frankiethull/kuzco/kuzco-custom

kuzco-custom

Run an arbitrary computer-vision prompt against an image using a local vision language model, exposed as an ellmer tool by the kuzco R package. This is the escape hatch for image tasks the specialized skills do not cover.

When to use

The task needs something from an image that classification, OCR, object recognition, sentiment analysis, or alt text do not capture — e.g. "count the windows", "is anyone wearing a red hat?", "describe the chart axes". Prefer a specialized skill when one fits; use this one otherwise.

Setup

library(kuzco)
library(ellmer)
 
chat <- chat_ollama(model = "qwen2.5vl")
chat$register_tool(tool_llm_image_custom())

Tool arguments (LLM-controlled)

• image (required): local file path to a jpeg, jpg, or png image.
• image_prompt (required): the prompt describing what to extract or describe about

the image. Be specific: state exactly what to look for and what shape the answer should take.

• system_prompt (optional): persona/style steering, defaults to a terse computer

vision assistant.

Result

Compact JSON text with the model's structured response to the custom prompt. Present it directly; do not reformat structured fields the user asked for.

Example

User: "How many bicycles are in img/street.jpg?"

1. Call kuzco_image_custom with image = "img/street.jpg" and

image_prompt = "Count the bicycles visible in the image. Return the count and briefly describe where each one is.".

2. Report the count and locations.