Vision (images)

Images come in as ImagePart (bytes + mime) and are translated to each SDK's native format. Always use a model with vision.

from jangada_ai import LLM, Image

llm = LLM("openai", "gpt-4o-mini")

# by path
llm.complete("What's shown here?", images=["photo.jpg"])

# by bytes or base64
img = Image.from_bytes(upload_bytes, "image/png")   # or Image.from_base64
receipt = llm.parse("Extract the total.", Receipt, images=[img]).parsed

Translation per provider

Only bytes circulate (use Image.from_path/bytes/base64). Combine with Structured output by passing images= to parse().

Label each image (shortcut)

To label each image without building messages by hand, pass (label, image) tuples in images= — the library interleaves a TextPart with the label before each image, in the given order:

llm.parse(
    "Compare the documents.", Comparison,
    images=[("front", Image.from_path("a_front.jpg")),
            ("back",  Image.from_path("a_back.jpg"))],
)

Unlabeled items (a plain ImagePart/path) keep working in the same list — mix labeled and plain freely.

Messages and interleaved multimodal

The images=[...] shortcut puts the prompt text first and the images after. For full control of the order (alternating multiple text and image blocks, multi-turn), build the messages by hand with Message, TextPart and ImagePart (all importable from jangada_ai) and pass them via history=.

from jangada_ai import LLM, Message, TextPart, Image

llm = LLM("openai", "gpt-4o-mini")

# text and images interleaved, each one labeled
msg = Message("user", [
    TextPart("Document A (front):"),
    Image.from_path("a_front.jpg"),       # Image.* returns an ImagePart
    TextPart("Document A (back):"),
    Image.from_path("a_back.jpg"),
    TextPart("Are both images from the same document? Compare them."),
])

resp = llm.complete(None, history=[msg])   # prompt=None: content comes from the messages
print(resp.text)

Message(role, content) takes role in "system" | "user" | "assistant" | "tool" and content as a str or a list of parts (TextPart / ImagePart). For multi-turn, just accumulate the previous messages in history:

resp1 = llm.complete("What's the capital of Brazil?")
resp2 = llm.complete(
    "And its population?",
    history=[
        Message("user", "What's the capital of Brazil?"),
        Message("assistant", resp1.text),
    ],
)

ImagePart carries only bytes (data + mime_type), so it works the same across every provider; use the Image.from_path/bytes/base64 factories instead of building the ImagePart by hand.

Vision vs. Documents

For docx/pdf/csv/xlsx, prefer Documents: by default jangada extracts the text locally (cheaper, works on models without vision) and only uses vision when you force mode="vision" or when the PDF is scanned.

Example

examples/vision_example.py — runnable script.