How the agent answers
The Datastat agent answers questions from the reviewed documentation available to the person asking. It finds the evidence first, then uses a language model to explain that evidence in plain language.
The agent does not treat model memory as company documentation.
What happens after you ask
- Confirm access. The agent identifies the user and the information they are allowed to receive.
- Search the documentation. It searches titles, descriptions, product areas, headings, keywords, and page text.
- Retrieve the smallest useful sections. It brings only the evidence needed for the question into the answer.
- Prefer the best source. Published, recently reviewed, directly relevant pages rank above older or indirect material.
- Check for disagreement. When several pages apply, the agent compares them before answering.
- Explain with a language model. The model turns the retrieved evidence into a concise response without changing its meaning.
- Cite and qualify. The answer links to the source and names any problem with the evidence.
The result is an answer that is easier to use than a list of links but still leads back to the maintained source.
Where the language model fits
The language model is the explanation layer. It can summarize, connect related steps, adapt wording to the question, and ask for missing context.
The retrieval and access layers decide what evidence the model receives. A fast, lower-cost model handles ordinary questions. The agent selectively uses a more capable model for difficult multi-source reasoning or an unclear request. Changing models does not change the documentation contract: every factual answer still needs authorized evidence and a citation.
The model may know general facts from training, but that knowledge does not override a Datastat source. When documentation is absent, the agent says that the answer is not documented instead of presenting a guess as policy or procedure.
What an answer contains
| Part | Purpose |
|---|---|
| Answer | The concise result or next action. |
| Based on | A link to the document and, when useful, the exact section. |
| Reviewed | The source's last review date. |
| Caution | A warning when evidence is stale, incomplete, or conflicting. |
The agent can cite more than one page when a workflow crosses products. It does not hide a disagreement by choosing whichever statement is easier to use.
How uncertainty is handled
| Evidence state | Agent behavior |
|---|---|
| Current and consistent | Answer and cite the relevant section. |
| Missing | Say that the answer is not documented and identify what source is needed. |
| Stale | Use it only as qualified context and show the review warning. |
| Conflicting | Name the conflict, cite both sources, and send it for owner review. |
| Unauthorized | Do not quote, summarize, or reveal the protected material. |
The agent stops at the boundary of the evidence. This is a useful result: it shows where the documentation or approval process needs attention.
Access comes before retrieval
Classification labels describe content; they do not grant access. Authorization must happen before protected text reaches a language model.
- Public indexes contain only reviewed public material.
- Internal material is retrieved only for an authenticated, authorized user.
- Unauthorized content is not included in prompts, excerpts, logs, or summaries.
- Secrets do not belong in Markdown, generated indexes, or environment-facing prompts.
A simple example
If someone asks, “Can two JudoNet computers send data online at the same time?”, the agent finds The master switches, explains that only one JudoNet may enable the downstream sending switches, and cites that section. It does not rely on a remembered answer when the maintained page is available.
For the source lifecycle behind those answers, read Publish and maintain documentation.