-
Notifications
You must be signed in to change notification settings - Fork 4
Closed
Labels
bugSomething isn't workingSomething isn't workingphase-0Phase 0: Critical blocking fixesPhase 0: Critical blocking fixespriority:criticalCritical priority - blocks productionCritical priority - blocks productionrag-qualityRAG system quality and accuracy issuesRAG system quality and accuracy issues
Description
Problem
The reranking system is completely broken, causing the RAG to return irrelevant document chunks despite having 731 chunks from the IBM annual report containing revenue data.
Error: ValueError: Template is required for batch generation
Duration: 4.21 seconds wasted (23:52:44.939 → 23:52:49.155)
Impact: BLOCKING - No accurate retrieval without working reranking
Root Cause
backend/rag_solution/retrieval/reranker.py
The reranker calls generate_text() in batch mode but fails to pass the required template parameter.
Fix Strategy
- Fetch RERANKING template from PromptTemplateService
- Pass template to
generate_text()for batch operations - Handle batch scoring properly
Success Criteria
- ✅ No
ValueError: Template is requirederrors - ✅ Reranking completes in <1 second
- ✅ Top 5-10 results have relevance scores >0.8
- ✅ Revenue-related chunks appear in top results
Files to Modify
backend/rag_solution/retrieval/reranker.py(~20-30 LOC)
Related
- Part of Phase 0: Critical Blocking Fixes
- Reference: RAG_ISSUES.md Issue Add requirements.txt #1
- Timeline: Day 1
Metadata
Metadata
Assignees
Labels
bugSomething isn't workingSomething isn't workingphase-0Phase 0: Critical blocking fixesPhase 0: Critical blocking fixespriority:criticalCritical priority - blocks productionCritical priority - blocks productionrag-qualityRAG system quality and accuracy issuesRAG system quality and accuracy issues