Prompt Details
Model
(claude-4-5-sonnet)
Token size
886
Example input
[CONTEXT]: The company’s Q2 report shows a 15% revenue increase compared to the previous quarter, primarily due to mid-market expansion.
[RAG_OUTPUT]: The company reported a 22% revenue growth in Q3 driven by strong enterprise sales.
[QUERY]: What was the company’s revenue growth in Q3?
Example output
```json
{
"status": "misaligned",
"failure_type": "wrong_source",
"evidence": {
"problematic_claim": "The company reported a 22% revenue growth in Q3 driven by strong enterprise sales.",
"context_says": "The company's Q2 report shows a 15% revenue increase compared to the previous quarter, primarily due to mid-market expansion.",
"mismatch_reason": "RAG_OUTPUT claims Q3 data (22% growth, enterprise sales) but CONTEXT only contains Q2 data (15% growth, mid-market expansion). The quarter, percentage, and sales driver are all mismatched."
},
"hidden_assumptions": {
"temporal_projection": "Model assumed Q3 data exists or extrapolated from Q2 trends without evidence",
"source_confusion": "Model may have conflated Q2 context with Q3 query requirements",
"data_fabrication": "Specific metrics (22%, enterprise sales) appear invented rather than retrieved"
},
"recommended_fix": {
"retrieval_adjustment": "Verify retrieval query targets Q3 documents specifically; current retrieval returned Q2 data only. Check document timestamp filters and semantic search accuracy for temporal queries.",
"generation_constraint": "Add prompt constraint: 'Answer ONLY using the provided context. If the context does not contain information for the requested time period, respond: \"The provided context covers [actual period] but does not contain [requested period] data.\"'",
"priority": "critical"
},
"confidence": 0.95
}
```
By purchasing this prompt, you agree to our terms of service
CLAUDE-4-5-SONNET
A structured RAG debugging system that analyzes retrieval-augmented generation failures, identifies evidence-bound reasoning issues, and proposes engineer-reviewed corrective recommendations grounded strictly in retrieved context.
Designed for teams auditing hallucinations, source mismatches, or retrieval errors in high-stakes RAG workflows.
...more
Added over 1 month ago
