[Feature] Add image input (multimodal/vision) support

What feature would you like to see?

Summary
Please add native image input (vision) support so the model can understand and respond to screenshots, scans, charts, formulas, and handwritten content.

Motivation

In real workflows, a lot of information arrives as images: error screenshots, UI captures, whiteboard photos, scanned PDFs, plots/diagrams.

Today we must run OCR/layout tools first and then feed text to the model. That pipeline is brittle, loses structure (tables/figures/layout), and increases integration cost.

First-class image input would improve quality, reduce glue code, and unlock more use cases.

Representative use cases

Diagnose and fix issues from exception/log screenshots.

Extract key numbers from chart/dashboard screenshots and analyze them.

Convert phone photos of notes/equations into editable text/LaTeX.

Understand UI screenshots to generate test cases or automation steps.

Parse architecture/flow diagrams into entities/relations and draft docs.

Proposal (high level)

Extend current APIs to accept an image content type (URL or base64), including multiple images and optional captions.

Support common formats (PNG/JPEG; optionally first/specified page of PDF). Document size/resolution limits.

Account for usage via separate “vision tokens” or similar.

Optionally return structured data (e.g., tables, detected boxes/coordinates) alongside text.

Are you interested in implementing this feature?

_No response_

Additional information

_No response_