Citations not displaying despite structured_output_enabled flag

[manavgup]

Problem

The citations feature is not displaying in the frontend despite all infrastructure being in place and the structured_output_enabled flag being correctly sent from frontend to backend.

Expected Behavior

When user checks "Enable Citations (Structured Output)" checkbox and sends a query:

1. Frontend sends metadata.config_metadata.structured_output_enabled: true
2. Backend receives config and uses structured output generation
3. Backend returns metadata.search_metadata.structured_answer.citations
4. Frontend displays "X citations" button in MessageMetadataFooter
5. User clicks button to see CitationsAccordion with citation details

Actual Behavior

• ✅ Frontend correctly sends structured_output_enabled: true in payload
• ✅ Backend MessageProcessingOrchestrator extracts and merges config_metadata
• ❌ Backend continues using Chain of Thought (CoT) instead of structured output
• ❌ No citations generated or returned
• ❌ Frontend only shows "5 sources" button, no "X citations" button

Evidence

Frontend payload (confirmed correct):

{
  "session_id": "3fb379e1-66e9-4c48-93fd-c46e8d9df6a5",
  "content": "what was IBM revenue?",
  "role": "user",
  "message_type": "question",
  "metadata": {
    "config_metadata": {
      "structured_output_enabled": true,
      "cot_enabled": false,
      "show_cot_steps": false
    }
  }
}

Backend logs:

• Shows "Using CoT-generated answer" (should use structured output)
• Debug logs from MessageProcessingOrchestrator not appearing
• Config may not be reaching generation_stage.py

Root Cause Analysis

Suspected Issue: Config metadata not propagating from MessageProcessingOrchestrator → SearchService → generation_stage.py

Files Involved:

1. backend/rag_solution/services/message_processing_orchestrator.py - Extracts config ✅
2. backend/rag_solution/services/search_service.py - Should pass config to generation
3. backend/rag_solution/generation/generation_stage.py - Needs to check structured_output_enabled flag

Current Logic in generation_stage.py (suspected issue):

# Line ~120-140: CoT takes precedence
if search_output.cot_output and search_output.cot_output.final_answer:
    final_answer = search_output.cot_output.final_answer  # ← Always used if CoT exists
    logger.info("Using CoT-generated answer")
elif search_output.structured_answer:
    final_answer = search_output.structured_answer.answer  # ← Never reached
    logger.info("Using structured answer")

Problem: Need to respect structured_output_enabled flag to skip CoT when structured output requested.

Proposed Solution

Option 1: Check Flag in generation_stage.py (Recommended)

# In generation_stage.py, check config_metadata first
structured_output_enabled = config_metadata.get("structured_output_enabled", False)

if structured_output_enabled and search_output.structured_answer:
    # User explicitly requested structured output - use it
    final_answer = search_output.structured_answer.answer
    logger.info("Using structured answer (user preference)")
elif search_output.cot_output and search_output.cot_output.final_answer:
    # Fall back to CoT if available
    final_answer = search_output.cot_output.final_answer
    logger.info("Using CoT-generated answer")
elif search_output.structured_answer:
    # Fall back to structured answer
    final_answer = search_output.structured_answer.answer
    logger.info("Using structured answer (fallback)")

Components Already Working ✅

1. Frontend UI: CitationsAccordion component, toggle checkbox, MessageMetadataFooter
2. Frontend API: apiClient.sendConversationMessage() sends correct payload structure
3. Backend Orchestrator: Extracts and merges user config_metadata
4. Backend Schemas: ConversationMessageInput accepts metadata with config_metadata
5. WatsonX Provider: Structured output generation with JSON schema validation

Related Issues

• PR Implement Structured Output with JSON Schema Validation #604 #626 - Structured output with JSON schema validation (parent PR)
• Issue Backend leaking internal processing data to frontend (10-20x payload bloat) #628 - Backend data leak (separate issue for payload bloat)

Priority

MEDIUM - Feature incomplete but not blocking other work. Citations infrastructure complete, just needs config propagation fix.

Estimated Effort

2-4 hours - Small change to generation_stage.py logic, testing, verification