AI / ML · Educational project
MiniLLM
A small LLM application framework built to see what agent libraries do.
In plain words. A small learning project that rebuilds the basic parts of an AI chatbot from scratch (memory, tools, looking things up in documents) to show what really happens inside popular AI tools.
The problem
Frameworks hide how messages are assembled, how tool results re-enter context and what memory actually stores.
What it does
A raw HTTP model client, prompt builder, in-memory conversation store, tool registry with JSON-schema generation, a bounded ReAct agent loop and a FAISS-backed RAG pipeline, driven from a CLI and a Streamlit UI.
Key decisions
- Transport and prompt construction are separate, so each can be tested alone.
- The agent loop has a hard iteration cap.
Limits and status
- It wraps an external model API. It does not train a model.
- Memory is in-process only; the vector index is local.