Hybrid search

MarkDB indexes memory two ways and combines them at query time.

Vector search

Summaries and pages are embedded (via the same proxy, using gemini-embedding-001 by default) and stored in pgvector. Vector search finds semantically related material even when the words don't match.

Keyword search

The same content is indexed in Meilisearch for fast, typo-tolerant full-text search. Keyword search wins when you remember an exact identifier, error string, or filename.

Hybrid retrieval

A search blends both signals: semantic recall when you mean it, exact matches when you need them. Internal pages (raw turn/transcript artifacts) are excluded from the index; summaries and durable artifacts are what you retrieve.

Consistency

Query-time embeddings use the same model and signature as the indexed pages, so the vector spaces match. If the embedder is unavailable, search degrades gracefully to keyword-only rather than failing.

You can search from the dashboard, the API, or an MCP client.