Ursa Copilot

Ursa Copilot is a tool to help data engineers make sense of healthcare claims data packages. It sets up a three-way conversation between the data engineer, the LLM, and the procedural code of Ursa Copilot.

Ursa Copilot drives the conversation based on a series of questions in an editable prompts file. It poses the questions to the LLM, which has the option of running a SQL query, asking the data engineer for clarification, or attempting to answer the question with an assertion that can be accepted or rejected from the data engineer.

graph TD;
    style User fill:#71bf00,stroke:#333,stroke-width:2px;
    style Ursa_Copilot fill:#71bf00,stroke:#333,stroke-width:2px;
    style Database fill:#999,stroke:#333,stroke-width:2px;
    style LLM fill:#71bf00,stroke:#333,stroke-width:2px;
    User([Data Engineer]) --> |Reviews Work<br>Answers Clarifications<br>Accepts/Rejects Assertions| Ursa_Copilot([Ursa Copilot]);
    Ursa_Copilot --> |Forwards Clarifications<br>Forwards Assertions| User;
    Ursa_Copilot --> |Executes SQL| Database[(Database)];
    Ursa_Copilot --> |Structures Conversation<br>Forwards SQL Results | LLM([LLM]);
    Database --> |Returns SQL Results| Ursa_Copilot;
    LLM --> |Writes SQL<br>Asks Clarifications<br>Makes Assertions | Ursa_Copilot;

How to use

There are three ways to use Ursa Copilot.

Run Ursa Copilot locally

The git repo of Ursa Copilot ships with a sample implementation in Next.js, which can be run locally on your workstation, as follows:

 $ (cd engine && npm install)
 $ cd nextjs-app
 $ npm install
 $ npm run dev

You will want to set DATABASE_TYPE and TARGET_DATABASE_URL in /nextjs-app/.env.local. You can edit the prompts in the prompts directory to fit your workflow.

Integrate Ursa Copilot into your application

The npm package of Ursa Copilot contains just the engine, without the Next.js sample implementation, and it can be incorporated into an existing application.

 $ npm install ursa-copilot

You can look at the nextjs-app directory as a reference implementation, but usage will look something like:

const options = {
    log: (contents) => {
        sendLogToUser(contents);
    },
    prompts: {
        "investigate-pharmacy": "Override prompt text"
    },
    promptUser: async (message) => {
        const userResponse = await waitForSocketResponse();
        return userResponse;
    },
    query: async (sql, params) => {
        const result = await executeDatabaseQuery(sql, params);
        return result;
    },
    sendMessage: (message: string, type: string) => {
        sendMessageToUser(type || "message", { text: message });
    }
}
await investigate("pharmacy", tableName, existingDocumentation, options);

Use Ursa Studio

If you're looking to use Ursa Copilot in the context of a battle-tested enterprise SaaS platform which covers the full breadth of healthcare data needs, contact [email protected] and ask about Ursa Studio, which has Ursa Copilot as one of its features.

HIPAA concerns

Yes, PHI is absolutely going to be flowing through this system. What makes Ursa Copilot HIPAA-compliant is that it uses AWS Bedrock and leverages your existing AWS account. You'll need to have an AWS account, with the BAA paperwork and all the other compliance steps in place.