Biblicus

Make your documents usable by your assistant, then decide later how you will search and retrieve them.

Granite Docling-258M Extractor

Extractor ID: docling-granite

Category: Vision-Language Models (VLM)

Overview

The Granite Docling-258M extractor uses IBM Research's Granite Docling vision-language model for state-of-the-art document understanding. It achieves superior accuracy on technical content including tables, code blocks, and mathematical equations.

Granite Docling-258M is a 258-million parameter VLM optimized for high-accuracy document processing. It outperforms SmolDocling on complex document structures with F1 scores of 0.988 for code, 0.992 for tables, and 0.975 for equations.

Installation

Transformers Backend (Cross-Platform)

pip install "biblicus[docling]"
pip install "biblicus[docling-mlx]"

The MLX backend provides 2-3x faster inference on Apple Silicon (M1/M2/M3/M4) with lower memory usage.

Supported Media Types

The extractor automatically skips text items (text/plain, text/markdown) and audio items.

Configuration

Config Schema

class DoclingGraniteExtractorConfig(BaseModel):
    output_format: str = "markdown"  # markdown, text, or html
    backend: str = "mlx"              # mlx or transformers

Configuration Options

Option Type Default Description
output_format str markdown Output format: markdown, text, or html
backend str mlx Inference backend: mlx (Apple Silicon) or transformers (cross-platform)

Output Formats

Usage

Command Line

Basic Usage

# Extract using Granite Docling with defaults (markdown, MLX)
biblicus extract my-corpus --extractor docling-granite

Custom Configuration

# Use Transformers backend with HTML output
biblicus extract my-corpus --extractor docling-granite \
  --config output_format=html \
  --config backend=transformers

Configuration File

extractor_id: docling-granite
config:
  output_format: markdown
  backend: mlx
biblicus extract my-corpus --configuration configuration.yml

Python API

from biblicus import Corpus

# Load corpus
corpus = Corpus.from_directory("my-corpus")

# Extract with defaults
results = corpus.extract_text(extractor_id="docling-granite")

# Extract with custom config
results = corpus.extract_text(
    extractor_id="docling-granite",
    config={
        "output_format": "html",
        "backend": "transformers"
    }
)

In Pipeline

Fallback Chain

extractor_id: select-text
config:
  extractors:
    - docling-granite  # Highest accuracy
    - docling-smol     # Faster fallback
    - ocr-rapidocr     # Traditional OCR

High-Accuracy Override

extractor_id: select-smart-override
config:
  default_extractor: docling-smol
  overrides:
    - media_type_pattern: "application/pdf"
      extractor: docling-granite  # Use Granite for PDFs

Examples

Academic Papers with Equations

Extract research papers with mathematical equations:

biblicus extract papers-corpus --extractor docling-granite \
  --config output_format=markdown

Technical Documentation

Process documentation with code blocks:

biblicus extract docs-corpus --extractor docling-granite \
  --config output_format=html

Complex Tables

Extract spreadsheets and documents with complex tables:

biblicus extract tables-corpus --extractor docling-granite \
  --config backend=mlx

High-Accuracy Pipeline

Prioritize accuracy over speed:

from biblicus import Corpus

corpus = Corpus.from_directory("important-docs")

# Use Granite for maximum accuracy
results = corpus.extract_text(
    extractor_id="docling-granite",
    config={
        "output_format": "markdown",
        "backend": "mlx"
    }
)

Performance

Benchmarks

Comparison with SmolDocling

Metric Granite-258M SmolDocling-256M
Tables F1 0.992 0.985
Code F1 0.988 0.980
Equations F1 0.975 0.970
Speed ~7 sec/page 6.15 sec/page

When to Use Granite vs SmolDocling

Use Granite Docling-258M when:

Use SmolDocling-256M when:

Error Handling

Missing Dependency

If the Docling library is not installed:

ExtractionRunFatalError: DoclingGranite extractor requires an optional dependency.
Install it with pip install "biblicus[docling]".

Missing MLX Support

If MLX backend is configured but not available:

ExtractionRunFatalError: DoclingGranite extractor with MLX backend requires MLX support.
Install it with pip install "biblicus[docling-mlx]".

Empty Output

Documents that cannot be processed produce empty extracted text and are counted in extracted_empty_items statistics.

Per-Item Errors

Processing errors for individual items are recorded in the extraction snapshot but don't halt the entire extraction. Check errored_items in extraction statistics.

Use Cases

Research Papers

Granite excels at academic papers with:

Source Code Documentation

Ideal for technical documentation with:

Financial Reports

Handles business documents with:

Processes legal content with:

Same Category

Alternatives

Pipeline Utilities

See Also