Agentic RAG
What I built
Planning, repeated retrieval, sufficiency checks, query reformulation, synthesis, routing, evidence windows and traces.
Result
Improved answer accuracy from 36.8% to 50% on 38 synthetic questions using iterative retrieval. The gain came with higher latency; run conditions are below.
Evaluation setup and selected results
On 38 synthetic questions, repeated retrieval increased answer accuracy from 36.8% to 50%. In the routed configuration, 95% of measured requests finished within 67.4 seconds, compared with 19.0 seconds for a single retrieval pass. This records the accuracy and response-time tradeoff.
Compare the evaluation evidence →
About the evidence
Based on source review and saved project artifacts. Recorded measurements apply to the stated model, dataset and configuration; they are not claims of a current production deployment.