01
Curated, not crawled
pasclaw kb add ~/docs/ registers documents in place into workspace/kb.db. They are never copied. The schema supports markdown, plain text, HTML, and common source-code formats.
// Knowledgebase
RAG over operator-curated documents.
The knowledgebase is a curated, local RAG corpus separate from session memory. Operators register documents with pasclaw kb add, and the agent gets two read-only tools — kb_search and kb_get — that auto-register once at least one source exists.
FTS5 BM25Hybrid + vectorIn-place indexingNative PDF ingest
02
Hybrid retrieval
FTS5 BM25 ranking is the default. When the local embedding runtime is provisioned with pasclaw memory provision (sqlite-vec + ONNX Runtime + MiniLM), kb_search auto-upgrades to hybrid keyword + vector with Reciprocal Rank Fusion.
03
Explicit refresh
Unlike memory_search, the KB does not re-index on every query. pasclaw kb sync is the explicit refresh, keeping kb_search fast on large corpora.
04
Native PDF text
Pure-Pascal PDF ingest parses text via FlateDecode + /ToUnicode lookup, so pasclaw kb add file.pdf works without Poppler or pdftotext. Scanned PDFs still need external OCR.
From the documentation
Technical details
Implementation notes drawn from the PasClaw repository documentation.
Two read-only tools
kb_search and kb_get auto-register once at least one source is added. Indexing happens in workspace/kb.db; documents stay in place.
- FTS5 BM25 default, hybrid + vector once memory provision is run.
- pasclaw kb sync is the explicit refresh — searches stay fast on large corpora.
- Supports markdown, text, HTML, common source code, and native PDF text.
Source documentation
Common questions
Frequently asked questions
Keep exploring
See how the rest of PasClaw fits together.