pgchat

A chat agent for your Postgres database, powered by Inferable.

Features

• Chat with your postgres database without exposing connection strings
• Agent dynamically learns your database schema as you chat
• Supports complex queries and data-modifying operations
• Approval mode enables you to review and approve queries before they are executed
• Privacy mode allows you to return data to your terminal directly, without sending it to the model
• Works locally, or deployed in a private subnet

Quickstart

This will connect a local postgres database to the pgchat agent.

npx pgchat postgresql://user:pass@localhost:5432/db

Usage

npx pgchat <POSTGRES_CONNECTION_STRING> [options]

Options:
      --approval-mode  Approval mode: "always" (all queries), "mutate" (only
                      data-modifying queries), or "off"
               [string] [choices: "always", "mutate", "off"] [default: "always"]
      --privacy-mode   Enable privacy mode. All data will be returned as blobs
                      (not sent to the model)         [boolean] [default: false]
      --schema        Database schema to use        [string] [default: "public"]
      --secret        Inferable API cluster secret                      [string]
      --cluster-id    Inferable API cluster ID                          [string]
      --endpoint      Inferable API endpoint                            [string]
  -h, --help          Show help                                        [boolean]

Rate Limits

pgchat is powered using Powered by Inferable. API usage limits apply, these can be be increased by providing an Inferable API secret --secret / Cluster ID --cluster-id.

Note: When using pgchat without providing an --api-secret, it will use an ephemeral Inferable cluster which is automatically cleaned up after 24 hours.

Inferable is also self-hostable for complete control over your data and privacy, pgchat can be configured to connect to a self-hosted instance of Inferable using the --endpoint flag.

FAQ

How does it work without exposing connection strings?

All pgchat queries are proxied via a function that runs on your local machine. As a result, the execution environment is completely isolated.

How do I know it's not going to drop my database?

By default, pgchat runs in approval mode, which means that all queries are sent to the model for review before execution. You will see the query in the terminal before it is executed.

How does it work without connection tunneling?

pgchat uses Inferable's Node.js SDK to ferry commands from your local machine to the Inferable Control Plane. The SDK works with long polling the agent.

Does the LLM see my data?

Yes by default. However, you can opt out of this by running with --privacy-mode. The tradeoff is that you won't get the benefits of the LLM's reasoning on your data. (This uses Inferable's blob storage to return data to your terminal without sending it to the model.)

What's Inferable?

Inferable is an open-source platform for building AI agents that interact with your internal systems that don't have public APIs (like a small service that wraps a Postgres database). For more, see Inferable. pgchat uses Inferable's free tier, chat state management, durable execution engine and the Re-Act agent to bring this to life.