Mercury enables users to train custom GPT models on their own documents and websites, offering a Perplexity-style UI for interactive chat with cited sources. It supports a range of file types including PDFs, DOCX, TXT, and image files, with plans to add CSV and PPTX support.

How It Works

The system leverages OpenAI's embedding API (text-embedding-ada-002) to convert text from uploaded files or scraped webpages into vector embeddings. These embeddings are stored in a Pinecone vector database. When a user queries the system, their prompt is also embedded, and a similarity search is performed against the Pinecone index. The most relevant document chunks are retrieved and used to construct a prompt for GPT-3, which generates a streamed response with source citations.

Quick Start & Requirements

• Install dependencies: npm i
• Requires Pinecone account and API key, OpenAI API key.
• Pinecone index must have dimensions 1536.
• Configure .env.local with API keys and Pinecone details.
• Start the app: npm run dev
• View at http://localhost:3000
• Official Docs: https://github.com/Jordan-Gilliam/ai-template

Highlighted Details

• Utilizes Next.js API Routes with Edge runtime for streaming responses.
• Integrates Tailwind CSS, Radix UI Primitives, and Lucide icons for a modern UI.
• Employs OpenAI's text-embedding-ada-002 for embeddings and GPT-3 for responses.
• Stores embeddings in Pinecone for efficient similarity search.

Maintenance & Community

• Inspired by projects like gpt4-pdf-chatbot-langchain and gpt3.5-turbo-pgvector.
• No explicit community links (Discord/Slack) or roadmap mentioned in the README.

Licensing & Compatibility

• The README does not specify a license.

Limitations & Caveats

The project is presented as a template, implying it may require further development for production use. Support for .csv and .pptx files is listed as "Coming Soon."