Private research prototype · recorded results
RAG Evaluation Explorer
Compare how retrieval providers did on retrieval, answers, citations and abstention in a pre-registered evaluation, and see how a single question counts towards each measure.
Recorded results · not live queries · no evaluated provider qualified
The result
No evaluated provider met every pre-registered quality target.
Every number on this page is a recorded result, copied from the archived reports of a private research prototype and checked against their archived hashes. Nothing is re-run or queried here. The questions were asked over short fictional corpora, so these are not claims about real documents. The prototype is not a production assistant, and it is separate from the RAG Knowledge Assistant architecture notes.
About this explorer
The business problem
An assistant that answers from company documents can sound right and be wrong. Before trusting one, you need to measure whether it finds the right passage, cites it, and says “I don't know” when the documents don't say.
What to try
Pick a recorded evaluation and a view, and read which targets each provider met or missed. Then step through the teaching examples to see why a question counts as a hit or a miss.
Recorded and illustrative
Recorded: the measurements, targets and pass or miss marks, exactly as archived.
Illustrative: the teaching examples, which are invented for this page.
Using this on a real business
Your own documents and questions, targets agreed before measuring, a held-out set nobody tuned on, and repeat runs when anything changes. A provider that misses a target is not ready, however good the demo looks.
1. Recorded results
Status Showing the qualifying v3 held-out set.
Met and missed marks are the ones recorded in each report. The injection measure was redefined in v3, so v1–v2 and v3 injection figures are not comparable.
2. What each measure means
3. Teaching examples: how one question counts
Teaching examples, invented for this page. Not from the recorded evaluation; nothing is retrieved or answered live.
Question and retrieved passages
Question
Response