DeepTutor

Why it's ranked here

Breadth here is genuine, and so is the weight that comes with it. A plain install already pulls an API server, vector search, PDF handling, several model provider SDKs and a database client, and the supported Python range is capped below 3.14 because compiled wheels break above it. Against that, the pinning is unusually well reasoned: many version bounds carry a comment naming the exact failure they prevent, with issue numbers. Judge it as a learning platform you operate, not a library you drop into an existing app.

What's good

Operational care shows in small places. Optional engines import lazily and degrade instead of crashing: vector search falls back to a brute force store when the accelerated backend is absent, and the memory probe falls back to reading kernel process files. The container drops both the backend and frontend processes to an unprivileged user with a fixed numeric id. Startup refuses to boot when a capability manifest names a tool the registry does not have. Chat history import reads folders in the browser, so those transcripts never transit the server.

Tradeoffs

Authentication is off by default, and in that mode the API accepts any origin through a wildcard pattern. Tightening it means enabling auth and listing origins explicitly. The subprocess sandbox that runs generated code is also on by default; the compose setup diverts execution to a hardened runner, a plain container build does not. Building the image compiles from source: it installs a Rust toolchain, build tooling and graphics libraries, and keeps a git client in the production stage so the tutor can install command line apps at runtime.

How to use it well

Pick this if you want a private study stack over your own documents, with your own model keys, and you are comfortable running containers plus an optional database sidecar. It pays off when the material is yours: course notes, papers, question banks you want retrieved and quizzed against. It is not a coding assistant, though it can consult external agent command line tools you already have. It also does not replace a course management system: enrolment, grading records and institutional reporting sit outside what it handles.

Observed

License

Apache-2.0, declared in pyproject.toml

Runtime

Python backend restricted to 3.11 through 3.13, with a Next.js web frontend

Interfaces

Typer command line app, HTTP and WebSocket API, browser UI, and an MCP client

Install surface

pip package with extras for server, partners, matrix, math animation, GraphRAG, LightRAG, parsing engines and dev tooling

Deployment

Multi-stage Dockerfile; supervisord runs backend and frontend as a non-root user with UID 1000

Retrieval

LlamaIndex with BM25 and a FAISS vector store by default, GraphRAG and LightRAG as opt-in engines

Auth default

Authentication disabled out of the box; that mode also enables a wildcard CORS origin regex

Chat channels

Partner extra bundles Telegram, Slack, Lark, DingTalk, WeCom, QQ and Zulip SDKs; Matrix is a separate extra