RAG Framework Connectors#

Docling ships official integration packages for three RAG ecosystems: LangChain, LlamaIndex, and Open WebUI. Each connector exposes Docling's document parsing and chunking pipeline within the host framework's abstractions, so engineers can swap in Docling without changing the rest of their retrieval stack.

The Docling integration stubs live in docs/integrations/ 1 2 3 and point out to the full documentation maintained by each framework.

LangChain — langchain-docling#

Package: langchain-docling · GitHub: docling-project/docling-langchain · Official docs: python.langchain.com/docs/integrations/providers/docling 1

Install with:

pip install langchain-docling

The integration exposes one primary class, DoclingLoader 4, which handles conversion and chunking in a single call and returns standard LangChain Document objects.

Two export modes — choose at construction time 5:

from langchain_docling import DoclingLoader
from langchain_docling.loader import ExportType

For a full RAG pipeline example, see docs/examples/rag_langchain.ipynb.

LlamaIndex — llama-index-readers-docling + llama-index-node-parser-docling#

Official docs: LlamaIndex step-by-step guide 2

Install with:

pip install llama-index-readers-docling llama-index-node-parser-docling

The integration is split into two cooperating components 6:

from llama_index.readers.docling import DoclingReader
from llama_index.node_parser.docling import DoclingNodeParser

Two export paths 7:

• Markdown (DoclingReader()) — simple, lossy. Each document's .text is Markdown. Works without DoclingNodeParser.
• JSON + NodeParser (DoclingReader(export_type=DoclingReader.ExportType.JSON) + DoclingNodeParser()) — lossless; nodes carry page_no, bbox, headings, and other Docling metadata. Recommended for rich RAG metadata.

Component references: DoclingReader PyPI · DoclingNodeParser PyPI

Open WebUI — Plugin via docling-serve#

Official docs: docs.openwebui.com/features/rag/document-extraction/docling 3

Unlike the Python-library connectors above, Open WebUI delegates document extraction to a running docling-serve instance over HTTP 8. Configure it in Admin Panel → Documents → Content Extraction Engine.

Three environment keys drive the integration 9:

Open WebUI uses the async conversion path (/v1/convert/source/async) internally 10. Key gotcha: use ocr_preset (not the deprecated ocr_engine) in DOCLING_PARAMS. For Tesseract, ocr_lang uses + as separator (e.g. "eng+deu"), not an array.

For version-specific issues (e.g. the v1.25.0 rocm72 regression) and tenant-isolation 404s, see the Open WebUI Integration KB article.

Choosing the Right Connector#

Both LangChain and LlamaIndex integrations pass documents through Docling's HybridChunker under the hood — the tokenizer used for chunking must match the embedding model's tokenizer to avoid token-boundary mismatches 11.