feat: Oracle AI Vector Search DocumentStore

[fede-kamel]

Summary

Oracle Database 23ai/26ai ships a native VECTOR data type with vector_distance() SQL, HNSW/IVF indexing, and ANN search — all included in the database license (including the Free Tier on OCI). There is currently no Oracle integration in haystack-core-integrations, which means Oracle users must leave the Haystack ecosystem to do RAG.

This issue tracks adding OracleDocumentStore and OracleEmbeddingRetriever as a first-class integration.

Proposed package: oracle-haystack

Located at integrations/oracle/, the package provides:

• OracleDocumentStore — full Haystack DocumentStore protocol
  • write_documents with all three DuplicatePolicy modes (INSERT / MERGE)
  • filter_documents with Haystack filter grammar → JSON_VALUE SQL
  • delete_documents, count_documents, async variants
  • to_dict / from_dict with Secret-wrapped credentials
• OracleEmbeddingRetriever — @component with filter merging and arun
• create_hnsw_index() — CREATE VECTOR INDEX … ORGANIZATION INMEMORY NEIGHBOR GRAPH
• ADB-S wallet support — Oracle Autonomous Database thin-mode connections (no Instant Client)

Oracle AI Vector Search highlights

• VECTOR(dim, FLOAT32) column type, stored alongside relational data
• FETCH APPROX FIRST N ROWS ONLY activates the HNSW index automatically
• COSINE, EUCLIDEAN, DOT distance metrics
• JSON column type for metadata with JSON_VALUE path expressions
• Available in Oracle DB 23ai Free (on-prem) and Oracle Autonomous Database Free Tier (OCI)

Validation

Tested end-to-end against Oracle AI Database 26ai on OCI Free Tier:

• 200 SQuAD Wikipedia passages written at 181 docs/sec
• HNSW index built in 0.7s
• Retrieval latency < 700ms (including ADB network round-trip)
• 34 unit tests + 13 integration tests passing

PR

A ready-to-review PR is attached to this issue.

[julian-risch]

Thanks for the initiative @fede-kamel !

[fede-kamel]

#3096

[fede-kamel]

@julian-risch FYI - this is the PR - Thank you for your interest!