Framework Integrations

Relevant source files

• .github/workflows/checks.yml
• README.md
• docs/examples/minimal_vlm_pipeline.py
• docs/index.md
• docs/usage/index.md
• mkdocs.yml

This page documents how Docling integrates with popular AI frameworks and libraries for downstream tasks such as Retrieval-Augmented Generation (RAG), document processing, and agentic applications. The focus is on the integration patterns, export mechanisms, and framework-specific adaptations that enable Docling to work seamlessly with LangChain, LlamaIndex, Haystack, and other AI development frameworks.

For document chunking strategies specific to RAG applications, see Document Chunking. For the Model Control Protocol (MCP) server integration with AI agents, see MCP Server Integration. For general export format details, see Export Formats.

Integration Architecture

Docling serves as a document parsing and conversion layer that produces structured DoclingDocument objects. These documents can be exported to various formats and consumed by downstream AI frameworks. The integration pattern is designed to be framework-agnostic, with Docling focusing on high-quality document understanding and frameworks focusing on their specific AI/ML workflows.

Sources: mkdocs.yml109-148 README.md29-43

Core Integration Pattern

The standard integration pattern involves three stages:

1. Document Conversion: Use DocumentConverter to parse input documents into DoclingDocument objects
2. Export/Transformation: Export the DoclingDocument to a format compatible with the target framework
3. Framework Consumption: Feed the exported data into framework-specific loaders, parsers, or converters

DoclingDocument as Integration Hub

The DoclingDocument class from the docling_core library serves as the central integration point. It provides a unified representation with:

• Structured Content: Body items (text, tables, pictures, sections) organized hierarchically
• Provenance Tracking: Bounding boxes and page numbers for spatial grounding
• Rich Metadata: Document origin, formatting, and semantic labels
• Export Methods: Multiple serialization formats for framework compatibility

Sources: README.md34-36 docs/index.md40

Framework-Specific Integrations

LangChain Integration

LangChain integration typically uses Docling as a document loader to populate Document objects for downstream processing. The integration leverages:

• Markdown Export: Convert DoclingDocument to Markdown text for LangChain's Document class
• Metadata Preservation: Transfer document metadata to LangChain's metadata fields
• Text Splitting: Combine Docling's chunking with LangChain's text splitters
• Vector Store Indexing: Feed processed documents into LangChain vector stores

Integration Pattern:

Sources: mkdocs.yml110-111 README.md38

LlamaIndex Integration

LlamaIndex integration uses Docling for document parsing with integration into LlamaIndex's Document and node structure:

• Document Reader Pattern: Docling acts as a custom document reader
• Node Creation: Convert DoclingDocument content items into LlamaIndex nodes
• Hierarchical Relationships: Preserve section hierarchy in node relationships
• Multimodal Support: Handle text, tables, and images through different node types

Integration Pattern:

Sources: mkdocs.yml112 README.md38

Haystack Integration

Haystack integration positions Docling as a document converter in Haystack's pipeline architecture:

• Converter Component: Docling as a preprocessing component
• Document Store Compatibility: Export to formats compatible with Haystack document stores
• Pipeline Integration: Embed Docling conversion in Haystack processing pipelines
• Metadata Handling: Map Docling metadata to Haystack's document metadata schema

Integration Pattern:

Sources: mkdocs.yml110 README.md38

Crew AI Integration

Crew AI integration uses Docling as a tool within agent workflows:

• Tool Integration: Docling converter as a Crew AI tool
• Document Processing Tasks: Agents can invoke Docling to parse documents
• Result Formatting: Export to agent-friendly formats
• Multi-Document Processing: Batch processing for complex agent tasks

Sources: mkdocs.yml141 README.md38

Other Framework Integrations

Additional frameworks supported through the Docling integration ecosystem:

Framework	Integration Type	Primary Use Case
txtai	Document loader	Semantic search and indexing
Semantica	Document processor	Knowledge graph construction
Bee Agent Framework	Agent tool	Document processing in agent workflows
Langflow	Node component	Visual workflow integration
Hector	Document parser	Domain-specific document processing
Apify	Actor integration	Web scraping and document processing
Data Prep Kit	Data pipeline	Large-scale document preparation
InstructLab	Knowledge base	Synthetic data generation
spaCy	NLP pipeline	Entity extraction and NLP tasks
Prodigy	Annotation tool	Document annotation workflows

Sources: mkdocs.yml137-166 README.md116-117

Common Integration Patterns

Document Loading Pipeline

Sources: README.md73-81 docs/usage/index.md5-20

RAG Pipeline Integration

Sources: mkdocs.yml109-113 README.md38

Multi-Framework Processing Workflow

This pattern demonstrates how a single DoclingDocument can feed multiple frameworks simultaneously:

Sources: README.md34-36

Export and Serialization for Framework Integration

Supported Export Formats

The DoclingDocument class provides multiple export methods optimized for different framework requirements:

Export Method	Output Format	Primary Use Case	Framework Compatibility
export_to_markdown()	Markdown text	Text-based processing, RAG	LangChain, LlamaIndex, Haystack
export_to_json()	JSON string	Structured data exchange	All frameworks
export_to_dict()	Python dictionary	Programmatic access	Custom integrations
export_to_html()	HTML string	Web rendering	Web applications
export_to_doctags()	DocTags format	Semantic structure	Specialized NLP pipelines

Export Options and Customization

Each export method supports options to control output formatting:

• Markdown Export: Control heading levels, include/exclude images, table formatting
• JSON Export: Include/exclude provenance, metadata, formatting details
• HTML Export: Custom CSS classes, image handling, table styling
• DocTags Export: Semantic tag preservation, hierarchical structure

Metadata Mapping

Framework integrations typically require mapping Docling metadata to framework-specific schemas:

Sources: README.md36 docs/index.md41

Chunking Integration for RAG

For RAG applications, document chunking is a critical integration point. Docling provides chunking capabilities that work with framework-specific chunkers:

• Hybrid Chunking: Combine semantic (section-based) and size-based chunking
• Provenance Preservation: Maintain source document references in chunks
• Hierarchical Chunks: Preserve document structure in chunk organization
• Framework Compatibility: Generate chunks compatible with LangChain, LlamaIndex chunkers

See Document Chunking for detailed information on chunking strategies and implementation.

Sources: mkdocs.yml103-106 docs/index.md78

Integration Examples

The Docling repository includes working examples demonstrating framework integrations:

Example	Framework	File Location	Description
RAG with Haystack	Haystack	docs/examples/rag_haystack.ipynb	Complete RAG pipeline with Haystack
RAG with LangChain	LangChain	docs/examples/rag_langchain.ipynb	Document loading and RAG with LangChain
RAG with LlamaIndex	LlamaIndex	docs/examples/rag_llamaindex.ipynb	Query engine integration with LlamaIndex
RAG with Milvus	Vector Store	docs/examples/rag_milvus.ipynb	Vector store integration with Milvus
RAG with Qdrant	Vector Store	docs/examples/retrieval_qdrant.ipynb	Retrieval pipeline with Qdrant
RAG with Weaviate	Vector Store	docs/examples/rag_weaviate.ipynb	Document indexing with Weaviate
RAG with MongoDB	Vector Store	docs/examples/rag_mongodb.ipynb	Atlas Vector Search integration
RAG with Azure Search	Vector Store	docs/examples/rag_azuresearch.ipynb	Azure Cognitive Search integration
RAG with OpenSearch	Vector Store	docs/examples/rag_opensearch.ipynb	OpenSearch vector store integration
Visual Grounding	Multimodal	docs/examples/visual_grounding.ipynb	Spatial provenance for multimodal RAG

Sources: mkdocs.yml109-135

Best Practices for Framework Integration

Document Preparation

1. Choose Appropriate Pipeline: Use StandardPdfPipeline for structured documents, VlmPipeline for complex layouts
2. Configure OCR Settings: Enable OCR for scanned documents to ensure text extraction
3. Enable Enrichments: Use picture classification, table structure, and code detection as needed
4. Set Timeout Limits: Configure appropriate timeouts for production environments

Export Strategy

1. Select Appropriate Format: Use Markdown for text-heavy workflows, JSON for structured data needs
2. Preserve Metadata: Include document origin and provenance for traceability
3. Handle Large Documents: Consider chunking before export for memory efficiency
4. Validate Exports: Verify export quality matches framework requirements

Framework Adaptation

1. Map Metadata Correctly: Ensure Docling metadata aligns with framework schemas
2. Handle Multimodal Content: Plan for tables, images, and formulas in framework pipelines
3. Preserve Structure: Maintain document hierarchy when supported by the framework
4. Test Integration Points: Validate end-to-end workflows with representative documents

Performance Optimization

1. Batch Processing: Use DocumentConverter for multiple documents efficiently
2. Model Caching: Leverage model caching for repeated conversions
3. GPU Acceleration: Enable GPU support for ML models when available
4. Pipeline Selection: Choose the fastest pipeline that meets quality requirements

Sources: README.md84-85 docs/usage/index.md22-24