JevRAG

RAG pipelines constantly make mid-flight calls — is this chunk a coherent unit, is the evidence gathered enough, which passages earn a context slot, is a cached answer safe to reuse, does the answer stay grounded — usually with a hardcoded threshold buried in code. JevRAG turns each into a first-class, measured decision.

Jev is the first backend wired behind a swappable Decision interface: hand it the relevant state, get back a typed answer with calibrated confidence, and let the calling code decide what to do with that confidence. A second logprob backend demonstrates the abstraction generalizes beyond Jev.

Five primitives: evidence-sufficiency (val n=300, test n=700), chunk-boundary (generalization hash-proven), context-selection (wrap and from-scratch rebuild on two real benchmarks), answer-abstain (benchmarked against a free alternative, n=300, tie) and cache-trust (extracted from prior art, independently corroborated). The real pipeline chains all five in two live end-to-end runs.

`pip install jevrag`, or `uv add jevrag` in a uv project, or `uvx jevrag --help` to run the CLI once without installing. It requires Python 3.11 or newer, and the README's quickstart covers keys and a first run.

It applies a per-decision threshold, and the README reports that one real bug was found in it, fixed, and disclosed. Shadow/observe mode complements it by running a challenger against production in a real what-if comparison.

No. It is a shared abstraction and evaluation harness for the decisions inside a RAG pipeline; the backend behind the Decision interface can be Jev, a different model, a trained classifier or a simple rule. Retrieval itself remains yours.

All results are the author's own, the README's summary table points to per-primitive caveats in docs/primitives.md that we did not review, and the answer-abstain primitive tied a free alternative rather than beating it. The project is at version v0.3.0 with tests running in CI.