Skip to content

Multimodal Vision Input

Agentic Chat accepts images as first-class, native multimodal input: attach a photo, ask a question, and a vision-capable model sees the actual pixels - not a text description of them. The pipeline handles resizing, EXIF metadata, storage, and model-capability checks so the conversation stays fast and private.

The chat input with an attached image chip above the prompt field - the chip shows a thumbnail with a remove button and the 1 / 5 counter

Three ways to attach

Entry point How
Attach button Click the picture icon in the prompt field and pick one or more images
Drag and drop Drop image files anywhere on the prompt field
Clipboard paste Copy an image (or a screenshot) and paste it into the prompt field

All three feed the same pipeline. Accepted formats: PNG, JPEG, WebP, GIF (10MB per image before optimization). Each attached image appears as a chip above the prompt - click the x on a chip to remove it before sending. Up to 5 images ride along with one message: the chip bar shows a N / 5 counter and the attach button disables at the cap (5 is a safe ceiling across local vision models - Ollama backends accept between 1 and 8 images per request depending on the model).

What happens in the browser

Before anything reaches the server, the browser optimizes the image the same way LibreChat, Open WebUI, and the OpenAI clients do:

  1. EXIF first - camera metadata (capture time, GPS, camera model, exposure, ISO, orientation) is extracted from the original file, because the resize step strips it.
  2. Orientation fix - the image is decoded with EXIF rotation applied.
  3. Resize - anything longer than 2048px on its longest edge is scaled down (never up), preserving aspect ratio.
  4. Re-encode - images with transparency stay PNG; everything else becomes JPEG at quality 0.92. Animated GIFs are passed through untouched.

The optimized bytes plus the extracted EXIF go to the server as the attachment. Nothing is uploaded to any third party - with a local Ollama model the image never leaves your machine.

Content-addressed storage

Every attached image is stored in the conversation's own workspace under the playground home at workspace/<conversation>/images/, keyed by the SHA-256 of its bytes:

workspace/Chat-a1b2c3/images/
├── 4e73726bd2c9...d62c25.jpg     # the optimized image
└── 4e73726bd2c9...d62c25.json    # metadata: original file name, MIME type, size, EXIF

Attaching the same image twice in a conversation stores it once, and images never leak across conversations - the same per-conversation isolation as the filesystem tools' workspace. The hash doubles as a stable reference the model can use later (see describeImage).

Sending and rendering

Attached images ride along with your next message as Spring AI Media on the user message - the provider-native multimodal path (Ollama images, OpenAI image_url). The sent images render as thumbnails inside your message bubble, and the image references persist with the conversation: reopening a chat restores the thumbnails from the image store. Only the references are saved in the conversation file - the bytes stay deduplicated in the conversation's images/ directory.

A user message with the attached image thumbnail and the vision model's answer reading the title text and describing the earthquake map

Vision capability check

Not every model can see. The playground checks before you send:

  • OpenAI - GPT-5 family chat models are multimodal; no warning.
  • Ollama - the model's manifest is inspected for actual vision tensors, not just the advertised vision capability flag. This matters because some conversions (notably mlx models on Apple Silicon) advertise vision while shipping no vision tensors - they silently hallucinate instead of reading the image. The playground detects this case and warns specifically.
State What you see
Vision tensors present No warning - attach and go
No vision capability "This model can't process images. Select a vision-capable model."
Capability flag without tensors (mlx) "Reports vision support but ships no vision tensors... pick a non-mlx GGUF vision model."

The warning fires at attach time but never blocks the send - you stay in control.

Apple Silicon: use a non-mlx model for vision

The auto-selected -mlx variants are great for text but do not carry vision tensors. For image analysis pick the standard GGUF build - for example qwen3.5:4b rather than qwen3.5:4b-mlx. If the model list in the settings drawer only offers -mlx builds, type the model name directly - custom names are accepted.

Errors and timeouts

If a non-vision model is sent an image anyway, provider errors are translated into actionable messages ("This model can't process images. Pick a vision-capable model...") instead of raw HTTP bodies. A stream that goes silent is cut off after 5 minutes with a retry hint - local vision models can take a minute or two before the first token on a cold start, so the window is generous.

describeImage: re-reference an earlier image

Attached images do not need to be re-sent to be discussed again. The describeImage default tool lets the model re-summon any stored image mid-conversation:

  • "What was in the background of that photo I sent earlier?" - the model calls describeImage, the chat loads the stored image by its hash, and attaches it to the next model call. This works even when the original turn has fallen out of the context window.
  • If several images match, the chat opens a chooser dialog; if none exist, it opens an upload dialog - the same human-in-the-loop seam as tool approval and requestFileUpload.

The tool reference is in Default Tool Examples. The Image analyst preset packages the whole loop - analyze an attached image, tabulate findings with renderTable, read the EXIF sidecars to map geotagged photos with plotPointsOnMap, and export tables as CSV (Tutorial 14 walks the map scenario end to end).

Privacy notes

  • Images are optimized and stored locally under the playground home; with a local provider they are never transmitted anywhere else.
  • EXIF metadata - including GPS location if the photo has it - is preserved in the local metadata store so the model can answer "where was this taken?". Nothing strips it, so treat conversations with remote providers accordingly.
  • Conversation files store only image references (hash, file name, MIME type). Deleting a conversation offers an "Also delete workspace file(s)" checkbox - checked by default, it removes the stored images with the chat; uncheck it to keep them on disk.
Was this page helpful? Ask on Discussions · Report a docs issue