QA: Multimodal Attachments — PydanticAI

Prerequisites

• Demo deployed and accessible at /demos/multimodal
• Agent backend healthy (check /api/health)
• OPENAI_API_KEY set
• Bundled public/demo-files/sample.png and public/demo-files/sample.pdf present (copied over from the langgraph-python reference assets)

Test Steps

1. Basic Functionality

• Navigate to /demos/multimodal
• Verify the header "Multimodal attachments" is visible
• Verify the sample-row is visible with two buttons: "Try with sample image" and "Try with sample PDF"
• Verify <CopilotChat /> renders a message composer

2. Sample image round-trip

• Click "Try with sample image"
• Within 10 seconds, an attachment chip labelled sample.png appears in the composer
• Type "Describe this image" and click send
• Within 90 seconds, an assistant response renders that references the CopilotKit / logo / image content

3. Sample PDF round-trip

• Click "Try with sample PDF"
• Within 10 seconds, an attachment chip labelled sample.pdf appears in the composer
• Type "Summarize this document" and click send
• Within 90 seconds, an assistant response renders that references the document contents (should mention "CopilotKit")

Known Limitations vs. langgraph-python port

• PydanticAI's AG-UI bridge does not expose a LangChain-style before_model middleware. The equivalent behaviour is implemented via a PydanticAI history_processors hook that rewrites incoming binary parts into image_url (for GPT-4o vision) or extracted text (for PDFs via pypdf) before each model call. Functionally equivalent to the langgraph-python reference.
• The frontend's onRunInitialized shim (rewriting modern image/document parts into legacy binary parts) is framework-agnostic and kept intact so the Python-side history processor sees the same wire shape the langgraph-python reference does.

Expected Results

• Images flow through to GPT-4o vision natively.
• PDFs flatten to inline text; the agent can answer questions about the document contents.
• Bundled samples work without mic / file-picker permission dialogs.