Docling GLM-OCR Plugin
A Docling OCR plugin that delegates text recognition to a remote GLM-OCR model served via vLLM.
GitHub Repository: DCC-BS/docling-glm-ocr
Features
- Remote Delegation: Offloads OCR processing to a remote vLLM server hosting
zai-org/GLM-OCR. - Markdown Output: The model returns Markdown-formatted text, preserving headings, tables, and formulas.
- Concurrent Processing: Sends page crops as base64-encoded images via concurrent API requests with retry logic.
- Skips Born-Digital Pages: Pages with PDF text are not sent to the model as a whole; only regions that need OCR are.
- Clean Output: Discards repetition loops, caps tokens per crop by its size (
GLMOCR_REMOTE_OCR_MAX_TOKENS_PER_MEGAPIXEL), strips code fences and echoed prompts, and turns HTML tables into plain text rows. - Engine Key: Registers under the
"glm-ocr-remote"engine key.
Installation
- Using uv (recommended):
uv add docling-glm-ocr - Using pip:
pip install docling-glm-ocr
Usage
Configure the DocumentConverter with GlmOcrRemoteOptions:
python
from docling_glm_ocr import GlmOcrRemoteOptions
pipeline_options.ocr_options = GlmOcrRemoteOptions(
api_url="http://localhost:8001/v1/chat/completions",
model_name="zai-org/GLM-OCR"
)Environment Variables
| Variable | Description | Default |
|---|---|---|
GLMOCR_REMOTE_OCR_API_URL | vLLM chat completion URL | http://localhost:8001/v1/chat/completions |
GLMOCR_REMOTE_OCR_API_KEY | Bearer token for the Authorization header | unset |
GLMOCR_REMOTE_OCR_MODEL_NAME | Model name sent to vLLM | zai-org/GLM-OCR |
GLMOCR_REMOTE_OCR_PROMPT | Prompt sent with each crop | Markdown OCR prompt |
GLMOCR_REMOTE_OCR_LANG | Comma-separated language hints | en |
GLMOCR_REMOTE_OCR_TIMEOUT | HTTP timeout per crop (s) | 120 |
GLMOCR_REMOTE_OCR_MAX_TOKENS | Max tokens per completion | 16384 |
GLMOCR_REMOTE_OCR_MAX_TOKENS_PER_MEGAPIXEL | Token budget per megapixel of a crop (0 disables) | 2500 |
GLMOCR_REMOTE_OCR_SCALE | Crop rendering scale | 3.0 |
GLMOCR_REMOTE_OCR_MAX_IMAGE_PIXELS | Pixel budget per crop | 4500000 |
GLMOCR_REMOTE_OCR_MAX_CONCURRENT_REQUESTS | Concurrent API requests | 10 |
GLMOCR_REMOTE_OCR_MAX_RETRIES | Retries on HTTP errors | 3 |
GLMOCR_REMOTE_OCR_RETRY_BACKOFF_FACTOR | Exponential backoff factor | 2.0 |