Hi HN,

I built StudyWithMiku, an open-source AI study assistant that:

Ingests PDFs automatically (drop into a folder)

Embeds them into ChromaDB

Uses RAG for grounded answers

Maintains conversational memory with LangGraph

Outputs synthesized character voice using Coqui TTS + DiffSinger

Runs locally (Ollama supported) or with cloud LLMs

The goal was to combine:

• Retrieval-augmented generation • Stateful agent orchestration • Tool execution (web/system tools) • Persistent memory • Character-specific TTS

into one cohesive system.

Engineering challenges included:

Chunking strategy and retrieval accuracy

Avoiding recursive agent tool loops

Managing CUDA/PyTorch/protobuf conflicts

Keeping latency reasonable with local models + TTS

It’s Linux-first (Ubuntu recommended) and GPU helps, but CPU works.

I’m particularly interested in feedback on:

RAG optimization approaches

LangGraph/agent architecture patterns

TTS expressiveness tuning

Local LLM deployment tradeoffs

Happy to answer technical questions.