DebateLearning

DebateLearning is a local learning-podcast generator. It takes uploaded lecture slides or PDFs, extracts the core knowledge with a local Ollama model, turns that knowledge into a structured debate-style transcript, and then generates spoken audio with F5-TTS.

App Structure

• backend/ - FastAPI + SQLite backend on port 8100
• website/ - Next.js frontend on port 3100
• voices/ - local reference voices used by the audio pipeline

The frontend proxies /api/* requests to the FastAPI backend through website/next.config.mjs.

Processing Flow

DebateLearning uses a multi-stage AI pipeline: multimodal slide extraction, batch knowledge synthesis, semantic retrieval with pgvector, transcript validation, and multi-speaker TTS generation.

flowchart TD
    A[Upload PDF or PPTX] --> B[Create job record]
    B --> C{File type}
    C -->|PPTX| D[Convert to PDF with LibreOffice]
    C -->|PDF| E[Skip conversion]
    D --> F[Render each page and extract text]
    E --> F

    F --> G[Slide-level knowledge extraction]
    G --> H[Batch-level knowledge creation]
    H --> I[Merge into one knowledge document]

    I --> J[Search related prior knowledge with pgvector]
    J --> K[Generate draft debate transcript]
    K --> L[Check transcript against source knowledge]
    L --> M[Polish transcript for spoken delivery]
    M --> N[Generate voice clips with F5-TTS]
    N --> O[Combine clips into final audio]
    O --> P[Clean temporary clips]

    I --> Q[Save knowledge.txt]
    M --> R[Save transcript.json]
    O --> S[Save output.wav]
    I --> T[Index knowledge chunks in pgvector]

Loading

High-level artifact flow:

PDF/PPTX source
-> extracted slide and batch knowledge
-> merged knowledge document
-> related prior knowledge retrieval
-> checked and polished transcript
-> generated audio episode

Each job stores artifacts under:

backend/outputs/<job_id>/
  knowledge.txt
  transcript.json
  output.wav

Backend

Run from Anaconda Prompt if the F5-TTS environment only works there:

conda activate f5-tts
cd /d C:\Users\pojwm\repos\f5-tts-sandbox\backend
python -m uvicorn main:app --reload --port 8100

Required local services/tools:

• Ollama running at http://localhost:11434
• Ollama model qwen3.5:4b
• LibreOffice installed for PPTX conversion
• F5-TTS available in the active Python environment

If PPTX conversion fails, confirm this exists:

C:\Program Files\LibreOffice\program\soffice.exe

PDF uploads skip the LibreOffice conversion step.

Frontend

cd /d C:\Users\pojwm\repos\f5-tts-sandbox\website
npm run dev

Open:

http://localhost:3100

The job detail page polls the backend while processing and exposes links for:

• knowledge extraction output
• transcript JSON
• generated audio

Main API Routes

POST /api/jobs
GET  /api/jobs
GET  /api/jobs/{job_id}
GET  /api/jobs/{job_id}/knowledge
GET  /api/jobs/{job_id}/transcript
GET  /api/jobs/{job_id}/audio