Knowledge Base

The Knowledge Base lets you ingest documents (PDFs, markdown, text) into a searchable vector store that agents can query. It powers RAG (Retrieval-Augmented Generation) — agents get relevant context from your documents when answering questions.

How It Works

Documents → Chunking → Embedding → SQLite Vector Store
                                          │
                              Agent query → Semantic search → Context injection

1. Ingest — Upload documents or point to files/URLs
2. Chunk — Documents are split into semantic chunks
3. Embed — Chunks are converted to vector embeddings
4. Store — Vectors stored in SQLite with full-text search
5. Search — Agents query the knowledge base; relevant chunks are injected into context

Agent Tools

Agents interact with the knowledge base through 4 tools:

knowledge_ingest

Add documents to the knowledge base:

knowledge({
  action: "ingest",
  source: "/path/to/document.pdf",
  metadata: {
    category: "product-specs",
    version: "2.0"
  }
})

knowledge_search

Search for relevant documents:

knowledge({
  action: "search",
  query: "What is the pricing for enterprise plans?",
  limit: 5
})

Returns ranked chunks with relevance scores:

{
  "results": [
    {
      "content": "Enterprise plans start at $499/month...",
      "source": "pricing-guide.md",
      "score": 0.92,
      "metadata": { "category": "product-specs" }
    }
  ]
}

knowledge_list

List all ingested documents:

knowledge({ action: "list" })

knowledge_delete

Remove a document and its chunks:

knowledge({ action: "delete", documentId: "doc_abc123" })

Deletion cascades — removing a document deletes all associated chunks and embeddings.

Chunking Strategy

The semantic chunker splits documents intelligently:

• Respects headers — Markdown headers create natural chunk boundaries
• Paragraph-aware — Doesn't split mid-paragraph
• Size limits — Chunks target ~500-1000 tokens for optimal retrieval
• Overlap — Adjacent chunks overlap slightly to preserve context at boundaries

Deduplication

The ingestion pipeline deduplicates documents by content hash. Re-ingesting the same document updates metadata without creating duplicate chunks.

Storage

Knowledge is stored in SQLite at ~/.reeve/memory/memory.db alongside Semantic Memory. The schema includes:

• documents table — Source files with metadata
• chunks table — Document chunks with embeddings
• Full-text search index — For keyword fallback