LLMs with type-safe generation

Pass images with images= alongside the text context. The served model must be a vision-language model, such as Qwen/Qwen3.8-27B.

result = client.generate(
    context="Read the attached photo of a restaurant receipt.",
    images=["receipt.jpg"],
    questions={
        "first_item": {"type": "string", "maxLength": 40,
                       "instructions": "Name of the first line item, exactly as printed."},
        "item_lines": {"type": "integer", "instructions": "How many line items are listed?"},
        "subtotal": {"type": "number", "instructions": "Subtotal as a plain number."},
        "discount": {"type": "number", "instructions": "Discount as a plain number."},
        "paid_in_cash": {"type": "boolean", "instructions": "Was the bill paid in cash?"},
        "discount_percent": {"type": "number", "depends_on": ["subtotal", "discount"],
                             "instructions": "Discount as a percentage of the subtotal."},
    },
)

A crumpled receipt read into typed fields: item names as strings, counts as integers and amounts as numbers, then a discount percentage, a change check and an expense note computed through depends_on. — A photographed receipt read into 14 typed fields with Qwen3.8-27B; the last three use `depends_on`. See the receipt example for the script and recorded results. Receipt from CORD v2, CC BY 4.0.

Accepted images

images is a list. Each entry can be a local file path, an http(s) URL, a data: URI, raw bytes or a PIL image. Local files are read by the client, so the SGLang server does not need access to your filesystem.

Behavior

Images come before the text in the first user turn, in list order.
Every request in the call carries the images, across batch, sequential and DAG execution, permutations, numeric decoding and thinking. Images apply to one generate() call only.
TypeLLM reads the image placeholder from the model's chat template. If the template does not render image content, generate() raises an error before sending any request.
With thinking, the budget uses the server's prompt-token count, because each image expands to many tokens on the server.