llama-index-node-parser-xberg 1.0.1

Structure-aware LlamaIndex node parser that turns xberg native chunks and elements into nodes

LlamaIndex Node Parser Xberg

Structure-aware LlamaIndex node parser for xberg-extracted documents. It turns xberg's native chunks into nodes, and falls back to structural elements when chunks are absent.

Installation

pip install llama-index-node-parser-xberg

Requires llama-index-core>=0.14.23,<0.15. This package does not depend on xberg directly — xberg is a dependency of the reader (llama-index-readers-xberg), which produces the documents this parser splits.

Prerequisites

This parser requires documents with _xberg_chunks or _xberg_elements metadata. These are produced by XbergReader. Prefer native chunking; use element-based extraction when you want one node per structural element. Documents carrying neither pass through unchanged with a warning.

from xberg import ChunkingConfig, ExtractionConfig
from llama_index.readers.xberg import XbergReader

# Preferred: native semantic chunks with heading path and page span.
reader = XbergReader(
    extraction_config=ExtractionConfig(chunking=ChunkingConfig(max_characters=1000, overlap=200))
)
documents = reader.load_data("report.pdf")

Features

• Chunk-aware splitting — each xberg native chunk becomes a node, carrying chunk_type, heading_path, and page span
• Element fallback — when no chunks are present, headings, paragraphs, tables, and code blocks each become a node
• Source and prev/next relationships tracked via NodeRelationship
• Graceful degradation — documents without chunk or element metadata pass through with a warning
• Composes with other transformations (e.g., SentenceSplitter)
• Async support via aget_nodes_from_documents
• Serialization support (to_dict / from_dict)

Usage

Basic

Full reader-to-nodes flow:

from xberg import ChunkingConfig, ExtractionConfig
from llama_index.readers.xberg import XbergReader
from llama_index.node_parser.xberg import XbergNodeParser

reader = XbergReader(
    extraction_config=ExtractionConfig(chunking=ChunkingConfig(max_characters=1000, overlap=200))
)
documents = reader.load_data("report.pdf")

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

IngestionPipeline

Chain with SentenceSplitter to further split any oversized nodes:

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

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

VectorStoreIndex

Using the transformations parameter:

from llama_index.core import VectorStoreIndex

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

Async

nodes = await parser.aget_nodes_from_documents(documents)

Behavior Notes

• Chunks take priority over elements. When a document carries both _xberg_chunks and _xberg_elements, the parser splits on chunks.
• Documents without either metadata key pass through unchanged with a warning. This is intentional — silently falling back would hide that you are not getting structure-aware splitting.
• Empty or whitespace-only chunks and elements are automatically skipped.