How answers are built
The path from a question to a cited answer: analysis, retrieval by keyword, meaning and graph, ranking, context, generation, and the guard that refuses when evidence is missing.
An answer in Breno is assembled from your own material and checked against it. This page explains each step, so that when an answer surprises you, you know where to look.
The path#
- Analyze. The question is read to decide how to search for it.
- Retrieve. Breno looks for evidence in three ways at once, and only among items you are allowed to read:
- Meaning. The question is turned into an embedding and compared with the embeddings of every passage, using a vector index.
- Exact words. A full-text search ranks passages that contain your words, with a fuzzy fallback for identifiers such as
INC-4471that full-text search splits into pieces. - The graph. Entities and relationships connected to the question add context, filtered to items you can read.
- Rank. The results are fused into one ranking with fixed weights, and the top 15 passages are kept. Filters (space, source, type, date) are applied inside each search, so the top results are chosen after filtering, never before.
- Build context. The best passages, with their provenance, are assembled for the model.
- Generate. A model writes the answer from those passages and cites them by number.
- Check. The answer is checked against its evidence. If the passages do not support an answer, the evidence-first refusal guard makes Breno say so rather than answer anyway.
Why it can refuse#
A system that always answers will sometimes invent. Breno's guard reads how strong the retrieved evidence is, and below a floor it declines. That is why an out-of-scope question gets "I could not find this in your knowledge" and not a plausible paragraph.
What the citations are#
Each number in an answer points to a retrieved passage carrying its item, source, author, date and a link to the original. The passage is the text the model was given, so you can read exactly what the claim rests on.
Confidence#
The figure shown with an answer says how strongly the evidence supports it. Treat a low figure as a prompt to open the sources, not as an error.
What each step uses a model for#
| Step | Uses a model? |
|---|---|
| Retrieve by meaning | An embedding model (to embed the question and, at capture time, each passage). |
| Retrieve by words and the graph | No. |
| Rank | No. |
| Access control | Never. It is a rule, not a model's judgement. |
| Generate and check | A generation model. If you added your own account, answers use it first. See Use your own AI model. |
How well it works#
Retrieval is measured, not assumed. A test set of 55 questions over 60 documents runs on every change: the right document is among the top three results for 83.3% of questions (floor 80%), with a mean reciprocal rank of 0.757 (floor 0.65), and the build fails below those floors. Retrieval is also measured nightly against a real model provider. The measure for faithfulness of the generated text is not implemented yet. See Known limits.