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

# Turn on deep research for an endpoint

> Give AI agents a second tool that runs several searches across your sources and returns one answer with numbered, linked citations.

Search answers a question with one retrieval. Some questions need more: comparing coverage over time, building a list or timeline, counting something, or answering a question with several parts. Deep research handles those. It plans a set of searches over the endpoint's sources, checks whether the evidence covers the question, searches again where it falls short, and writes one answer with numbered citations that link back to your content.

Deep research is slower than search. A run takes up to about three minutes.

> [!NOTE]
>
> Deep research is available on request. [Contact us](https://www.raily.ai/contact) to turn it on for your Account. Until it is on, the **Deep research** switch does not appear on your endpoints.

## Before you begin

* Deep research is on for your Account.
* The endpoint serves search over at least one collection. The switch only appears on endpoints that have a search tool.
* You have permission to edit endpoints in the Organization that owns the endpoint.

## Turn it on

**1. Open the endpoint**

In **Search Rail**, select the endpoint under **Your audience**. Its **Endpoint settings** page opens.

**2. Switch on Deep research**

Find the **Deep research** card and turn the switch on.

**3. Save**

Click **Save Changes**.

You can also turn on **Deep research** when you create an endpoint. The switch sits under **Search results with UI** on the **New endpoint** form.

Agents pick up the new tool the next time they load the endpoint's tool list. In Claude or ChatGPT, that usually means reconnecting or starting a new chat.

To turn it off, switch **Deep research** off and click **Save Changes**. The tool disappears from the endpoint. An agent that still has the old tool list cached can try to call it until it reconnects, and that call fails.

## What agents see

With deep research on, the endpoint offers two tools: your search tool and `deep_research`. Both read the same sources with the same search config.

The agent decides which one to call from the tool descriptions. Raily writes those for you:

* The `deep_research` description tells the agent to choose it for comparisons across time or sources, how something evolved, lists and timelines, multi-part questions, counting, and anything that needs more than 8 results. It then includes your search tool's description, so the agent knows what the content covers.
* The search description gains one sentence that points to `deep_research` for those questions.
* Both tell the agent to use search for a single fact, a specific item, or the latest news, and to use search when unsure.

## What an answer looks like

The answer depends on how the caller signs in.

**AI agents signed in through the endpoint's sign-in (Claude, ChatGPT).** The agent gets a written answer with inline markers like `[1]` and `[2]`, followed by a **Sources** list. Each source line shows its number and title, linked to the original item when your content has a link for it:

```text theme={null}
...the policy changed twice in 2025 [1], and coverage shifted after the March ruling [2].

Sources:
- [1] [Article title](https://example.com/article-1)
- [2] [Another title](https://example.com/article-2)
```

**Apps that call with an API key.** The app gets typed data instead of prose formatting, with a short status line (`1 answer`):

| Field | What it holds |
| - | - |
| `content` | The written answer, with `[n]` markers. |
| `sources` | One entry per cited source: `number`, `title`, `url` (or `null` when the item has no link), and `collection`. |
| `coverage_score` | A number from 0 to 1: how well the gathered evidence covers the question. |
| `provisional` | `true` when the research stopped before the evidence was judged sufficient, for example at the time limit. |

Links come from the link field in your collection's display settings. Titles come from its title field.

## Time limit and provisional answers

A run has about three minutes. If time runs out while the answer is being written, deep research returns what it found instead of failing:

> I ran out of time before I could write the answer. These are the most relevant sources I found.

That reply lists the top five sources. If the research stopped early but an answer was written, API-key apps see `provisional: true`.

When no supporting sources turn up at all, the reply says so plainly and suggests a narrower question or search. Deep research never writes an answer without evidence.

## Hourly limit

Each Account can run 30 deep research calls per hour by default. The limit is shared across every endpoint in the Account and with the **Deep research this** button in **Test your endpoint**. [Contact us](https://www.raily.ai/contact) if you need a different limit.

When the limit is reached, the agent gets this message instead of an answer:

> You've reached the deep research limit for now. Try again later, or use search.

The count is approximate. Calls that run at the same moment can briefly go over it.

## Try it in Test your endpoint

On an endpoint with deep research on, **Test your endpoint** shows a **Deep research this** button under each answer that ran a search. Click it to run deep research on the same question. The panel shows the answer and its numbered, linked sources, headed **Deep research · N sources cited**.

These runs count toward the hourly limit.

## See deep research in Evals

Deep research calls are graded in [Evals](https://docs.raily.ai/features/evals.md) like searches, using the sources the answer cited. Each one carries a **Deep research** badge. To see only these calls, set the question filter at the top of the list to **Deep research**.

## Next steps

> [card] [Connect to Claude](https://docs.raily.ai/connect/claude.md)
>
> Add an endpoint to Claude and sign in

> [card] [Evaluations](https://docs.raily.ai/features/evals.md)
>
> Check how well searches find the right content


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.