V Vant Docs

RAG-powered memory reranking and compression for LLM context.

Overview

Vant’s rerank module provides Retrieval-Augmented Generation capabilities:

This is distinct from search (semantic BM25/Vector) - rerank does keyword matching on results.

CLI Usage

# Rerank memories against query
vant rerank "lessons learned"
vant rerank "security fixes" -k 10

# Compress a file to token budget
vant rerank compress lessons.md -t 2000

# Refine: rerank + compress
vant rerank refine "security fixes" -k 10 -t 2000

# Stats
vant rerank -s

(There is no vant rerank pipeline CLI verb - the pipeline shape lives in the lib as rerank-then-compress.)

Options

Flag Description Default
-k, --top-k Top K results 5
-t, --max-tokens Max tokens for compression 2000
-s, --stats Show rerank statistics -
-v, --verbose Verbose output -

MCP Tools

The registered vant_rerank tool takes { query, docs } and reranks the docs array:

await mcp.call('vant_rerank', {
    query: 'lessons learned',
    docs: ['doc one...', 'doc two...']
});
// { query, results } - reranked docs

(The mode/topK/maxTokens parameter surface was fiction - the registered schema is exactly query + docs; compression is the lib’s compress function.)

Programmatic Usage

Rerank lives in lib/search.js (there is no vant.rerank getter):

const rerank = require('./lib/search');

// Get memories from brain
const memories = [
    { id: '1', title: 'Security', content: '...', date: '2026-01-01' },
    { id: '2', title: 'Lessons', content: '...', date: '2026-01-02' }
];

// Rerank against query
const results = rerank.rerank(memories, 'security fixes', 5);

// Compress to token budget
const compressed = rerank.compress(results, 2000);

// Refine: rerank + compress
const refined = rerank.refine(memories, 'security', { topK: 10, maxTokens: 4000 });

(The pipeline export name was fiction - the combined operation is refine.)

How It Works

Rerank

  1. Extract query terms (words > 2 chars)
  2. Score each memory by:
    • Title match (+10)
    • Query term in content (+1 each)
    • Query term in title (+2 each)
    • Recency boost (+0.5)
  3. Return top-K sorted by score

Compress

  1. Strip markdown fluff (headers, bold, empty lines)
  2. Truncate to token budget
  3. Mark truncated entries

Pipeline

Runs rerank -> compress in sequence, returns stats.

Integration

Rerank is separate from search. Use it to:

Search already hooks into rerank: vant search <query> -r reranks and compresses results (live since pass 134’s search verification).