ADR-0010 — RAG as a contract, implementations by the developer

• Status: Accepted
• Issue: #88

Context

Agents could only read whole files (MCP) or call APIs. Every project embeds and stores documents differently, so a built-in embeddings model or vector store would be the wrong default.

Decision

• The core defines RagConnector (retrieve, optional index); a @Rag class implements it (k required); agents bind it with a required mode — tool or context.
• The core does the wiring, cost reporting (category retrieval) and tracing; context retrieval fails open.
• The reference implementation is SQLite FTS5 (BM25): no dependencies, no cost, deterministic in tests.

Consequences

• Semantic search is another connector under the same contract; swapping touches one file.
• The reference finds words, not synonyms.