Load Docugami Documents into Knowledge Graph

The Docugami Loader takes in IDs of PDF, DOCX, or DOC files that have been processed by Docugami and returns nodes in a Document XML Knowledge Graph for each document - a rich representation capturing the semantic and structural characteristics of the document's chunks as an XML tree. When entire sets of documents are processed, the result is a forest of these XML semantic trees, one per document.

Using it requires a Docugami workspace first: documents are uploaded and Docugami ingests and clusters them into sets of similar documents (for example NDAs, Lease Agreements, or Service Agreements - the clustering is not fixed and depends on your own documents, and doc set assignments can be changed later). An access token is created via the Developer Playground, and the Docugami API can then be explored to find docset IDs and the document IDs within them. DocugamiReader is called with load_data(docset_id=..., document_ids=...), where the document_ids array is optional - omitting it loads every document in the docset.

The source frames Docugami's chunking as distinct from simpler techniques (whitespace-based splitting, recursive chunking by character length) across four dimensions. Intelligent Chunking breaks each document into a hierarchical semantic XML tree of chunks ranging from single words to entire sections, following the document's own semantic contours rather than an arbitrary length cutoff. Structured Representation adds attributes marking headings, paragraphs, lists, tables, and other structural elements consistently across formats - including scanned PDFs or DOCX files - while correctly handling page headers/footers and multi-column layouts for clean text extraction. Semantic Annotations tag chunks coherently across an entire document set, so a query for something like the Landlord, Tenant, or Renewal Date works consistently across a set of differently-formatted lease agreements, down to more complex information like sub-lease wording or jurisdiction-specific termination exceptions. Additional Metadata attaches further annotations (for Docugami users) that support high-accuracy document QA without hitting context window limits.

Use it when you need to index a set of related documents (contracts, agreements, or similar structured document families) with chunking that respects their actual semantic and structural layout, rather than naive length-based splitting - particularly when you need consistent cross-document queries like comparing a clause across many lease agreements. Do not use it for one-off documents with no meaningful structure or cross-document comparison need; the value here is specifically in the semantic/structural richness and cross-document consistency Docugami's clustering and XML tree provide.

load_data fetches a Docugami-processed docset (optionally scoped to specific document IDs) and returns each document as a semantic XML knowledge graph node, with structural attributes (headings, paragraphs, lists, tables), cross-document-consistent semantic tags, and optional additional metadata for high-accuracy QA.

pip install llama-index-readers-docugami

Requires a Docugami workspace with documents already uploaded, clustered into docsets, and an access token created via the Developer Playground.

Developers indexing sets of related structured documents - contracts, agreements, or similar families - who need chunking and cross-document semantic consistency that simple length-based splitting cannot provide.