> ## Documentation Index
> Fetch the complete documentation index at: https://partners.centaur.io/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Assistant

> How the in-app Centaur assistant uses tool-backed data.

The in-app assistant answers questions about Centaur concepts and live Centaur data.

## How it works

Conceptual answers can be answered directly. Any live data claim must be grounded in one of the assistant's Centaur API tools.

The assistant can use tool calls for:

* the activity feed
* trader and asset discovery
* events
* messages
* generated narrative summaries
* positions and open positions
* trader and asset stats
* trader rankings
* activity summaries

Visible tool calls show when the assistant queried Centaur data.

## Time handling

The assistant uses a single UTC server timestamp as the time anchor for each response.

Common relative periods resolve as:

| User phrase   | Resolution                                                                  |
| ------------- | --------------------------------------------------------------------------- |
| **last 24h**  | Rolling 24 hours ending at the current UTC server time.                     |
| **today**     | Current UTC calendar day through the current server time.                   |
| **this week** | Current UTC week starting Monday 00:00:00Z through the current server time. |
| **recent**    | Rolling 7 days unless the user names a different duration.                  |

For live or recent market-data requests, the assistant should pass explicit `startTime` and `endTime` values instead of relying on omitted defaults.

## Guardrails

The assistant should:

* use the feed for "what is happening" and recent-activity questions; feed rows are already curated and are presented as returned
* use `list_traders` and `list_assets` to resolve unknown IDs before stat calls
* use open-position reads for current positioning and open-position skew
* use generated narrative summaries for themes, narratives, drivers, asset sentiment, and risks
* use trader rankings and activity summaries for "most active" or "best performing" questions instead of counting raw rows
* avoid using generated summaries as evidence for exact trade counts, activity rankings, or current open-position skew
* skip assumed or auto-generated system events in raw event listings
* qualify answers that depend on partial pages, failed tools, or unavailable data

The assistant keeps recent chat state in the browser session and sends only the recent message window needed for a new response.