About grounded answers

Context

Every hosted site includes an assistant visitors can ask in plain language. It answers only from this workspace index, cites the page behind every claim, and says so when it cannot ground an answer.

Why

Index first, model second. Strong matches can answer with zero AI tokens. Harder questions use a short or full model call with citations and spend workspace credits. That split keeps common questions cheap and keeps every claim checkable.

Trade-offs

The assistant will not invent an answer from general world knowledge. That is less fluent than an unconstrained chatbot, and more honest. Public widget chat only retrieves public content. Private dashboard chat applies permission filters.

Further reading

*(see How-to: How to connect a source)* *(see Reference: Analytics)* *(see How-to: How to expose docs over MCP)*

Grounded in your docs

Your workspace knowledge index is filled when you publish a package and when you sync connected sources. When a visitor asks a question, the assistant retrieves the most relevant passages from that index and answers from those. It does not answer from general world knowledge, and it cannot see another workspace's content.

Index first, AI when needed

We search your knowledge index before calling a model. Strong matches can answer with zero AI tokens (index-only or cache). Harder questions use short or full AI with citations, and those turns use workspace credits. In the dashboard, Chat quality and AI usage show index-only rates for the last 30 days so you can see the mix.

Finding the right page, two ways at once

Retrieval searches your content by meaning and by exact wording at the same time, and the two rankings are blended before the best passages are chosen. That matters at both ends of how people ask: a question using your docs' precise terminology matches on the words themselves, and a loosely phrased question still lands on the page that actually answers it. Page titles count extra, so the authoritative page for a topic rises above pages that merely mention it.

When helpful, the assistant can expand one hop through your docs knowledge graph so related pages enter the candidate set for multi-hop questions, still scoped to the same workspace and published version.

Public chat and private chat

The embedded widget on your docs site is public and only retrieves content marked public. Workspace members who are signed in can use private docs chat from the dashboard, which applies permission filters so restricted source content is only available to callers with matching grants. See also How to connect a source.

Meet Doxie

Doxie is our dachshund mascot. On your published docs site the chat launcher uses the agent sparkles icon, multi-session history stays on the visitor's device, and the panel footer says Powered by Doxie once (the Doxie mark is decorative; light and dark safe). The company behind the product remains Dev Docs LLC; Doxie is the friendly face of grounded answers.

Every claim is cited, and checked

Each claim in an answer links back to the page it came from, so anyone reading it can open the source and verify it in one click. Before an answer is shown, every citation is checked mechanically against the retrieved source text: the cited passage must be a real, exact excerpt of the page it points to. A citation that fails the check is dropped, and a statement with no verified citation behind it is suppressed rather than shown. Citations are the point, not a decoration: they turn an answer into something a reader can check rather than something they have to trust.

It says so when it cannot ground

The honest failure mode is the important one. Before an answer is shown, its claims are checked against your published documentation, and anything that cannot be verified is left out rather than presented as fact. If the docs do not cover a question, the assistant says it does not know instead of filling the gap with a guess.

Quality is measured, not assumed

We hold the assistant to its grounding promise with an automated evaluation that runs every night against the live product: a fixed set of questions with known-correct outcomes, scored for whether answers were grounded when they should be, declined when they should be, and cited accurately. The run fails loudly if any score drops below its threshold, so a quality regression is caught within a day instead of drifting silently.

Turn gaps into a to-do list

When the assistant cannot ground an answer, that question is logged for you automatically and ranked by how often it was asked. In the dashboard Content gaps view you can search the index for each question, copy it, open Sources to re-sync or add coverage, or open Sites when the answer is ready to promote. See also Analytics.

Your data is not training data

We do not train, fine-tune, or otherwise improve any model on your documentation or your visitors' questions, and our agreement with our model provider does not permit them to either. Thumbs up or down feedback from readers goes only to your analytics, never anywhere that trains a model. For the full statement, see Security.