agentic_vector 0.1.0 copy "agentic_vector: ^0.1.0" to clipboard
agentic_vector: ^0.1.0 copied to clipboard

Vector store abstraction for the agentic framework: records, metadata filtering, similarity search and a portable store port, with an in-process implementation and a Qdrant adapter.

Changelog #

0.1.0 #

Initial release of the vector layer.

Added #

  • ValuesVectorRecord, VectorQuery and VectorMatch. Identifiers are caller-supplied, which is what makes re-indexing a changed document replace its chunks instead of duplicating them. Search results omit vectors unless asked for; VectorRecord.hasVector says which you have.
  • Filtering — a sealed MetadataFilter hierarchy (equals, notEquals, inValues, greaterThan, lessThan, exists, and, or, not) that every adapter must translate in full, so a new filter kind is a compile error rather than a silently ignored predicate.
  • SimilaritySimilarityMetric covering cosine, dot product and Euclidean, all scored higher-is-better so that topK ordering and minScore never invert when the metric changes; plus cosineSimilarity, dotProduct, euclideanDistance and normalise over raw lists.
  • PortVectorStore with upsert, search, get, delete, deleteWhere, count and clear, and VectorStoreOperations supplying upsertAll (batched and cancellable), searchVector, exists, deleteOne and the checkDimensions / checkNamespace guards adapters call.
  • In-process storeInMemoryVectorStore, exact by construction, with namespaces, oldest-first eviction under maxRecords, and snapshot / fromJson so an offline-first app embeds once and restores thereafter.
  • Qdrant adapterQdrantVectorStore, including deterministic UUID derivation for identifiers Qdrant will not accept, score-direction conversion for Euclidean collections in both directions, full filter translation, and ensureCollection / dropCollection.
  • CompositionEmbeddingIndex binds an embedding model to a store, batching to each side's limit, using the document and query encoders correctly, and refusing a width mismatch at construction rather than after ingestion.
  • DecoratorsDelegatingVectorStore, ObservableVectorStore (traces, logs and events) and NamespacedVectorStore (a handle that cannot reach another tenant's partition).
  • EventsVectorUpsertCompleted, VectorSearchCompleted (carrying requested against returned, the number worth alerting on), VectorDeleteCompleted and VectorOperationFailed.
2
likes
0
points
235
downloads

Documentation

Documentation

Publisher

unverified uploader

Weekly Downloads

Vector store abstraction for the agentic framework: records, metadata filtering, similarity search and a portable store port, with an in-process implementation and a Qdrant adapter.

Repository (GitHub)
View/report issues

Topics

#ai #vector-database #embeddings #rag #agentic

License

unknown (license)

Dependencies

agentic_core, agentic_llm, crypto, http, meta

More

Packages that depend on agentic_vector