vision

name	vision
description	Analyze images using a Vision-Language Model through the wu CLI. Use when the user wants to describe an image, caption a picture, ask questions about visual content, inspect renders or screenshots, compare visual evidence, or produce text or JSON answers with public VLM backends.
version	0.1.0
author	NVIDIA Content Agents
tags	["content-agents","vision","vlm","cli"]
tools	["Shell","Filesystem","wu"]
compatibility	Requires the wu CLI, readable image files, and provider credentials for the selected VLM backend such as NVIDIA_API_KEY, OPENAI_API_KEY, ANTHROPIC_API_KEY, or GOOGLE_API_KEY.

Vision

Analyze images with wu vision and a configured VLM backend.

When to Use

• Use when the user asks to describe, caption, inspect, or ask questions about an image.
• Use when the user wants JSON output for downstream automation.
• Use when a render, screenshot, or reference image needs quick VLM feedback.
• Use the repository image attachment viewer first when the user needs visual inspection inside this conversation rather than a CLI/backend call.

Limitations

• The command sends image content to the selected model provider. Confirm that the user is comfortable with that provider before sending sensitive images.
• Backend behavior and image limits vary by provider and model.
• The CLI analyzes one image path per invocation.
• JSON output is the tool response shape, not a guaranteed strict schema for arbitrary prompts.

Prerequisites

• Activate the repo Python environment and confirm wu is on PATH.
• Ensure the image file exists locally.
• Set the credential for the selected backend: NVIDIA_API_KEY, OPENAI_API_KEY, ANTHROPIC_API_KEY, or GOOGLE_API_KEY.

Instructions

1. Confirm the image path and the user's question or captioning goal.
2. Choose the backend. Use nim by default unless the user asks for another supported provider.
3. Use a specific --model only when the user requests it or the task needs a known model.
4. Use --format json for programmatic handoff.
5. Summarize the VLM answer and include the exact command used.

Command Reference

wu vision <image_path> [OPTIONS]

Option	Description
--prompt, -p	Prompt or question. Default asks for a detailed description.
--backend, -b	VLM backend: nim, gemini, openai, or anthropic.
--model, -m	Optional model name.
--system-prompt	System prompt for the VLM.
--temperature, -t	Sampling temperature. Default is 0.7.
--max-tokens	Maximum response tokens. Default is 1024.
--format, -f	text or json.
--verbose, -v	Enable debug logging.

Common Workflows

# Describe an image.
wu vision image.png

# Ask a specific question.
wu vision image.png -p "What objects are visible in this image?"

# JSON output.
wu vision image.png --format json

# Use OpenAI.
wu vision image.png --backend openai

# Use a specific public NIM model.
wu vision image.png --backend nim --model "qwen/qwen3.5-397b-a17b"

Output Format

Report:

• Command executed, image path, backend, model when provided, and output format.
• The answer or a concise summary of the answer.
• Any confidence caveats from the model output.
• Any blocker such as missing file, missing API key, unsupported backend, or provider timeout.

Troubleshooting

Symptom	Cause	Fix
API key required	Credential for the selected backend is missing.	Set the matching provider key in the local environment or .env.
File not found	Image path is wrong or relative to another directory.	Use an absolute path or rerun from the correct working directory.
Slow response	Large image, cold provider, or high token budget.	Try a smaller image, lower --max-tokens, or another backend.
Weak answer	Prompt is too broad or the image lacks detail.	Ask a more specific question or provide a higher-quality image.