Axon Docs

End-to-end PDF intelligence app: upload and view PDFs, extract AI insights, and generate a podcast—all behind a single container on port 8080.

• Frontend: Next.js (App Router) + Tailwind • Backend: Express (file upload, related content, Python orchestration) • Reverse proxy: Nginx (serves everything on port 8080) • Python: local models + PDF/search pipeline (round_1b)

Open http://localhost:8080 after running the container.

Screenshots

Add your screenshots to docs/screenshots/ with these filenames to render below. Each image has a short caption to match the requested order.

1. Landing Page

2. PDF Viewer with highlighted sections and subsection analysis (Allows click to jump to that specific page)

3. AI Insights page

4. Podcast page

5. Upload PDF page

6. Related Findings for selected texts (Allows click to jump to that specific page)

Quick Start (Docker)

Build (linux/amd64):

docker build --platform linux/amd64 -t yourimageidentifier .

Run (env vars inline):

docker run --rm --name axon-docs \
  -e ADOBE_EMBED_API_KEY=8f94cb51c25d45b6ad598687c409fdcb \
  -e LLM_PROVIDER=gemini \
  -e GEMINI_MODEL=gemini-2.5-flash \
  -e GEMINI_API_KEY=<GEMINI_API_KEY> \
  -e TTS_PROVIDER=azure \
  -e AZURE_TTS_KEY=<AZURE_TTS_KEY> \
  -e AZURE_TTS_ENDPOINT=<AZURE_TTS_ENDPOINT> \
  -p 8080:8080 \
  yourimageidentifier

Then open http://localhost:8080

What you get

• Upload PDFs: Drag-and-drop or pick files to place under frontend/public/pdfs for serving.
• PDF Viewer: In-browser view with highlighted sections and sub-section analysis.
• AI Insights: Gemini-based structured insights across single or multiple PDFs.
• Podcast: Gemini-created two-speaker script synthesized via Azure TTS to MP3.

Architecture

• Nginx on 8080
  • Proxies Next.js on 3000 and Express on 5001
  • Serves static assets and /runtime-env.js (browser runtime configuration)
• Next.js (App Router)
  • Pages under frontend/app
  • API routes for AI features:
    • POST /api/generate-insights
    • POST /api/generate-overview-insights
    • POST /api/generate-podcast
    • POST /api/generate-overview-podcast
• Express backend (5001)
  • Uploads, related content, Python process orchestration
• Python pipeline (round_1b/)
  • Local models + PDF/text embedding, ranking, and inference

Environment variables

Required for core features:

• ADOBE_EMBED_API_KEY: Adobe PDF Embed (client-side)
• LLM_PROVIDER: Must be gemini for current implementation
• GEMINI_MODEL: e.g., gemini-2.5-flash
• GEMINI_API_KEY: Google Generative AI key
• TTS_PROVIDER: Must be azure
• AZURE_TTS_KEY: Azure Cognitive Services Speech key
• AZURE_TTS_ENDPOINT: e.g., https://eastus.tts.speech.microsoft.com/cognitiveservices/v1
  • Alternatively, set AZURE_TTS_REGION (e.g., eastus) and omit endpoint

Runtime config is exposed to the browser via /runtime-env.js (generated at container start). To match the Adobe-style pattern, window.__ENV includes keys like ADOBE_EMBED_API_KEY, GEMINI_API_KEY, and others.

Project structure (high level)

backend/                 # Express server (5001)
frontend/                # Next.js app (3000)
  app/
    api/                 # Next.js route handlers (AI endpoints)
    pdfviewer/, upload/  # Main pages
    components/          # UI components (Header, Sidebar, PDFViewer, etc.)
  public/pdfs/           # PDFs served to the browser
nginx/                   # Nginx config (reverse proxy)
round_1b/                # Python models, requirements, and scripts
Dockerfile               # Multi-stage build and runtime setup

How it works (flow)

1. Browser sends API calls to /api/* on port 8080.
2. Nginx routes generate-* to Next.js; other /api/* to Express.
3. Next.js APIs call Gemini to generate insights/scripts; Podcast endpoints call Azure TTS.
4. Express endpoints handle uploads/related content and optionally call Python.
5. Static PDFs are read from frontend/public/pdfs when needed.