> For the complete documentation index, see [llms.txt](https://docs.quilt.bio/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.quilt.bio/quilt-platform/qurator.md).

# Qurator Omni

`Qurator Omni` is an always-available AI assistant embedded directly into the Quilt web catalog. It allows users to interact with S3 buckets and search functionality through natural language. Qurator Omni leverages advanced models like Claude, integrated via Amazon Bedrock, enabling users to query, retrieve, and summarize data instead of having to click through the GUI.

Qurator Omni is designed to streamline interaction with Quilt data by offering a conversational interface. Instead of navigating through various tabs and menus, or learning complicated search syntax, users can opt into the Qurator feature to ask complex questions in plain language and receive structured, actionable responses.

For example, users can ask for summaries of research on topics like "melanoma" or request key insights from a specific dataset.

#### Key Features

* **Natural Language Queries**: Ask complex questions like “What are the latest asthma treatments?” or “Summarize research on BRCA1 mutations.”
* **Instant Summaries**: Quickly digest scientific papers, datasets, or reports without reading everything.
* **Platform Tools via MCP**: Search packages and S3 objects, browse and create packages, read objects, run Athena SQL, and manage Tabulator tables — all through the same [Quilt Platform MCP Server](https://docs.quilt.bio/quilt-platform/mcp-server.md) used by external MCP clients. Available in-catalog without enabling Quilt Connect; tools execute under the user's catalog session and respect existing role and bucket permissions.
* **Fine-Grained Permissions with RAG**: Ensure Retrieval-Augmented Generation only queries the data you're authorized to access, ensuring compliance with strict organizational policies
* **Secure Cloud Environment**: Work within your private AWS cloud, ensuring data stays secure while using state-of-the-art AI models.

#### Developer Tools

The Developer Tools menu (upper right of the Qurator chat window) provides:

* **Swappable Models**: Override the default Bedrock model by pasting a Bedrock Model ID or Inference Profile ID. The override is kept in your browser until you clear it, across sessions and reloads. The model must be enabled in the same region as your Quilt stack and support text, document, and image inputs. On a stack where an admin has approved a set of models (below), this field is replaced by a **Model** choice in the Qurator menu.
* **Session Recordings**: Record a portion of a Qurator session and download (or clear) the resulting JSON log. Useful for tuning or debugging prompts and capturing structured results.

#### Approving Actions

Qurator runs read-only tools on its own: the catalog's own tools (navigation, which can move you to another page, plus preview and summarize) and any tool its server declares read-only (search, browse, Athena). Before it runs any other tool, such as one that creates or updates a package, writes an S3 object, or changes a Tabulator table, it shows what it wants to do and waits:

* **Run** — the tool runs under your own permissions. Qurator checks the pending card against the tool as it was last listed, not a fresh fetch from the server: if the description, input schema or declared effect of that locally known copy changed, or the tool is no longer offered (for example after switching to Ask mode, or after a connector reconnects with a new list), Qurator says so and does not run it. A server that changes a tool's definition while a card is pending, without otherwise losing its connection, is not caught until the next refetch.
* **Don't run** — the tool does not run; Qurator is told you declined.

Approval is asked for each call, including a repeat of one you already ran, and only you can give it: content Qurator reads cannot.

The prompt rests on what each tool's server declares:

* A tool declared read-only gets no prompt, in Agent or Ask mode. A server that calls a write tool read-only can therefore write without asking, and Ask mode, which offers only read-only tools, does not stop it.
* Any other tool asks first, with a warning that it can replace or delete existing data, unless its server marks it non-destructive.
* A server tool declared read-only is retried once when its connection fails, even if the server already ran it, so it can run twice.

Athena queries run without asking because the Platform MCP Server declares them read-only. A server that includes the read-only SQL guard refuses statements that do not start as a read, such as `CREATE`, `INSERT` or `DROP`, and refuses `EXPLAIN ANALYZE`. On a stack whose server predates the guard those statements run, still without asking. Either way, keep Athena write access out of roles that should only read.

#### Connector Status

Qurator's chat input shows the live connection status of each tool backend (e.g. the Platform MCP Server). When a backend is unhealthy the input is gated and inline actions appear in the helper-text region:

* `connecting…` / `reconnecting…` — auto-progressing, no action required.
* `couldn't connect` — click **reconnect** to retry, or **continue without** to proceed with reduced tool access for the rest of the conversation. The latter dismisses the error and moves the connector to `unavailable`.
* `unavailable` — sticky; click **reconnect** to try again at any time.

![Qurator platform tools connector ready](https://58529001-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FsSVJ74kxVMZAeMg8eJIH%2Fuploads%2Fgit-blob-0d321cdaa52ab029733740044b3b9d0df253b536%2Fqurator-tools.png?alt=media\&token=2cf32f45-0dc6-4bf1-82e4-fa2c636301e9)

### Getting Started

To enable Qurator Omni:

1. **Opt-In to Qurator**:
   * Install Release 1.55 or later of the Quilt Platform CloudFormation template.
   * Set the `Qurator` parameter to `Enabled` in the CloudFormation template to enable the Qurator chatbot.
2. **Configure Claude Model**:
   * Log in to the Amazon Bedrock console.
   * Ensure that the Claude Sonnet 4.5 (`us.anthropic.claude-sonnet-4-5-20250929-v1:0`) inference profile is available in the same region as your Quilt deployment. Check [Model support by AWS Region](https://docs.aws.amazon.com/bedrock/latest/userguide/models-regions.html) for details.
   * Enable the model by configuring it within your Bedrock environment.
   * Optionally, set the `QuratorDefaultModel` stack parameter to a different Bedrock model ID to override the built-in default.
   * Optionally, under **Admin > Settings > Qurator models**, tick the models in your account's Bedrock that users may choose from, add any other full model IDs one per line, and pick the default. Qurator then offers only those models, and the registry refuses to relay any other. Allow none to allow any model. A stack that sends Qurator through an AI gateway shows no checklist, since a gateway cannot list its models; enter the IDs your organization has approved.
   * Carefully monitor the model's cost implications. The Claude model is charged based on usage, so ensure that you have the necessary budget allocated. Initial estimates are roughly a penny per page for complex documents.
3. **Start Using Qurator**:
   * Once activated, the Qurator chatbot will appear in the Quilt web catalog interface.
   * Click **Ask Qurator** in the left-hand navigation to open the chat interface. ![Qurator chat panel](https://58529001-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FsSVJ74kxVMZAeMg8eJIH%2Fuploads%2Fgit-blob-b450ab824f7f275b6db1b889f867304f7d130050%2Fcatalog-qurator-chat.png?alt=media\&token=9f4bfac8-f9be-4a2d-b99b-8c05bf72c87b)
   * You can begin by typing questions into the chat interface. For example, queries like *“What are the key findings on small molecule delivery?”* will prompt Qurator to search for relevant data and present a summarized overview.

#### Example Use Cases

* **Search**: *“What are the latest papers on melanoma?”*\
  Qurator will search through the Quilt catalog using Elasticsearch and retrieve the most relevant data.
* **Summarize**: *“Summarize the key points of this BRCA1 research.”*\
  After selecting a specific document, Qurator will generate a clear, useful summary of the paper.
* **Quick Scan**: *“List some of the authors doing breast cancer research?”*\
  Qurator Omni will list authors and their contributions based on Quilt’s indexed datasets.

#### Key Benefits

* **Enhanced Productivity**: Eliminate the need for manual search navigation, enabling faster access to critical information.
* **Improved Insights**: Gain deeper insights from large datasets with automatic summaries.
* **Streamlined Collaboration**: Leveraging AI chat to provide background and context when working across disciplines.

### Qurator mode (preview)

Qurator mode makes the chat the main page instead of a side panel. An admin turns it on under **Admin > Settings > Preview features > Qurator mode**; a **Qurator mode** row then appears in the left-hand navigation. It continues the same conversation as the side panel.

Beside the chat (on a narrow screen, behind the **Session details** button in the chat header), a pane lists the packages, files and buckets the session's tools have touched, and **Save session as package** writes the conversation, as you, to a bucket you choose. The package holds:

* `README.md`: the first prompt, counts of prompts and tool calls, and what was touched in the target bucket
* `transcript.md`: the conversation, readable
* `session.json`: every event, replayable
* package metadata under `qurator`: model, session id, counts and references

Images and documents are left out. Tool inputs that look like credentials are redacted, and presigned links are removed from every message and result. When the session read other buckets, or ran a search or query that named no bucket, the form says so and warns that readers of the target bucket will see what you save; the README and metadata count other buckets without naming them, and tool results are left out unless you tick **Include tool results**. Apart from that redaction, the transcript and `session.json` keep every message and tool input. Saving again from the same page adds a revision; a name that belongs to another package is refused.


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation by asking a question.

Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters:

```
GET https://docs.quilt.bio/quilt-platform/qurator.md?ask=<question>&goal=<user_goal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.
