[FEATURE] Add image input support for Vision-Language models (e.g. Qwen3-VL)

[AbdulShahzeb]

Summary

Add support for processing image inputs in text/transcript mode, enabling CAAL to answer questions about images, identify objects, read text from photos, etc., through VLMs like Qwen3-VL and LLaVA.

Proposed Solution

1. Image Input Pipeline

• Add image upload button to chat interface
• Display uploaded images in conversation history

2. Provider Support

• Detect VLM models
• Format image messages for Ollama API:

{
  "role": "user",
  "content": "What's in this image?",
  "images": ["base64_encoded_image_data"]
}

Use Case

1. Vision Assistance

• "What's in this image?" - General image description
• "Read the text from this receipt" - OCR functionality

2. Smart Home Integration

• "What's in my fridge?" - Using camera input
• "Is there a package on my doorstep?" - Using doorbell camera

Additional Context

Performance:

• Longer inference time for vision models
• Image preprocessing (resize to 512x512 or 768x768 for faster inference)

User experience:

• Display image thumbnails in chat history
• Visual feedback when image is being processed
• Error handling for unsupported formats

Compatibility

• Text-only models continue to work unchanged
• VLM support is opt-in (ideally auto-detected)
• If manually enabled, image messages should be ignored by non-VLM providers
• Should work with any VL model (similar to LM Studio)

[cmac86]

@AbdulShahzeb, love the idea. I will start investigating this and will probably have more questions about your use case. I want to implement a couple easier features first before tackling this one.

[AbdulShahzeb]

Sounds good, appreciate you taking a look at it. Happy to answer any questions or help test it out, but I agree that this may be a lot of effort for a feature that's exclusive to text-mode.