When reprocessing a document whose PDF is already in the store,
copyFile() would fail with 'open /srv/docsys/inbox/...: no such file
or directory' because the upload wrote to a temp inbox path that was
already cleaned up by the time async OCR completed.
The store is keyed by content hash so if the file is already there,
the copy is a no-op — skip it rather than error out.
pdftoppm zero-pads the page number based on total page count:
- <10 pages: page-1.png
- <100 pages: page-01.png
- <1000 pages: page-001.png
The code hardcoded 'page-1.png' and 'page-N.png', which fails for any
multi-page document. Use filepath.Glob('page-*.png') to find the actual
output regardless of padding width.
Fixed in both ConvertToImage() (first-page preview) and the multi-page
OCR loop in ProcessDocument().
Replaces kimi-k2p5 for all vision tasks. K2.5 was outputting chain-of-thought
reasoning instead of JSON for non-English docs, requiring a fallback path.
qwen3-vl works first try, no retry needed, preserves original language correctly.
- Add 'DO NOT translate, preserve original language' to vision prompts
- Shorter/tighter JSON prompt to reduce K2.5 reasoning verbosity
- Fallback: when AnalyzeWithVision returns no JSON, do AnalyzePageOnly (plain text) then AnalyzeText (classify)
- Fallback to AnalyzePageOnly for single-page PDFs with empty/placeholder full_text
- Switch model back to kimi-k2p5 (only vision model on this Fireworks account)
- Build with CGO_ENABLED=1 -tags fts5 (required for SQLite FTS5)
James