Skip to main content

PGMate Copilot POC

· 2 min read
Marco Pegoraro
Marco Pegoraro
founder

Copilot brings real-time context-aware AI to PGMate.
It enables the chat with your data scenario with in-line code-blocks execution.

There is a lot of fuzz around the CHAT WITH YOUR DATA eldorado.

After reading an article about QueryGPT at Uber I found myself thinking: "How hard could that be?" 🤓
So I've decided to give it a go.

This demo showcases what is already available to all users that pull the PGMate image from DockerHub.

Setup:​

You simply have to provide an OpenAI Api Key as environment variable and you are good to go:

docker run \
    -p 8080:8080 \
    -e PGMATE_OPENAPI_KEY=xxx \
    pgmate/pgmate:latest-full

Postgres LLM Context​

For this first POC, Copilot uses a one shot prompt approach to try solving the user's request.

What makes it working so well is the real-time context that is injected into the conversation.

I'm having a lot of brainstorms about how to optimize this context for accuracy and economical convenience.

🤓 Because that's the real challange.

Any fool can copy/paste the full context into ChatGPT and let it crunch by a powerful model. But this could end up being very costly, and also may lead to inaccuracies. Imagine overflowing the model with tens of tables metadata when the business request is only affecting 2 of those.

👉 Too much information is both useless and harmful.

Plans for the Future​

For the engine, the goal is to move to a multi-step agenting oriented approach:

  1. A first agent receives a comprehensive but shallow context and is tasked to figure out which objects are needed to solve the request
  2. A second agent receives a deep context on the resources previously identified and it is tasked with solving the reuquest

I plan also to introduce a goal-detection agent that would be used to navigate a prompt-tree and specialize prompts based on intentions.

Last but not least, I plan to work on the conversation history introducing an agent that will digest a "compact memory" of cold messages with the goal of balancing memory and token consumptions.

On the User Experience side, the idea is to move the chat to be stored in the support db (now it's local storage) and enable a classic multi-conversation experience.