debezium/dbz#2156 feat: implement unified metadata filtering API and …

[KMohnishM]

…adapter translations

Description

This Pull Request implements the first half of GSoC Week 6 objectives under tracking issue debezium/dbz#2156. It introduces structured metadata filtering capabilities across the Chroma, PGVector, and Milvus vector store backends, providing a unified Python dict interface to query filtered subsets of synchronized documents.

Key Changes

1. Interface Contract Update:
   • Extended as_retriever signature in VectorStoreAdapter (base.py) to accept an optional metadata_filter: dict[str, Any] keyword argument.
2. Adapter Translations:
   • Chroma: Maps dictionaries to Chroma filter formats. Combines with pre-existing search filter kwargs using Chroma's logical $and operator.
   • PGVector: Maps filters directly to Postgres filter dictionaries.
   • Milvus: Converts filters into boolean query expression strings (e.g. (category == 'books') and (in_stock == true)), merging them safely with existing expr query strings.
3. Robust Verification & Testing:
   • Added unit/integration tests for filtering behavior across all three adapters.
   • Refactored the test suite to ensure compatibility across varying LangChain/Milvus document schema formats.

Verification Results

• Ruff Quality Check: Passed (All checks passed!).
• MyPy Type Safety: Passed (Success: no issues found in 39 source files).
• Test Suite: Run and passed 86 tests: