Docling for structured document conversion

How it works

Docling represents parsed content in a document model. Its conversion pipeline can preserve relationships between text, tables and pages for downstream export or chunking.

What you need

Prepare sample files that reflect the layouts your app sees and pick an output format that preserves the needed structure. A document export and a chunking pipeline are separate choices.

What to try first

Run a report with a two-column page and a table. Compare reading order and table rows in the exported result with the original.

Compare captions with nearby figures and tables after conversion. Preserving the wording without that relationship can put important context in the wrong chunk.

Where to look for failures

OCR and layout processing can need additional models and compute. Inspect the document representation before deciding how to split it for retrieval.

Related Franklin pages

Compare converted output on a few documents with different structures. Choose according to the information downstream consumers need to recover, such as tables, source locations or heading boundaries.

• Turn documents into typed elements that a retrieval pipeline can filter and chunk. Unstructured for partitioning source documents
• Build ingestion, indexing and query workflows around external data. LlamaIndex for connecting documents to model queries

For how recognizing characters fits into the larger document processing workflow, see Franklin’s glossary. Optical character recognition

Project documentation

• Current project README
• README at the reviewed revision