
board_id_int = int(board_id)
stack_id_int = int(stack_id)
card_id_int = int(doc_id)
except (TypeError, ValueError) as e:
logger.warning(
"Non-numeric deck metadata for card %s "
"(board_id=%r, stack_id=%r): %s; keeping result",
doc_id,
board_id,
stack_id,
e,
)
accessible.add(doc_id)
return
async with semaphore:
try:
await client.deck.get_card(
board_id=board_id_int,
stack_id=stack_id_int,
card_id=card_id_int,
)
accessible.add(doc_id)
except HTTPStatusError as e:
if _is_definitive_404_or_403(e):
return
logger.warning(
"Transient error verifying deck card %s: %s %s; keeping result",
doc_id,
e.response.status_code,
e,
)
accessible.add(doc_id)
except Exception as e:
logger.warning(
"Unexpected error verifying deck card %s: %s; keeping result",
doc_id,
e,
)
accessible.add(doc_id)
async with anyio.create_task_group() as tg:
for r in results:
tg.start_soon(check, r)
return accessible
async def _verify_news_items(
client: NextcloudClientProtocol,
results: list[SearchResult],
semaphore: anyio.Semaphore,
) -> set[int | str]:
"""Batch-verify news items with a single fetch.
The Nextcloud News API has no per-item endpoint, so ``news.get_item`` is
implemented as a fetch-all + filter — which would be O(N × all_items) if
called per id. Instead we fetch once and intersect. Only one slot of
the shared semaphore is consumed per search (rather than one per id),
but that slot is held for the full ``get_items`` round-trip; see the
in-body comment for the backpressure rationale.
"""
doc_ids = [r.id for r in results]
# Semaphore lifetime: this slot is held for the duration of ONE
# deduplicated News fetch (≤1 per search), not per-id. That is the
# correct backpressure behaviour — a single user's news verification
# must not hammer Nextcloud with concurrent fetch-all requests, and
# any other verifiers running in parallel for the same search share
# the same semaphore. Latency of this fetch is proportional to the
# user's full news corpus; see the News caveat in
# docs/configuration.md for production guidance.
#
# Multi-user note: the "≤1 per search" bound is per-search, not
# per-process. If N users simultaneously search news content, all N
# hold a slot for the duration of their respective fetches, each
# consuming 1/max_concurrent of the shared verification budget. A
# single news-heavy user can therefore hold their slot for seconds.
async with semaphore:
try:
# TODO(perf): if profiling shows this fetch dominates query latency
# for news-heavy users, cache the per-request item set or push for
# a per-item News API endpoint. The shared semaphore protects
# against runaway concurrent fetches, but the payload itself can
# be large (News auto-purge cap is in the thousands of items).
#
# NOTE: ``batch_size`` is intentionally unbounded (-1). A numeric
# ceiling here would silently *break correctness*: any item beyond
# the cap would be missing from ``present_ids`` and incorrectly
# dropped from the result set. The fail-open contract requires
# fetching every item the user has access to. See the news caveat
# in docs/configuration.md (Verify-on-Read) for the latency
# tradeoff and follow-up paths.
news_fetch_start = anyio.current_time()
items = await client.news.get_items(batch_size=-1, get_read=True)
logger.debug(
"News fetch for verification took %.2fs (%d item(s) returned)",
anyio.current_time() - news_fetch_start,
len(items),
)
except HTTPStatusError as e:
# If the News API itself is gone (app disabled, user lost access),
# treat *all* requested items as inaccessible. Eviction will reclaim.
if _is_definitive_404_or_403(e):
# News app commonly disabled/uninstalled — debug-level keeps
# this off operator dashboards; transient errors below stay
# at warning because they're unexpected.
logger.debug(
"News API returned %s for user %s; treating all %d news_items as inaccessible",
e.response.status_code,
client.username,
len(doc_ids),
)
return set()
logger.warning(
"Transient error fetching news items for verification: %s %s; keeping all results",
e.response.status_code,
e,
)
return set(doc_ids)
except Exception as e:
logger.warning(
"Unexpected error fetching news items for verification: %s; keeping all results",
e,
)
return set(doc_ids)
# Build present_ids from the API response. Granularity is intentionally
# asymmetric with the per-item loop below:
#
# * Here (structural failure): if the API response itself is corrupt
# — even one item with a non-numeric id — we cannot reliably build
# `present_ids`, so every requested doc_id fails open. The batch
# is the only safe blast radius when the source-of-truth payload
# can't be trusted.
# * Below (data failure): a single non-numeric *stored* doc_id is a
# local data issue. Failing the whole batch open would let one bad
# row in Qdrant mask real revocations for every other item, so we
# scope the fail-open to that one id.
try:
present_ids = {
int(item.get("id")) for item in items if item.get("id") is not None
}
except (TypeError, ValueError) as e:
logger.warning(
"Non-numeric id in news API response (sample=%r): %s; keeping all results",
items[:3] if items else items,
e,
)
return set(doc_ids)
# Per-item check: a single non-numeric *stored* doc_id is fail-open
# for THAT item only — not the whole batch. Mirrors the per-item
# shape of the notes/files/deck verifiers. See the granularity note
# above for why this is narrower than the API-response failure path.
accessible: set[int | str] = set()
for d in doc_ids:
try:
if int(d) in present_ids:
accessible.add(d)
except (TypeError, ValueError):
logger.debug("Non-numeric news doc_id %r; keeping (cannot verify)", d)
accessible.add(d)
return accessible
_VERIFIERS: dict[str, BatchVerifier] = {
"note": _verify_notes,
"file": _verify_files,
"deck_card": _verify_deck_cards,
"news_item": _verify_news_items,
}
def get_supported_doc_types() -> set[str]:
"""Return the set of doc_types that have registered verifiers.
Used by CI guards and tests to ensure every indexed doc_type has a
verifier (see ADR-019 implementation checklist).
"""
return set(_VERIFIERS.keys())
# ---------------------------------------------------------------------------
# Public entry point
# ---------------------------------------------------------------------------
async def verify_search_results(
client: NextcloudClientProtocol,
results: list[SearchResult],
*,
evict_on_missing: bool = True,
max_concurrent: int | None = None,
eviction_task_group: TaskGroup | None = None,
) -> tuple[list[SearchResult], int]:
"""Filter search results to those the user can currently access.
Deduplicates by ``(doc_id, doc_type)`` before verifying, so multiple
chunks from the same document cost a single check. Verifiers run
concurrently per doc_type and concurrently per id within each verifier,
bounded by a shared semaphore (``max_concurrent``).
When ``evict_on_missing=True``, points for documents that fail verification
are deleted from Qdrant. If ``eviction_task_group`` is provided (the
lifespan-owned task group from ``app.py::VectorSyncState``), eviction is
fire-and-forget — the search response returns immediately and Qdrant
deletes happen in the background. If no task group is provided (unit
tests, modes without vector sync), eviction falls back to running inline
in a local task group. Eviction failures are logged but never propagated.
Args:
client: Authenticated NextcloudClient (must expose ``username``).
results: SearchResult list from the algorithm layer (may include
multiple chunks per document).
evict_on_missing: Schedule lazy eviction for inaccessible docs.
max_concurrent: Cap on concurrent verification round-trips against
Nextcloud. When ``None`` (the default), resolved from
``Settings.verification_concurrency`` (env var
``VERIFICATION_CONCURRENCY``, default 20).
eviction_task_group: Optional long-lived task group on which to
spawn fire-and-forget eviction. Pass
``ctx.request_context.lifespan_context.eviction_task_group``
from FastMCP tools.
Returns:
Tuple of ``(kept_results, dropped_count)`` where ``kept_results`` is
the filtered list preserving the original order and ``dropped_count``
is the number of unique ``(doc_id, doc_type)`` pairs that failed
verification (ghost records).
"""
if not results:
return results, 0
user_id: str = client.username
if max_concurrent is None:
max_concurrent = get_settings().verification_concurrency
# Group unique (doc_id, doc_type) by doc_type so each verifier sees a
# deduplicated batch. We pick one SearchResult per (id, doc_type) to carry
# metadata (path, board_id/stack_id) into the verifier — chunks of the
# same document share these fields, so any chunk works.
by_type: dict[str, dict[int | str, SearchResult]] = {}
for r in results:
by_type.setdefault(r.doc_type, {}).setdefault(r.id, r)
# Shared semaphore bounds total Nextcloud round-trips across all
# per-id verifiers. Without it, a 50-result mostly-notes page could fan
# out 50 concurrent get_note calls and exhaust the connection pool.
semaphore = anyio.Semaphore(max_concurrent)
# Concurrency note: ``accessible_by_type`` is mutated by multiple
# ``run_verifier`` tasks running under the task group below. This is
# safe without an explicit lock because (a) anyio uses cooperative
# multitasking — a task only yields at ``await`` points, never
# mid-statement; (b) each task is dispatched once per ``doc_type``
# by the loop ``for doc_type, ... in by_type.items()`` further down,
# so two tasks never write to the same key; and (c) Python dict key
# assignment is not an await point, so two tasks cannot race on the
# same write. Adding a lock would be dead weight; using ``anyio.Lock``
# here would force serialization on a path that is intentionally
# parallel.
accessible_by_type: dict[str, set[int | str]] = {}
async def run_verifier(doc_type: str, unique_results: list[SearchResult]) -> None:
verifier = _VERIFIERS.get(doc_type)
if verifier is None:
logger.warning(
"No verifier registered for doc_type=%r; keeping %d result(s) unverified",
doc_type,
len(unique_results),
)
accessible_by_type[doc_type] = {r.id for r in unique_results}
return
try:
accessible_by_type[doc_type] = await verifier(
client, unique_results, semaphore
)
except Exception as e:
# Verifier itself blew up (not per-id) — fail open.
logger.error(
"Verifier for doc_type=%s raised: %s; keeping all %d result(s) unverified",
doc_type,
e,
len(unique_results),
exc_info=True,
)
accessible_by_type[doc_type] = {r.id for r in unique_results}
async with anyio.create_task_group() as tg:
for doc_type, id_to_result in by_type.items():
tg.start_soon(run_verifier, doc_type, list(id_to_result.values()))
# Compute (doc_id, doc_type) pairs that failed verification
inaccessible: set[tuple[int | str, str]] = set()
for doc_type, id_to_result in by_type.items():
# The .get() default is defensive only — run_verifier always populates
# accessible_by_type[doc_type], either with the verifier's result or
# with all ids on verifier crash (fail-open).
accessible = accessible_by_type.get(doc_type, set(id_to_result.keys()))
for doc_id in id_to_result.keys():
if doc_id not in accessible:
inaccessible.add((doc_id, doc_type))
if inaccessible:
# Tag ids with their type (int vs str) so ghost-record logs are
# unambiguous: int 42 and str "42" both render as "42" otherwise.
logger.info(
"Verification dropped %d inaccessible document(s): %s",
len(inaccessible),
sorted((f"{type(d).__name__}:{d}", t) for d, t in inaccessible),
)
# Filter results, preserving order. All chunks of an inaccessible document
# are dropped together (dedup happened before verification, but the result
# list still contains all chunks).
kept = [r for r in results if (r.id, r.doc_type) not in inaccessible]
# Lazy eviction.
#
# Preferred path: spawn evict() on the lifespan-owned task group via
# `start_soon`, which returns immediately — the search response is not
# blocked on Qdrant deletes. If the server is shutting down, the task
# group is cleared back to None (see app.py) and we fall through to the
# inline path. Cancellation mid-eviction is fine: the next query will
# re-verify and re-attempt (self-healing per ADR-019).
#
# Fallback path: when no task group is supplied (unit tests, deployment
# modes without vector sync), run eviction inline in a local task group.
# This preserves prior behaviour for tests that rely on eviction being
# complete by the time `verify_search_results` returns.
if evict_on_missing and inaccessible:
async def evict(doc_id: int | str, doc_type: str) -> None:
try:
await delete_document_points(doc_id, doc_type, user_id)
except Exception as e:
logger.warning(
"Failed to evict %s_%s from Qdrant: %s", doc_type, doc_id, e
)
if eviction_task_group is not None:
for doc_id, doc_type in inaccessible:
# Guard against the lifespan task group having exited between
# the getattr() capture in server/semantic.py and this call —
# start_soon raises RuntimeError on a closed group, which
# would otherwise surface as a search error. Eviction is
# best-effort: the next query re-verifies and re-attempts.
try:
eviction_task_group.start_soon(evict, doc_id, doc_type)
except RuntimeError:
logger.debug("Eviction task group closed; will retry on next query")
else:
async with anyio.create_task_group() as tg:
for doc_id, doc_type in inaccessible:
tg.start_soon(evict, doc_id, doc_type)
return kept, len(inaccessible)
nextcloud_mcp_server/server/semantic.py
+225 -76
View File
@@ -32,6 +32,7 @@ from nextcloud_mcp_server.observability.metrics import (
)
from nextcloud_mcp_server.search.bm25_hybrid import BM25HybridSearchAlgorithm
from nextcloud_mcp_server.search.context import get_chunk_with_context
from nextcloud_mcp_server.search.verification import verify_search_results
from nextcloud_mcp_server.vector.placeholder import get_placeholder_filter
from nextcloud_mcp_server.vector.qdrant_client import get_qdrant_client
@@ -86,15 +87,29 @@ def configure_semantic_tools(mcp: FastMCP):
context_chars: Number of characters to include before/after matched chunk (default: 300)
Returns:
SemanticSearchResponse with matching documents ranked by fusion scores
SemanticSearchResponse with matching documents ranked by fusion scores.
Verification fields (ADR-019 verify-on-read):
- verified_chunk_count: chunk rows that passed access checks
(sized in chunks; counted before trimming to ``limit``, so it
can exceed ``len(results)`` when a doc has multiple matching
chunks).
- dropped_document_count: unique ``(doc_id, doc_type)`` pairs
evicted as ghost records during this search (sized in
documents, not chunks).
"""
settings = get_settings()
client = await get_client(ctx)
username = client.username
logger.info(
f"BM25 hybrid search: query='{query}', user={username}, "
f"limit={limit}, score_threshold={score_threshold}, fusion={fusion}"
"BM25 hybrid search: query=%r, user=%s, "
"limit=%d, score_threshold=%s, fusion=%s",
query,
username,
limit,
score_threshold,
fusion,
)
# Check that vector sync is enabled
@@ -119,48 +134,122 @@ def configure_semantic_tools(mcp: FastMCP):
if doc_types is None:
# Cross-app search: search all indexed types
# Get unverified results from Qdrant
# Get unverified results from Qdrant.
#
# NOTE (ADR-019): Over-fetch by 2× to absorb ghost-record drops
# during verify-on-read. When ghost density is high (e.g. a
# large board share was just revoked) this budget can still
# under-deliver against the requested ``limit``; the index
# self-heals via lazy eviction so subsequent searches recover.
# The 2× factor is a deliberate v1 trade-off — raising it
# costs Nextcloud round-trips on every search. Trim to
# ``limit`` happens AFTER verification.
# TODO(ADR-019): expose VERIFICATION_OVERFETCH so operators
# with persistent high ghost density can tune this without a
# code change.
unverified_results = await search_algo.search(
query=query,
user_id=username,
limit=limit * 2, # Get extra for access filtering
limit=limit * 2,
doc_type=None, # Signal to search all types
score_threshold=score_threshold,
)
all_results.extend(unverified_results)
else:
# Search specific document types
# For each requested type, execute search and combine results
# Search specific document types.
#
# Per-Qdrant-query cost: this branch issues ONE query per
# requested doc_type, each capped at `limit * 2`. With N
# types in `doc_types`, the pre-merge result pool is
# therefore N × `limit * 2`, NOT `limit * 2`. That is more
# Qdrant work than the cross-app branch above (which makes a
# single multi-type query returning `limit * 2` total).
#
# The post-merge trim below clamps the pool back down to
# `limit * 2` so verification (and the Nextcloud round-trips
# it triggers) sees the same budget as the cross-app branch.
# The per-type Qdrant cost remains higher; pre-trim cost
# scales linearly with len(doc_types).
for dtype in doc_types:
unverified_results = await search_algo.search(
query=query,
user_id=username,
limit=limit * 2, # Get extra for combining and filtering
limit=limit * 2,
doc_type=dtype,
score_threshold=score_threshold,
)
all_results.extend(unverified_results)
# Sort combined results by score
# Sort combined results by score, then cap to `limit * 2` to
# match the cross-app branch's over-fetch budget. Without this
# cap, N requested doc_types × `limit * 2` results would all
# flow into verification, multiplying the Nextcloud round-trip
# cost by N.
all_results.sort(key=lambda r: r.score, reverse=True)
all_results = all_results[: limit * 2]
# Note: BM25HybridSearchAlgorithm already deduplicates at chunk level
# (doc_id, doc_type, chunk_start, chunk_end), which allows multiple
# chunks from the same document while preventing duplicate chunks.
# No additional deduplication needed here - multiple chunks per document
# are valuable for RAG contexts.
# Qdrant already filters by user_id for multi-tenant isolation.
# Sampling tool will verify access when fetching full content.
search_results = all_results[
:limit
] # Final limit after chunk-level dedup in algorithm
# ADR-019: Verify-on-read. The vector index is a recall layer;
# Nextcloud is the source of truth for access. Filter out ghost
# records (deleted/unshared docs not yet reconciled by webhooks)
# BEFORE trimming to `limit`, so we don't lose accessible results
# to the limit slot that ghosts would otherwise occupy. We also
# run this BEFORE context expansion to avoid re-fetching docs that
# are about to be dropped. Pass the lifespan-owned task group so
# eviction of dropped points is fire-and-forget (does not block
# the response).
# Direct attribute access — both AppContext and OAuthAppContext
# expose ``eviction_task_group`` as a @property (see app.py),
# reading dynamically from the module-level VectorSyncState
# singleton. A defensive ``getattr(..., None)`` here would mask
# typos; if a future lifespan-context type forgets the property,
# AttributeError surfaces during the first search rather than
# silently degrading to inline eviction for the life of the
# process.
eviction_task_group = (
ctx.request_context.lifespan_context.eviction_task_group
)
verification_start = anyio.current_time()
verified_results, dropped_count = await verify_search_results(
client,
all_results,
eviction_task_group=eviction_task_group,
)
verified_chunk_count = len(verified_results)
logger.debug(
"Verification completed in %.2fs: kept %d chunk(s), dropped %d doc(s)",
anyio.current_time() - verification_start,
verified_chunk_count,
dropped_count,
)
search_results = verified_results[:limit]
# Convert SearchResult objects to SemanticSearchResult for response
# Convert SearchResult objects to SemanticSearchResult for response.
# SearchResult.id is typed `int | str` for forward-compat with future
# doc_types, but every currently indexed type uses numeric ids and
# the MCP response model narrows to `int`. Casting here makes the
# narrowing explicit and surfaces any future string-id type as a
# loud failure at the boundary instead of silently widening the
# public API.
results = []
for r in search_results:
try:
narrowed_id = int(r.id)
except (TypeError, ValueError) as e:
# Re-raise with explicit context so the outer handler logs
# something operators can act on (the generic "Search
# failed: invalid literal for int()" is opaque).
raise TypeError(
f"SemanticSearchResult.id must be int-convertible, "
f"got {r.id!r} (type={type(r.id).__name__}) for "
f"doc_type={r.doc_type!r}. This indicates a doc_type "
f"with non-numeric ids has been indexed but the "
f"public response model has not been widened. Add "
f"the doc_type to the SemanticSearchResult.id type "
f"or convert at the verifier layer."
) from e
results.append(
SemanticSearchResult(
id=r.id,
id=narrowed_id,
doc_type=r.doc_type,
title=r.title,
category=r.metadata.get("category", "") if r.metadata else "",
@@ -181,12 +270,20 @@ def configure_semantic_tools(mcp: FastMCP):
# Expand results with surrounding context if requested
if include_context and results:
logger.info(
f"Expanding {len(results)} results with context "
f"(context_chars={context_chars})"
"Expanding %d results with context (context_chars=%d)",
len(results),
context_chars,
)
# Fetch context for all results in parallel
# Limit concurrent requests to prevent connection pool exhaustion
# Fetch context for all results in parallel.
# Limit concurrent requests to prevent connection pool exhaustion.
#
# Intentionally distinct from settings.verification_concurrency:
# that knob bounds Nextcloud round-trips during access
# verification (ADR-019); this one bounds context-expansion
# fetches that run only when ``include_context=True``. Operators
# tuning one rarely want the other in lockstep, so they share
# the default value (20) but not the env var.
max_concurrent = 20
semaphore = anyio.Semaphore(max_concurrent)
expanded_results = [None] * len(results)
@@ -240,20 +337,27 @@ def configure_semantic_tools(mcp: FastMCP):
has_after_truncation=chunk_context.has_after_truncation,
)
logger.debug(
f"Expanded context for {result.doc_type} {result.id}"
"Expanded context for %s %s",
result.doc_type,
result.id,
)
else:
# Context expansion failed, keep original result
expanded_results[index] = result
logger.debug(
f"Failed to expand context for {result.doc_type} {result.id}, "
"keeping original result"
"Failed to expand context for %s %s, "
"keeping original result",
result.doc_type,
result.id,
)
except Exception as e:
# Context expansion failed, keep original result
expanded_results[index] = result
logger.warning(
f"Error expanding context for {result.doc_type} {result.id}: {e}"
"Error expanding context for %s %s: %s",
result.doc_type,
result.id,
e,
)
# Run all context fetches in parallel using anyio task group
@@ -264,16 +368,19 @@ def configure_semantic_tools(mcp: FastMCP):
# Replace results with expanded versions
results = [r for r in expanded_results if r is not None]
logger.info(
f"Context expansion completed: {len(results)} results with context"
"Context expansion completed: %d results with context",
len(results),
)
logger.info(f"Returning {len(results)} results from BM25 hybrid search")
logger.info("Returning %d results from BM25 hybrid search", len(results))
return SemanticSearchResponse(
results=results,
query=query,
total_found=len(results),
search_method=f"bm25_hybrid_{fusion}",
verified_chunk_count=verified_chunk_count,
dropped_document_count=dropped_count,
)
except ValueError as e:
@@ -293,14 +400,14 @@ def configure_semantic_tools(mcp: FastMCP):
ErrorData(code=-1, message=f"Network error during search: {str(e)}")
)
except Exception as e:
logger.error(f"Search error: {e}", exc_info=True)
logger.error("Search error: %s", e, exc_info=True)
raise McpError(ErrorData(code=-1, message=f"Search failed: {str(e)}"))
@mcp.tool(
title="Search with AI-Generated Answer",
annotations=ToolAnnotations(
readOnlyHint=True, # Search doesn't modify data
openWorldHint=False, # Searches only indexed Nextcloud data
openWorldHint=True, # Calls into Nextcloud via nc_semantic_search
),
)
@require_scopes("semantic.read")
@@ -352,6 +459,14 @@ def configure_semantic_tools(mcp: FastMCP):
Note: Requires MCP client to support sampling. If sampling is unavailable,
the tool gracefully degrades to returning documents with an explanation.
The client may prompt the user to approve the sampling request.
Latency profile: For each note in the result page, this tool fetches
the full note body via ``client.notes.get_note`` after upstream
verify-on-read has already round-tripped to the same endpoint as a
race guard (ADR-019). Expect one additional Nextcloud round-trip per
note result; raising ``limit`` above the default of 5 amplifies this
cost roughly linearly. File / news / deck results do not pay this
cost — they reuse the verified excerpt.
"""
# 1. Retrieve relevant documents via existing semantic search
search_response = await nc_semantic_search(
@@ -366,7 +481,7 @@ def configure_semantic_tools(mcp: FastMCP):
# 2. Handle no results case - don't waste a sampling call
if not search_response.results:
logger.debug(f"No documents found for query: {query}")
logger.debug("No documents found for query: %r", query)
return SamplingSearchResponse(
query=query,
generated_answer="No relevant documents found in your Nextcloud content for this query.",
@@ -383,22 +498,25 @@ def configure_semantic_tools(mcp: FastMCP):
# Log capability check result for debugging
logger.info(
f"Sampling capability check: client_has_sampling={client_has_sampling}, "
f"query='{query}'"
"Sampling capability check: client_has_sampling=%s, query=%r",
client_has_sampling,
query,
)
if hasattr(ctx.session, "_client_params") and ctx.session._client_params:
client_caps = ctx.session._client_params.capabilities
logger.debug(
f"Client advertised capabilities: "
f"roots={client_caps.roots is not None}, "
f"sampling={client_caps.sampling is not None}, "
f"experimental={client_caps.experimental is not None}"
"Client advertised capabilities: "
"roots=%s, sampling=%s, experimental=%s",
client_caps.roots is not None,
client_caps.sampling is not None,
client_caps.experimental is not None,
)
if not client_has_sampling:
logger.info(
f"Client does not support sampling (query: '{query}'), "
f"returning {len(search_response.results)} documents"
"Client does not support sampling (query: %r), returning %d documents",
query,
len(search_response.results),
)
return SamplingSearchResponse(
query=query,
@@ -414,14 +532,24 @@ def configure_semantic_tools(mcp: FastMCP):
success=True,
)
# 4. Fetch full content for notes in parallel (also verifies access)
# Use anyio task group for concurrent fetching with semaphore to prevent
# connection pool exhaustion
# 4. Fetch full content for notes in parallel.
# Access verification has already happened upstream in
# nc_semantic_search via verify_search_results (ADR-019), so any
# exception here is a sub-second race (doc deleted between
# verification and this fetch) — drop the result in that case.
client = await get_client(ctx)
accessible_results = [None] * len(search_response.results)
full_contents = [None] * len(search_response.results)
# Limit concurrent requests to prevent connection pool exhaustion
# Limit concurrent requests to prevent connection pool exhaustion.
#
# Intentionally distinct from settings.verification_concurrency:
# that knob bounds Nextcloud round-trips during access
# verification (ADR-019). This one bounds the answer tool's
# full-content fetch — a separate request phase tied to RAG
# answer generation. Operators tuning one rarely want the other
# in lockstep, so they share the default value (20) but not the
# env var.
max_concurrent = 20
semaphore = anyio.Semaphore(max_concurrent)
@@ -429,26 +557,30 @@ def configure_semantic_tools(mcp: FastMCP):
"""Fetch full content for a single document (parallel with semaphore)."""
async with semaphore:
if result.doc_type == "note":
# SemanticSearchResult.id is typed `int` (Pydantic enforces
# at construction); no defensive cast is needed here. The
# catch-all below covers only the verify-then-delete race.
try:
note = await client.notes.get_note(result.id)
# Note is accessible, store result and full content
content = note.get("content", "")
accessible_results[index] = result
full_contents[index] = content
logger.debug(
f"Fetched full content for note {result.id} "
f"(length: {len(content)} chars)"
"Fetched full content for note %s (length: %d chars)",
result.id,
len(content),
)
except Exception as e:
# Note might have been deleted or permissions changed
# Leave as None to filter out later
# Race window after verify_search_results — drop result.
logger.debug(
f"Note {result.id} not accessible: {e}. "
f"Excluding from results."
"Note %s disappeared between verification and "
"content fetch: %s. Excluding from results.",
result.id,
e,
)
else:
# Non-note document types (future: calendar, deck, files)
# For now, keep them with excerpts
# Non-note types (file, news_item, deck_card) keep the
# excerpt — already access-verified upstream.
accessible_results[index] = result
# full_contents[index] remains None (will use excerpt)
@@ -466,7 +598,9 @@ def configure_semantic_tools(mcp: FastMCP):
# Check if we filtered out all results
if not accessible_results:
logger.warning(f"All search results became inaccessible for query: {query}")
logger.warning(
"All search results became inaccessible for query: %r", query
)
return SamplingSearchResponse(
query=query,
generated_answer="All matching documents are no longer accessible.",
@@ -508,9 +642,12 @@ def configure_semantic_tools(mcp: FastMCP):
)
logger.info(
f"Initiating sampling request: query_length={len(query)}, "
f"documents={len(search_response.results)}, "
f"prompt_length={len(prompt)}, max_tokens={max_answer_tokens}"
"Initiating sampling request: query_length=%d, documents=%d, "
"prompt_length=%d, max_tokens=%d",
len(query),
len(search_response.results),
len(prompt),
max_answer_tokens,
)
# 6. Request LLM completion via MCP sampling with timeout
@@ -543,13 +680,15 @@ def configure_semantic_tools(mcp: FastMCP):
# Handle non-text responses (shouldn't happen for text prompts)
generated_answer = f"Received non-text response of type: {sampling_result.content.type}"
logger.warning(
f"Unexpected content type from sampling: {sampling_result.content.type}"
"Unexpected content type from sampling: %s",
sampling_result.content.type,
)
logger.info(
f"Sampling successful: model={sampling_result.model}, "
f"stop_reason={sampling_result.stopReason}, "
f"answer_length={len(generated_answer)}"
"Sampling successful: model=%s, stop_reason=%s, answer_length=%d",
sampling_result.model,
sampling_result.stopReason,
len(generated_answer),
)
return SamplingSearchResponse(
@@ -565,8 +704,10 @@ def configure_semantic_tools(mcp: FastMCP):
except TimeoutError:
logger.warning(
f"Sampling request timed out after {sampling_timeout_seconds} seconds for query: '{query}', "
f"returning search results only"
"Sampling request timed out after %d seconds for query: %r, "
"returning search results only",
sampling_timeout_seconds,
query,
)
return SamplingSearchResponse(
query=query,
@@ -588,18 +729,20 @@ def configure_semantic_tools(mcp: FastMCP):
if "rejected" in error_msg.lower() or "denied" in error_msg.lower():
# User explicitly declined - this is normal, not an error
logger.info(f"User declined sampling request for query: '{query}'")
logger.info("User declined sampling request for query: %r", query)
search_method = "semantic_sampling_user_declined"
user_message = "User declined to generate an answer"
elif "not supported" in error_msg.lower():
# Client doesn't support sampling - also normal
logger.info(f"Sampling not supported by client for query: '{query}'")
logger.info("Sampling not supported by client for query: %r", query)
search_method = "semantic_sampling_unsupported"
user_message = "Sampling not supported by this client"
else:
# Other MCP protocol errors
logger.warning(
f"MCP error during sampling for query '{query}': {error_msg}"
"MCP error during sampling for query %r: %s",
query,
error_msg,
)
search_method = "semantic_sampling_mcp_error"
user_message = f"Sampling unavailable: {error_msg}"
@@ -620,8 +763,10 @@ def configure_semantic_tools(mcp: FastMCP):
except Exception as e:
# Truly unexpected errors - these SHOULD have tracebacks
logger.error(
f"Unexpected error during sampling for query '{query}': "
f"{type(e).__name__}: {e}",
"Unexpected error during sampling for query %r: %s: %s",
query,
type(e).__name__,
e,
exc_info=True,
)
@@ -670,11 +815,15 @@ def configure_semantic_tools(mcp: FastMCP):
)
try:
# Get document receive stream from lifespan context
# Get document receive stream from lifespan context. Direct
# attribute access matches the eviction_task_group pattern at
# ``nc_semantic_search`` (see comment there): both AppContext
# and OAuthAppContext define ``document_receive_stream``, so a
# missing attribute is a typo that should fail loudly. The
# value itself can legitimately be ``None`` before sync starts,
# which the check below handles.
lifespan_ctx = ctx.request_context.lifespan_context
document_receive_stream = getattr(
lifespan_ctx, "document_receive_stream", None
)
document_receive_stream = lifespan_ctx.document_receive_stream
if document_receive_stream is None:
logger.debug(
@@ -705,7 +854,7 @@ def configure_semantic_tools(mcp: FastMCP):
indexed_count = count_result.count
except Exception as e:
logger.warning(f"Failed to query Qdrant for indexed count: {e}")
logger.warning("Failed to query Qdrant for indexed count: %s", e)
# Continue with indexed_count = 0
# Determine status
@@ -719,7 +868,7 @@ def configure_semantic_tools(mcp: FastMCP):
)
except Exception as e:
logger.error(f"Error getting vector sync status: {e}")
logger.error("Error getting vector sync status: %s", e)
raise McpError(
ErrorData(
code=-1,
nextcloud_mcp_server/vector/eviction.py
+61
View File
@@ -0,0 +1,61 @@
"""Lazy eviction of stale documents from the vector index.
Used by the verify-on-read path (ADR-019) to remove points for documents that
have been deleted or unshared in Nextcloud but not yet reconciled by the
webhook/scanner sync loop. Eviction is fire-and-forget from the search hot
path; failures are logged but never propagated, since the next query will
simply re-verify and re-attempt.
"""
import logging
from qdrant_client.models import FieldCondition, Filter, MatchValue
from nextcloud_mcp_server.config import get_settings
from nextcloud_mcp_server.vector.qdrant_client import get_qdrant_client
logger = logging.getLogger(__name__)
async def delete_document_points(
doc_id: str | int,
doc_type: str,
user_id: str,
) -> None:
"""Remove all Qdrant points for a single document.
Deletes both real chunk points and any leftover placeholder points for the
given (user_id, doc_id, doc_type) tuple. Safe to call when the document is
not present — Qdrant returns successfully with zero points affected.
Args:
doc_id: Document ID (int for notes/files/cards/news, str otherwise)
doc_type: Document type (note, file, deck_card, news_item)
user_id: Owner of the points being evicted
Raises:
Exception: If the underlying Qdrant client raises. Callers in the
search hot path should catch and log; eviction failures must not
block search responses.
"""
qdrant_client = await get_qdrant_client()
settings = get_settings()
await qdrant_client.delete(
collection_name=settings.get_collection_name(),
points_selector=Filter(
must=[
FieldCondition(key="user_id", match=MatchValue(value=user_id)),
FieldCondition(key="doc_id", match=MatchValue(value=doc_id)),
FieldCondition(key="doc_type", match=MatchValue(value=doc_type)),
]
),
)
logger.info(
"Evicted Qdrant points for %s_%s (user=%s); "
"document was inaccessible at verification time",
doc_type,
doc_id,
user_id,
)
nextcloud_mcp_server/vector/processor.py
+17
View File
@@ -658,6 +658,23 @@ async def _index_document(
indexed_at = int(time.time())
points = []
# Surface deck card data quality issues at indexing time rather than
# only at verification time (where _verify_deck_cards falls through to
# legacy-data pass-through when board_id/stack_id are missing). This is
# logged once per document — not per chunk — to avoid log spam.
if doc_task.doc_type == "deck_card":
missing_deck_fields = [
field for field in ("board_id", "stack_id") if not file_metadata.get(field)
]
if missing_deck_fields:
logger.warning(
"Indexing deck_card %s for user %s with missing metadata: %s; "
"verification will fall back to legacy-data pass-through",
doc_task.doc_id,
doc_task.user_id,
missing_deck_fields,
)
for i, (chunk, dense_emb, sparse_emb) in enumerate(
zip(chunks, dense_embeddings, sparse_embeddings)
):
nextcloud_mcp_server/vector/scanner.py
+10
View File
@@ -29,6 +29,16 @@ from nextcloud_mcp_server.vector.qdrant_client import get_qdrant_client
logger = logging.getLogger(__name__)
# Single source of truth for which doc_types this scanner indexes. The verifier
# registry in `search/verification.py` must cover every type listed here
# (enforced by `tests/unit/search/test_verification.py`). Add a verifier in the
# same PR that adds a new indexed doc_type, or accept ghost-record exposure for
# that type (see ADR-019).
INDEXED_DOC_TYPES: frozenset[str] = frozenset(
{"note", "file", "deck_card", "news_item"}
)
@dataclass
class DocumentTask:
"""Document task for processing queue."""
tests/integration/test_rag.py
+12 -3
View File
@@ -38,6 +38,7 @@ from typing import Any, AsyncGenerator
import anyio
import pytest
from httpx import HTTPStatusError
from mcp import ClientSession
from nextcloud_mcp_server.providers.base import Provider
@@ -130,10 +131,18 @@ async def indexed_manual_pdf(nc_client, nc_mcp_client):
logger.info(f"Setting up indexed manual PDF: {manual_path}")
# Get file info to verify file exists and get file ID
file_info = await nc_client.webdav.get_file_info(manual_path)
# Get file info to verify file exists and get file ID. After the
# round-7 contract widening, get_file_info raises HTTPStatusError on
# 404 instead of returning None — so wrap and skip on a definitive
# not-found.
try:
file_info = await nc_client.webdav.get_file_info(manual_path)
except HTTPStatusError as e:
if e.response.status_code == 404:
pytest.skip(f"Manual PDF not found at '{manual_path}'")
raise
if not file_info:
pytest.skip(f"Manual PDF not found at '{manual_path}'")
pytest.skip(f"Manual PDF unreadable at '{manual_path}' (malformed PROPFIND)")
file_id = file_info["id"]
logger.info(f"Found manual PDF: {manual_path} (file_id={file_id})")
tests/integration/test_verify_on_read.py
+172
View File
@@ -0,0 +1,172 @@
"""Integration tests for verify-on-read access checks (ADR-019).
These tests exercise ``verify_search_results`` against a real Nextcloud
instance — the verification path's whole purpose is to consult Nextcloud as
the source of truth, so unit-level mocks don't catch protocol or status-code
mismatches between our verifier and the real API.
**Coverage**: only the ``note`` verifier is exercised against real Nextcloud
here. The ``file`` (WebDAV PROPFIND), ``deck_card`` (Deck app), and
``news_item`` (News app) verifiers are unit-tested with mocked HTTP
responses in ``tests/unit/search/test_verification.py``. Adding integration
coverage for those types is tracked as a follow-up — it requires fixture
data (tagged PDFs in user files, a Deck board with cards, a News feed) that
is non-trivial to seed from CI. The mocked unit tests are accurate for
status-code semantics but won't catch payload-shape regressions in those
Nextcloud apps; the trade-off is documented here so future readers know
which suite owns which verifier.
Qdrant is mocked out (``delete_document_points`` and the payload-resolution
helpers) so these tests don't require a running vector database. The unit
suite in ``tests/unit/search/test_verification.py`` covers the Qdrant-side
behaviour separately.
"""
import logging
import uuid
import pytest
from httpx import HTTPStatusError
from nextcloud_mcp_server.client import NextcloudClient
from nextcloud_mcp_server.search import verification
from nextcloud_mcp_server.search.algorithms import SearchResult
from nextcloud_mcp_server.search.verification import verify_search_results
logger = logging.getLogger(__name__)
pytestmark = pytest.mark.integration
def _result_for_note(note_id: int) -> SearchResult:
return SearchResult(
id=note_id,
doc_type="note",
title=f"note_{note_id}",
excerpt="...",
score=0.9,
)
async def test_verify_keeps_accessible_note(
nc_client: NextcloudClient, temporary_note: dict, mocker
):
"""A note that exists in Nextcloud must be kept by verification."""
spy_evict = mocker.AsyncMock()
mocker.patch.object(verification, "delete_document_points", spy_evict)
note_id = temporary_note["id"]
results = [_result_for_note(note_id)]
kept, dropped_count = await verify_search_results(nc_client, results)
assert [r.id for r in kept] == [note_id]
assert dropped_count == 0
spy_evict.assert_not_awaited()
async def test_verify_drops_deleted_note_and_schedules_eviction(
nc_client: NextcloudClient, mocker
):
"""The core ghost-record scenario.
Create a note, delete it via the API (no webhook delivery), then run
verification with a SearchResult still pointing at the gone-but-indexed
document. verify-on-read must drop it and schedule eviction.
"""
spy_evict = mocker.AsyncMock()
mocker.patch.object(verification, "delete_document_points", spy_evict)
# Create a note we'll delete to simulate a ghost record
unique_suffix = uuid.uuid4().hex[:8]
created = await nc_client.notes.create_note(
title=f"verify-on-read ghost {unique_suffix}",
content="This note will be deleted before verification runs.",
category="VerifyOnReadTest",
)
note_id = created["id"]
# Delete via API directly. In production a webhook *should* fire and
# evict from Qdrant — but the whole point of ADR-019 is that we cannot
# rely on this. Verification must catch the drift independently.
await nc_client.notes.delete_note(note_id=note_id)
# Confirm the note is really gone before running verification, so the
# test fails fast if the API behaves unexpectedly.
with pytest.raises(HTTPStatusError) as exc_info:
await nc_client.notes.get_note(note_id)
assert exc_info.value.response.status_code == 404
kept, dropped_count = await verify_search_results(
nc_client, [_result_for_note(note_id)]
)
assert kept == [], "deleted note must not pass verification"
assert dropped_count == 1
spy_evict.assert_awaited_once_with(note_id, "note", nc_client.username)
async def test_verify_mixed_accessible_and_deleted(
nc_client: NextcloudClient, temporary_note: dict, mocker
):
"""Verification must drop only the inaccessible result, keep the rest."""
spy_evict = mocker.AsyncMock()
mocker.patch.object(verification, "delete_document_points", spy_evict)
# temporary_note stays alive for the duration of the test.
accessible_id = temporary_note["id"]
# Make a second note and immediately delete it to create a ghost id.
unique_suffix = uuid.uuid4().hex[:8]
ghost = await nc_client.notes.create_note(
title=f"verify-on-read ghost mix {unique_suffix}",
content="ghost",
category="VerifyOnReadTest",
)
ghost_id = ghost["id"]
await nc_client.notes.delete_note(note_id=ghost_id)
results = [
_result_for_note(accessible_id),
_result_for_note(ghost_id),
]
kept, dropped_count = await verify_search_results(nc_client, results)
assert [r.id for r in kept] == [accessible_id]
assert dropped_count == 1
spy_evict.assert_awaited_once_with(ghost_id, "note", nc_client.username)
async def test_verify_dedupes_chunks_of_same_document(
nc_client: NextcloudClient, temporary_note: dict, mocker
):
"""Multiple chunks of the same note must produce ONE Nextcloud round-trip."""
spy_evict = mocker.AsyncMock()
mocker.patch.object(verification, "delete_document_points", spy_evict)
# Spy through to the real notes client to count round-trips
real_get_note = nc_client.notes.get_note
spy_get_note = mocker.AsyncMock(side_effect=real_get_note)
mocker.patch.object(nc_client.notes, "get_note", spy_get_note)
note_id = temporary_note["id"]
# Three chunks of the same note (chunk_index varies)
results = [
SearchResult(
id=note_id,
doc_type="note",
title="note",
excerpt=f"chunk {i}",
score=0.9 - i * 0.1,
chunk_index=i,
)
for i in range(3)
]
kept, dropped_count = await verify_search_results(nc_client, results)
# All three chunks kept (they're all from the same accessible note)
assert len(kept) == 3
assert dropped_count == 0
# ...but verification only fetched the note ONCE
assert spy_get_note.await_count == 1
tests/unit/client/test_webdav.py
+11 -6
View File
@@ -164,8 +164,14 @@ async def test_get_file_info_returns_file_details(mocker):
@pytest.mark.unit
async def test_get_file_info_returns_none_for_missing_file(mocker):
"""Test that get_file_info returns None for missing files."""
async def test_get_file_info_raises_on_404(mocker):
"""get_file_info now raises HTTPStatusError on 404 (was: returned None).
The contract was widened so verify-on-read can distinguish a definitive
404 from an ambiguous malformed-PROPFIND response. Callers that want
"absent → None" semantics should catch HTTPStatusError and check the
status code themselves.
"""
from httpx import HTTPStatusError, Response
mock_http_client = AsyncMock()
@@ -180,11 +186,10 @@ async def test_get_file_info_returns_none_for_missing_file(mocker):
)
)
# Call get_file_info
result = await client.get_file_info("nonexistent.pdf")
with pytest.raises(HTTPStatusError) as exc_info:
await client.get_file_info("nonexistent.pdf")
# Verify result is None
assert result is None
assert exc_info.value.response.status_code == 404
@pytest.mark.unit
tests/unit/search/test_verification.py
+1125
View File
File diff suppressed because it is too large Load Diff
third_party/astrolabe
Vendored
+1 -1
Submodule third_party/astrolabe updated: 2bf4a70599...d9e641b93e