Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.perceptron.inc/llms.txt

Use this file to discover all available pages before exploring further.

​
Overview

Focus gives Isaac 0.2 a tool that lets the model zoom into a region of the image and invoke itself again on the cropped view. When enabled, the model can crop into interesting regions, call itself on those zooms, and fuse the extra context back into the final answer. It effectively allocates more compute to the most relevant patches—great for tiny details, dense scenes, or hard-to-spot objects.

​
When to use Focus

  • Use Focus for pixel-level scrutiny (serial numbers, fine print, tiny defects), busy multi-object scenes where exhaustive search matters, or shelf-audit / parts-list flows where localized views improve recall.
  • Skip Focus for whole-image Q&A on clean scenes, time-sensitive workloads (Focus adds extra tool-call rounds → higher latency), and pure spatial detection where you only need point/box/polygon outputs.

​
How it works

  • Focus is modeled as a tool call — Isaac 0.2 (isaac-0.2-2b-preview, isaac-0.2-1b) can choose to invoke focus when the <hint>FOCUS</hint> system message is set.
  • The model selects regions, re-runs itself on cropped views, and combines those findings with the original context.

​
Usage (Python SDK)

import perceptron
from perceptron import perceive, image, text

perceptron.configure(provider="perceptron", api_key="YOUR_API_KEY")

# Let Isaac 0.2 zoom into the photo to recover fine details.
@perceive(model="isaac-0.2-2b-preview", focus=True)
def describe_scene(photo):
  return image(photo) + text("List every object you can see in this scene with colors.")

result = describe_scene("https://raw.githubusercontent.com/perceptron-ai-inc/perceptron/main/cookbook/_shared/assets/capabilities/qna/studio_scene.webp")
print("answer:", result.text)

​
Tips

  • Keep prompts concise; Focus decides where to zoom and how many views to take.
  • Use low temperature for consistent zoom choices when automating QA checks.
  • Combine with structured expects (expects="box" or "point") if you need precise localization alongside detailed text.

​
API usage

curl https://api.perceptron.inc/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $PERCEPTRON_API_KEY" \
  -d '{
    "model": "isaac-0.2-2b-preview",
    "messages": [
      { "role": "system", "content": "<hint>FOCUS</hint>" },
      { "role": "user",
        "content": [
          { "type": "image_url", "image_url": { "url": "<image-url>" } },
          { "type": "text", "text": "List every object you can see in this scene with colors." }
        ]
      }
    ]
  }'