🚨 Improve Red Flags Detection with Enhanced LLM Prompt

[pranaya-mathur]

Problem

LLM was not consistently detecting and listing red flags in suspicious claims. Even when identifying issues in fraud_explanation, the red_flags array remained empty.

Solution

Enhanced LLM Prompt

Added explicit red flag checklist with clear examples:

IMPORTANT - RED FLAGS to check:
- "delayed-report" - mentions filing later ("will file later", "baad mein")
- "no-police-report" - no police report filed ("no FIR", "nahi karwayi")
- "vague-details" - lacks specific information
- "short-narrative" - unusually brief (under 50 chars)
- "inconsistent-timeline" - events don't follow logic
- "excessive-claim" - disproportionate amount
- "suspicious-timing" - timing raises questions

Key Improvements:

1. Specific Instructions - LLM now knows EXACTLY what to look for
2. Bilingual Support - Includes Hindi patterns ("baad mein", "nahi karwayi")
3. Fraud Risk Calibration - Rule: "If multiple red flags present, fraud_risk should be 0.4 or higher"
4. Role Clarity - Changed from "claim evaluator" to "fraud detection expert"

Updated Fallback Heuristic:

Also improved local_fallback() function to detect Hindi patterns:

• Added "baad mein" detection
• Added "nahi karwayi" detection

Testing Required:

Test Case 1 - Delayed Report (Hindi):

{
  "narrative": "Kal raat accident hua. FIR abhi tak nahi karwayi. Baad mein file kar dunga."
}

Expected:

• red_flags: ["delayed-report", "no-police-report"]
• fraud_risk: >= 0.4

Test Case 2 - Vague + Short:

{
  "narrative": "Accident. Will file later."
}

Expected:

• red_flags: ["short-narrative", "delayed-report", "vague-details"]
• fraud_risk: >= 0.5

Test Case 3 - Legitimate Claim:

{
  "narrative": "Car accident on highway. Police report filed immediately with FIR number 123. Full damage documentation attached."
}

Expected:

• red_flags: []
• fraud_risk: < 0.2

Impact:

• ✅ More accurate fraud detection
• ✅ Better transparency (users see specific red flags)
• ✅ Consistent results across similar cases
• ✅ Bilingual pattern recognition

---

Ready for testing - merge after validation!