Supersimple CoBrain

Inspiration & prior art

I had previously trained a language model that answers questions from databases and precisely lays out its thought process (part of our product at Supersimple).

Connecting this model with the rest of a company's knowledge was the natural next step!

What it does

A ChatGPT Plugin that lets the user ask questions across databases and internal tools like Notion and Linear. One question can span multiple data sources & answering it can take multiple steps.

Any database queries are accurately explained via a no-code UI (via Supersimple), and any information from Notion/Linear is backed up by specific reference links.

How we built it

Used a couple of different pre-existing APIs as data sources:

• Supersimple (which in turn connects to data warehouses, Mixpanel, Google Ads etc)
• Notion
• Linear

The plugin itself is a small FastAPI server.

The Notion integration & embedding-generation is done as a separate background job; Weaviate's is used as our vector database for semantic search. The most recent Linear issues are searched by ChatGPT itself, and Supersimple is queried through natural language that Supersimple's own AI handles.

Challenges we ran into

Notion's API wasn't great for our use case – web scraping would have been significantly easier, but we did eventually build a proper integration that normalizes their blocks & generates reasonable embeddings for them.

Accomplishments that we're proud of

The way multiple data sources can be used together to answer complex questions.

What we learned

This was my first ChatGPT Plugins rodeo, so – how that works! Also got to go deeper into some new vector databases.

What's next for Supersimple CoBrain

• Getting the plugin released
• Making the generated database explorations even more in-depth & flexible
• Adding more pieces of a company's knowledge, such as calendars and Slack

Built With

• fastapi
• linear
• notion
• python
• supersimple
• weaviate