rag-ci

Regression testing and configuration sweeps for RAG pipelines.

Status: usable. Build a golden set, measure against it, gate pull requests on the result, sweep configurations to find what actually wins, and score answer grounding — including as a GitHub Action. See the design document.

The problem

You changed the chunk size. Did retrieval get better?

Most teams cannot answer that. They ship a change, eyeball a few queries, and move on. The 2026 literature suggests this is expensive: Beyond the Reranker finds that many retrieval enhancements stop contributing anything once a strong reranker is present. People are stacking techniques that do nothing and paying for them in latency and tokens.

Existing evaluation tools do not close the gap. They report a bare score with no confidence interval — so a recall@10 moving from 0.71 to 0.74 on 200 questions looks like progress when it is indistinguishable from noise. None of them run as a gate on a pull request, and none of them sweep configurations.

Meanwhile the academic side settled this years ago: paired bootstrap tests with 10,000 resamples and 95% confidence intervals are standard in T2-RAGBench and HetDocQA. That rigor has not reached the tools practitioners actually run.

Commands

uvx rag-ci init            # scaffold an adapter for your pipeline   ✅
uvx rag-ci golden gen      # generate candidate questions            ✅
uvx rag-ci golden review   # accept / edit / reject, then commit     ✅
uvx rag-ci run             # measure, with confidence intervals      ✅
uvx rag-ci gate            # fail the PR only on a real regression   ✅
uvx rag-ci sweep           # find the configuration that wins        ✅
uvx rag-ci run --judge     # score answer grounding, not just recall ✅
uvx rag-ci judge calibrate # check the judge before trusting it      ✅

Generating questions needs the optional extra and an ANTHROPIC_API_KEY (uvx "rag-ci[generate]" golden gen …). Everything else installs and runs with no model dependency at all. See docs/golden-sets.md and docs/sweeps.md, and docs/judging.md.

Use it in CI

name: rag-ci
on: pull_request

permissions:
  contents: read
  pull-requests: write

jobs:
  retrieval:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v7
      - uses: Nokimalos/rag-ci@v0.4.1
        with:
          adapter: ragci_adapter.py
          golden: tests/golden.jsonl

The gate exits 1 on a regression and 2 when it cannot make a trustworthy comparison — an invalid run, or a baseline recorded against a different golden set. See docs/github-action.md for recording your first baseline and tuning min-effect.

A worked example lives in examples/reference/: an adapter, its golden set, and a committed baseline. rag-ci gates it on every pull request to this repository.

What makes it different

• Ground truth anchored to document passages, never to chunks. Change the chunk size and your golden set still works. This is what makes sweeping possible at all.
• Statistics as the core mechanism. Every metric carries a confidence interval. The gate uses a paired bootstrap test and blocks only when a regression is both statistically significant and large enough to matter.
• Built for real corpora. Stratified sampling for question generation, successive halving instead of grid search, index-time and query-time parameters separated so sweeps rebuild indexes as rarely as possible, and a recall-vs-pool-size curve so results measured on a sub-corpus are extrapolated honestly rather than quietly overstated.
• Agnostic by design. No RAG framework as a dependency. You write one adapter file; your stack stays yours. The adapter contract captures multi-step retrieval trajectories, so agentic and multi-hop pipelines are first-class.

License

MIT