v2.6.6

Added

• Multimodal figure/image extraction for Content Understanding (Azure/GPT-RAG#446): When using Content Understanding as the document analysis backend (USE_DOCUMENT_INTELLIGENCE=false), the multimodal chunker now extracts figures from documents, uploads them to the documents-images blob container, generates captions using a vision-capable model, and populates relatedImages, imageCaptions, and captionVector fields in the search index — achieving full multimodal parity with the Document Intelligence path. Supports PDF (PyMuPDF page rendering with bounding-box crop), DOCX (word/media/ ZIP extraction), and PPTX (ppt/media/ ZIP extraction). The ContentUnderstandingClient now parses and returns figure and page metadata from the API response instead of discarding it. New dependencies: PyMuPDF, python-docx, python-pptx.

Changed

• Bumped gpt-rag-ingestion to v2.3.3.

Tested Service Versions

The following component versions were validated together for this release:

Component	Version
gpt-rag-ui	v2.3.1
gpt-rag-orchestrator	v2.6.2
gpt-rag-ingestion	v2.3.3
infra (landing zone)	v1.0.7

v2.6.5

Fixed

• OpenTelemetry version pinning (orchestrator): Pinned azure-monitor-opentelemetry==1.8.7, azure-monitor-opentelemetry-exporter==1.0.0b49, opentelemetry-instrumentation-httpx==0.61b0, and opentelemetry-instrumentation-fastapi==0.61b0 in requirements.txt. Unpinned versions caused non-deterministic Docker builds where an older exporter (referencing the removed LogData class) could be paired with opentelemetry-sdk>=1.39.0, crashing the container on startup with ImportError: cannot import name 'LogData' from 'opentelemetry.sdk._logs'. (#445)
• Permission trimming header format (orchestrator): Removed erroneous Bearer prefix from the x-ms-query-source-authorization header value in both the REST API path (search.py) and the SDK path (search_context_provider.py). Azure AI Search expects the raw OBO token without the prefix; including it caused 400 Invalid header errors when permissionFilterOption was enabled on the search index. (#447)

Changed

• Bumped gpt-rag-orchestrator to v2.6.2.

Tested Service Versions

The following component versions were validated together for this release:

Component	Version
gpt-rag-ui	v2.3.1
gpt-rag-orchestrator	v2.6.2
gpt-rag-ingestion	v2.3.2
infra (landing zone)	v1.0.7

v2.6.4

Fixed

• Restored missing parent_id field in the RAG search index template (config/search/search.j2), which was accidentally removed during the v2.6.0 merge. This caused gpt-rag-ingestion blob storage and SharePoint indexers to fail with Could not find a property named 'parent_id' errors.

Changed

• Updated infra submodule to bicep-ptn-aiml-landing-zone tag v1.0.7, fixing Log Analytics provisioning failure in Sweden Central caused by forceCmkForQuery default.

Tested Service Versions

The following component versions were validated together for this release:

Service	Tested Version
gpt-rag-ui	v2.3.1
gpt-rag-orchestrator	v2.6.1
gpt-rag-ingestion	v2.3.2
infra (landing zone)	v1.0.7

v2.6.3

Changed

• Updated infra submodule to bicep-ptn-aiml-landing-zone tag v1.0.6.
• Parametrized Container App CPU and memory per app entry with fallback defaults (0.5 CPU / 1.0Gi).
• Increased dataingest Container App resources to 1.0 CPU and 3.0Gi memory.
• Increased text-embedding-3-large deployment capacity from 40 to 100.
• Bumped gpt-rag-ingestion from v2.2.5 to v2.3.2.

What's New in gpt-rag-ingestion (v2.3.0 → v2.3.2)

v2.3.0 — Admin Dashboard, Content Understanding & Retry Tracking

• Admin dashboard: React frontend at /dashboard with paginated job/file tables, search, filters, and unblock action.
• Content Understanding integration: New default document analysis path using Azure AI Foundry prebuilt-layout instead of Document Intelligence (~69% cost reduction per page).
• Per-file retry tracking: Files exceeding MAX_FILE_PROCESSING_ATTEMPTS (default 3) are automatically blocked. Applies to blob storage and SharePoint indexers.
• Scheduled log cleanup: Automatic old run-summary blob cleanup via APScheduler.

v2.3.1 — Processing Insights, Large PDF Handling & Memory Safety

• Processing timings breakdown: Per-phase timing data (download, analysis, chunking + embeddings, index upload) displayed as a stacked color bar in the dashboard.
• Per-file cost estimation: Cost broken down by service (analysis per page, embeddings per token, completions per token) with configurable unit prices.
• Automatic PDF splitting: PDFs exceeding the analysis service page limit (default 300) are split automatically, preventing InputPageCountExceeded errors.
• Memory guard: Checks file size against available container memory before download, skipping oversized files instead of risking OOM crashes.
• Temp file download: PDFs >10 MB downloaded to disk instead of memory, keeping peak usage bounded.
• Fixed dashboard unresponsive during large file processing (async event loop was blocked).
• Fixed stale error field retained on successful re-processing.
• Fixed _as_datetime NameError crashing indexer runs.

v2.3.2 — Stability Fixes

• Default INDEXER_MAX_CONCURRENCY lowered to 2 (reduced memory pressure and rate-limit contention).
• Fixed stale running jobs stuck forever after container crash/restart (auto-detected after 2 hours).
• Fixed dashboard retries column showing inflated count.
• Fixed cost estimate displayed with excessive decimal places.
• Fixed 429 rate-limit display issues.

---

Tested Service Versions

The following component versions were validated together for this release:

Component	Version
gpt-rag-ui	v2.3.1
gpt-rag-orchestrator	v2.6.1
gpt-rag-ingestion	v2.3.2
infra (landing zone)	v1.0.6

v2.6.2

Changed

• Bumped gpt-rag-orchestrator to v2.6.1.

Tested Service Versions

The following service versions were validated together for this release:

Service	Tested Version
gpt-rag-ui	v2.3.1
gpt-rag-orchestrator	v2.6.1
gpt-rag-ingestion	v2.2.5

v2.6.1

Fixed

• Fixed Zero Trust provisioning failure caused by jumpbox Custom Script Extension using incorrect release tag. Replaced install_script URL field with ailz_tag in manifest.json, allowing the install script URL and release parameter to be derived from the landing zone tag.

Changed

• Updated infra submodule to bicep-ptn-aiml-landing-zone tag v1.0.5.
• Bumped gpt-rag-ui to v2.3.1.
• Bumped gpt-rag-ingestion to v2.2.5. (Fixed #436)

Tested Service Versions

The following service versions were validated together for this release:

Service	Tested Version
gpt-rag-ui	v2.3.1
gpt-rag-orchestrator	v2.6.0
gpt-rag-ingestion	v2.2.5

v2.6.0

What's Changed

• Update component versions (orchestrator v2.6.0, UI v2.3.0, ingestion v2.2.4), bump infra submodule to v1.0.4
• Add explicit partition keys to Cosmos DB container definitions (conversations uses /principal_id)
• Add conversation-documents storage container and conversationId field to search index
• Remove standalone MCP Container App from default deployment

Tested Service Versions

The following service versions were validated together for this release:

Service	Tested Version
gpt-rag-ui	v2.3.0
gpt-rag-orchestrator	v2.6.0
gpt-rag-ingestion	v2.2.4

Full Changelog: v2.5.3...v2.6.0

v2.5.3

What's Changed

• chore: update default chat model to gpt-5-nano, bump infra submodule to v1.0.3, update component versions, add copilot-instructions

Tested Service Versions

The following service versions were validated together for this release:

Service	Tested Version
gpt-rag-ui	v2.2.3
gpt-rag-mcp	v0.3.5
gpt-rag-orchestrator	v2.5.0
gpt-rag-ingestion	v2.2.3

Full Changelog: v2.5.2...v2.5.3

v2.5.2

What's Changed

• fix: make postProvision.sh venv cleanup non-fatal — closes #426 by @sihbher in #427
• feat: skip clone when component repo already exists (#428) by @sihbher in #429

Tested Service Versions

The following service versions were validated together for this release:

Service	Tested Version
gpt-rag-ui	v2.2.2
gpt-rag-mcp	v0.3.5
gpt-rag-orchestrator	v2.4.2
gpt-rag-ingestion	v2.2.3

Full Changelog: v2.5.1...v2.5.2

New Contributors

• @sihbher made their first contribution in #427

v2.5.1

Changed

• Updated infra submodule to external bicep-ptn-aiml-landing-zone tag v1.0.1.
• Improved runtime performance by upgrading the Orchestrator and UI components to v2.4.2 and v2.2.2, respectively:
  • Bumped gpt-rag-orchestrator to v2.4.2.
  • Bumped gpt-rag-ui to v2.2.2.
• Bumped gpt-rag-ingestion to v2.2.3.

Tested Service Versions

The following service versions were validated together for this release:

Service	Tested Version
gpt-rag-ui	v2.2.2
gpt-rag-mcp	v0.3.5
gpt-rag-orchestrator	v2.4.2
gpt-rag-ingestion	v2.2.3

Full Changelog: v2.5.0...v2.5.1