Boof 🍑

Local-first document processing: PDF → markdown → RAG index → token-efficient analysis.

Documents stay local. Only relevant chunks go to the LLM. Maximum knowledge absorption, minimum token burn.

Powered by opendataloader-pdf — #1 in PDF parsing benchmarks (0.90 overall, 0.93 table accuracy). CPU-only, no GPU required.

Quick Reference

Convert + index a document

bash {SKILL_DIR}/scripts/boof.sh /path/to/document.pdf

Convert with custom collection name

bash {SKILL_DIR}/scripts/boof.sh /path/to/document.pdf --collection my-project

Query indexed content

qmd query "your question" -c collection-name

Core Workflow

1. Boof it: Run boof.sh on a PDF. This converts it to markdown via opendataloader-pdf (local Java engine, no API, no GPU) and indexes it into QMD for semantic search.

1. Query it: Use qmd query to retrieve only the relevant chunks. Send those chunks to the LLM — not the entire document.

1. Analyze it: The LLM sees focused, relevant excerpts. No wasted tokens, no lost-in-the-middle problems.

When to Use Each Approach

"Analyze this specific aspect of the paper" → Boof + query (cheapest, most focused)

"Summarize this entire document" → Boof, then read the markdown section by section. Summarize each section individually, then merge summaries. See advanced-usage.md.

"Compare findings across multiple papers" → Boof all papers into one collection, then query across them.

"Find where the paper discusses X" → qmd search "X" -c collection for exact match, qmd query "X" -c collection for semantic match.

Output Location

Converted markdown files are saved to knowledge/boofed/ by default (override with --output-dir).

Setup

If boof.sh reports missing dependencies, see setup-guide.md for installation instructions (Java + opendataloader-pdf + QMD).

Environment

• ODL_ENV — Path to opendataloader-pdf Python venv (default: ~/.openclaw/tools/odl-env)
• QMD_BIN — Path to qmd binary (default: ~/.bun/bin/qmd)
• BOOF_OUTPUT_DIR — Default output directory (default: ~/.openclaw/workspace/knowledge/boofed)