Repository navigation
Introduce Oracle Database embedding store extension - #2494
Conversation
This comment has been minimized.
This comment has been minimized.
Use Oracle Database 23ai AI Vector Search as an embedding store, backed by langchain4j-oracle. Follows the pgvector conventions: a non-final wrapper bean, build-time/runtime config split, named stores via @EmbeddingStoreName, Agroal datasource, and Dev Services. * Fix quarkiverse#2492 Signed-off-by: Matheus Oliveira <[email protected]>
db6d678 to
0e6719b
Compare
Status for workflow
|
|
Thanks!
Is this an official image? Also cc @jmartisk |
It is used by Quarkus core and also Spring Data and others (see https://hub.docker.com/r/gvenzl/oracle-free#users-of-these-images) so I guess it should be fine |
|
Thanks for having a look! |
|
Thanks everyone! |
Describe your changes
Adds a new
quarkus-langchain4j-oracleextension that uses Oracle Database 23ai AI Vector Search as an embedding store, backed by thelangchain4j-oraclelibrary.It follows the conventions established by the existing stores (closest to pgvector):
QuarkusOracleEmbeddingStoreis a non-final wrapper that delegates to langchain4j's finalOracleEmbeddingStore, so the bean can be proxied by ArC.default-store-enabled, per-storedatasource) and runtime (table, columns,exact-search, IVFvector-index.*, and JSONmetadata-indexes).@EmbeddingStoreNamequalifier, each able to target a different Agroal datasource.gvenzl/oracle-free:23-slim) automatically; the image is overridable.Unlike pgvector, there is no
dimensionproperty: Oracle stores embeddings in a nativeVECTOR(*, FLOAT32)column and infers the dimension at insert time. The vector index and metadata indexes are created independently, so metadata filtering can be accelerated without an IVF index on small tables that rely on exact search.Configuration
Default store:
Named store backed by a separate datasource:
Injection