LectureNoteAgent — Full Lecture Note Generator

This project builds a full AI agent that takes:

• course slides (.pdf, .pptx, .md, .txt)
• class transcript (.txt / .md)

and generates a comprehensive DOCX lecture note that:

• covers slide content + teacher speech
• keeps a clean, structured format
• highlights special mentions from instructor
• embeds extracted slide images directly in the DOCX file
• applies slide-image cropping metadata (for PPTX images) before embedding
• extracts and preserves formulas
• validates coverage and repairs missing points automatically

Features

• Multi-source ingestion for slides and transcripts
• Model-only OCR for PDFs (automatic whole-file + page fallback)
• Coverage checklist generation (atomic items with IDs)
• Lecture-note drafting with references like [S3], [T17]
• DOCX export with attached images embedded in the final file
• Strict validation pass with JSON audit
• Auto-repair loop to fill dropped/missing content
• Artifacts export (checklist.md, audit.json, source_bundle.json)
• Light Streamlit UI for easy upload/run/download flow

Project Structure

• src/lecture_note_agent/io_utils.py — parsing slides/transcript and source payload build
• src/lecture_note_agent/prompts.py — generation/audit/repair prompts
• src/lecture_note_agent/agent.py — orchestration pipeline + iterative validation
• src/lecture_note_agent/cli.py — command line interface
• src/lecture_note_agent/ui.py — lightweight Streamlit web UI
• Dockerfile + docker-compose.yml — one-command containerized run

Setup

1. Install dependencies:

   pip install -r requirements.txt

2. Configure .env:

   • OPENAI_API_KEY=your_openai_api_key_here
   • OPENAI_BASE_URL=https://openrouter.ai/api/v1
   • OPENAI_MODEL=openai/gpt-5.4 (fallback)
   • OPENAI_MODEL_OCR=openai/gpt-5.4
   • OPENAI_MODEL_CHECKLIST=openai/gpt-5.4
   • OPENAI_MODEL_DRAFT=openai/gpt-5.4
   • OPENAI_MODEL_AUDIT=openai/gpt-5.4
   • OPENAI_MODEL_REPAIR=openai/gpt-5.4
   • MAX_REPAIR_LOOPS=3
   • MAX_MODEL_CALLS=6
   • MAX_OUTPUT_TOKENS=3500
   • FAST_MODE=false (set true for faster runs)
   • PDF_OCR_MODE=auto (whole is usually faster than auto)

The app now uses an OpenAI-compatible client only. Keep credentials in .env only.

Usage

Run from project root:

python -m lecture_note_agent --course-name "Data Structures" --slides ./input/week1.pdf --transcript ./input/week1_transcript.txt --output ./output/week1_lecture_notes.docx --artifacts-dir ./artifacts/week1

Or after editable install (pip install -e .):

slideagent --course-name "Data Structures" --slides ./input/week1.pdf --transcript ./input/week1_transcript.txt --output ./output/week1_lecture_notes.docx --artifacts-dir ./artifacts/week1

PDF OCR strategy (automatic)

• whole: uploads the full PDF once and asks the model to return per-page JSON text.
• page: uploads one-page PDFs and extracts each page separately.
• auto: tries whole first, then falls back to page for weak/missing pages.

This strategy is always used for PDFs; no OCR toggles are required in UI/CLI/env.

Speed tuning

If runs feel slow, use one or more of these:

• Enable FAST_MODE=true (skips audit/repair loop, disables continuation calls, uses whole-PDF OCR)
• Set PDF_OCR_MODE=whole for faster OCR on large PDFs
• Reduce MAX_REPAIR_LOOPS and MAX_OUTPUT_TOKENS
• Use a faster model for draft/checklist phases

Multi-model by phase

SlideAGENT can route each phase to a different model:

• OCR phase → OPENAI_MODEL_OCR
• Checklist phase → OPENAI_MODEL_CHECKLIST
• Draft phase → OPENAI_MODEL_DRAFT
• Audit phase → OPENAI_MODEL_AUDIT
• Repair phase → OPENAI_MODEL_REPAIR

If a phase model is not provided, OPENAI_MODEL is used as fallback.

Web UI

Run locally:

streamlit run src/lecture_note_agent/ui.py

Or with script (after pip install -e .):

slideagent-ui

Docker / Compose

Build and run UI with Docker Compose:

docker compose up --build

Then open http://localhost:8501.

Output Quality Contract

The generated DOCX is designed to include:

1. Full lecture structure (headings/subheadings)
2. All concepts from slides and transcript
3. Special instructor instructions/reminders
4. Embedded slide images with exact refs from source
5. Formula sheet with exact formula text
6. Inline source references for traceability

Validation ensures high coverage, then repair loop attempts to fix any missing items before final DOCX output is written.

Notes

• For best results, provide clean transcript text (timestamps/speaker names are supported).
• PDF image extraction depends on available image metadata in the PDF.
• PPTX image references use shape names from slides.