feat: add opt-in MCP decomposition hook points (design §10)

Adds the seven §10.2 hook-point modules + five env vars so Astrolabe Cloud can
offload document processing to the external document-processor / embedding
gateway. Purely additive: with every setting unset the server behaves exactly
as today, so self-hosters are unaffected (Deck #92).

Hook points (all default to current monolith behavior):
- config: EMBEDDING_PROVIDER, INGEST_MODE, STATUS_BACKEND,
  COLLECTION_METADATA_SOURCE, FACT_EVENT_EMITTER (+ supporting settings),
  validated in Settings.__post_init__ (fail-fast STATUS_BACKEND=local with
  INGEST_MODE=external); shared canonical.py.
- vector/payload_keys.py + acl_hash.py: cross-impl NAMESPACE/point_id (§2.2)
  and BLAKE2b-128 ACL hash (§11), pinned by fixtures shared with the
  document-processor repo.
- embedding/gateway_client.py: OpenAI-compatible GatewayProvider authenticating
  via M2M OIDC client-credentials (separate realm); manual-only registry entry.
- vector/collection_metadata.py: sentinel-point / API metadata source with env
  fallback.
- vector/queue/: hexagonal ingest producer ports + memory/NATS adapters
  (Postgres seam); INGEST_MODE=external publishes mcp.ingest.requested.{tenant}
  instead of the in-memory stream and skips the in-process processor pool. The
  lifespan becomes a composition root across both deployment branches.
- vector/queue/status.py: STATUS_BACKEND=bus subscriber feeding a StatusStore
  the vector-sync status endpoint reads.
- admin/payload_backfill.py: POST /api/v1/admin/payload-backfill (admin scope);
  processor writes the new payload keys; query-side ACL pre-filter gated behind
  ACL_PREFILTER_ENABLED (default off).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Chris Coutinho
2026-05-29 13:13:25 +02:00
co-authored by Claude Opus 4.8
parent c7da612f20
commit d883052fb8
37 changed files with 2534 additions and 55 deletions
+66
View File
@@ -0,0 +1,66 @@
"""Canonical ACL hash (design §11) — a cross-implementation contract.
The external document-processor and the local processor inside this server both
write the ``acl_hash`` Qdrant payload independently; the query path
(``search/verification.py``) builds the matching accessible set. All three must
agree byte-for-byte, so the canonicalization, hash, and accessible-set rules are
pinned here and exercised by an identical ``tests/fixtures/acl_hash_corpus.json``
in both repos (§11.5).
Key invariants:
- ``principal_id`` is NFC-normalized; **case is preserved, not folded** (§11.2).
- ``BLAKE2b-128`` (``digest_size=16`` → 32 hex chars), NOT the 64-byte default
(§11.3).
- Public-link shares are excluded by the caller — they are bearer secrets not
bound to an identity and never reachable via authenticated MCP queries (§11.1).
"""
from __future__ import annotations
import hashlib
import unicodedata
from collections.abc import Iterable
from nextcloud_mcp_server.canonical import canonical_json
PRINCIPAL_TYPES = frozenset({"user", "group", "public"})
# The world-readable principal, included in every requester's accessible set
# (§11.4) and written on any document Nextcloud marks world-readable.
PUBLIC_PRINCIPAL: tuple[str, str] = ("public", "public")
def compute_principal_hash(principal_type: str, principal_id: str) -> str:
"""Hash one ``(principal_type, principal_id)`` tuple (§11.211.3)."""
if principal_type not in PRINCIPAL_TYPES:
raise ValueError(
f"principal_type must be one of {sorted(PRINCIPAL_TYPES)}; "
f"got {principal_type!r}"
)
# principal_type is ASCII by construction; only the id is normalized.
normalized_id = unicodedata.normalize("NFC", principal_id)
canonical = canonical_json([principal_type, normalized_id])
return hashlib.blake2b(canonical, digest_size=16).hexdigest()
def compute_acl_hash(share_set: Iterable[tuple[str, str]]) -> list[str]:
"""Per-principal hash array for a document's share-set (§11.2).
One element per share-set entry, in input order (Qdrant treats the array
with set-semantics, so order is irrelevant at query time). The caller must
have already dropped public-link shares (§11.1).
"""
return [compute_principal_hash(ptype, pid) for ptype, pid in share_set]
def accessible_hash_set(username: str, groups: Iterable[str] = ()) -> set[str]:
"""Accessible hash set for an authenticated requester (§11.4).
Derived entirely from OIDC claims: the requester's own ``(user, username)``,
each ``(group, <name>)`` they hold, and unconditionally ``(public, public)``.
"""
hashes = {compute_principal_hash("user", username)}
for group in groups:
hashes.add(compute_principal_hash("group", group))
hashes.add(compute_principal_hash(*PUBLIC_PRINCIPAL))
return hashes
+1
View File
@@ -0,0 +1 @@
"""Admin-only REST endpoints (``/api/v1/admin/*``)."""
@@ -0,0 +1,104 @@
"""One-shot payload backfill admin endpoint (design §10.2).
``POST /api/v1/admin/payload-backfill`` walks the collection and adds default
values for any *missing* decomposition payload keys (so existing corpora gain
them without a re-index), then upserts the collection-metadata sentinel.
Scope: this backfills the cheap, deployment-level scalar keys
(``processor_version``, ``parsed_at``, ``pipeline_tier``, ``embedding_identity``)
only. It deliberately does NOT synthesize ``acl_hash`` — a correct value needs
per-document share enumeration (a separate job), and writing a placeholder
``acl_hash`` would be unsafe to pre-filter on. The query-side ACL pre-filter
therefore stays disabled until a real ACL backfill runs (see
``ACL_PREFILTER_ENABLED``).
"""
from __future__ import annotations
import logging
from qdrant_client import models
from starlette.requests import Request
from starlette.responses import JSONResponse
from nextcloud_mcp_server.api.management import (
AdminScopeRequired,
require_admin_scope,
)
from nextcloud_mcp_server.config import get_settings
from nextcloud_mcp_server.embedding import get_embedding_service
from nextcloud_mcp_server.vector import payload_keys
from nextcloud_mcp_server.vector.collection_metadata import (
env_default_metadata,
upsert_sentinel,
)
from nextcloud_mcp_server.vector.qdrant_client import get_qdrant_client
logger = logging.getLogger(__name__)
async def handle_payload_backfill(request: Request) -> JSONResponse:
try:
await require_admin_scope(request)
except AdminScopeRequired:
return JSONResponse({"error": "admin scope required"}, status_code=403)
except Exception:
return JSONResponse({"error": "unauthorized"}, status_code=401)
settings = get_settings()
if not settings.vector_sync_enabled:
return JSONResponse({"error": "vector sync disabled"}, status_code=404)
client = await get_qdrant_client()
collection = settings.get_collection_name()
meta = env_default_metadata(settings)
# Deployment-level scalar defaults (safe to set only where missing).
defaults = {
payload_keys.PROCESSOR_VERSION: "backfill",
payload_keys.PIPELINE_TIER: "fast",
payload_keys.EMBEDDING_IDENTITY: meta["embedding_identity"],
}
applied: dict[str, str] = {}
for key, value in defaults.items():
try:
await client.set_payload(
collection_name=collection,
payload={key: value},
# Only points missing this key (don't clobber existing values).
points=models.Filter(
must=[
models.IsEmptyCondition(is_empty=models.PayloadField(key=key))
]
),
wait=True,
)
applied[key] = "set-where-missing"
except Exception as e:
logger.warning("payload backfill failed for key %s: %s", key, e)
applied[key] = f"error: {e}"
# Upsert the collection-metadata sentinel so query-path metadata reads work
# for this collection even without a control plane.
sentinel_ok = True
try:
dimension = get_embedding_service().get_dimension()
await upsert_sentinel(
client,
collection,
embedding_identity=meta["embedding_identity"],
chunking_config=meta["chunking_config"],
dimension=dimension,
)
except Exception as e:
sentinel_ok = False
logger.warning("sentinel upsert failed during backfill: %s", e)
return JSONResponse(
{
"status": "ok",
"collection": collection,
"keys_applied": applied,
"sentinel_upserted": sentinel_ok,
}
)
+44
View File
@@ -122,6 +122,25 @@ async def validate_token_and_get_user(
return user_id, validated
class AdminScopeRequired(Exception):
"""Raised when an authenticated caller lacks the ``admin`` scope."""
async def require_admin_scope(request: Request) -> str:
"""Authenticate the caller and require the ``admin`` scope.
This is the first ``/api/v1/admin/*`` guard; it establishes the pattern
(auth via :func:`validate_token_and_get_user`, then an explicit scope check).
Raises ``ValueError`` on auth failure and :class:`AdminScopeRequired` on a
valid token that lacks ``admin``; callers map these to 401 / 403.
"""
user_id, validated = await validate_token_and_get_user(request)
scopes = validated.get("scopes") or []
if "admin" not in scopes:
raise AdminScopeRequired("admin scope required")
return user_id
def _sanitize_error_for_client(error: Exception, context: str = "") -> str:
"""
Return a safe, generic error message for clients.
@@ -274,6 +293,31 @@ async def get_vector_sync_status(request: Request) -> JSONResponse:
)
try:
# Bus status backend (INGEST_MODE=external): there is no in-process
# queue; pending/terminal state comes from the NATS status subscriber's
# store. indexed_documents stays the mode-independent Qdrant count.
if settings.status_backend == "bus":
store = getattr(request.app.state, "status_store", None)
indexed_count = 0
try:
qdrant_client = await get_qdrant_client()
count_result = await qdrant_client.count(
collection_name=settings.get_collection_name(),
count_filter=Filter(must=[get_placeholder_filter()]),
)
indexed_count = count_result.count
except Exception as e:
logger.warning("Failed to query Qdrant for indexed count: %s", e)
return JSONResponse(
{
"status": "idle",
"indexed_documents": indexed_count,
"pending_documents": 0,
"status_backend": "bus",
"recent_states": store.counts() if store is not None else {},
}
)
# Get document receive stream from app state (set by starlette_lifespan in app.py)
document_receive_stream = getattr(
request.app.state, "document_receive_stream", None
+169 -37
View File
@@ -31,6 +31,7 @@ from starlette.staticfiles import StaticFiles
from starlette.types import ASGIApp, Receive, Send
from starlette.types import Scope as StarletteScope
from nextcloud_mcp_server.admin.payload_backfill import handle_payload_backfill
from nextcloud_mcp_server.api import (
create_webhook,
delete_app_password,
@@ -132,7 +133,13 @@ from nextcloud_mcp_server.vector.oauth_sync import (
from nextcloud_mcp_server.vector.placeholder import sweep_orphan_placeholders
from nextcloud_mcp_server.vector.processor import processor_task
from nextcloud_mcp_server.vector.qdrant_client import get_qdrant_client
from nextcloud_mcp_server.vector.scanner import scanner_task
from nextcloud_mcp_server.vector.queue import (
MemoryTaskProducer,
TaskProducer,
build_external_producer,
)
from nextcloud_mcp_server.vector.queue.status import NatsStatusSubscriber, StatusStore
from nextcloud_mcp_server.vector.scanner import DocumentTask, scanner_task
from nextcloud_mcp_server.vector.webhook_receiver import handle_nextcloud_webhook
logger = logging.getLogger(__name__)
@@ -324,6 +331,13 @@ class VectorSyncState:
document_send_stream: MemoryObjectSendStream | None = None
document_receive_stream: MemoryObjectReceiveStream | None = None
# Ingest producer the scanner/webhook send to: the in-memory send stream
# (local mode) or the NATS bus producer (external mode). The webhook reads
# this; in local mode it is the same object as document_send_stream.
task_producer: "TaskProducer | None" = None
# Bus status store (STATUS_BACKEND=bus): populated by the NATS status
# subscriber, read by the vector-sync status endpoint. None in local mode.
status_store: "StatusStore | 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
@@ -336,6 +350,28 @@ class VectorSyncState:
_vector_sync_state = VectorSyncState()
async def _build_status_subscriber(
settings: "Settings",
) -> "tuple[StatusStore | None, NatsStatusSubscriber | None]":
"""Build the bus status store + subscriber when STATUS_BACKEND=bus.
Returns ``(None, None)`` for local status (the status endpoint reads the
in-memory stream buffer instead). __post_init__ guarantees that bus status
only pairs with external ingest, so ingest_bus_url/tenant_id are set.
"""
if not (settings.ingest_mode == "external" and settings.status_backend == "bus"):
return None, None
assert settings.ingest_bus_url is not None
assert settings.tenant_id is not None
store = StatusStore(max_size=settings.vector_sync_queue_max_size)
subscriber = await NatsStatusSubscriber.connect(
url=settings.ingest_bus_url,
tenant_id=settings.tenant_id,
store=store,
)
return store, subscriber
@dataclass
class AppContext:
"""Application context for BasicAuth mode."""
@@ -344,6 +380,7 @@ class AppContext:
storage: "RefreshTokenStorage | None" = None
document_send_stream: MemoryObjectSendStream | None = None
document_receive_stream: MemoryObjectReceiveStream | None = None
task_producer: "TaskProducer | None" = None
shutdown_event: anyio.Event | None = None
scanner_wake_event: anyio.Event | None = None
@@ -369,6 +406,7 @@ class OAuthAppContext:
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
task_producer: "TaskProducer | None" = None
shutdown_event: anyio.Event | None = None
scanner_wake_event: anyio.Event | None = None
@@ -1643,22 +1681,47 @@ def get_app(transport: str = "streamable-http", enabled_apps: list[str] | None =
# Orphan-sweep before scanner starts — card #101.
await _sweep_orphan_placeholders_if_enabled()
# Initialize shared state
send_stream, receive_stream = anyio.create_memory_object_stream(
max_buffer_size=settings.vector_sync_queue_max_size
)
# Initialize shared state. INGEST_MODE selects the transport
# (design §10.1): local uses the in-memory anyio stream + the
# in-process processor pool; external publishes to NATS and runs no
# in-process consumer (the document-processor consumes).
external = settings.ingest_mode == "external"
shutdown_event = anyio.Event()
scanner_wake_event = anyio.Event()
# Store in app state for access from routes (ADR-007)
send_stream = None
receive_stream = None
task_producer: TaskProducer
if external:
task_producer = await build_external_producer(settings)
logger.info(
"Ingest mode external: publishing to %s", settings.ingest_bus_url
)
else:
send_stream, receive_stream = anyio.create_memory_object_stream[
DocumentTask
](max_buffer_size=settings.vector_sync_queue_max_size)
task_producer = MemoryTaskProducer(send_stream)
# Bus status backend: subscribe to mcp.document.* into a store the
# status endpoint reads (STATUS_BACKEND=bus; external mode only).
status_store, status_subscriber = await _build_status_subscriber(settings)
# Store in app state for access from routes (ADR-007). In external
# mode there is no in-memory stream, so document_send/receive_stream
# stay None; task_producer is the bus producer.
app.state.document_send_stream = send_stream
app.state.document_receive_stream = receive_stream
app.state.task_producer = task_producer
app.state.status_store = status_store
app.state.shutdown_event = shutdown_event
app.state.scanner_wake_event = scanner_wake_event
# Also store in module singleton for FastMCP session lifespans
_vector_sync_state.document_send_stream = send_stream
_vector_sync_state.document_receive_stream = receive_stream
_vector_sync_state.task_producer = task_producer
_vector_sync_state.status_store = status_store
_vector_sync_state.shutdown_event = shutdown_event
_vector_sync_state.scanner_wake_event = scanner_wake_event
logger.info("Vector sync state stored in module singleton")
@@ -1669,6 +1732,8 @@ def get_app(transport: str = "streamable-http", enabled_apps: list[str] | None =
browser_app = cast(Starlette, route.app)
browser_app.state.document_send_stream = send_stream
browser_app.state.document_receive_stream = receive_stream
browser_app.state.task_producer = task_producer
browser_app.state.status_store = status_store
browser_app.state.shutdown_event = shutdown_event
browser_app.state.scanner_wake_event = scanner_wake_event
logger.info("Vector sync state shared with browser_app for /app")
@@ -1676,26 +1741,33 @@ def get_app(transport: str = "streamable-http", enabled_apps: list[str] | None =
# Start background tasks using anyio TaskGroup
async with anyio.create_task_group() as tg:
# Start scanner task
# Start scanner task (publishes to task_producer)
await tg.start(
scanner_task,
send_stream,
task_producer,
shutdown_event,
scanner_wake_event,
client,
username,
)
# Start processor pool (each gets a cloned receive stream)
for i in range(settings.vector_sync_processor_workers):
await tg.start(
processor_task,
i,
receive_stream.clone(),
shutdown_event,
client,
username,
)
# The in-process processor pool runs only in local mode; in
# external mode the document-processor service is the consumer.
if not external:
assert receive_stream is not None
for i in range(settings.vector_sync_processor_workers):
await tg.start(
processor_task,
i,
receive_stream.clone(),
shutdown_event,
client,
username,
)
# Bus status subscriber (STATUS_BACKEND=bus).
if status_subscriber is not None:
await tg.start(status_subscriber.run, shutdown_event)
# Expose this long-lived task group to request-path code that
# wants to spawn background work (e.g. ADR-019 verify-on-read
@@ -1704,8 +1776,9 @@ def get_app(transport: str = "streamable-http", enabled_apps: list[str] | None =
_vector_sync_state.eviction_task_group = tg
logger.info(
"Background sync tasks started: 1 scanner + %s processors",
settings.vector_sync_processor_workers,
"Background sync tasks started: 1 scanner + %s processors (ingest=%s)",
0 if external else settings.vector_sync_processor_workers,
settings.ingest_mode,
)
# Run MCP session manager and yield
@@ -1718,6 +1791,12 @@ def get_app(transport: str = "streamable-http", enabled_apps: list[str] | None =
shutdown_event.set()
# Request path must not spawn into a cancelling group.
_vector_sync_state.eviction_task_group = None
# Drain the shared bus connection (external mode only).
_drain = getattr(task_producer, "drain", None)
if external and _drain is not None:
await _drain()
if status_subscriber is not None:
await status_subscriber.aclose()
await client.close()
# TaskGroup automatically cancels all tasks on exit
@@ -1825,25 +1904,51 @@ def get_app(transport: str = "streamable-http", enabled_apps: list[str] | None =
except Exception as e:
logger.warning("App password cleanup failed (non-fatal): %s", e)
# Initialize shared state
send_stream, receive_stream = anyio.create_memory_object_stream(
max_buffer_size=settings.vector_sync_queue_max_size
)
# Initialize shared state. INGEST_MODE selects the transport
# (design §10.1): local uses the in-memory anyio stream + the
# in-process processor pool; external publishes to NATS and runs
# no in-process consumer (the document-processor consumes).
external = settings.ingest_mode == "external"
shutdown_event = anyio.Event()
scanner_wake_event = anyio.Event()
# User state tracking for user manager
user_states: dict = {}
send_stream = None
receive_stream = None
task_producer: TaskProducer
if external:
task_producer = await build_external_producer(settings)
logger.info(
"Ingest mode external: publishing to %s",
settings.ingest_bus_url,
)
else:
send_stream, receive_stream = anyio.create_memory_object_stream[
DocumentTask
](max_buffer_size=settings.vector_sync_queue_max_size)
task_producer = MemoryTaskProducer(send_stream)
# Bus status backend: subscribe to mcp.document.* into a store
# the status endpoint reads (STATUS_BACKEND=bus; external only).
status_store, status_subscriber = await _build_status_subscriber(
settings
)
# Store in app state for access from routes (ADR-007)
app.state.document_send_stream = send_stream
app.state.document_receive_stream = receive_stream
app.state.task_producer = task_producer
app.state.status_store = status_store
app.state.shutdown_event = shutdown_event
app.state.scanner_wake_event = scanner_wake_event
# Also store in module singleton for FastMCP session lifespans
_vector_sync_state.document_send_stream = send_stream
_vector_sync_state.document_receive_stream = receive_stream
_vector_sync_state.task_producer = task_producer
_vector_sync_state.status_store = status_store
_vector_sync_state.shutdown_event = shutdown_event
_vector_sync_state.scanner_wake_event = scanner_wake_event
logger.info("Vector sync state stored in module singleton")
@@ -1854,6 +1959,8 @@ def get_app(transport: str = "streamable-http", enabled_apps: list[str] | None =
browser_app = cast(Starlette, route.app)
browser_app.state.document_send_stream = send_stream
browser_app.state.document_receive_stream = receive_stream
browser_app.state.task_producer = task_producer
browser_app.state.status_store = status_store
browser_app.state.shutdown_event = shutdown_event
browser_app.state.scanner_wake_event = scanner_wake_event
logger.info(
@@ -1870,10 +1977,12 @@ def get_app(transport: str = "streamable-http", enabled_apps: list[str] | None =
# `token_broker` constructed above is still used by the
# management API revoke endpoint (via app.state.oauth_context).
async with anyio.create_task_group() as tg:
# Start user manager task (supervises per-user scanners)
# Start user manager task (supervises per-user scanners).
# Each per-user scanner clones task_producer; for the bus
# producer clone() returns the shared connection.
await tg.start(
user_manager_task,
send_stream,
task_producer,
shutdown_event,
scanner_wake_event,
token_storage,
@@ -1882,15 +1991,22 @@ def get_app(transport: str = "streamable-http", enabled_apps: list[str] | None =
tg,
)
# Start processor pool (each gets a cloned receive stream)
for i in range(settings.vector_sync_processor_workers):
await tg.start(
oauth_processor_task,
i,
receive_stream.clone(),
shutdown_event,
nextcloud_host_for_sync,
)
# In-process processor pool runs only in local mode; in
# external mode the document-processor service consumes.
if not external:
assert receive_stream is not None
for i in range(settings.vector_sync_processor_workers):
await tg.start(
oauth_processor_task,
i,
receive_stream.clone(),
shutdown_event,
nextcloud_host_for_sync,
)
# Bus status subscriber (STATUS_BACKEND=bus).
if status_subscriber is not None:
await tg.start(status_subscriber.run, shutdown_event)
# Expose this long-lived task group to request-path code
# that wants to spawn background work (e.g. ADR-019
@@ -1899,8 +2015,9 @@ def get_app(transport: str = "streamable-http", enabled_apps: list[str] | None =
_vector_sync_state.eviction_task_group = tg
logger.info(
"Background sync tasks started: 1 user manager + %s processors",
settings.vector_sync_processor_workers,
"Background sync tasks started: 1 user manager + %s processors (ingest=%s)",
0 if external else settings.vector_sync_processor_workers,
settings.ingest_mode,
)
# Run MCP session manager and yield
@@ -1913,6 +2030,12 @@ def get_app(transport: str = "streamable-http", enabled_apps: list[str] | None =
shutdown_event.set()
# Request path must not spawn into a cancelling group.
_vector_sync_state.eviction_task_group = None
# Drain the shared bus connection (external only).
_drain = getattr(task_producer, "drain", None)
if external and _drain is not None:
await _drain()
if status_subscriber is not None:
await status_subscriber.aclose()
# Close token broker HTTP client
if token_broker._http_client:
await token_broker._http_client.aclose()
@@ -2114,6 +2237,15 @@ def get_app(transport: str = "streamable-http", enabled_apps: list[str] | None =
settings.enable_multi_user_basic_auth and settings.enable_offline_access
)
if enable_authenticated_management_apis:
# Admin: one-shot payload backfill (design §10.2). Requires the `admin`
# scope (enforced inside the handler via require_admin_scope).
routes.append(
Route(
"/api/v1/admin/payload-backfill",
handle_payload_backfill,
methods=["POST"],
)
)
routes.append(
Route(
"/api/v1/users/{user_id}/session",
+26
View File
@@ -0,0 +1,26 @@
"""Canonical JSON encoding shared across cross-implementation hashes.
The Astrolabe Cloud decomposition (design §2.3) fixes a single canonical JSON
encoding so hashes computed here match those computed independently by the
external document-processor and embedding-gateway services. Any drift in
separators, key ordering, or unicode handling would break NATS dedup keys,
Qdrant point-ID idempotency, and ACL-hash compatibility.
"""
from __future__ import annotations
import json
from typing import Any
def canonical_json(obj: Any) -> bytes:
"""Encode ``obj`` to canonical JSON bytes.
Deterministic across implementations: sorted keys, no inter-token
whitespace, non-ASCII preserved (UTF-8). Consumers: the NATS
``Nats-Msg-Id`` dedup header (vector/queue/nats.py), Qdrant point IDs
(vector/payload_keys.py), and ACL hashes (acl_hash.py).
"""
return json.dumps(
obj, sort_keys=True, separators=(",", ":"), ensure_ascii=False
).encode("utf-8")
+147
View File
@@ -161,6 +161,38 @@ _DEFAULTS: dict[str, Any] = {
# Nextcloud system tag names. Files/folders carrying any of these tags
# are hidden from WebDAV MCP tools. Empty = feature off.
"excluded_tags": "",
# MCP decomposition hook points (design §10). Every default reproduces
# the current monolithic behavior; self-hosters who set none are
# unaffected. See docs/architecture/mcp-decomposition.md (sibling repo).
"embedding_provider": "autodetect", # autodetect | gateway
"ingest_mode": "local", # local | external
"status_backend": "local", # local | bus
"collection_metadata_source": "qdrant", # qdrant | api
# CP base URL for COLLECTION_METADATA_SOURCE=api (e.g. http://control-plane).
# Required only when the source is api.
"collection_metadata_api_url": None,
"fact_event_emitter": "none", # none | nats | stdout
"ingest_bus_url": None, # required when ingest_mode=external
"embedding_gateway_url": None, # required when embedding_provider=gateway
# Logical model the gateway routes on (e.g. mistral-embed → Mistral for
# the MVP). Only consulted when embedding_provider=gateway.
"embedding_gateway_model": "mistral-embed",
# Gateway auth: the MCP server is an OIDC *client* in the gateway's own
# M2M realm (parallel to, and distinct from, the tenant realm it already
# serves). It obtains a client-credentials token and the gateway maps the
# client-id → the tenant's underlying provider API key. All four unset =
# call the gateway unauthenticated (matches today's not-yet-authed gateway).
"embedding_gateway_token_url": None, # M2M token endpoint
"embedding_gateway_client_id": None,
"embedding_gateway_client_secret": None,
"embedding_gateway_scope": None, # e.g. astrolabe-embedding-gateway/embed
"tenant_id": None, # NATS per-tenant subject token (UUID form)
"ingest_bus_num_replicas": 1, # JetStream stream replicas (prod: 3)
# Query-side ACL pre-filter (design §11). OFF by default: a Qdrant
# `match any` on `acl_hash` excludes points missing the key, so enabling
# this before a real ACL backfill would silently drop legacy results.
# verify-on-read remains the correctness backstop regardless.
"acl_prefilter_enabled": False,
}
@@ -653,6 +685,26 @@ class Settings:
# are hidden from WebDAV MCP tools.
excluded_tags: str = ""
# MCP decomposition hook points (design §10, opt-in). All defaults
# reproduce the current monolith; validated in __post_init__.
embedding_provider: str = "autodetect" # autodetect | gateway
ingest_mode: str = "local" # local | external
status_backend: str = "local" # local | bus
collection_metadata_source: str = "qdrant" # qdrant | api
collection_metadata_api_url: str | None = None # CP URL when source=api
fact_event_emitter: str = "none" # none | nats | stdout
ingest_bus_url: str | None = None # required when ingest_mode=external
embedding_gateway_url: str | None = None # required when provider=gateway
embedding_gateway_model: str = "mistral-embed" # logical model gateway routes
# Gateway M2M OIDC client creds (separate realm; see _DEFAULTS comment).
embedding_gateway_token_url: str | None = None
embedding_gateway_client_id: str | None = None
embedding_gateway_client_secret: str | None = None
embedding_gateway_scope: str | None = None
tenant_id: str | None = None # NATS per-tenant subject token (UUID form)
ingest_bus_num_replicas: int = 1 # JetStream stream replicas (prod: 3)
acl_prefilter_enabled: bool = False # query-side ACL pre-filter (§11); OFF
def __post_init__(self):
"""Validate configuration and set defaults."""
logger = logging.getLogger(__name__)
@@ -733,6 +785,84 @@ class Settings:
self.document_chunk_size,
)
# --- MCP decomposition hook points (design §10) ---
# Normalize + validate the opt-in enum settings. Defaults reproduce
# the monolith, so deployments that set none of these pass through.
_enum_fields = {
"embedding_provider": {"autodetect", "gateway"},
"ingest_mode": {"local", "external"},
"status_backend": {"local", "bus"},
"collection_metadata_source": {"qdrant", "api"},
"fact_event_emitter": {"none", "nats", "stdout"},
}
for _field, _allowed in _enum_fields.items():
_val = (getattr(self, _field) or "").strip().lower()
setattr(self, _field, _val)
if _val not in _allowed:
raise ValueError(
f"{_field.upper()} must be one of {sorted(_allowed)}; got {_val!r}"
)
# Fail-fast: external ingest sources its status from the bus. With the
# in-process state machine empty, STATUS_BACKEND=local would leave
# status streams silently empty — crash loudly instead (design §10.1).
if self.status_backend == "local" and self.ingest_mode == "external":
raise RuntimeError(
"STATUS_BACKEND=local is incompatible with INGEST_MODE=external; "
"set STATUS_BACKEND=bus"
)
# Conditional-required settings for the active hook points.
if self.ingest_mode == "external":
if not self.ingest_bus_url:
raise ValueError("INGEST_BUS_URL is required when INGEST_MODE=external")
if not self.tenant_id:
raise ValueError("TENANT_ID is required when INGEST_MODE=external")
if self.embedding_provider == "gateway" and not self.embedding_gateway_url:
raise ValueError(
"EMBEDDING_GATEWAY_URL is required when EMBEDDING_PROVIDER=gateway"
)
if (
self.collection_metadata_source == "api"
and not self.collection_metadata_api_url
):
raise ValueError(
"COLLECTION_METADATA_API_URL is required when "
"COLLECTION_METADATA_SOURCE=api"
)
# Gateway M2M OIDC creds are all-or-nothing: a partial set (e.g. a
# client_id with no token endpoint) is a misconfiguration that would
# silently fall back to unauthenticated calls. scope is optional.
_gw_creds = (
self.embedding_gateway_token_url,
self.embedding_gateway_client_id,
self.embedding_gateway_client_secret,
)
if any(_gw_creds) and not all(_gw_creds):
raise ValueError(
"EMBEDDING_GATEWAY_TOKEN_URL, EMBEDDING_GATEWAY_CLIENT_ID, and "
"EMBEDDING_GATEWAY_CLIENT_SECRET must be set together (M2M OIDC "
"client-credentials) or all left unset (unauthenticated gateway)"
)
# TENANT_ID is a NATS subject token; '.', '*', '>', and whitespace are
# reserved/illegal there and would silently break subscriptions (§3.4).
if self.tenant_id and (
any(c in self.tenant_id for c in ".*>")
or any(c.isspace() for c in self.tenant_id)
):
raise ValueError(
"TENANT_ID must not contain '.', '*', '>', or whitespace "
"(it is used as a NATS subject token)"
)
if self.ingest_bus_num_replicas < 1:
raise ValueError(
f"INGEST_BUS_NUM_REPLICAS must be >= 1; "
f"got {self.ingest_bus_num_replicas}"
)
# --- ADR-022 follow-up: deployment mode is the single source of truth ---
# The ENABLE_MULTI_USER_BASIC_AUTH and ENABLE_LOGIN_FLOW env vars were
# removed in favour of MCP_DEPLOYMENT_MODE. We do TWO things here:
@@ -1128,6 +1258,23 @@ def get_settings() -> Settings:
"log_level": "LOG_LEVEL",
"log_include_trace_context": "LOG_INCLUDE_TRACE_CONTEXT",
"excluded_tags": "EXCLUDED_TAGS",
# MCP decomposition hook points (design §10)
"embedding_provider": "EMBEDDING_PROVIDER",
"ingest_mode": "INGEST_MODE",
"status_backend": "STATUS_BACKEND",
"collection_metadata_source": "COLLECTION_METADATA_SOURCE",
"collection_metadata_api_url": "COLLECTION_METADATA_API_URL",
"fact_event_emitter": "FACT_EVENT_EMITTER",
"ingest_bus_url": "INGEST_BUS_URL",
"embedding_gateway_url": "EMBEDDING_GATEWAY_URL",
"embedding_gateway_model": "EMBEDDING_GATEWAY_MODEL",
"embedding_gateway_token_url": "EMBEDDING_GATEWAY_TOKEN_URL",
"embedding_gateway_client_id": "EMBEDDING_GATEWAY_CLIENT_ID",
"embedding_gateway_client_secret": "EMBEDDING_GATEWAY_CLIENT_SECRET",
"embedding_gateway_scope": "EMBEDDING_GATEWAY_SCOPE",
"tenant_id": "TENANT_ID",
"ingest_bus_num_replicas": "INGEST_BUS_NUM_REPLICAS",
"acl_prefilter_enabled": "ACL_PREFILTER_ENABLED",
}
# Only pass values that dynaconf actually has; omit unset keys so
@@ -0,0 +1,137 @@
"""OpenAI-compatible embedding provider targeting the Astrolabe Cloud embedding
gateway (design §10.2).
Active only when ``EMBEDDING_PROVIDER=gateway``. Registered *manually* in
``providers/registry.py`` — never part of the autodetect chain — so self-hosters
who don't opt in are unaffected.
**Auth model.** The MCP server is an OIDC *client* in the gateway's own
machine-to-machine realm — a realm *parallel to, and distinct from*, the tenant
realm the MCP server already serves as a client (Nextcloud user_oidc). It
obtains a ``client_credentials`` token and presents it as a Bearer; the gateway
maps the token's client-id → the tenant's underlying provider API key. This
mirrors the control-plane CLI's ``fetch_m2m_token`` pattern
(astrolabe-cloud-website ``services/control-plane/.../cli/_common.py``). When no
M2M creds are configured the client calls the gateway unauthenticated — matching
the gateway's current (not-yet-authenticated) state.
The gateway speaks the OpenAI ``/v1/embeddings`` wire format and routes by model
name (e.g. ``mistral-embed`` → Mistral for the MVP). Embeddings-only: ``generate``
is disabled (inherited ``NotImplementedError``).
"""
from __future__ import annotations
import logging
import time
import httpx
from ..providers.openai import OpenAIProvider
logger = logging.getLogger(__name__)
# Refresh the cached token this many seconds before its stated expiry, so a
# token never expires mid-flight (matches AstrolabeClient / CP CLI behavior).
_EARLY_REFRESH_SECONDS = 60
class GatewayTokenProvider:
"""Caches a gateway M2M access token via the ``client_credentials`` grant.
HTTP Basic client auth + form-encoded grant, mirroring the website's
``fetch_m2m_token``. Tokens are cached until ``_EARLY_REFRESH_SECONDS``
before expiry.
"""
def __init__(
self,
token_url: str,
client_id: str,
client_secret: str,
scope: str | None = None,
timeout: float = 10.0,
):
self.token_url = token_url
self.client_id = client_id
self.client_secret = client_secret
self.scope = scope
self.timeout = timeout
self._cache: tuple[str, float] | None = None # (token, expires_at)
async def get_token(self, *, force_refresh: bool = False) -> str:
if (
self._cache is not None
and not force_refresh
and time.time() < self._cache[1]
):
return self._cache[0]
data = {"grant_type": "client_credentials"}
if self.scope:
data["scope"] = self.scope
async with httpx.AsyncClient(
timeout=httpx.Timeout(self.timeout, connect=5.0)
) as client:
resp = await client.post(
self.token_url,
data=data,
auth=(self.client_id, self.client_secret),
)
resp.raise_for_status()
body = resp.json()
expires_in = body.get("expires_in", 3600)
self._cache = (
body["access_token"],
time.time() + expires_in - _EARLY_REFRESH_SECONDS,
)
logger.info("Obtained embedding-gateway M2M token (expires in %ss)", expires_in)
return self._cache[0]
class GatewayProvider(OpenAIProvider):
"""Embeddings-only OpenAI-compatible provider pointed at the gateway."""
def __init__(
self,
*,
base_url: str,
embedding_model: str,
token_provider: GatewayTokenProvider | None = None,
timeout: float = 120.0,
):
# AsyncOpenAI rejects an empty key; use a non-secret placeholder when
# the gateway is unauthenticated. When a token provider is configured,
# the real Bearer is set on the client before each request.
super().__init__(
api_key="gateway-unauthenticated",
base_url=base_url,
embedding_model=embedding_model,
generation_model=None, # gateway never generates
timeout=timeout,
)
self._token_provider = token_provider
logger.info(
"Initialized gateway embedding provider: base_url=%s, model=%s, auth=%s",
base_url,
embedding_model,
"oidc-m2m" if token_provider else "none",
)
async def _ensure_bearer(self) -> None:
"""Refresh the OIDC M2M token onto the OpenAI client (no-op when
unauthenticated). AsyncOpenAI reads ``api_key`` per request to build
the Authorization header, so updating it here applies to the next call.
"""
if self._token_provider is not None:
self.client.api_key = await self._token_provider.get_token()
async def embed(self, text: str) -> list[float]:
await self._ensure_bearer()
return await super().embed(text)
async def embed_batch(self, texts: list[str]) -> list[list[float]]:
await self._ensure_bearer()
return await super().embed_batch(texts)
@@ -50,6 +50,44 @@ class ProviderRegistry:
"""
settings = get_settings()
# 0. Gateway (manual-only; never autodetected). When
# EMBEDDING_PROVIDER=gateway, bypass the autodetect chain entirely
# and route embeddings through the Astrolabe Cloud embedding gateway
# (design §10.1/§10.2). Lazy import to avoid an import cycle
# (embedding/gateway_client → providers/openai → ... → registry).
if settings.embedding_provider == "gateway":
from ..embedding.gateway_client import ( # noqa: PLC0415
GatewayProvider,
GatewayTokenProvider,
)
# Settings.__post_init__ guarantees the URL is set when the
# provider is gateway; assert narrows the type for the checker.
assert settings.embedding_gateway_url is not None
# __post_init__ enforces all-or-nothing on the M2M creds, so
# checking one is enough to know the full triple is present.
token_provider = None
if settings.embedding_gateway_client_id:
assert settings.embedding_gateway_token_url is not None
assert settings.embedding_gateway_client_secret is not None
token_provider = GatewayTokenProvider(
token_url=settings.embedding_gateway_token_url,
client_id=settings.embedding_gateway_client_id,
client_secret=settings.embedding_gateway_client_secret,
scope=settings.embedding_gateway_scope,
)
logger.info(
"Using embedding gateway provider: url=%s, model=%s, auth=%s",
settings.embedding_gateway_url,
settings.embedding_gateway_model,
"oidc-m2m" if token_provider else "none",
)
return GatewayProvider(
base_url=settings.embedding_gateway_url,
embedding_model=settings.embedding_gateway_model,
token_provider=token_provider,
)
# 1. Bedrock
if (
settings.aws_region
+17 -1
View File
@@ -3,8 +3,9 @@
import logging
from typing import Any
from qdrant_client.models import FieldCondition, Filter, MatchValue
from qdrant_client.models import FieldCondition, Filter, MatchAny, MatchValue
from nextcloud_mcp_server.acl_hash import accessible_hash_set
from nextcloud_mcp_server.config import get_settings
from nextcloud_mcp_server.embedding import get_embedding_service
from nextcloud_mcp_server.observability.metrics import record_qdrant_operation
@@ -13,6 +14,7 @@ from nextcloud_mcp_server.search.algorithms import (
SearchResult,
build_search_result_from_point,
)
from nextcloud_mcp_server.vector.payload_keys import ACL_HASH
from nextcloud_mcp_server.vector.placeholder import get_placeholder_filter
from nextcloud_mcp_server.vector.qdrant_client import get_qdrant_client
@@ -112,6 +114,20 @@ class SemanticSearchAlgorithm(SearchAlgorithm):
)
)
# ACL pre-filter (design §11), opt-in via ACL_PREFILTER_ENABLED and OFF
# by default. Additive `must` condition — it can only narrow results,
# never broaden them, and verify-on-read remains the correctness
# backstop. Only enable after a real acl_hash backfill: a MatchAny on
# acl_hash excludes points missing the key (legacy docs), so enabling
# it on an un-backfilled collection would silently drop results.
if settings.acl_prefilter_enabled:
# Groups are not yet threaded into the search signature; user +
# public principals are covered. Group support is a follow-up.
accessible = accessible_hash_set(user_id)
filter_conditions.append(
FieldCondition(key=ACL_HASH, match=MatchAny(any=sorted(accessible)))
)
# Search Qdrant
qdrant_client = await get_qdrant_client()
try:
@@ -0,0 +1,155 @@
"""Per-collection metadata: embedding identity + chunking config (design §10.1).
The query path needs to know which embedding produced a collection's vectors
(``embedding_identity``) and how it was chunked (``chunking_config``) so it can
request the matching embedding at lookup time. Two sources, selected by
``COLLECTION_METADATA_SOURCE``:
- ``qdrant`` — a sentinel point (deterministic UUID, normalisable non-zero dense
vector) stored inside the collection. Works for any Qdrant deployment, so
self-hosters benefit even without a control plane.
- ``api`` — an HTTP GET against the control plane
(``/v1/qdrant-collections/{name}/metadata``).
On a missing/unreadable sentinel the query path logs a warning and falls back to
the environment-configured defaults — matching today's monolith behavior, so
query availability is preserved (design §10.1).
"""
from __future__ import annotations
import logging
from typing import Any
import httpx
from qdrant_client import AsyncQdrantClient, models
from ..config import Settings, get_settings
from .payload_keys import EMBEDDING_IDENTITY
logger = logging.getLogger(__name__)
# Deterministic sentinel point id (design §10.1). Carries collection metadata
# and never matches a search (no user_id/doc_id/doc_type payload to match).
SENTINEL_POINT_ID = "00000000-0000-0000-0000-000000000000"
# Sentinel payload keys.
CHUNKING_CONFIG = "chunking_config"
IS_SENTINEL = "is_sentinel"
def build_embedding_identity(settings: Settings | None = None) -> str:
"""The embedding identity for locally-produced vectors: the model name.
The gateway and query path route on this name; for the monolith it is the
active embedding model (matching the collection-name derivation).
"""
s = settings or get_settings()
return s.get_embedding_model_name()
def env_default_metadata(settings: Settings | None = None) -> dict[str, Any]:
"""Metadata derived purely from environment config — the fallback when no
sentinel/API metadata is available."""
s = settings or get_settings()
return {
"embedding_identity": build_embedding_identity(s),
"chunking_config": {
"chunk_size": s.document_chunk_size,
"chunk_overlap": s.document_chunk_overlap,
},
}
def _sentinel_dense(dimension: int) -> list[float]:
# Cosine distance is undefined for the zero vector (and Qdrant Cloud strict
# mode rejects it), so use one tiny non-zero element — mirrors the doc-id
# backfill sentinel in qdrant_client.py.
return [1e-9] + [0.0] * (dimension - 1)
async def upsert_sentinel(
client: AsyncQdrantClient,
collection_name: str,
*,
embedding_identity: str,
chunking_config: dict[str, Any],
dimension: int,
) -> None:
"""Idempotently write the metadata sentinel point for a collection."""
point = models.PointStruct(
id=SENTINEL_POINT_ID,
vector={
"dense": _sentinel_dense(dimension),
"sparse": models.SparseVector(indices=[], values=[]),
},
payload={
EMBEDDING_IDENTITY: embedding_identity,
CHUNKING_CONFIG: chunking_config,
IS_SENTINEL: True,
},
)
await client.upsert(collection_name=collection_name, points=[point], wait=True)
logger.debug("Upserted metadata sentinel on '%s'", collection_name)
async def _read_from_qdrant(
client: AsyncQdrantClient, collection_name: str
) -> dict[str, Any] | None:
points = await client.retrieve(
collection_name=collection_name,
ids=[SENTINEL_POINT_ID],
with_payload=True,
)
if not points:
return None
payload = points[0].payload or {}
if EMBEDDING_IDENTITY not in payload:
return None
return {
"embedding_identity": payload.get(EMBEDDING_IDENTITY),
"chunking_config": payload.get(CHUNKING_CONFIG),
}
async def _read_from_api(api_url: str, collection_name: str) -> dict[str, Any] | None:
url = f"{api_url.rstrip('/')}/v1/qdrant-collections/{collection_name}/metadata"
async with httpx.AsyncClient(timeout=httpx.Timeout(10.0, connect=5.0)) as client:
resp = await client.get(url)
if resp.status_code == 404:
return None
resp.raise_for_status()
return resp.json()
async def read_collection_metadata(
client: AsyncQdrantClient,
collection_name: str,
settings: Settings | None = None,
) -> dict[str, Any]:
"""Read collection metadata from the configured source, falling back to env
defaults on any miss/error (preserves query availability — §10.1)."""
s = settings or get_settings()
meta: dict[str, Any] | None = None
try:
if s.collection_metadata_source == "api":
assert s.collection_metadata_api_url is not None
meta = await _read_from_api(s.collection_metadata_api_url, collection_name)
else:
meta = await _read_from_qdrant(client, collection_name)
except Exception:
logger.warning(
"Collection metadata read failed for '%s' (source=%s); using env defaults",
collection_name,
s.collection_metadata_source,
exc_info=True,
)
if not meta or not meta.get("embedding_identity"):
logger.warning(
"Collection metadata missing for '%s' (source=%s); using env defaults",
collection_name,
s.collection_metadata_source,
)
return env_default_metadata(s)
return meta
+5 -7
View File
@@ -24,16 +24,14 @@ from dataclasses import dataclass, field
import anyio
from anyio.abc import TaskGroup, TaskStatus
from anyio.streams.memory import (
MemoryObjectReceiveStream,
MemoryObjectSendStream,
)
from anyio.streams.memory import MemoryObjectReceiveStream
from httpx import BasicAuth, HTTPStatusError
from nextcloud_mcp_server.auth.storage import RefreshTokenStorage
from nextcloud_mcp_server.client import NextcloudClient
from nextcloud_mcp_server.config import get_settings
from nextcloud_mcp_server.vector.processor import process_document
from nextcloud_mcp_server.vector.queue.ports import TaskProducer
from nextcloud_mcp_server.vector.scanner import DocumentTask, scan_user_documents
logger = logging.getLogger(__name__)
@@ -101,7 +99,7 @@ async def get_user_client_basic_auth(
async def user_scanner_task(
user_id: str,
send_stream: MemoryObjectSendStream[DocumentTask],
send_stream: TaskProducer,
shutdown_event: anyio.Event,
wake_event: anyio.Event,
nextcloud_host: str,
@@ -330,7 +328,7 @@ oauth_processor_task = multi_user_processor_task
async def _run_user_scanner_with_scope(
user_id: str,
cancel_scope: anyio.CancelScope,
send_stream: MemoryObjectSendStream[DocumentTask],
send_stream: TaskProducer,
shutdown_event: anyio.Event,
wake_event: anyio.Event,
nextcloud_host: str,
@@ -358,7 +356,7 @@ async def _run_user_scanner_with_scope(
async def user_manager_task(
send_stream: MemoryObjectSendStream[DocumentTask],
send_stream: TaskProducer,
shutdown_event: anyio.Event,
wake_event: anyio.Event,
refresh_token_storage: "RefreshTokenStorage",
@@ -0,0 +1,45 @@
"""Qdrant payload-key constants + the shared point-ID namespace (design §2.2).
These names and the ``NAMESPACE`` UUID are a cross-implementation contract: the
external document-processor (astrolabe-cloud-website) computes identical chunk
point IDs and writes the same payload keys. Divergence in ``NAMESPACE`` would
break Qdrant upsert idempotency and duplicate chunks at the Phase 2 cutover, so
the literal is pinned by a fixture checked into both repos
(``tests/fixtures/namespace_uuid.txt``) and asserted equal in each repo's suite.
Even in the default local mode the MCP server writes these keys on upsert, so a
later migration to the external processor is friction-free (design §10.2).
"""
from __future__ import annotations
import uuid
from nextcloud_mcp_server.canonical import canonical_json
# Payload keys introduced by the decomposition (design §10.2).
EMBEDDING_IDENTITY = "embedding_identity"
ACL_HASH = "acl_hash"
PROCESSOR_VERSION = "processor_version"
PARSED_AT = "parsed_at"
PIPELINE_TIER = "pipeline_tier"
# Fixed platform namespace for deterministic chunk point IDs (design §2.2).
# Derived once from ``uuid5(NAMESPACE_DNS, "astrolabe.cloud/mcp/point-id/v1")``
# and pinned here as a literal so neither repo recomputes it. DO NOT CHANGE —
# see the module docstring for the cross-implementation contract.
NAMESPACE = uuid.UUID("b050b8ac-c6aa-5566-9584-506b39c1c096")
def point_id(tenant_id: str, doc_id: str, chunk_index: int) -> str:
"""Deterministic chunk point ID, identical across MCP server + processor.
``uuid5(NAMESPACE, canonical_json({...}))`` per design §2.2. The canonical
JSON name (sorted keys, no whitespace) must match the processor
byte-for-byte so re-indexing the same chunk upserts in place rather than
duplicating.
"""
name = canonical_json(
{"tenant_id": tenant_id, "doc_id": doc_id, "chunk_index": chunk_index}
).decode("utf-8")
return str(uuid.uuid5(NAMESPACE, name))
+17
View File
@@ -14,6 +14,7 @@ from anyio.streams.memory import MemoryObjectReceiveStream
from httpx import HTTPStatusError
from qdrant_client.models import FieldCondition, Filter, MatchValue, PointStruct
from nextcloud_mcp_server.acl_hash import compute_acl_hash
from nextcloud_mcp_server.client import NextcloudClient
from nextcloud_mcp_server.config import get_settings
from nextcloud_mcp_server.document_processors import get_registry
@@ -25,6 +26,7 @@ from nextcloud_mcp_server.observability.metrics import (
)
from nextcloud_mcp_server.observability.tracing import trace_operation
from nextcloud_mcp_server.search.pdf_highlighter import PDFHighlighter
from nextcloud_mcp_server.vector import payload_keys
from nextcloud_mcp_server.vector.document_chunker import DocumentChunker
from nextcloud_mcp_server.vector.html_processor import html_to_markdown
from nextcloud_mcp_server.vector.placeholder import delete_placeholder_point
@@ -672,6 +674,15 @@ async def _index_document(
indexed_at = int(time.time())
points = []
# Decomposition payload keys (design §10.2) — written even in local mode so
# a future migration to the external processor is friction-free. Computed
# once per document (not per chunk). The local processor has no triage, so
# PIPELINE_TIER is "fast"; ACL hash records at least the owner principal
# (full share enumeration is a follow-up — a missing/partial acl_hash is
# safe because the query-side pre-filter only applies when present + enabled).
_embedding_identity = get_settings().get_embedding_model_name()
_acl_hash = compute_acl_hash([("user", doc_task.user_id)])
# 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
@@ -719,6 +730,12 @@ async def _index_document(
"chunk_start_offset": chunk.start_offset,
"chunk_end_offset": chunk.end_offset,
"metadata_version": 2, # v2 includes position metadata
# Decomposition payload keys (design §10.2), additive.
payload_keys.PROCESSOR_VERSION: "monolith-v1",
payload_keys.PARSED_AT: indexed_at,
payload_keys.PIPELINE_TIER: "fast",
payload_keys.EMBEDDING_IDENTITY: _embedding_identity,
payload_keys.ACL_HASH: _acl_hash,
# File-specific metadata (PDF, etc.)
**(
{
@@ -0,0 +1,7 @@
"""Ingest-path ports & adapters (design §10, hexagonal)."""
from .factory import build_external_producer
from .memory import MemoryTaskProducer
from .ports import TaskProducer
__all__ = ["MemoryTaskProducer", "TaskProducer", "build_external_producer"]
@@ -0,0 +1,46 @@
"""Composition root for the ingest producer (design §10).
``build_external_producer`` is called by the lifespan only when
``INGEST_MODE=external``; local mode uses the in-memory stream directly (it
already satisfies :class:`TaskProducer`). The transport under ``external`` is
selected from the ``INGEST_BUS_URL`` scheme (``nats://`` now, ``postgres://``
later) so moving the external processor to Postgres needs no new INGEST_MODE.
"""
from __future__ import annotations
from urllib.parse import urlsplit
from ...config import Settings
from .ports import TaskProducer
def _transport_for(url: str) -> str:
scheme = urlsplit(url).scheme.lower()
if scheme.startswith("postgres"):
return "postgres"
return "nats"
async def build_external_producer(settings: Settings) -> TaskProducer:
"""Build the external-ingest producer for the configured transport.
Precondition: ``settings.ingest_mode == "external"`` (so __post_init__ has
guaranteed ``ingest_bus_url`` and ``tenant_id`` are set).
"""
assert settings.ingest_bus_url is not None
assert settings.tenant_id is not None
transport = _transport_for(settings.ingest_bus_url)
if transport == "postgres":
from .postgres import PostgresTaskProducer # noqa: PLC0415
return await PostgresTaskProducer.connect(settings)
from .nats import NatsTaskProducer # noqa: PLC0415
return await NatsTaskProducer.connect(
url=settings.ingest_bus_url,
tenant_id=settings.tenant_id,
num_replicas=settings.ingest_bus_num_replicas,
)
@@ -0,0 +1,45 @@
"""In-memory ``TaskProducer`` — the default (local) ingest transport.
A thin adapter over anyio's ``MemoryObjectSendStream`` so the local path is an
explicit :class:`TaskProducer` (rather than relying on structural typing of a
third-party class). Semantics are identical to using the stream directly:
``send`` enqueues, ``clone`` yields an independent per-user handle, ``async
with`` / ``aclose`` close the (cloned) send end so the processor pool's
receivers observe end-of-stream.
"""
from __future__ import annotations
from types import TracebackType
from typing import TYPE_CHECKING
from anyio.streams.memory import MemoryObjectSendStream
if TYPE_CHECKING:
from ..scanner import DocumentTask
class MemoryTaskProducer:
def __init__(self, stream: MemoryObjectSendStream[DocumentTask]):
self._stream = stream
async def send(self, task: DocumentTask, /) -> None:
await self._stream.send(task)
def clone(self) -> MemoryTaskProducer:
return MemoryTaskProducer(self._stream.clone())
async def __aenter__(self) -> MemoryTaskProducer:
await self._stream.__aenter__()
return self
async def __aexit__(
self,
exc_type: type[BaseException] | None,
exc: BaseException | None,
tb: TracebackType | None,
) -> None:
await self._stream.__aexit__(exc_type, exc, tb)
async def aclose(self) -> None:
await self._stream.aclose()
+147
View File
@@ -0,0 +1,147 @@
"""NATS JetStream ``TaskProducer`` — external ingest transport (design §3.4).
Publishes ``mcp.ingest.requested.{tenant_id}`` for the external
document-processor to consume. Translates the in-process ``DocumentTask`` into
the wire ``IngestMessage`` schema (mirrored in astrolabe-cloud-website's
``bus/messages.py``), with the JetStream ``Nats-Msg-Id`` dedup header per §3.4.
This server is only the *producer* on this transport; the document-processor
owns the consumer. ``nats-py`` is imported lazily so deployments that never
enable external ingest don't pay the import.
"""
from __future__ import annotations
import hashlib
import logging
from datetime import datetime, timezone
from types import TracebackType
from typing import TYPE_CHECKING, Any
from ...canonical import canonical_json
if TYPE_CHECKING:
from ..scanner import DocumentTask
logger = logging.getLogger(__name__)
STREAM_NAME = "mcp"
INGEST_SUBJECT_PREFIX = "mcp.ingest.requested"
def _modified_at_rfc3339(modified_at: int) -> str:
"""DocumentTask.modified_at is an epoch int (0 for deletes)."""
return datetime.fromtimestamp(int(modified_at), tz=timezone.utc).isoformat()
def _content_hash(task: DocumentTask) -> str:
"""etag is the change-detection token; fall back to modified_at when it is
absent (e.g. deletes, or sources whose etag we don't thread through)."""
return task.etag or str(task.modified_at)
def msg_id(tenant_id: str, doc_id: str, modified_at_rfc3339: str) -> str:
"""JetStream dedup header per §3.4. SHA-256 over canonical JSON (NOT the
BLAKE2b helper) — it is an opaque external header, not a stored field."""
return hashlib.sha256(
canonical_json(
{
"tenant_id": tenant_id,
"doc_id": doc_id,
"modified_at": modified_at_rfc3339,
}
)
).hexdigest()
class NatsTaskProducer:
"""Publishes ingest requests to NATS JetStream."""
def __init__(self, nc: Any, js: Any, tenant_id: str):
self._nc = nc
self._js = js
self.tenant_id = tenant_id
@classmethod
async def connect(
cls, *, url: str, tenant_id: str, num_replicas: int = 1
) -> NatsTaskProducer:
import nats # noqa: PLC0415 (lazy: optional dependency for external mode)
nc = await nats.connect(url)
js = nc.jetstream()
await cls._ensure_stream(js, num_replicas)
logger.info("Connected NATS ingest producer: url=%s, tenant=%s", url, tenant_id)
return cls(nc, js, tenant_id)
@staticmethod
async def _ensure_stream(js: Any, num_replicas: int) -> None:
# noqa: PLC0415 — nats.js types are only importable once nats-py is present.
from nats.js.api import RetentionPolicy, StreamConfig # noqa: PLC0415
config = StreamConfig(
name=STREAM_NAME,
subjects=["mcp.>"],
retention=RetentionPolicy.LIMITS,
num_replicas=num_replicas,
)
try:
await js.add_stream(config=config)
logger.info("nats.stream_created stream=%s", STREAM_NAME)
except Exception as exc:
# add_stream is idempotent in spirit but errors when the stream
# already exists; treat as benign (mirrors the processor's
# ensure_stream). A genuinely broken broker surfaces on publish.
logger.info("nats.stream_exists_or_unavailable detail=%s", exc)
def ingest_message(self, task: DocumentTask) -> dict[str, Any]:
"""DocumentTask → wire IngestMessage dict (mirrors the sibling schema)."""
return {
"tenant_id": self.tenant_id,
"doc_id": task.doc_id,
"content_hash": _content_hash(task),
"modified_at": _modified_at_rfc3339(task.modified_at),
"doc_type": task.doc_type,
"operation": task.operation,
"user_id": task.user_id,
"file_path": task.file_path,
}
async def send(self, task: DocumentTask) -> None:
message = self.ingest_message(task)
subject = f"{INGEST_SUBJECT_PREFIX}.{self.tenant_id}"
headers = {
"Nats-Msg-Id": msg_id(self.tenant_id, task.doc_id, message["modified_at"])
}
await self._js.publish(subject, canonical_json(message), headers=headers)
# The scanner/oauth_sync use the producer as a clone-able async context
# manager (memory-stream semantics). The bus connection is owned by the
# lifespan, so cloning shares it and __aexit__ is a no-op (close happens via
# aclose() on shutdown).
def clone(self) -> NatsTaskProducer:
return self
async def __aenter__(self) -> NatsTaskProducer:
return self
async def __aexit__(
self,
exc_type: type[BaseException] | None,
exc: BaseException | None,
tb: TracebackType | None,
) -> None:
return None
async def aclose(self) -> None:
# Per-handle close (e.g. a per-user scanner clone exiting). The bus
# connection is shared and owned by the lifespan, so this is a no-op;
# the connection is torn down once via ``drain()`` on shutdown.
return None
async def drain(self) -> None:
"""Drain + close the shared NATS connection (lifespan shutdown only)."""
try:
await self._nc.drain()
except Exception:
logger.warning("NATS drain on shutdown failed", exc_info=True)
@@ -0,0 +1,59 @@
"""Ingest-path ports (design §10, hexagonal).
A ``TaskProducer`` is where the scanner + webhook receiver send a
``DocumentTask``. The transport behind it is swappable:
- the in-process anyio ``MemoryObjectSendStream`` (local ingest — the default),
- ``NatsTaskProducer`` (external ingest → the document-processor), and
- a future Postgres-queue producer (seam only; the *external* processor owns the
consume side — see ``postgres.py``).
The protocol is exactly the surface the scanner/oauth_sync already use on the
memory stream (``send`` + ``clone`` + ``async with``), so both adapters drop in
with only a type-annotation change at the call sites. There is intentionally NO
consumer port: the MCP server's only in-process consumer is the memory stream;
when ingest is external the document-processor is the consumer, not this server.
"""
from __future__ import annotations
from types import TracebackType
from typing import TYPE_CHECKING, Protocol, runtime_checkable
if TYPE_CHECKING:
from ..scanner import DocumentTask
@runtime_checkable
class TaskProducer(Protocol):
"""Sink for scanner/webhook ``DocumentTask``s (see module docstring)."""
# Positional-only so anyio's MemoryObjectSendStream.send(item) structurally
# satisfies this protocol (its parameter is named "item", not "task").
async def send(self, task: DocumentTask, /) -> None: ...
def clone(self) -> TaskProducer:
"""Return a producer handle for one user's scanner (multi-user mode).
For the memory stream this is a real clone (each closed independently);
for the bus it returns ``self`` (one shared connection).
"""
...
async def __aenter__(self) -> TaskProducer: ...
async def __aexit__(
self,
exc_type: type[BaseException] | None,
exc: BaseException | None,
tb: TracebackType | None,
) -> None: ...
async def aclose(self) -> None:
"""Close *this* handle (e.g. a per-user clone when its scanner exits).
For the memory stream this closes the clone; for the shared bus
connection it is a no-op (the connection is owned by the lifespan,
which drains it once on shutdown).
"""
...
@@ -0,0 +1,49 @@
"""Postgres-queue ``TaskProducer`` — documented seam, not implemented.
The external document-processor may later drain a Postgres-backed queue instead
of NATS to limit NATS operational overhead. Processing stays *external*; only
the transport changes — so on this server it would be a drop-in producer swap.
The consume side + the queue-table migration belong to that processor-side
refactor (cross-repo), NOT here. This stub exists so the transport value and the
``TaskProducer`` Protocol conformance are testable today.
"""
from __future__ import annotations
from types import TracebackType
from typing import TYPE_CHECKING, Any
if TYPE_CHECKING:
from ..scanner import DocumentTask
_NOT_IMPLEMENTED = (
"Postgres ingest transport is a documented seam. The external "
"document-processor owns the Postgres-drain refactor (transport swap only; "
"processing stays external). Use INGEST_BUS_URL=nats://… for now."
)
class PostgresTaskProducer:
@classmethod
async def connect(cls, settings: Any) -> PostgresTaskProducer:
raise NotImplementedError(_NOT_IMPLEMENTED)
async def send(self, task: DocumentTask) -> None: # pragma: no cover
raise NotImplementedError(_NOT_IMPLEMENTED)
def clone(self) -> PostgresTaskProducer: # pragma: no cover
return self
async def __aenter__(self) -> PostgresTaskProducer: # pragma: no cover
return self
async def __aexit__(
self,
exc_type: type[BaseException] | None,
exc: BaseException | None,
tb: TracebackType | None,
) -> None: # pragma: no cover
return None
async def aclose(self) -> None: # pragma: no cover
return None
+163
View File
@@ -0,0 +1,163 @@
"""Status surface for ingest jobs (design §10.1, ``STATUS_BACKEND``).
- ``local``: in-process job state — the memory-stream buffer (today's behavior,
read directly by the status endpoint).
- ``bus``: a background subscriber consumes
``mcp.document.{ready,failed,reparsed}.{tenant_id}`` into a bounded in-process
:class:`StatusStore` that the status endpoint / ``nc_get_vector_sync_status``
read.
**Honest constraint (design §10.2 / decision):** MCP progress notifications
(``ctx.report_progress``) can only be emitted inside an *active tool-call
request*; a background subscriber has no ``ctx`` and the MCP SDK exposes no
out-of-band push. So "surface events as MCP progress notifications" is delivered
via this store (polled by the status endpoint / a tool), not an unsolicited
server push. True server-initiated progress / SSE is a follow-up — the
``on_event`` callback seam is left in place for it.
"""
from __future__ import annotations
import logging
from collections import OrderedDict
from typing import TYPE_CHECKING, Any, Callable
if TYPE_CHECKING:
import anyio
from anyio.abc import TaskStatus
logger = logging.getLogger(__name__)
# Terminal/intermediate document states carried on mcp.document.* subjects.
_VALID_STATES = {"ready", "failed", "reparsed"}
class StatusStore:
"""Bounded LRU of recent document states keyed by ``doc_id``."""
def __init__(self, max_size: int = 10_000):
self._entries: OrderedDict[str, dict[str, Any]] = OrderedDict()
self._max = max_size
def record(
self,
doc_id: str,
state: str,
*,
content_hash: str | None = None,
transitioned_at: str | None = None,
) -> None:
self._entries[doc_id] = {
"state": state,
"content_hash": content_hash,
"transitioned_at": transitioned_at,
}
self._entries.move_to_end(doc_id)
while len(self._entries) > self._max:
self._entries.popitem(last=False)
def get(self, doc_id: str) -> dict[str, Any] | None:
return self._entries.get(doc_id)
def counts(self) -> dict[str, int]:
out: dict[str, int] = {}
for entry in self._entries.values():
out[entry["state"]] = out.get(entry["state"], 0) + 1
return out
def __len__(self) -> int:
return len(self._entries)
def state_from_subject(subject: str) -> str | None:
"""``mcp.document.<state>.<tenant_id>`` → ``<state>`` (or None if unknown)."""
parts = subject.split(".")
if len(parts) >= 4 and parts[0] == "mcp" and parts[1] == "document":
state = parts[2]
if state in _VALID_STATES:
return state
return None
class NatsStatusSubscriber:
"""Consumes ``mcp.document.*.{tenant_id}`` into a :class:`StatusStore`."""
def __init__(
self,
nc: Any,
js: Any,
tenant_id: str,
store: StatusStore,
on_event: Callable[[str, str], None] | None = None,
):
self._nc = nc
self._js = js
self.tenant_id = tenant_id
self.store = store
# on_event(doc_id, state) — seam for a future SSE / progress bridge.
self._on_event = on_event
def handle_message(self, subject: str, data: bytes) -> None:
"""Parse one status message into the store. Unit-testable without NATS."""
import json # noqa: PLC0415
state = state_from_subject(subject)
if state is None:
logger.warning("status.unknown_subject subject=%s", subject)
return
try:
payload = json.loads(data)
doc_id = payload["doc_id"]
except Exception:
logger.warning("status.bad_message subject=%s", subject, exc_info=True)
return
self.store.record(
doc_id,
state,
content_hash=payload.get("content_hash"),
transitioned_at=payload.get("transitioned_at"),
)
if self._on_event is not None:
self._on_event(doc_id, state)
@classmethod
async def connect(
cls, *, url: str, tenant_id: str, store: StatusStore
) -> NatsStatusSubscriber:
import nats # noqa: PLC0415
nc = await nats.connect(url)
js = nc.jetstream()
return cls(nc, js, tenant_id, store)
async def run(
self,
shutdown_event: anyio.Event,
*,
task_status: TaskStatus = None, # type: ignore[assignment]
) -> None:
"""Durable pull-consumer loop. Requires a live broker (integration)."""
import anyio # noqa: PLC0415
subject = f"mcp.document.*.{self.tenant_id}"
sub = await self._js.pull_subscribe(
subject, durable=f"mcp-status-{self.tenant_id}"
)
if task_status is not None:
task_status.started()
while not shutdown_event.is_set():
try:
msgs = await sub.fetch(batch=16, timeout=5)
except Exception:
# fetch timeout when idle — loop and re-check shutdown.
await anyio.sleep(0)
continue
for msg in msgs:
self.handle_message(msg.subject, msg.data)
await msg.ack()
async def aclose(self) -> None:
try:
await self._nc.drain()
except Exception:
logger.warning("NATS status subscriber drain failed", exc_info=True)
+12 -5
View File
@@ -13,7 +13,6 @@ from typing import cast
import anyio
from anyio.abc import TaskStatus
from anyio.streams.memory import MemoryObjectSendStream
from qdrant_client import AsyncQdrantClient
from qdrant_client.models import FieldCondition, Filter, MatchValue, Record
@@ -31,6 +30,7 @@ from nextcloud_mcp_server.vector.placeholder import (
write_placeholder_point,
)
from nextcloud_mcp_server.vector.qdrant_client import get_qdrant_client
from nextcloud_mcp_server.vector.queue.ports import TaskProducer
logger = logging.getLogger(__name__)
@@ -108,6 +108,11 @@ class DocumentTask:
metadata: dict[str, int | str] | None = (
None # Additional metadata (e.g., board_id/stack_id for deck_card)
)
# Change-detection token (Nextcloud etag). Used as the external ingest
# content_hash when present; the bus producer falls back to modified_at
# when it is None (deletes, or sources whose etag isn't threaded). Harmless
# in local mode — the in-process processor reads its own etag.
etag: str | None = None
# Track documents potentially deleted (grace period before actual deletion)
@@ -179,7 +184,7 @@ async def get_last_indexed_timestamp(user_id: str) -> int | None:
async def scanner_task(
send_stream: MemoryObjectSendStream[DocumentTask],
send_stream: TaskProducer,
shutdown_event: anyio.Event,
wake_event: anyio.Event,
nc_client: NextcloudClient,
@@ -233,7 +238,7 @@ async def scanner_task(
async def scan_user_documents(
user_id: str,
send_stream: MemoryObjectSendStream[DocumentTask],
send_stream: TaskProducer,
nc_client: NextcloudClient,
initial_sync: bool = False,
):
@@ -339,6 +344,7 @@ async def scan_user_documents(
doc_type="note",
operation="index",
modified_at=modified_at,
etag=note.get("etag", ""),
)
)
queued += 1
@@ -405,6 +411,7 @@ async def scan_user_documents(
doc_type="note",
operation="index",
modified_at=modified_at,
etag=note.get("etag", ""),
)
)
queued += 1
@@ -740,7 +747,7 @@ async def scan_user_documents(
async def scan_news_items(
user_id: str,
send_stream: MemoryObjectSendStream[DocumentTask],
send_stream: TaskProducer,
nc_client: NextcloudClient,
initial_sync: bool,
scan_id: int,
@@ -928,7 +935,7 @@ async def scan_news_items(
async def scan_deck_cards(
user_id: str,
send_stream: MemoryObjectSendStream[DocumentTask],
send_stream: TaskProducer,
nc_client: NextcloudClient,
initial_sync: bool,
scan_id: int,
@@ -41,8 +41,9 @@ def _warn_missing_secret_once() -> None:
async def handle_nextcloud_webhook(request: Request) -> JSONResponse:
"""Receive a Nextcloud webhook and queue a DocumentTask for vector sync.
Returns quickly so NC's webhook worker is not blocked. The send-stream is
read from ``request.app.state.document_send_stream``; when vector sync
Returns quickly so NC's webhook worker is not blocked. The task producer is
read from ``request.app.state.task_producer`` (the in-memory send stream in
local mode, or the NATS bus producer in external mode); when vector sync
isn't running we return 503 so NC retries delivery.
When ``WEBHOOK_SECRET`` is set, the request must carry
@@ -93,8 +94,8 @@ async def handle_nextcloud_webhook(request: Request) -> JSONResponse:
status_code=200,
)
send_stream = getattr(request.app.state, "document_send_stream", None)
if send_stream is None:
producer = getattr(request.app.state, "task_producer", None)
if producer is None:
logger.warning(
"Webhook received but vector sync is not running; rejecting so NC retries"
)
@@ -105,7 +106,7 @@ async def handle_nextcloud_webhook(request: Request) -> JSONResponse:
try:
with anyio.fail_after(1.0):
await send_stream.send(task)
await producer.send(task)
except TimeoutError:
# Queue is saturated (default 10 000 tasks). Returning 503 lets NC
# retry rather than pinning this handler until its outbound timeout