RAG Tutor — Marshall & McAdam Dynamics

Live Python · Streamlit · LangChain · ChromaDB · Ollama · Docker

Python Streamlit LangChain ChromaDB Ollama RAG Docker

How it works

01 — INGEST

PDF textbook chunked & embedded into ChromaDB via Ollama

02 — RETRIEVE

Semantic search returns the most relevant textbook passages

03 — GENERATE

Local LLM synthesises grounded answers from retrieved context

04 — ASSESS

Study cycle grades text, speech, and handwritten answers

Live Demo

← Back to Projects

Build Journal

Summary

A fully local AI tutor combining LangChain RAG over ChromaDB with Ollama-hosted LLMs. Structures sessions into timed Reading, Recall, and Problem-Solving phases and can grade answers submitted as text, speech, or a webcam photo of handwritten work.

Implementation

PDF ingested with PyMuPDF; chunks embedded via Ollama into ChromaDB (HttpClient mode)
LangChain RetrievalQA chain retrieves top-k passages and feeds them as context to the LLM
Study cycle state machine drives timed phase transitions with auto-advance and local history
SpeechRecognition transcribes verbal answers; LLM grades the transcript against the expected answer
OpenCV detects equation regions in webcam frames; vision LLM scores the handwritten solution

Troubleshooting

Streamlit WebSocket connection broke behind a sub-path proxy — resolved by setting baseUrlPath to match the reverse-proxy route
ChromaDB switched to remote HttpClient in 0.5.x, breaking the embedded-mode import — updated client instantiation throughout
PyAudio build fails in slim Python images — split into a builder stage so the runtime image only ships the compiled library
Ollama model fallback ladder raised different exceptions across versions — normalised error handling to catch both ResponseError and ConnectionError

To Do

[ ]Show page number and paragraph alongside each retrieved source chunk
[ ]Add SM-2 spaced repetition scheduling to the question bank
[ ]Export completed study session as a PDF report
[ ]Swap default embedding model for a retrieval-optimised alternative
[ ]Build progress table generation for the Progress tab