Feature Request: Vector Similarity Search for Routine Registry
Context: We are building an LLM orchestration pipeline (Gemma + DeepSeek) that
automatically discovers, evaluates, and synthesizes routines. To prevent
overflowing the LLM context window and to minimize network latency, we need
the Registry API to handle semantic filtering natively.
Requirements:
New Endpoint: Implement a semantic search endpoint (e.g., POST
/routines/search or a query parameter on GET /routines).
Input: The endpoint should accept a natural language query string (the
specification or need) and an optional limit integer (defaulting to 20).
Embedding Generation: Integrate a lightweight local embedding model (e.g.,
all-MiniLM-L6-v2 or nomic-embed-text) into the registry backend.
Indexing Strategy: Upon creation/update (PUT /routines/:name), the registry
must concatenate the routine's name, description (from spec), and parameter
types into a single text block, generate its vector embedding, and store it.
Execution: When the search endpoint is hit, embed the incoming query, perform
a cosine similarity scan against the stored routine vectors, and return the
spec.json definitions of the highest-scoring matches, sorted by confidence.
Goal: The AI pipeline should only ever receive the top 10-20 most semantically
relevant candidate specs, bypassing the need to fetch the entire mesh
inventory.