Skip to content

RAG & Knowledge Base

Incremental Indexing

Indexes .cs, .ts, .js, .py, .go, .md, .json files. Only re-indexes changed files (5-minute full re-index interval).

  • File watcher triggers incremental updates
  • Content chunked by semantic boundaries (functions, classes)
  • Embeddings generated via local or API embedding models
  • Vector storage in SQLite with FAISS-compatible format

Context Budget

Max 700 characters injected per step, 2 chunks per step. Symbol extraction (class, method, property names) as chunk metadata.

  • Relevance scoring combines vector similarity + symbol overlap
  • Deduplication prevents duplicate context
  • Priority: current file > recent edits > semantic matches

Knowledge Service

Persistent knowledge base of SOPs and learned patterns. Stored as markdown in knowledge/. Agents can add new items at runtime.

  • knowledge.list - List all knowledge items
  • knowledge.read - Read a knowledge item by key
  • knowledge.add - Create new knowledge item (SOP)
  • Auto-indexed for RAG retrieval