The tools tell an agent to prefer BM25 for a literal error code, semantic for a concept it cannot word exactly, and the count when only a number is wanted — guidance that keeps a search from being run five times the wrong way. Embedding stays on the machine, so a private corpus can be searched by meaning without shipping the query anywhere.
A retrieval server over an Infino catalog, which can sit in a local directory, in your own S3 or Azure bucket, or on hosted Infino Cloud. Five retrieval styles sit side by side — BM25, vector similarity, a fused hybrid ranking, unranked filters and full SQL — and each tool says when to reach for it instead of the others.
- The catalog discoverable from inside the conversation: tables listed, then a table's columns and types described so a search targets the right column
- Semantic search that finds paraphrases and synonyms, with an optional pushdown pre-filter that narrows by keyword before the ranking runs
- BM25 keyword search for identifiers, error codes and exact phrases, matching tokens and their stems and returning a relevance score
- Hybrid search that fuses BM25 and vector similarity in a single pass, so rows matching both the literal terms and the intent come out on top
- Two unranked filters for when order does not matter — the set of rows containing a token, and rows whose column exactly equals a value
- A count that tallies matches without fetching or ranking them, for when the number is the whole answer
- SQL for structural questions — joins, GROUP BY, aggregates — with the search functions callable as table-valued relations, so one query can rank and aggregate at once
- Query embedding with a local model and no API key: nothing leaves the machine to embed, and after the first run the model is cached and embedding works offline
- Read-only by default — writes and DDL/DML through SQL only appear when a single environment flag is set
Node.js and an MCP client that launches a stdio server; the package is @infino-ai/mcp-server on npm, usually run through npx. INFINO_MCP_URI points at the data — a local path, an s3:// or az:// bucket, or a hosted https endpoint, which additionally needs INFINO_API_KEY. Cloud credentials come from the standard provider variables rather than any of its own, so ambient IAM or managed identity works. INFINO_MCP_ENABLE_WRITES turns on infino_add_documents and lets SQL write. The first run downloads the local embedding model, about 90 MB, once. INFINO_MCP_EMBED_MODEL must match the model that produced a table's stored vectors — a mismatch gives meaningless similarity or a dimension error, while keyword and SQL search are unaffected. INFINO_MCP_VALIDATE probes the store at startup so bad credentials fail then rather than on the first search.
One command — claude mcp add infino --scope user -e INFINO_MCP_URI=/Users/me/.infino/memory -- npx -y @infino-ai/mcp-server