# Week 3 Day 5 — RAG Workflow Trace

## Where Day 5 Fits in RAG

A normal RAG workflow is:

```txt
Index
→ Retrieve
→ Augment
→ Generate
→ Verify
→ Return

In ClaimFlow AI:

Policy documents were already indexed as chunks with embeddings.
Day 4 handled retrieval.
Day 5 adds answer generation + citation verification.

So Day 5 is not “just call Gemini.”

Day 5 is:

retrieved policy chunks
+ reviewed claim context
+ user question
→ grounded coverage answer
→ citation guardrails
→ saved CoverageQuestion

Example Input

User asks:

{
  "question": "Is this claim covered under the policy?"
}

Run context after human review:

{
  "claimSource": "reviewDecision.correctedJson",
  "lossType": "theft",
  "firNumber": "FIR-2026-001",
  "policeReport": true,
  "missingFieldsJson": [],
  "requiredEvidenceJson": []
}

This means the claim is no longer using the stale extraction result. It is using the reviewed claim JSON.

Step 1 — Query Starts as User Question

Original question:

Is this claim covered under the policy?

This question alone is too generic.

The system combines it with claim context:

question + reviewed claim JSON