Inspiration
When I was interviewing at a ~20 person startup and reading through their introduction doc, I read that the co-founders would often spend lots of nights just catching up on docs from the prior week. To me, the last thing any employee, much less a founder, should be spending their time on is writing docs. Ground Truth lets you simply review and accept changes to your docs based on what you have been coding.
What it does
Each time you make a commit to a specific repo, a new entry is made
How we built it
Chroma
Docs can be pretty large. Far too large to pass multiple pages into a context window. To account for this, we embedded each page of the docs into a chromaDB using their built in functions, and we query the Vector DB we created to find the most similar doc so we can pass that as context.
Groq
We used Groq for basically all of our small context text completion. This was especially helpful when passing a code diff into the llama model and getting a description of what was being updated. This allowed us to consistently retrieve the right documentation that we would then update
Reflex.dev
Since most of our initialization and ChromaDB stuff was already being done in Python, it made sense to continue using it and go ahead with Reflex for the full frontend and backend.
Challenges we ran into
The biggest issue for us was finding good contenders for projects. Firstly, they had to be open source, non negotiable for us to test with. Secondly, they had to have a developer program or some need for docs that we could reasonably update.
Accomplishments that we're proud of
This was all of our first times working with any of these technologies (RAG anything, really) and we're proud to have put it all together in a somewhat attractive way over the 36 hours. The approach we took probably was a good 10 hours of ideation so it's nice to have it built and working
What we learned
Pretty much everything about RAG and Reflex. We all knew very little about the domain coming into this.
What's next for Ground Truth
This project has real potential as a company but is an incredibly hard engineering problem. If we stay excited about building it, we could make it widespread and very modular.
Log in or sign up for Devpost to join the conversation.