# How answers work

## Grounded in your pages

When a visitor asks a question, Docora searches only your project's indexed pages. It combines meaning-based search with keyword search, picks the best passages and gives them to the model as evidence. The model is told to answer only from that evidence.

Citations are built from the retrieved pages. The model never writes a link: a marker like `[1]` in an answer is matched to a retrieved page, and a marker that matches nothing is dropped.

## When the docs do not say

If no passage is relevant enough, Docora does not call the model for an answer. The visitor sees your `fallbackMessage`, in their own language, and the stream ends with outcome `no_evidence`. Questions like this are listed for you in the dashboard so you can see what your docs are missing.

## Any language

Visitors can ask in any language about docs written in another. Search is multilingual, and the answer is written in the language of the question. The sources keep their original titles and text.

## Documentation is data, not instructions

Pages you index may contain text that tries to give the assistant orders. Docora treats all retrieved text as evidence only. Questions that try to override the assistant's rules, reveal its instructions or change its role are declined before any search, in many languages, and the answer is also checked so that internal instructions cannot leak.

## Conversations

A conversation keeps its history on the server, so a client cannot rewrite earlier turns. Each conversation has a turn limit (see [limits](/guides/limits)); after that the visitor starts a new chat.