The hot-path classifier escalated to OCR purely on character count, so a
scanned/handwritten PDF with a low-quality embedded text layer (>16 chars/page
but garbled) routed `fast` and indexed the junk -- e.g. Student 147.pdf's
"Little Acoms Primary"/"0110912020", which pollutes the vector and demotes the
doc in search (Deck #207).
- classifier: recalibrate `_text_quality` with a long-token-fraction term that
detects word-merging (dropped inter-word spaces) -- the dominant junk-layer
failure the old whitespace/overlong(>20) terms missed. Measured: the Student
147 scan ~0.42 (60% pages junk) vs >=0.94 for clean digital docs.
- classify_from_text now routes on quality + scan: a page is OCR-worthy if
near-empty OR low text-quality OR (when OCR + scan detection are enabled) it's
mostly a raster image. New `image_coverage_per_page` re-opens the PDF for the
scan signal, so that cost is paid only by OCR-opted-in tenants. Thresholds are
passed in from per-tenant settings (keyword-only).
- config: 4 per-tenant settings -- DOCUMENT_OCR_MIN_TEXT_QUALITY (0.5),
DOCUMENT_OCR_PAGE_FRACTION (0.5), DOCUMENT_OCR_MIN_PAGE_CHARS (16),
DOCUMENT_OCR_DETECT_SCANNED (true) -- with range validators.
- metrics: new astrolabe_document_ocr_page_fraction histogram (the value the
page-fraction threshold acts on) alongside document_text_quality, so operators
can tune the OCR escalation per tenant (quality vs cost).
Escalation gate, OCR backends, and off-by-default behavior unchanged (#858).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Address PR #858 review round 2:
- OcrProcessor backend resolution is now guarded by an anyio.Lock (lazy-init,
double-checked) so a burst of concurrent first-OCR calls resolves the backend
once instead of each fetching its own gateway M2M token.
- The document_tier1_engine=pymupdf rollback now logs a warning when it falls
back to the fast processor (no 'structured' registered) instead of silently
using the very engine the operator opted out of.
- classify_from_text defaults ocr_frac to 0.0 (not 1.0) for a zero-page PDF, so
the recorded classification metric is "fast" (no OCR evidence) rather than a
misleading "ocr"; the no_text_layer/bad_text_layer flags are gated on having
sampled at least one page.
New tests: zero-page classify routes fast, rollback-fallback warning.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Address PR #858 review:
- 🔴 OcrProcessor now resolves its backend once and reuses it. Rebuilding per
call created a fresh GatewayTokenProvider each time -- discarding its M2M-token
cache, so every OCR'd document fetched a new token -- and a new Mistral client.
- 🔴 build_ocr_backend uses explicit ValueError (not assert, which is stripped
under `python -O`) for the gateway M2M triple.
- PIPELINE_TIER in the Qdrant payload now reflects the tier that actually
produced the doc: the registry stamps result.metadata["pipeline_tier"] and the
processor reads it (was hardcoded "fast", wrong for OCR/structured).
- Escalation now requires classification.page_count > 0, so a zero-page
(empty/corrupt) PDF isn't pointlessly sent to OCR; documented that a fast
FAILURE (encrypted/unopenable) is a hard failure and is not OCR-escalated.
- Documented the OCR page_boundaries separator-attribution choice.
- Downgraded the per-document page-boundary / page-assignment INFO logs to debug.
New tests: zero-page no-escalation, pipeline_tier stamping.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
OCR is an enhancement, not a gate. Previously, escalating a scanned doc to the
OCR tier returned the OCR result unconditionally -- so with DOCUMENT_OCR_ENABLED
=true but no backend configured (no gateway URL / no MISTRAL_API_KEY) the OCR
processor returned success=False and the whole document was marked failed and
skipped: strictly worse than leaving OCR off (where it would at least index the
tier-1 text).
Now the registry keeps the tier-1 fast result when the OCR escalation doesn't
succeed (no backend, API down, empty output), logging a warning. A
misconfiguration degrades gracefully instead of dropping scanned docs.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Replaces single-engine pymupdf4llm extraction with a tiered pipeline (Deck #205,
follows the tier-0 classifier #855). pypdfium2 becomes the default and only
hot-path PDF extractor; pymupdf4llm is deprecated to a rollback toggle.
Why: pymupdf4llm's O(n^2) find_tables drove the OOM (#852) and the form-PDF
parse timeouts (#856), carries AGPL/commercial licensing liability, and -- per
the benchmarks -- recovers near-zero usable tables on the real corpus. pypdfium2
(Apache/BSD) extracts the same text far faster (Student 1a.pdf: 120s timeout ->
0.2s) with no table-detection bomb.
- document_processors/pypdfium2_fast.py: tier-1 "fast" processor emitting text +
exact page_boundaries (the pdf_highlighter contract). pymupdf processor is now
tier "structured" (the rollback engine), registered but not default.
- registry: tiered routing in ProcessorRegistry. tier-1 fast extracts, then
classification is DERIVED from that text (classifier.classify_from_text -- no
PDF re-open), records the classification metrics, and escalates scanned /
no-text-layer docs to the "ocr" tier when document_ocr_enabled (default off;
no provider yet, so fast is terminal). Wires record_document_escalation + the
real "escalated" span attribute (was hardcoded False).
- Removes the separate _shadow_classify pass from vector/processor.py -- it
re-opened every PDF and re-extracted text (~0.5-1.3s/doc of pure duplicated
CPU that lowered throughput); classification now rides the tier-1 extraction.
- Settings: document_tier1_engine ("pypdfium2" default | "pymupdf" rollback,
enum-validated), document_ocr_enabled (default false).
Tests: pypdfium2 extractor, registry tiering (fast routing, rollback, classify
recording, OCR escalation on/off), classify_from_text. Full unit suite green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
