Vue d’ensemble
GLM-OCR combines layout analysis with recognition. Its complete pipeline uses PP-DocLayoutV3 to identify document regions before recognition. The SDK can forward images and PDF documents to a hosted service or connect to self-hosted inference. In a knowledge-base workflow, it sits before indexing: convert pages into inspectable text and structure, then chunk and index those results while preserving page relationships. The OCR model has approximately 0.9B parameters; the complete service also includes layout processing.
Fonctionnalités clés
- Handles document text, tables and formulas.
- Includes layout analysis in the complete pipeline.
- Hosted mode returns Markdown and JSON layout details.
- Provides the glmocr parse command-line interface.
- The Python API accepts individual images and multiple pages.
- Offers hosted, vLLM and SGLang deployment paths.
Prérequis, installation et démarrage rapide
2. Configure pipeline.maas.enabled and an API key according to the official example, keeping credentials in local configuration or deployment secrets.
3. For the self-hosted layout pipeline, install pip install "glmocr[selfhosted]" and configure the model endpoint.
4. Run glmocr parse image.png --output ./results/ on a sample.
5. In Python, call parse("image.png") and result.save(output_dir="./results").
6. Inspect the extracted text and structure before processing directories or multi-page documents.
Utilisation
Implementation notes
Blur, skew, stamps and dense tables can introduce recognition errors. Sample-check numbers and fields that affect downstream retrieval or analysis, and validate extracted values against application rules.
Compatibilité des modèles et cas d’usage
Hosted mode needs credentials and network access and sends documents to the provider. Self-hosting requires a compatible inference environment. A GPU-free client does not remove the compute requirements of the server.
Licence et notes sur les risques
Repository code uses Apache-2.0, the GLM-OCR model uses MIT, and PP-DocLayoutV3 in the full pipeline uses Apache-2.0.