The most common fear about AI in regulated work is hallucination - an answer that sounds authoritative but cites something that isn't there. For a tool used on contracts, claims, and medical records, a confident wrong answer is worse than no answer. Here is how InVerba is designed to avoid it.
Grounding before generation
InVerba does not answer from a model's general memory. Every answer is built from passages retrieved out of your own documents, and the model is constrained to use that retrieved material. If a fact is not in the retrieved sources, it does not belong in the answer. This is the single most important defence against hallucination: the answer has nowhere to come from except your documents.
Confidence thresholds
Retrieval is not always successful. Sometimes the archive simply does not contain what was asked. When the best retrieved passages are not relevant enough to support an answer, InVerba does not lower its standards to produce something fluent - it reports that it could not find a grounded answer. Treating "I don't know" as a valid, correct response is what keeps the tool honest.
Citations you can check
Every answer points back to the document, section, and page it relied on. This is not decoration - it is the mechanism that lets a human verify the output in seconds. An answer you can check is an answer you can use in a filing or an audit.
- Answers are constructed only from retrieved passages of your documents.
- Low-confidence questions return an honest "not found" rather than a guess.
- Every claim is tied to a named source for verification.
The trade-off we chose
A system tuned this way will occasionally decline to answer a question that a looser system would have answered - sometimes correctly, sometimes not. For everyday consumer use that caution can feel conservative. For law, insurance, and medicine, it is exactly the right default. We would rather return less and be trusted than return more and be wrong.