Misa — AI Study Partner

Overview

Misa is a study-partner chatbot: a Django back end that proxies a language model, a plain HTML/JavaScript front end that renders replies as they stream in, and a sidebar for keeping several independent study threads. It is deliberately small — the whole back end is a single view module of about 140 lines — and it was a way to understand the full request path of a streaming chat app without a framework doing it for me.

It began as a ten-line script calling an LLMLLMLarge language model — a neural network trained on large amounts of text to predict the next token, used for generation, summarisation, classification and more. API, moved into Django, and grew one feature per pull request. The repository is private, so this page describes the design.

What It Does

• Streaming replies: tokens appear as they are generated, using server-sent events from a Django StreamingHttpResponse
• Multiple chats: a sidebar to create, switch, clear and delete study threads, each with its own context
• Session memory: conversation history is stored in the Django session, so a page reload restores the chat without accounts or a custom model
• Markdown rendering of answers, including code blocks
• Persona: a system prompt tuned for clear, encouraging study help that stays concise on simple greetings
• Landing page at / with the app at /app/
• Deployed on Render with a build script and a render.yaml

How It Works

1. The browser POSTs a message and a chat id to /chat/.
2. The view appends it to that chat's history in request.session and opens a streaming request to OpenRouter.
3. Each streamed chunk is parsed and re-emitted to the browser as an SSE data: line.
4. When the stream ends, the assembled reply is appended to the history and the session is saved explicitly.

Tech Stack

Back end: Python, Django 6, requests, python-dotenv, Gunicorn

Front end: HTML, CSS and vanilla JavaScript with a Markdown renderer

AI: OpenRouter with free-tier models (the model was changed several times as free limits and availability shifted)

Hosting: Render

Honest Notes

This is a learning prototype, and I would not run it as-is for real users. Known gaps: some endpoints are CSRF-exempt, debug mode and a permissive host list are still on in settings, the session grows without a cap, and there is no retrieval or document grounding despite the repository's original name. A hardening pass is described in the accompanying blog post.