Fix/llm handler improvements

[fbs617]

Fix: LLM Handler Improvements & Automated Testing Suite

Contributors

• @emirakdag0 - LLM handler improvements
• @fbs617 - Automated testing suite

---

Summary

Major improvements to LLM handler robustness with comprehensive two-tier testing framework for quality assurance.

---

Changes

LLM Handler Improvements (@emirakdag0 )

Prompt Engineering:

• Improved system prompts for better DREF field extraction
• Enhanced multi-document context handling
• Reduced hallucination with clearer instructions

Response Handling:

• Stricter JSON parsing with error recovery
• Field validation and type coercion (e.g., "5,000" → 5000)
• Retry logic with exponential backoff for failures

Integration:

• Seamless conflict resolution integration
• Source tracking for all extracted fields
• Confidence scoring for uncertain extractions

---

Testing Suite (@fbs617 )

Tier 1: Pytest (19 tests)

• 5 blocker tests (conflict detection, security)
• Runs automatically in CI on every push/PR
• Fast (<30s), deterministic checks

Tier 2: Promptfoo (34 tests)

• 12 categories with LLM-as-judge evaluation
• Ground truth scoring across all edge cases
• Comprehensive quality metrics

Infrastructure:

• Test helpers: InputBuilder, FormStateFactory, assertions
• Results viewer: python review.py
• GitHub Actions CI workflow