Semantic Search

MUR v2 uses LanceDB for vector storage combined with BM25 keyword scoring for hybrid ranked results.

How It Works

When you run mur search "error handling", MUR:

Semantic search — Converts your query to an embedding vector, finds similar patterns by cosine similarity
BM25 keyword search — Traditional text matching for exact keyword hits
Hybrid ranking — Combines both scores for the final ranked results

Setup

OpenAI (Recommended)

Best quality, very low cost (~$0.001 per 200 patterns):

export OPENAI_API_KEY=sk-...

# ~/.mur/config.yaml
search:
  provider: openai
  model: text-embedding-3-small
  api_key_env: OPENAI_API_KEY

Ollama (Free, Local)

No API key needed. Runs entirely on your machine:

ollama pull qwen3-embedding

# ~/.mur/config.yaml
search:
  provider: ollama
  model: qwen3-embedding

Building the Index

After configuring a search provider, build (or rebuild) the index:

mur reindex

This processes all YAML pattern files and creates vector embeddings in ~/.mur/index/.

The index is fully rebuildable — YAML files are always the source of truth.

Using Search

# Natural language queries
mur search "how to handle authentication errors"
# → error-handling-auth (0.84)
# → retry-with-backoff  (0.71)

# Technical queries
mur search "Swift Testing @Test macro"
# → swift-testing-macro (0.92)

Context Injection

For hook integration, MUR can inject relevant patterns based on the current context:

# Auto-detect project from current directory
mur context

# Compact output
mur context --compact

# Explicit query
mur inject --query "fix SwiftUI layout bug" --project my-app