Docling (PDF Processing)

GPU-accelerated PDF processing service using the Docling VLM (Vision Language Model). Extracts text, tables, figures, and markdown from PDF documents.

Port: 8080
Model: GraniteDocling-258M

What it Does

Docling processes PDFs and extracts structured content:

Text extraction with document structure preservation
Table detection and structured output
Figure extraction with bounding boxes
Markdown conversion for readable output
OCR support for scanned documents

API Endpoints

Process PDF (Base64)

Endpoint: POST /process

curl -X POST "http://docling-server:8080/process" \
  -H "Content-Type: application/json" \
  -d '{
    "pdf_base64": "<base64-encoded-pdf>",
    "source_name": "document.pdf",
    "collection_name": "my-collection",
    "return_markdown": true
  }'

Request:

Parameter	Type	Description
`pdf_base64`	string	Base64-encoded PDF bytes
`source_name`	string	Original filename
`collection_name`	string	Collection for organizing
`chunk_size`	number	Text chunk size (default: 1000)
`chunk_overlap`	number	Overlap between chunks (default: 100)
`return_markdown`	boolean	Include markdown output

Response:

{
  "success": true,
  "chunks": [
    {"type": "text", "page_no": 1, "chunk_id": 0, "content": "..."}
  ],
  "tables": [...],
  "figures": [...],
  "markdown": "# Document Title\n\nContent...",
  "metadata": {
    "source_name": "document.pdf",
    "status": "completed"
  }
}

Process PDF (File Upload)

Endpoint: POST /process/upload

curl -X POST "http://docling-server:8080/process/upload" \
  -F "file=@document.pdf" \
  -F "collection_name=my-collection"

Response:

{
  "success": true,
  "markdown": "# Document Title...",
  "metadata": {
    "source_name": "document.pdf",
    "docling_pipeline_enabled": true,
    "ocr_mode": "auto"
  },
  "page_count": 5,
  "processing_time": 2.34
}

Health Check

curl http://docling-server:8080/health

Usage Example

import requests
import base64
 
# Read PDF file
with open("contract.pdf", "rb") as f:
    pdf_b64 = base64.b64encode(f.read()).decode()
 
# Process document
response = requests.post(
    "http://docling-server:8080/process",
    json={
        "pdf_base64": pdf_b64,
        "source_name": "contract.pdf",
        "return_markdown": True
    }
)
 
result = response.json()
print(result["markdown"])

Configuration

Variable	Default	Description
`DOCLING_ENABLE_PIPELINE`	`true`	Full VLM pipeline; `false` for fast text-only
`DOCLING_USE_CUDA`	`true`	Enable GPU acceleration
`DOCLING_CUDA_USE_FLASH_ATTENTION2`	`true`	Flash Attention 2

Fast Mode

Set DOCLING_ENABLE_PIPELINE=false for fast text-only extraction (no OCR, no AI):

docker run -e DOCLING_ENABLE_PIPELINE=false docling-server

ColPali (Document Vision)FastEmbed (Sparse Vectors)