OCR & document extraction

Coding work often touches non-code artifacts — scanned specs, screenshots, PDFs, diagrams with formulas. OCR/extraction tooling lets Pi pull structured text out of them and reason over it.

Alternatives

1. pi-ocr

A multi-backend OCR extension. Extract text, formulas, and tables from images and PDFs, with three backends: MinerU (free cloud), Ollama (local GPU, including LaTeX formulas), and Pix2Text (local Python). Works zero-config by default.

pi install npm:pi-ocr

Benefits

• Multiple backends — cloud or fully local, depending on privacy/compute needs.
• Handles formulas and tables, not just plain text.
• Zero-config out of the box.

Drawbacks

• Local backends require GPU (Ollama) or a Python setup (Pix2Text).
• Cloud backend sends documents off-machine — mind sensitive data.

2. pi-web-access (PDF extraction subset)

If your documents are mostly PDFs reachable by URL, pi-web-access already includes PDF extraction alongside web search and fetching — one package for research and PDFs.

pi install npm:pi-web-access

Benefits

• One install covers PDFs + web research.
• No separate OCR stack to manage.

Drawbacks

• PDF text extraction only — no image OCR, formulas, or tables.
• Struggles with scanned (image-only) PDFs.

3. An MCP document server via pi-mcp-adapter

For specialized pipelines, run a dedicated document/OCR MCP server and connect it with pi-mcp-adapter.

pi install npm:pi-mcp-adapter

Benefits

• Reuse best-in-class MCP document servers.
• Swap servers without changing your Pi setup.

Drawbacks

• You provision and run the MCP server yourself.
• More setup than a turnkey extension.

Which should I pick?

• Need real OCR (images, scans, formulas, tables): pi-ocr.
• Only need text from URL-based PDFs: pi-web-access.
• Have a preferred document pipeline already: wrap it as an MCP server and use pi-mcp-adapter.