Multimodal (Images & Video)

Multimodal AI

Work with images and videos in byLLM.

Prerequisites

• Completed: Structured Outputs
• Time: ~25 minutes

---

Installation

Images are supported in the default byLLM distribution:

pip install byllm

For video support, install with the video extra:

pip install byllm[video]

---

Working with Images

byLLM supports image inputs through the Image type. Images can be provided as input to any by llm() function or method.

Basic Example

import from byllm.lib { Model, Image }

glob llm = Model(model_name="gpt-4o");

"""Describe what you see in this image."""
def describe_image(img: Image) -> str by llm();

with entry {
    image = Image("photo.jpg");
    description = describe_image(image);
    print(description);
}

Structured Output from Images

Combine image input with structured outputs for powerful data extraction:

import from byllm.lib { Model, Image }

glob llm = Model(model_name="gpt-4o");

enum Personality {
    INTROVERT,
    EXTROVERT
}

sem Personality.INTROVERT = "Person who is shy and reticent";
sem Personality.EXTROVERT = "Person who is outgoing and socially confident";

obj Person {
    has full_name: str;
    has year_of_death: int;
    has personality: Personality;
}

"""Extract person information from the image."""
def get_person_info(img: Image) -> Person by llm();

with entry {
    image = Image("einstein.jpg");
    person = get_person_info(image);
    print(f"Name: {person.full_name}");
    print(f"Year of Death: {person.year_of_death}");
    print(f"Personality: {person.personality}");
}

Output:

Name: Albert Einstein
Year of Death: 1955
Personality: Personality.INTROVERT

---

Image Input Formats

The Image type accepts multiple input formats:

Format	Example
File path	Image("photo.jpg")
URL (http/https)	Image("https://example.com/image.png")
Google Cloud Storage	Image("gs://bucket/path/image.png")
Data URL	Image("data:image/png;base64,...")
PIL Image	Image(pil_image)
Bytes	Image(raw_bytes)
BytesIO	Image(bytes_io_buffer)
pathlib.Path	Image(Path("photo.jpg"))

In-Memory Usage

import from byllm.lib { Image }
import io;
import from PIL { Image as PILImage }

with entry {
    # Load with PIL
    pil_img = PILImage.open("photo.jpg");

    # From BytesIO buffer
    buf = io.BytesIO();
    pil_img.save(buf, format="PNG");
    img_from_buffer = Image(buf);

    # From raw bytes
    raw = buf.getvalue();
    img_from_bytes = Image(raw);

    # From PIL image directly
    img_from_pil = Image(pil_img);

    # From URLs
    img_from_url = Image("https://example.com/image.png");
    img_from_gs = Image("gs://bucket/path/image.png");
}

---

Working with Videos

byLLM supports video inputs through the Video type. Videos are processed by extracting frames at a specified rate.

Basic Example

import from byllm.lib { Model, Video }

glob llm = Model(model_name="gpt-4o");

"""Describe what happens in this video."""
def explain_video(video: Video) -> str by llm();

with entry {
    video = Video(path="sample_video.mp4", fps=1);
    explanation = explain_video(video);
    print(explanation);
}

Output:

The video features a large rabbit emerging from a burrow in a lush, green
environment. The rabbit stretches and yawns, seemingly enjoying the morning.
The scene is set in a vibrant, natural setting with bright skies and trees,
creating a peaceful and cheerful atmosphere.

Video Parameters

Parameter	Type	Description
path	str	Path to the video file
fps	int	Frames per second to extract (default: 1)

Lower fps values extract fewer frames, reducing token usage. Higher values provide more temporal detail.

import from byllm.lib { Video }

with entry {
    # Extract 1 frame per second (good for most cases)
    video = Video(path="video.mp4", fps=1);

    # Extract 2 frames per second (more detail)
    video = Video(path="video.mp4", fps=2);
}

---

Practical Examples

Receipt Analyzer

import from byllm.lib { Model, Image }

glob llm = Model(model_name="gpt-4o");

obj LineItem {
    has description: str;
    has quantity: int;
    has price: float;
}

obj Receipt {
    has store_name: str;
    has date: str;
    has items: list[LineItem];
    has total: float;
}

"""Extract all information from this receipt image."""
def parse_receipt(img: Image) -> Receipt by llm();

with entry {
    receipt_image = Image("receipt.jpg");
    receipt = parse_receipt(receipt_image);

    print(f"Store: {receipt.store_name}");
    print(f"Date: {receipt.date}");
    print("Items:");
    for item in receipt.items {
        print(f"  - {item.description}: ${item.price}");
    }
    print(f"Total: ${receipt.total}");
}

Math Problem Solver

import from byllm.lib { Model, Image }

glob llm = Model(model_name="gpt-4o");

obj MathSolution {
    has problem: str;
    has steps: list[str];
    has answer: str;
}

"""Solve the math problem shown in the image."""
def solve_math(img: Image) -> MathSolution by llm();

with entry {
    problem_image = Image("math_problem.png");
    solution = solve_math(problem_image);

    print(f"Problem: {solution.problem}");
    print("Solution steps:");
    idx = 0;
    for step in solution.steps {
        print(f"  {idx+1}. {step}");
        idx += 1;
    }
    print(f"Answer: {solution.answer}");
}

Video Content Analysis

import from byllm.lib { Model, Video }

glob llm = Model(model_name="gpt-4o");

obj VideoAnalysis {
    has summary: str;
    has key_events: list[str];
    has duration_estimate: str;
    has content_type: str;
}

"""Analyze this video and extract key information."""
def analyze_video(video: Video) -> VideoAnalysis by llm();

with entry {
    video = Video(path="presentation.mp4", fps=1);
    analysis = analyze_video(video);

    print(f"Summary: {analysis.summary}");
    print(f"Content Type: {analysis.content_type}");
    print("Key Events:");
    for event in analysis.key_events {
        print(f"  - {event}");
    }
}

---

Combining with Tools

Multimodal inputs work with tool calling:

import from byllm.lib { Model, Image }

glob llm = Model(model_name="gpt-4o");

"""Search for products matching the description."""
def search_products(query: str) -> list[str] {
    # Simulated product search
    return [f"Product matching '{query}' - $29.99"];
}

"""Look at the image and find similar products."""
def find_similar_products(img: Image) -> str by llm(
    tools=[search_products]
);

with entry {
    product_image = Image("shoe.jpg");
    results = find_similar_products(product_image);
    print(results);
}

---

Key Takeaways

Concept	Usage
Image input	Image("path.jpg") or Image(url)
Video input	Video(path="video.mp4", fps=1)
Structured output	Return objects/enums from images
Multiple formats	URLs, files, PIL, bytes all supported
Install video	pip install byllm[video]

---

Next Steps

• Agentic AI - Combine multimodal with tool calling
• byLLM Reference - Complete documentation
• Examples Gallery - More multimodal examples