v2.4.4

Added

• Blob storage metadata indexed as custom_metadata (issue Azure/GPT-RAG#487). The blob storage indexer now extracts every user-defined tag from blob.metadata (skipping reserved metadata_security_* keys) and stamps it onto each AI Search chunk as custom_metadata, a Collection(Edm.ComplexType) with {key, value} pairs. Completes the operator-facing ship of #487 (the matching index schema was added in gpt-rag v2.9.0).

Fixed

• Content Understanding multimodal ingestion regression: restored the format-based figure extraction path for MultimodalChunker when Content Understanding is used, including PDF page/region rendering and Office embedded image extraction. Brings back the behavior released in v2.3.3 without changing the /ingest-documents upload ACL fix.

Changed

• Warn when `` diverges from the azd environment during component deploy (Refs Azure/GPT-RAG#491). scripts/deploy.ps1 and `scripts/deploy.sh` now print a yellow warning when the shell and the azd env disagree.

Validation

• pytest tests/ (7 passed locally — blob metadata extraction tests).