Show what was extracted from V2 image uploads, index text-free images, and default image generation to 2025-04-01-preview

[Paul Lizer (paullizer)]

Refs #1583

This brings the Development fix from #1584 into React V2 and adds the V2-only chat fixes for uploaded images. Version 0.261.210.

Why

Two nearly identical photos of an empty lab floor were uploaded. Neither has any real text.

• The less overexposed photo worked only because Document Intelligence Layout misread a strip of floor tape as a page number (<!-- PageNumber="J" -->, confidence 0.18). That created a chunk for the vision description to ride on.
• The overexposed photo got no chunk, no title, and nothing chat could use, even though its vision analysis succeeded.
• A later upload returned a 404 because the vision model (gpt-5.6-luna) is hosted on a different AI connection, and ingestion always called the primary GPT endpoint.
• In V2, opening the upload showed File content not found, a cited image showed only the picture, and the upload card did not match a generated image.

Folded from Development (#1584)

• Images without OCR text are indexed from their vision description. A page that holds only Document Intelligence page-number or page-break annotations counts as no text. Chunks are embedded from the stored text, including the vision block. Failed or empty analyses are never indexed.
• Content Understanding image analyzer fallback. A standalone image the document analyzer returns nothing for goes to the image analyzer before Document Intelligence Layout, for workspace and direct chat uploads. The analyzer's Summary field is now read, which also fixes Office-embedded images that were indexed as ![image](pages/1).
• Vision analysis uses the AI connection that hosts the model. Ingestion, chat image uploads, and Test Vision Analysis resolve the model through find_enabled_model_endpoint_for_model_name / resolve_vision_model_endpoint, and fall back to the GPT connection otherwise.
• Image generation API version defaults to 2025-04-01-preview. A saved 2024-12-01-preview or blank legacy value is upgraded when settings load. APIM versions are never changed.
• CodeQL follow-up. Tests and docs only.

React V2 adjustments to the fold

• The shared connection client helper keeps V2's order: it finds the model and runs require_model_capability before any Key Vault lookup, so metadata extraction behaves as before. Vision also checks image support against the resolved connection's model.
• V2 refuses chunks that are too large for non-legacy embedding profiles. When only the appended vision block pushes a page over the budget, including a description-only chunk, the embedding input is cut to the budget and the stored text keeps the whole block.
• Image connection profiles imported with 2024-12-01-preview are upgraded; APIM profiles are not. Unmigrated legacy direct Images settings with no version fall back to 2025-04-01-preview. An unmigrated APIM route keeps sending 2024-12-01-preview. The V2 admin field default and the docs are updated to match.

V2-only

• /api/get_file_content for workspace-backed uploads returns the linked document's indexed chunks, read in that document's own scope through get_ordered_document_chunks. It also returns browser-safe details: engine and reason, vision analysis, title, summary, keywords, and status. There are explicit messages for processing, failed, and nothing-indexed files. Screening errors keep their public wording, and other failures return a fixed message. The new pure module functions_chat_upload_extraction.py shapes the response. The classic popup benefits too.
• Upload card parity. The image is drawn at its own size, like a generated image. Hover or focus reveals Show sources, Message details, Open file preview, Use as reference and Edit. Show sources opens an inspector panel with the extraction, and Uploaded file shows the same content.
• Cited images show What was extracted from this image (the cited passage) beneath the picture.

Tests

• New: functional_tests/test_chat_upload_extraction_details.py. It covers response shaping, fallback to the stored vision analysis, processing, failed and absent wording, and the route helper's scope and error handling.
• Folded and adapted for V2: test_image_description_indexing_without_ocr_text.py (9 passed). Its offline probe imports the real modules in normal and -O processes, and it adds V2 embedding-guard and capability-order cases. Also test_image_generation_api_version_default.py (21 passed), which adds connection-profile migration, persistence and APIM fallback cases.
• ui_tests/test_v2_chat_image_uploads.py: 16 passed. It checks card sizing against a 320×200 image, hover reveal, the Sources panel, the file preview, processing and failed states, and the cited-image passage. test_v2_image_reference_generation.py and test_v2_image_editor.py also pass.
• Related suites pass. The route policy suites, docs coverage and quality, and the BAC, XSS and swagger guardrails all pass.
• Failures that also occur on the base commit b0ed6ea4a, checked in a clean worktree and unchanged by this PR:
  • test_backend_documents_swagger_integration (3)
  • test_v2_tabular_parity::test_the_confirmation_precedes_the_send
  • test_content_understanding_extraction_engine::test_settings_and_admin_surface_contract
  • test_model_vision_capability_resolution::test_the_catalog_overrules_the_name_heuristic
  • test_image_generation_model_compatibility (2)
  • 3 V2 admin parity/schema tests
  • test_image_generation_sdk_http::test_installed_sdk_matches_existing_application_requirement, which fails because the local openai SDK version differs.

Deploying

After deploying, re-upload or reprocess the overexposed image so it gets indexed.

Docs: V2 Chat Upload Extraction and Image Card Fix, Image Description Indexing Without OCR Text Fix, Image Generation API Version Default Fix, and release notes under v0.261.210.