---

name: ocr description: "Routes LayoutParser OCR wrappers, response parsing, and text aggregation workflows." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

OCR

Use this sub-skill when the task is about extracting text from images or parsing saved OCR responses into LayoutParser objects.

What belongs here

• TesseractAgent and TesseractFeatureType
• GCVAgent and GCVFeatureType
• OCR response loading, saving, and aggregation
• live OCR setup, credentials, and language selection

What does not belong here

• Layout detection model inference: use layout-models
• Geometry or reading-order logic after OCR: use layout-objects
• Box/text rendering: use visualization
• PDF or JSON/CSV loading: use layout-io

Read these files

• references/guide.md for agent behavior, aggregation levels, and troubleshooting notes
• ../layout-objects/SKILL.md when you need to group or sort OCR blocks
• ../visualization/SKILL.md when you need to render OCR output

Fast path

1. Decide whether the task needs live OCR or only saved-response parsing.
2. Use Tesseract for local OCR and GCV for Google Cloud Vision workflows.
3. Confirm the backend package and host dependency before constructing the agent.
4. Aggregate at the required level before post-processing the OCR output.
5. Keep saved responses when you need deterministic follow-up tests.

Common user requests

• "Use Tesseract OCR on this image"
• "Parse a saved Google Vision response"
• "Group the OCR output by line"
• "Load OCR results back into LayoutParser"
• "Set the custom Tesseract binary path"

Minimal smoke

The environment inspector confirms the optional OCR backends and the local tesseract binary:

python ../../scripts/inspect_backends.py

Failure clues

• pkg_resources missing after installing GCV usually means the setuptools build is not compatible with google-cloud-vision==1.
• Missing tesseract binary means the Python wrapper is present but live OCR cannot run.
• Missing credentials mean the GCV live API path is not ready.
• If the output text changes across runs, use a saved response instead of live OCR for tests.

Output discipline

When answering, name the OCR engine, the language or credential assumption, and the desired aggregation level. If the task only needs saved-response parsing, say so explicitly so future agents do not over-install the runtime.