Zovda
Grounded systems fail in two places: the right passage was not retrieved, or it was retrieved and not used. The two failures look identical in the output and have completely different fixes.
Teams usually start with the second, because prompts feel like progress.
Measure the first stage alone
For a sample of questions, check whether the correct passage appears in the top ten results. That single number bounds everything downstream.
If recall at ten is weak, no prompt will save the answer, and the work belongs in the corpus.
Then measure whether it was used
For questions where the passage was present, check whether the answer cites it and agrees with it. Disagreement here is a generation problem.
Separating the two numbers is the difference between a week of prompt work and a week of corpus work, and the wrong choice costs a month.
Corpus problems look like model problems
- A passage that starts mid-sentence cannot support an answer
- A stale version outranks the current one
- Three duplicates split the ranking and none wins clearly
Each of these presents as the model being unhelpful.
Keep the passage in the record
When a system answers, store which passages were retrieved and which were cited. Without that, every failure investigation starts from zero.
Zovda builds these collections through Datuva, where passages keep their document and version. The question bank collection shows how answers stay tied to their evidence.
Tell us where your system gives up. We reply within one business day.