Brain Search

Ask the corpus a question, get the brain files that answer it.

Brain search is semantic retrieval over everything the brain holds. It is how an agent finds the lesson from three weeks ago without remembering which file it lives in.

CLI

Search the corpus:

vant search "why did the WAL replay fail"

The query runs through the embed provider, scores the corpus, and returns ranked brain files. This is the command an agent calls when context is missing and the answer probably exists in memory already.

Code

The search module exposes the query surface directly:

const search = require('./lib/search');
const results = await search.queryBrain('migration failure recovery');

Embeddings power the scoring. The provider is pluggable:

Check what is active:

vant embed info

Generate an embedding for a single text:

vant embed generate "the sentence to embed"

Rerank

Retrieval quality improves when a second pass reorders the first pass results against the query:

vant rerank "query" 

The rerank module (Rerank, Hybrid, Hyde in lib/search.js) implements relevance second-stage scoring. Use it when the corpus is large enough that top-5 by embedding alone starts missing.

Citations

Answers grounded in brain files can carry git-backed receipts. The citations module records sources and renders them as a commit footer:

const citations = require('./lib/citations');
citations.addSource('lessons.md#sync-race');
citations.getCommitFooter();

Details and the verification flow are in Citations.

Internals

The architecture, LTC freshness, and the settlement model behind the search index are documented in Search Architecture.