CourseChat

Having_free_Gemini_API_Key
Indexing_Many_Files
Being_able_to_index_all_course_notes
Preview_of_what_text_is_captured_from_an_uploaded_file
Answering_questions_one_by_one
Answering_big_questions
Solving_Midterm_Questions_requiring_Math
Copy_and_paste_well_formatted_text
Export_All_Chats
Delete_Individual_Chat
Chatbox_Expandability
Text_to_Speech
Speech_to_text

Inspiration

Every student has had the experience of searching through dozens of lecture slides at 2 am trying to find one formula or definition. Course materials are scattered across LMS platforms — PDFs, slides, scanned notes, past exams — and no search tool actually understands what's in them. We wanted to build something that reads your course materials like a study partner and gives you cited, trustworthy answers grounded in your actual syllabus.

What it does

CourseChat is a Chrome extension that lives in your browser's side panel. You upload your course materials — lectures, assignments, exams, tutorials, textbook chapters — and it indexes them server-side using Backboard.io. Then you ask questions in plain English, and it answers using only your course content with inline citations back to the exact source document, page, and section. Math-heavy answers render cleanly with Unicode notation instead of raw LaTeX. It works with any course, any subject, any school.

How we built it

Chrome Manifest V3 extension with a React-based side panel UI
PDF.js for client-side text extraction across 15+ file formats (PDF, PPTX, DOCX, Jupyter notebooks, source code, LaTeX, and more)
Drag & drop file upload — students can drop one or many files directly into the panel
Backboard.io handles all RAG orchestration: chunking, embeddings, vector search, and answer generation
BYOK (Bring Your Own Key) model — users provide their free Google Gemini API key, achieving $0 cost
Multi-model fallback chain (gemini-3.6-flash → gemini-3.5-flash-lite → gemini-2.5-flash-lite) with programmatic retry logic for HTTP 429/404 resilience
Direct-Gemini fallback path when Backboard is unreachable — answers still work using locally cached context
Property-based testing with fast-check: 97 tests covering the RAG engine, Gemini wrapper, indexing orchestrator, and Backboard client
GitHub Actions CI/CD pipeline that runs tests, builds, and auto-bumps the version on every successful push

Challenges we ran into

Handling the variety of document formats students actually use (multi-part chapter PDFs, PPTX with embedded content, DOC legacy formats, Jupyter notebooks)
Designing the fallback chain to be truly transparent — users never know which model answered their question or whether Backboard or direct-Gemini handled it
Getting LaTeX math to render cleanly in a side panel — Gemini outputs raw LaTeX notation ($2^{12}$, \times) which looks terrible in a chat UI, so we built a client-side LaTeX-to-Unicode converter
Ensuring course context isolation so indexed materials never leak across courses
Removing legacy constraints (80k character cap, 300-word answer limit) that made sense before Backboard but were choking real-world usage with 30+ file courses

Accomplishments that we're proud of

Zero-cost AI tutoring — students with a free Gemini API key get full RAG capabilities without paying anything
Every answer has inline citations tied to the exact file and page of the source material — right under each bullet point, not lumped at the bottom
No document size limit — students uploaded 36 files (240k+ characters) in testing without hitting any cap
15+ file format support covering CS, humanities, business, and STEM courses
Property-based testing with formal correctness guarantees: confidence thresholds, word limits, citation completeness, and course isolation are all verified programmatically
Clickable file tiles that let you preview the extracted text, so you can verify exactly what the AI is reading

What we learned

RAG systems need strict confidence thresholds — without them, the model hallucates "close enough" answers from unrelated chunks
Browser extension architecture (Manifest V3 service workers) has unique constraints around lifecycle and memory that don't exist in normal web apps
Backboard.io's API simplified what would have been weeks of vector database setup into a clean integration
LaTeX in chat UIs is a universal pain point — the solution was converting client-side at the rendering layer rather than fighting the model's output format
File format diversity is the real challenge — students don't just use PDFs; they have .docx lecture notes, .ipynb lab notebooks, .py starter code, and .pptx presentations all for one course

What's next for CourseChat

Chrome Web Store publication (in review)
Auto-detect course materials from LMS pages (Brightspace, Canvas) so students don't need to manually upload
AI Vision for diagrams — use Gemini's image understanding to read figures, flowcharts, and scanned pages that standard text extraction can't handle
Incremental textbook indexing — index 1500-page textbooks chapter by chapter across sessions on the free tier
Collaborative study groups — share a Course Context with classmates
Browser-native offline mode — cache recent answers locally for study without internet