Skip to content

RAG Evaluation Infrastructure

Project Summary

Client: Enterprise Platform Industry: Enterprise Search / Knowledge Management

Impact Metrics:

  • Replaced generic quality metrics with calibrated measurements, such as answer accuracy, retrieval accuracy, and stability
  • CI/CD regression testing for search quality on every deployment
  • Open-source models in the judge role, cutting the cost per evaluation run
  • A documented framework the team runs for ongoing quality monitoring

Challenge

The client had a production RAG-based search assistant and no reliable way to measure whether it was working. The same evaluation run twice gave different numbers, so teams couldn't tell if a change improved or degraded search quality. Without trustworthy evaluation, every deployment was still guesswork.

Approach

I built a systematic measurement layer designed for production reliability:

  • LLM-as-a-judge framework: evaluation against defined rubrics, with the judge calibrated on a labelled reference set
  • Open-source judges: OSS models in the judge role, which cut the recurring cost of every evaluation run
  • CI/CD integration: automated regression testing, so search quality is verified on every deployment
  • Metric calibration: metrics tied to the reference set, so a movement in the number reflects a movement in the system

What the client has now

  • A rubric-based judge, calibrated against a labelled reference set
  • A regression suite that runs in CI on every deployment
  • Evaluation runs cheap enough to repeat often, using open-source judges
  • Documented metric definitions, so the same number means the same thing next quarter

Tech Stack

  • Python
  • LLM-as-a-judge evaluation framework
  • Open-source language models for cost-efficient judging
  • CI/CD pipeline integration
  • Statistical calibration and metric design

My Role

Designed and built the entire evaluation infrastructure, from metric definition and judge calibration to CI/CD integration and production deployment.

The reasoning behind the layered judge is in The judge does not score. The judge diagnoses..

  • Need to measure your RAG system's quality?


    Tell me what your system does and where you suspect it drifts, and I'll say what measuring it would take.

    Get in touch