6 posts
Polecat 5.29 added keyword search to the Critter Stack's SQL Server store, and it does not use SQL Server's full-text engine to do it. The inverted index Polecat maintains itself, the LINQ operators over it, and BM25 ranking derived from the token rows.
SQL Server 2025 added a VECTOR type. Polecat 5.28 reaches it with a persisted computed column over the JSON document body, which means no write path changes and no backfill. Exact k-nearest-neighbour, three metrics that are all distances, and the one real limit.
Keyword search and vector search fail in different places, so a document both rank well beats either one's favourite. How Polecat 5.29 fuses the two rankings with RRF, why only the ranks are fused and never the scores, and why candidate depth has to exceed the limit.
A Fisher store is one SQLite file, which makes backup sound trivial. Copying that file while the application runs silently loses committed writes. Here is the demonstration, the SQLite primitive that gets it right, and a hosted service that snapshots, verifies, compresses and ships every database to S3.
Event sourcing in Fisher: a sale as a stream, committing in the same transaction as your documents, projections across the lifecycles, and the async daemon. Plus the two places SQLite genuinely changes the semantics, where the exclusive append methods fail instead of waiting and where contiguous sequences make the high-water mark trivial.
Fisher is the Marten and Polecat programming model on SQLite, running inside your own process with no database server at all. A tour of the document side: identity, sessions, LINQ including joins and paging, expression indexes against generated columns, patching, optimistic concurrency, and the single-writer constraint everything else follows from.