🔍Search Tool

semantic_search

Vector similarity search across all conversations using local embeddings. Find by concept, not keyword.

When to use this

Use semantic_search when you remember a concept but not the exact words. “That thing about chunking strategies for embeddings” will find relevant conversations even if you never used those exact words.

Quick Example

mcporter call brain.semantic_search query="bottleneck as amplifier"
Response
## Semantic Search: "bottleneck as amplifier"

Found 10 results (similarity > 0.65):

1. [0.89] "Bottleneck Thesis Deep Dive" (2024-08)
   "The bottleneck isn't the constraint — it's the
   amplifier. Everything that passes through it gets
   compressed into its most essential form..."
   ID: abc-123...

2. [0.82] "SEED Principle: Compression" (2024-09)
   "Compression through constraint. The narrower the
   channel, the more signal gets concentrated..."
   ID: def-456...

3. [0.78] "Monotropism as Cognitive Architecture" (2024-10)
   "Single-channel processing isn't a bug — it's a
   feature that amplifies depth at the cost of breadth..."
   ID: ghi-789...

Parameters

NameTypeRequiredDefaultDescription
querystringYesNatural language search query — conceptual, not just keywords
limitintegerNo10Maximum number of results to return

How It Works

  1. Your query is embedded locally using nomic-embed-text-v1.5 (768 dimensions)
  2. The embedding is compared against 82K+ pre-computed message vectors in LanceDB
  3. Results are ranked by cosine similarity and returned with context

All processing happens on your machine — no API calls, no data leaving your device.

Examples

Conceptual search

mcporter call brain.semantic_search query="tradeoffs between consistency and availability"

Technical search

mcporter call brain.semantic_search query="embedding chunking strategies for long documents"

More results

mcporter call brain.semantic_search query="auth patterns" limit=20

semantic_search vs search_conversations

semantic_searchsearch_conversations
MethodVector similarityKeyword matching (ILIKE)
Best forConcepts, ideas, fuzzy recallExact terms, names, code
SearchesEmbedded user messages onlyAll messages (user + assistant)