rag-context

# Retrieval-Augmented Context (Dynamic RAG) ## Trigger conditions - User mentions RAG, retrieval, context injection, knowledge base - Task requires info not in active conversation (code reference, API docs, config) - User asks "check X" or "look up Y" ## Retrieval strategy (ordered by cost) ### 1. Exact match (fastest) - Grep/RG for known symbols, filenames, error strings - Use glob for file patterns ### 2. Semantic (codebase) - Read file headers, imports, exports, type signatures - Grep for function/class definitions matching intent ### 3. Structural (docs/config) - Read README, package.json, tsconfig, config YAML - Scan directory tree for relevant paths ### 4. External (web) - Fetch docs URLs, search web for API references - Fetch GitHub raw content for library source ## Context injection rules - **Prefer excerpts over whole files** — extract 5-20 relevant lines - **Tag source origin**: `[src/path.ts:42]` or `[web: docs.url]` - **Summarize what you found** before quoting - **If nothing found**: say so immediately — don't fabricate ## Multi-turn RAG - Track what's already retrieved in-session - Don't re-fetch same content unless explicitly asked - On ambiguity: retrieve top 2-3 candidates, present choices