Derive attachment content type from MIME type

[Adam-D-Lewis]

Derive attachment content type from MIME type

Repo: nebari-chat-pack
Branch: fix/derive-attachment-mime-type
Base: main
Commit: 4efb6d2

Summary

Attachments in the chat input were always tagged with the AG-UI document
content type. This mislabels image/audio/video uploads, so a vision-capable
model never receives an uploaded image as an actual image. This PR derives the
AG-UI content discriminator from each file's MIME type instead.

Problem

In frontend/src/chat/chatinput.tsx, when attached files are converted to
FileInputContent, the discriminator was hardcoded:

return {
  type: "document",
  source: { type: "data", mimeType: file.type, value: ... },
  metadata: { name: file.name },
};

The downstream pipeline (Ravnar file handling → pydantic-ai AG-UI adapter) only
turns an ImageInputContent into an image part for the model. A document-typed
image is never converted to an image, so vision models can't see uploaded images
at all.

Change

Derive the discriminator from file.type:

• image/* → image
• audio/* → audio
• video/* → video
• otherwise → document (safe fallback)

The valid discriminator literals (image / audio / video / document) were
confirmed against the @ag-ui/core FileInputContent discriminated union
(frontend/src/api/files.ts → the agui.*InputContentSchema definitions).

Files changed

• frontend/src/chat/chatinput.tsx

Testing

• tsc --noEmit passes.
• Not yet exercised in a running browser.

Notes

• This is a prerequisite for image upload actually working, but on its own it is
  not user-visible: the file picker's accept attribute still restricts which
  files can be selected. The companion change (feat/capability-driven-uploads)
  makes the picker capability-driven and is stacked on top of this branch.

[jbouder]

LGTM, just need to resolve the conflict