XiaoyaoClaw KB Retriever
Local KB retriever — layered index navigation + progressive search. Zero-dep, zero-API-key, md/pdf/xlsx covered, sensitive data stays on-device.
Product Introduction
OpenClaw KB Retriever solves four local-KB pains: full-file reads blowing up tokens, blind searches without index, wrong-tool mistakes on PDF/Excel, and RAG solutions that depend on cloud embeddings or external API keys. It uses layered index navigation (a data_structure.md per dir, drill-down instead of full scan), progressive search (grep/Select-String → offset/limit windowed read → max 5 iterations, never load full file), learn-before-do (force-read references for PDF/Excel first), and fully-local zero-API-key (no vector index, no cloud API). Windows / macOS parity (PDF via Python pdfplumber). One-script index bootstrap. Answers ship with file + location citations — verifiable, not invented. Forked from ConardLi/garden-skills kb-retriever (MIT).
Core Features
- Layered index navigation — a data_structure.md per dir, drill down the index tree
- Progressive search — grep/Select-String locate → offset/limit windowed read → max 5 iterations
- Learn-before-do — force-read references for PDF/Excel before touching them
- Local-only zero API key — no vector index, no cloud API; only whitelisted Python pkgs on demand
- Cross-platform parity — Python pdfplumber, identical Windows / macOS behavior
- One-command bootstrap — python build_index.py generates the index skeleton
- Source citations — every answer carries file + location, verifiable
- md / pdf / xlsx — unified search across docs, sheets, and resource dirs