Semantic Search finds memos by meaning rather than keyword match — the associative half of 📝Library Search. Ask a question in natural language and MythOS retrieves the most relevant passages from a library using AI vector embeddings and hybrid scoring, surfacing the right memo even when the words don't line up. It's the retrieval engine behind AI features rather than the in-app search bar.
Key Capabilities
- Meaning-based retrieval — finds memos about a concept even when the exact words differ
- Hybrid scoring — blends 70% vector similarity with 30% keyword relevance for each chunk
- Chunk-level precision — returns the specific passages that matched, with scores and section
- Title and tag enrichment — each chunk is embedded with its memo's title and tags for context
- Permission-aware — every result is gated to what the requester is allowed to read
- Graceful degradation — if the vector index is unavailable, it falls back to text-only with a flag
How It Works
Every memo is split into roughly 500-token chunks, each enriched with the memo's title and tags, then embedded with OpenAI's text-embedding-3-small model and stored in a dedicated embeddings collection. A query is embedded the same way and run against two indexes in parallel — vector similarity for meaning and text search for keywords — then merged into one hybrid score (70% vector, 30% text) and aggregated per memo. Results respect the permission model: you see your own memos, collaborators see shared content, visitors see only public memos.
Getting Started
Semantic search is reached through AI, not the in-app search bar. Connect your library to an AI client via 📝MythOS MCP and call search_memos with mode: "semantic", or chat with a library — its RAG pipeline runs the same retrieval. See 📝How to Use Semantic Search in MythOS for step-by-step instructions.
FAQ
- How is it different from keyword search? 📝Keyword Search matches titles and tags; semantic search matches meaning across a memo's full content.
- Can I use it from the search bar? Not currently — semantic search runs via the
search_memostool, the internal API, and RAG-backed chat, not the in-app feed search. - Do I need an API key? Semantic retrieval needs an embedding provider configured; without one, the tool returns guidance to use substring mode instead.
- Are private memos exposed? No — permission filtering applies, so a requester only ever retrieves memos they're allowed to read.
Related
- 📝How to Chat with Your Library via MCP — RAG chat built on the same retrieval