feat: vision LLM description pipeline for image documents

[nv78]

Summary

When a user uploads an image file (JPEG, PNG, GIF, WebP, BMP, TIFF) as a document, instead of attempting extraction through Apache Tika (which produces nothing useful for images), the new vision pipeline generates a rich text description via GPT-4o or Claude 3.5 Sonnet and feeds it into the existing RAG chunking + embedding pipeline. The image becomes fully searchable.

New file: backend/services/vision_service.py

• describe_image(bytes, mime_type, prompt?) — main entry point; dispatches to the correct provider based on AgentConfig.DEFAULT_AGENT_MODEL_TYPE
• _describe_openai() — calls openai.OpenAI with an image_url content block using detail=high for better OCR accuracy
• _describe_anthropic() — calls anthropic.Anthropic with a source content block
• Uses a detailed _INDEXING_PROMPT that asks the model to transcribe text verbatim, describe charts/diagrams/objects/layout, and be thorough for search indexing
• Respects ENABLE_MULTIMODAL and MAX_IMAGE_BYTES config flags
• Never raises — returns a human-readable placeholder string on failure so the document record is always created

Updated: backend/api_endpoints/documents/handler.py

New elif category == "image" branch in IngestDocumentsHandler:

1. Reads raw bytes from the upload
2. Calls describe_image()
3. Stores the description as document_text (with media_type='image', mime_type)
4. Enqueues chunk_document_fn.remote() so the description is split, embedded, and indexed

Video and audio remain as binary-only stubs — to be wired up in PRs 3 & 4.

Test plan

• Upload a PNG screenshot → document record created with media_type='image'
• Chunks appear in the chunks table after the Ray task completes
• Ask a question about the image content → RAG retrieves the vision-generated text, LLM answers correctly
• Upload a PDF → existing Tika path unaffected (no regression)
• Upload a 25 MB image → describe_image returns a size-limit placeholder, document still created
• Set ENABLE_MULTIMODAL=false → placeholder text stored, no LLM call made

Depends on: claude/add-multimodal-support-QBQca (schema + add_document signature)

https://claude.ai/code/session_01C9mHttiQ4ZAaBbQecVV7uu