Question 1

What is jev-rag-guard?

Accepted Answer

It is a typed-decision accuracy layer for RAG pipelines. It does not retrieve and does not generate; it sits between the two stages, runs six small typed checks, and gives you a trace of what it decided and why.

Question 2

What are the six checks it runs?

Accepted Answer

Intent runs before retrieval and can stop safety, legal or privacy-sensitive queries. Relevance and clarify run after retrieval, dropping noise and asking for missing information. Sufficiency runs before generation, grounding after it, and caveat last to flag answers that depend on an unspoken condition or exception.

Question 3

How does jev-rag-guard relate to Jev?

Accepted Answer

Each check is a single bounded yes/no or closed-set judgment with a calibrated probability attached, which the README describes as the same shape used by typed decision models like TypeSafe's Jev over program state. An optional typesafe extra wires in a native typed-decision model instead of the built-in heuristic.

Question 4

How do I install it?

Accepted Answer

It requires Python 3.10+. Install the core with pip install -e ., which has zero required dependencies, and add an optional backend extra such as .[openai], .[anthropic] or .[typesafe] if you want a real decision model instead of the built-in heuristic.

Question 5

What does the sufficiency gate do?

Accepted Answer

It runs before generation and asks whether the kept context is enough to answer confidently and completely. On a no, generation is skipped and the pipeline returns an honest I don't know; on a partial match it generates with an instruction to answer only what the context supports.

Question 6

Does it replace my retriever or generator?

Accepted Answer

No, deliberately. The README says rag-guard does not retrieve and does not generate; it wraps the seam between the two with six checks and is meant to be added around an existing pipeline. The README even has a What this deliberately doesn't do section.

Question 7

What are its limits?

Accepted Answer

The README visible in the digest cuts off at the start of the Quickstart, so the API surface beyond the check table and install commands is not documented here, and no benchmarks or accuracy numbers are published. The default backend is a built-in heuristic, not a learned decision model.

Jev RAG Guard

Highlights

Quickstart

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community