NornicDB

Query embedder readiness ​

Query executors and protocol adapters share one cypher.QueryEmbedderBinding. A provider published by SetEmbedder becomes visible to later requests on existing and new connections. Before readiness, keyword fallback is allowed. Publishing a provider changes the cache generation, so startup fallbacks and vectors from a previous provider cannot serve later requests.

A request captures one provider, including an unavailable provider, for its entire lifetime. In-flight requests finish using that snapshot; the next request sees a replacement. This avoids mixing chunking, vector dimensions, or cached vectors from different providers.

Scoped Cypher executors ​

Bind a scoped executor during construction, before serving requests. ShareEmbedderFrom shares future publications; copying GetEmbedder() shares only the currently loaded provider.

go
base := db.GetCypherExecutor()
scoped := cypher.NewStorageExecutor(db.GetStorage())
scoped.ShareEmbedderFrom(base)
// Publishing on base also reaches scoped and executor clones.
base.SetEmbedder(provider) // provider implements cypher.QueryEmbedder

Protocol adapters that need direct access use base.QueryEmbedderBinding(). The zero value of cypher.QueryEmbedderBinding is unavailable; Load() returns nil. Store(provider) publishes a provider and resets its caches; Store(nil) makes later requests unavailable. The binding retains no registry or history of old providers.

Native text searches ​

Capture a per-database provider once when a text request begins, then use its methods for chunking and embedding:

go
request := db.QueryEmbeddingForDB("documents")
chunks, vectors, err := request.EmbedQueryChunks(ctx, "find related release notes")
if err != nil {
    return err
}
// Score each vector against this database, retaining the best chunk score.
_ = chunks
_ = vectors

The returned nornicdb.QueryEmbeddingSnapshot also provides ChunkQuery and EmbedQuery. Per-database model configuration and dimension validation retain their existing behavior. ID and numeric-vector searches do not need a text provider and should not resolve one.

For adapters using the global convenience methods, pass one snapshot context through every operation:

go
requestCtx := db.WithQueryEmbeddingSnapshot(ctx)
chunks, err := db.ChunkQuery(requestCtx, query)
if err != nil {
    return err
}
vector, err := db.EmbedQuery(requestCtx, query)
_ = chunks
_ = vector
return err

Keep that context within the request. A nested call on the same database preserves its outer snapshot, including nil during startup.

Analytics privacy