| name | unified-search |
| description | Cross-source retrieval combining conversation embeddings, GitHub commits, and markdown docs into one ranked list. Use for broad orientation queries. |
| layer | L1 |
| reads | ["L1.embeddings","L0.github_commits","L0.markdown_docs"] |
| writes | [] |
| citations | output-is-citation |
| determinism | pure-function |
| allowed-tools | mcp__my-brain__unified_search |
unified-search — L1 cross-source retrieval
Framework context
L1 retrieval that aggregates three substrate stores into one ranked list. No synthesis — just merge + sort. Fixed source-quotas (conversations: 5, github: 3, markdown: 3). Static score weights per source (0.4 github, 0.45 markdown) — conversation rows carry real cosine similarity.
When to invoke
- User asks an orientation question that could be answered by conversations OR code OR notes
- First-pass before committing to a specific retrieval mode
- When the user says "anywhere I mentioned X"
Input
query: str
limit: int = 15
Output contract
Grouped markdown by source type. Each row has source tag [conversation] | [github] | [markdown] + citation (conv_id / commit SHA / file path).
Does NOT do
- Rerank across sources with a cross-encoder (deprecated in fastembed migration)
- Dedupe semantically across sources — a GitHub commit and a conversation about that commit may both appear
- Hide missing substrates — if GitHub table absent, just skip GitHub rows; do not crash
Verification checklist
