Konverse AI

Inspiration

Everybody is using Gen-AI nowadays, and the LLMs behind Gen-AI comes pre-trained with no so recent data. Konverse lets users use their own documents and our blog posts to provide context for Gen-AI to enhance relevance, personalization and more accurate content.

What it does

Konverse leverages Convex vector search to find documents and blog posts that match a query and use them for context with Gemini API

Basically with Konverse you can chat with your document for easier content retrieval and increased productivity. Konverse leverages Gemini API to generate embeddings for vector search and content generation.

How we built it

I have built Konverse using Typescript and Python. Using Convex httpActions for FileStorage, Queries, Mutations and Actions.

To extract text from .pdf, .docx, .xlsx, .html i used Python adding embeddings to Convex Mutation directly using convex Python Module.

Challenges we ran into

Uploading large documents (text and Embeddings) using Convex Mutation returns error from Python Module if exceeds 1MB.

Using Convex Auth with httpActions not returning sessions for authenticated users. Convex request limit, somehow limits the size of document one can upload. Convex Vector search returns almost all documents even if does not fit, though still in dev, cant wait for the future roll outs.

Accomplishments that we're proud of

Konverse

What we learned

Though developing Konverse i continue to learn more about Convex and how to use it with other tech stack such as python Flask e.t.c

What's next for Konverse AI

KonverseAI is an MVP, still in active development.

Built With

• angular.js
• convex
• flask
• python
• typescript
• vercel