Vision & Multimodal Guide

In addition to text completions and image generation, KoboiLLM natively supports Vision Models — AI models capable of “seeing” and analyzing input images. Ideal applications include:

• Automated captioning and alt-text generation
• Document processing and simple OCR
• Product photo and screenshot inspection
• Visual Question Answering (Visual QA)

The API strictly adheres to the standard OpenAI Chat Completions specification using multimodal array content, making existing OpenAI and LiteLLM SDKs immediately compatible.

---

Table of Contents

1. Endpoint & Authentication
2. Available Vision Models
3. Quick Start — Sending Image URLs
4. Sending Local Images (Base64)
5. Multi-Image Inputs
6. Parameters (detail & format)
7. Checking Model Vision Support
8. Quick Spec Summary
9. Tips & Best Practices

---

1. Endpoint & Authentication

Base URL:

https://lite.koboillm.com/v1

(The alias https://api.koboillm.com/v1 is also available and identical in behavior).

Primary Endpoint: POST /v1/chat/completions — identical to standard chat completions, except that the user message content field is structured as an array containing text and image_url objects.

Authentication: Include the header `Authorization: Bearer *** Set your secret API key in your environment variables:

export LITELLM_API_KEY="sk-xxxxxx"

---

2. Available Vision Models

The following models natively support image understanding at KoboiLLM:

Provider	Endpoint Name	Best Used For
OpenAI	openai/gpt-4o-mini	Fast & budget-friendly vision — recommended default
OpenAI	openai/gpt-4o	High-reasoning visual analysis and complex layout reading
Google Gemini	gemini/gemini-2.5-flash	High-speed, low-cost multimodal processing
Google Gemini	gemini/gemini-2.5-pro	Comprehensive document analysis and detailed OCR

For the live model catalog, query /v1/models or check your dashboard at https://lite.koboillm.com.

---

3. Quick Start — Sending Image URLs

3.1 — cURL

curl -X POST https://lite.koboillm.com/v1/chat/completions \
  -H "Authorization: Bearer *** \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-4o-mini",
    "messages": [
      {
        "role": "user",
        "content": [
          { "type": "text", "text": "Describe what you see in this image in detail." },
          {
            "type": "image_url",
            "image_url": {
              "url": "https://docs.koboillm.com/koboi_llm_logo.png"
            }
          }
        ]
      }
    ]
  }'

3.2 — Python (OpenAI SDK)

from openai import OpenAI

client = OpenAI(
    base_url="https://lite.koboillm.com/v1",
    api_key="sk-xxxxxx",
)

response = client.chat.completions.create(
    model="openai/gpt-4o-mini",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "Describe what you see in this image."},
                {
                    "type": "image_url",
                    "image_url": {"url": "https://docs.koboillm.com/koboi_llm_logo.png"},
                },
            ],
        }
    ],
)

print(response.choices[0].message.content)

---

4. Sending Local Images (Base64)

For local images stored on disk, encode the image to base64 and supply it as a data URL:

import base64
from openai import OpenAI

client = OpenAI(
    base_url="https://lite.koboillm.com/v1",
    api_key="sk-xxxxxx",
)

def encode_image(image_path):
    with open(image_path, "rb") as image_file:
        return base64.b64encode(image_file.read()).decode('utf-8')

base64_image = encode_image("document.png")

response = client.chat.completions.create(
    model="openai/gpt-4o-mini",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "Extract all text from this document image."},
                {
                    "type": "image_url",
                    "image_url": {
                        "url": f"data:image/png;base64,{base64_image}"
                    },
                },
            ],
        }
    ],
)

print(response.choices[0].message.content)