LlamaIndex Node Parser Kreuzberg

Element-aware LlamaIndex node parser for kreuzberg-extracted documents.

Installation

pip install llama-index-node-parser-kreuzberg

Requires llama-index-core>=0.13.0,<0.15. This package does not depend on kreuzberg directly — the kreuzberg package is a dependency of the reader (llama-index-readers-kreuzberg), which is needed for producing documents with element metadata.

Prerequisites

This parser requires documents with _kreuzberg_elements metadata. These are produced by KreuzbergReader configured with element-based extraction. Install llama-index-readers-kreuzberg (which brings in kreuzberg) to use the full workflow.

from kreuzberg import ExtractionConfig
from llama_index.readers.kreuzberg import KreuzbergReader

reader = KreuzbergReader(
    extraction_config=ExtractionConfig(result_format="element_based")
)
documents = reader.load_data("report.pdf")

Features

• Element-aware splitting — headings, paragraphs, tables, and code blocks each become a node
• Element type metadata preserved on each node (element_type, page_number, element_index)
• Source document relationships tracked via NodeRelationship.SOURCE
• Graceful degradation — documents without elements pass through with a warning
• Composes with other transformations (e.g., SentenceSplitter for further chunking)
• Async support via aget_nodes_from_documents
• Serialization support (to_dict / from_dict)

Usage

Basic

Full reader-to-nodes flow:

from kreuzberg import ExtractionConfig
from llama_index.readers.kreuzberg import KreuzbergReader
from llama_index.node_parser.kreuzberg import KreuzbergNodeParser

reader = KreuzbergReader(
    extraction_config=ExtractionConfig(result_format="element_based")
)
documents = reader.load_data("report.pdf")

parser = KreuzbergNodeParser()
nodes = parser.get_nodes_from_documents(documents)

IngestionPipeline

Chain with SentenceSplitter for further chunking of large elements:

from llama_index.core.ingestion import IngestionPipeline
from llama_index.core.node_parser import SentenceSplitter

pipeline = IngestionPipeline(
    transformations=[
        KreuzbergNodeParser(),
        SentenceSplitter(chunk_size=512),  # Further split large elements
    ]
)
nodes = pipeline.run(documents=documents)

VectorStoreIndex

Using the transformations parameter:

from llama_index.core import VectorStoreIndex

index = VectorStoreIndex.from_documents(
    documents,
    transformations=[KreuzbergNodeParser()],
)

Async

nodes = await parser.aget_nodes_from_documents(documents)

Behavior Notes

• Documents without _kreuzberg_elements metadata pass through unchanged with a warning. This is intentional — silently falling back would prevent users from noticing they are not getting element-aware splitting.
• Empty or whitespace-only elements are automatically skipped.