An AI-powered Q&A platform that allows users to create rooms, record audio content, and ask questions that get answered based on the audio context using vector similarity search. (Simple app for study based)
Screenshots below 📷
- Runtime: Node.js with TypeScript
- Framework: Fastify with Zod validation
- Database: PostgreSQL with Drizzle ORM
- AI Integration: Google Gemini AI for transcription and embeddings
- Vector Search: PostgreSQL vector extension for similarity search
- Development: Biome for linting/formatting
Key Features:
- Audio transcription using Gemini AI
- Vector embeddings generation for semantic search
- Question answering based on audio context
- RESTful API with type-safe validation
- Framework: React 19 with TypeScript
- Build Tool: Vite
- Styling: Tailwind CSS with Radix UI components
- State Management: TanStack Query (React Query)
- Forms: React Hook Form with Zod validation
- Routing: React Router DOM
- Development: Biome for linting/formatting
Key Features:
- Modern, responsive UI with Tailwind CSS
- Real-time audio recording interface
- Type-safe API integration
- Form validation with Zod schemas
- Rooms: Store room information (id, name, description, timestamps)
- Questions: Store user questions and AI responses
- Audio Chunks: Store transcribed audio with vector embeddings for similarity search
- Audio Upload: Users record audio in rooms
- Transcription: Gemini AI transcribes audio to text
- Embedding Generation: Text is converted to vector embeddings
- Question Processing: When users ask questions, the system:
- Searches for similar content using vector similarity (threshold: 0.7)
- If similarity found: Answers based on audio context
- If no similarity: Answers using general AI knowledge
| Description | Screenshot |
|---|---|
| Rooms |  |
| Rooms |  |
| Room w/ no questions |  |
| When we record an audio |  |
| We separete it in chunks of audio trasncribing and extracting embeddings to handle searchs by vectors in postgresql |  |
| Then, if we make a question related to the audio mentioned (where similarity using the embeddings got at least 0.7), AI will reply it. |  |
| A future possible feature would be another person reply manually. |  |
| If it does not find similarity, it will reply based on common knowledge from ai. | |