AI agents

Debugging and measuring retrieval quality

• Detecting a Citation the Agent Never Used
• Measuring Retrieval Quality: Hit Rate, MRR and Recall@k for RAG
• Measuring Retrieval Recall After a Document Refresh
• Provenance Fields That Survive a Regeneration
• Reading a Retrieval Change as a Curve Shift, Not a Score
• Scoring Retrieval With No Labels: Clicks and Citations
• What Recall@k Misses: A Complementary Retrieval Check
• A Citation Check That Skips the Model
• A Click Log That Grades Retrieval: Fields and Joins
• Answer Quality vs Context Size: A Test You Can Run
• Calibrating a Retriever Score Against Reviewer Labels
• Clicks and Citations as a Label-Free Retrieval Signal
• Conflict-First Retrieval: Asking for the Disagreement
• Debugging Retrieval: Tracing Which Chunks Reached the Prompt
• Hit Rate per Query Type: A Retrieval Log Schema and a Dashboard
• Logging Retrieval Decisions for Post-Refresh Debugging
• Measuring Memory Quality: Staleness, Hit Rate and Answer Accuracy
• Proving a Restore Did Not Widen the Prompt: A Post-Release Check
• Quarantine Rates by Source: Finding the Ingest Sending Bad Chunks
• Recomputing the Cut Point After a Mix Shift
• Scoring Span Recall Instead of Chunk Recall
• Setting a Recall Target Before You Pick a Retriever
• Source Age in the Trace: Reconstructing a Stale Answer From Logs
• Splitting the Threshold by Query Class
• Tracing the Passage: A Log Schema for Citation Provenance
• Tracing Which Chunks Reached the Prompt
• Tracing Which Chunks Reached the Prompt: A Log Schema
• Turning a Checksum Mismatch Into a Live Alert
• Turning a Miss Into a Labeled Pair
• What a Merge Step Should Log
• When No Click Does Not Mean a Negative
• When Recall@k and MRR Disagree: Which Number to Trust
• When Recall@k Misses the Missing Facts
• Where the Score Cut Belongs: Retriever, Reranker or Gate