Knowledge Extractor
Extracts agent-accessible knowledge from document file trees into clean Markdown.
- Microsoft Word (.docx)
- Microsoft PowerPoint (.pptx)
- Microsoft Excel (.xlsx)
- PDF (.pdf)
- Images (.jpg, .jpeg, .png)
uv sync
cp .env.example .env # add your OpenRouter API key
uv run python main.py --input ./input --output ./output --temp ./temp --model google/gemini-2.5-flash
| Flag |
Default |
Description |
--input |
(required) |
Input directory with source documents |
--output |
./output |
Output directory for final Markdown files |
--temp |
./temp |
Intermediate data directory (debug artifacts) |
--model |
google/gemini-2.5-flash |
OpenRouter vision model for image analysis |
--dry-run |
false |
List which files would be processed and which would be skipped, then exit without processing |
| Variable |
Description |
OPENROUTER_API_KEY |
OpenRouter API key. If not set, AI steps are skipped and raw extractions are output. |
- Incremental processing — skips a source file if its output Markdown already exists; delete the output file (or the output dir) to re-run
- Intermediate results — per-file markdown with original image references saved to temp dir
- AI image analysis — converts diagrams to Mermaid, charts to descriptions (requires API key)
- OCR for scanned PDFs — auto-detects scanned/image-only PDFs and extracts text via AI vision OCR (requires API key)
- LaTeX formula extraction — detects mathematical formulas in DOCX/PPTX (OMML) and PDF (heuristic font/character analysis), converts to LaTeX via AI vision. Inline formulas wrapped with
$...$, display formulas with $$...$$. Requires API key; without it, a placeholder is inserted.
- Heuristic filtering — removes title slides, logo-only sections, repeated headers
- Markdown linting — auto-fixes formatting issues (trailing spaces, heading spacing, list indentation) via pymarkdownlnt
- Logging — console output +
output/extraction.log
output/index.md — flat index grouped by original folder structureoutput/**/*.md — one Markdown file per source document