Back to projects
CompletedRAG & AISolo
RepoInsight AI
End-to-end RAG system specialising in semantic analysis of source code, documentation and GitHub issues. It generates architectural diagnostics and code patches. Its Voice-First interface relies on a bidirectional Speech-to-Speech pipeline coupled with a real-time holographic avatar (lip-sync) via WebSockets.
95 %
Ragas relevance score
≤ 8 GB
VRAM (local GPU)
28.9 s
Average end-to-end latency
Architecture
- Voice input
- Silero VAD
- Faster-Whisper
- Next.js interface, WebSockets
- Code indexing
- tree-sitter AST parser
- Jina Embeddings
- Qdrant
- Retrieval
- Hybrid Dense + BM25
- Cross-Encoder Reranker (BAAI)
- Reciprocal Rank Fusion
- Generation
- Local LLM (Ollama)
- Diagnostics and code patches
- Output
- Kokoro-ONNX
- Holographic avatar (lip-sync)
Stack
PythonNext.jsFastAPIQdranttree-sitterOllamaFaster-WhisperKokoro-ONNXSilero VADJina EmbeddingsBAAI RerankerWebSocketsDocker
- Avoiding the function truncation inherent to standard text chunking
- Retrieving the right excerpt across code, documentation and issues
- Full-duplex WebSocket communication for real-time voice interaction
- AST parser based on tree-sitter to extract code by logical entities (full classes, methods)
- Hybrid search (Dense + BM25) and Cross-Encoder Reranker with Reciprocal Rank Fusion
- WebSocket infrastructure streaming Kokoro-ONNX audio, with real-time lip synchronisation on the avatar
- Decoupled architecture with async tasks to separate STT, LLM and TTS
- 95% Ragas relevance score
- Average end-to-end latency of 28.9 s on a local GPU, with at most 8 GB of VRAM
- Architectural diagnostics and code patches generated from the analysed repository
A question about this project?
Contact me