No-Code Langchain

• 

  Get an answer about your preferred library connecting to your preferred no-code environemtn

Inspiration

No-coders want to build, but they need some code to make it happen. Traditional no-code systems make a visual environment that provides a deterministic framework for managing layout and flows. LLMs and especially the chat metaphor are good at creating code based on slightly more vague instructions, which can be particularly useful for managing integrations and control flows.

However, ChatGPT gets things wrong a lot by hallucinating about code that does not exist. But people are not sitting in a VS COde environment with everything locally. This is the opportunity for the no-code AI helper.

What it does

This bot will cache the github repository of whatever interesting widget you want to integrate with webflow, and give you instructions specific to webflow and that repository. It avoids hallucination by using semantic search to find relevant pieces of code and present them to the prompt.

How we built it

We learned langchain, advanced OpenAI, pinecone on the AI side. We deployed with Netlify, Clerk and Convex on the back-end.

We initially built with langchain python, but converted to langchain TS to make it easier to integrate with node-based backend tech.

Challenges we ran into

Cloning the git repository was a lot more work than we had planned! Ordinary lambdas don't want to do this work. We wound up using a docker-based system that is part of statechange.ai to get a low-code solution fo this toughest part of the equation. Once we had cloned the data locally, indexing and prompting were straightforward.

On the front-end, our biggest challenge was streaming the output to give a chatgpt-style experience to customers. Props to convex for making this remarkably natural with their websockets-based interface.

Accomplishments that we're proud of

We cracked a lot of the hard parts in just a few days after knowing very little about langchain, vector databases, or these new backend vendors that we integrated with in this project. We built a functional prototype that proves out the potential for really useful script-tag based integrations coming out of this tech.

What we learned

We learned pretty much everything from the start. How Convex works. How langchain thinks. How Pinecone works. Pinecone namespaces to make more efficient searches. How to tune the prompt to get it to focus on our particular problem area.

What's next for No-Code Langchain

This is awesome. No-coders are looking for power without having to train up. They are willing to look for relevant packages, and then get a little help gluing them together. That's the opportunity in front of us - an AI agent that helps with that gluing. Connecting more no-code systems, as well as the intersection of multiple packages. And then overlaying with the human helping hand for when chatgpt becomes wronggpt. That's a high leverage model we're looking forward to exploring - tomorrow!

Built With

• clerk.dev
• cohere
• convex
• langchain
• netlify
• pinecone
• typescript