Embeddings

Optionally augment FTS search with vector similarity. When configured, mor find merges FTS and vector results using Reciprocal Rank Fusion (RRF) — combining rankings without manual weight tuning.

Configuration

Add to ~/.config/mor/config.json:

{
  "embedding": {
    "provider": "openai",
    "model": "text-embedding-3-small",
    "dimensions": 1536
  }
}

Then rebuild the index:

mor reindex

Embeddings are computed automatically on add and update. Only reindex is needed for the initial build or after changing providers.

Providers

OpenAI

{
  "embedding": {
    "provider": "openai",
    "model": "text-embedding-3-small",
    "dimensions": 1536
  }
}

Set OPENAI_API_KEY in your environment. Default base URL is https://api.openai.com/v1.

Azure OpenAI

{
  "embedding": {
    "provider": "azure-openai",
    "model": "text-embedding-3-small",
    "baseUrl": "https://your-resource.openai.azure.com"
  }
}

Set AZURE_OPENAI_API_KEY in your environment (falls back to OPENAI_API_KEY).

The deployment name defaults to the model name. Set "deployment": "my-deploy" to override. apiVersion defaults to 2024-10-21.

Ollama

{
  "embedding": {
    "provider": "ollama",
    "model": "nomic-embed-text",
    "dimensions": 768
  }
}

Default base URL is http://localhost:11434. No API key needed.

How it works

1. On add/update, the memory’s title + tags + content are concatenated and sent to the embedding provider
2. The resulting vector is stored in the embeddings table and indexed via sqlite-vec for fast KNN search
3. On find, the query is embedded and compared against stored vectors using cosine distance
4. FTS and vector rankings are combined using Reciprocal Rank Fusion (RRF)
5. Frequently accessed memories get a small ranking boost (~5% max)

When to use embeddings

Embeddings help when:

• You search for concepts rather than exact words (“error handling” finds retryWithBackoff.ts)
• FTS tokenization misses your query (searching “correct” finds AwaitTaskCorrect)

Embeddings may not help when:

• Your memory store is small (FTS + grep cover most cases)
• You search for exact strings (use grep instead)
• Score distributions are flat (all results score similarly)

For most personal use, FTS + grep is sufficient. Embeddings are worth trying if you have hundreds of memories and find yourself refining searches often.