This temporary demo transcribes the bundled 20-second video with Gemini and sends manually created OpenInference spans to Arize AX. The captured LLM input uses the stable, public GitHub video URL rather than Gemini's temporary Files API URI.

1. Create and activate a Python virtual environment.

2. Install the dependencies:

   pip install -r requirements.txt

3. Copy .env.example to .env, populate the variables, and export them into your shell.

4. Run the demo:

   python transcribe_video.py

The script uploads assets/video-demo.mp4 to Gemini, waits for processing, invokes gemini-3.8-flash, and deletes the temporary Gemini upload. It creates a root CHAIN span and child LLM span explicitly; no auto-instrumentor is used. Gemini still receives its temporary Files API URI, but the captured LLM input uses the public video JSON object.

• The message_content.type = "video" and nested message_content.video object, with video.mime_type = "video/mp4" and the public video.url, on the manually set OpenInference input-message part.
• The exported LLM input part: { "message_content.video": { "video.mime_type": "video/mp4", "video.url": "https://…/video-demo.mp4" } }.
• Text prompt and transcript as adjacent OpenInference message-content parts

The video is an authorized, public, derived 20-second/720p clip from the repository owner's source file. This repository is temporary and will be deleted after validation.