Problem
The worst silent failure in the codebase. pypdf's extract_text() returns an empty string for image-only pages, so a scanned 90-page contract flows through the happy path and produces char_count: 0, token_count: 0, error: None. The UI shows a completed analysis of an empty document. A user checking whether a scanned contract fits in context gets "yes, 0 tokens" and acts on it.
Where
parsing.py:32 — "\n".join(page.extract_text() or "" ...)service.py:56 — builds a normal AnalysisResult, error=None
Fix
Detect a PDF that parsed with pages but near-zero extractable text, and surface it as a distinct warning state ("this PDF appears to be scanned images, no text layer found"), not as a result. Deliberately not an error, since the file did parse. OCR is a separate enhancement.
Impact
Silent undercounting is the one failure mode this tool cannot afford, since avoiding a surprise bill is the whole point.
Problem
The worst silent failure in the codebase.
pypdf'sextract_text()returns an empty string for image-only pages, so a scanned 90-page contract flows through the happy path and produceschar_count: 0, token_count: 0, error: None. The UI shows a completed analysis of an empty document. A user checking whether a scanned contract fits in context gets "yes, 0 tokens" and acts on it.Where
parsing.py:32—"\n".join(page.extract_text() or "" ...)service.py:56— builds a normal AnalysisResult, error=NoneFix
Detect a PDF that parsed with pages but near-zero extractable text, and surface it as a distinct warning state ("this PDF appears to be scanned images, no text layer found"), not as a result. Deliberately not an error, since the file did parse. OCR is a separate enhancement.
Impact
Silent undercounting is the one failure mode this tool cannot afford, since avoiding a surprise bill is the whole point.