TurboSeek

An open source AI search engine. Powered by Together.ai.

Tech stack

• Next.js app router with Tailwind
• Together AI for LLM inference
• Mixtral 8x7B & Llama-3 for the LLMs
• Bing for the search API
• Helicone for observability
• Plausible for website analytics

How it works

1. Take in a user's question
2. Make a request to the bing search API to look up the top 6 results and show them
3. Scrape text from the 6 links bing sent back and store it as context
4. Make a request to Mixtral-8x7B with the user's question + context & stream it back to the user
5. Make another request to Llama-3-8B to come up with 3 related questions the user can follow up with

Cloning & running

1. Fork or clone the repo
2. Create an account at Together AI for the LLM
3. Create an account at SERP API or with Azure (Bing Search API)
4. Create an account at Helicone for observability
5. Create a .env (use the .example.env for reference) and replace the API keys
6. Run npm install and npm run dev to install dependencies and run locally

Future tasks

• Move back to the Together SDK + simpler streaming
• Add a tokenizer to smartly count number of tokens for each source and ensure we're not going over
• Add a regenerate option for a user to re-generate
• Make sure the answer correctly cites all the sources in the text & number the citations in the UI
• Add sharability to allow folks to share answers
• Automatically scroll when an answer is happening, especially for mobile
• Fix hard refresh in the header and footer by migrating answers to a new page
• Add upstash redis for caching results & rate limiting users
• Add in more advanced RAG techniques like keyword search & question rephrasing
• Add authentication with Clerk if it gets popular along with postgres/prisma to save user sessions

Inspiration

• Perplexity
• You.com
• Lepton search