LlamaIndex Readers Xberg

LlamaIndex reader for 107 file formats powered by xberg's Rust extraction engine.

Installation

pip install llama-index-readers-xberg

Requires Python ≥3.10, xberg>=1.0.0, and llama-index-core>=0.14.23,<0.15.

Features

• 107 file formats -- PDF, DOCX, PPTX, XLSX, HTML, images, emails, archives, and more (full list)
• Rich metadata -- quality scores, language detection, keywords, annotations
• Native chunking -- xberg semantic chunks with heading path and page span, forwarded to XbergNodeParser
• Element extraction -- structural elements for structure-aware RAG pipelines
• Image extraction -- base64-encoded image data with position, format, and OCR metadata
• Per-page splitting -- one Document per page for fine-grained retrieval
• Batch processing -- multiple files in a single call
• Raw bytes input -- extract from in-memory bytes with a MIME type
• Native async -- true async via xberg's Rust tokio runtime
• Error tolerance -- skip failed files with warnings, or raise on failure
• Full serialization -- custom ExtractionConfig round-trips through to_dict()/from_dict() for pipeline caching

Usage

Basic Extraction

from llama_index.readers.xberg import XbergReader

reader = XbergReader()
documents = reader.load_data("report.pdf")

# Each document carries rich metadata
print(documents[0].metadata["file_name"])       # "report.pdf"
print(documents[0].metadata["file_type"])        # "application/pdf"
print(documents[0].metadata["total_pages"])      # 12
print(documents[0].metadata["quality_score"])    # 0.95
print(documents[0].metadata["detected_languages"])  # ["en"]

OCR Configuration

force_ocr is a top-level ExtractionConfig option. Language and backend are set on OcrConfig.

from xberg import ExtractionConfig, OcrConfig

reader = XbergReader(
    extraction_config=ExtractionConfig(
        force_ocr=True,
        ocr=OcrConfig(language="deu", backend="tesseract"),
    )
)
documents = reader.load_data("scanned.pdf")

Per-Page Splitting

PageConfig is nested inside ExtractionConfig. Each page becomes its own Document with a page_number metadata field.

from xberg import ExtractionConfig, PageConfig

reader = XbergReader(
    extraction_config=ExtractionConfig(
        pages=PageConfig(extract_pages=True),
    )
)
documents = reader.load_data("multi_page.pdf")  # One Document per page

for doc in documents:
    print(f"Page {doc.metadata['page_number']}: {doc.text[:80]}...")

Native Chunking

Enable xberg's native chunker to attach semantic chunks — each carrying its heading path and page span — under _xberg_chunks. Feed the documents into XbergNodeParser (from llama-index-node-parser-xberg) to turn each chunk into a node. This is the preferred input for RAG.

from xberg import ChunkingConfig, ExtractionConfig

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

chunks = documents[0].metadata["_xberg_chunks"]

Element Extraction

Setting result_format="element_based" populates _xberg_elements in document metadata for structure-aware processing. XbergNodeParser uses these when no chunks are present.

from xberg import ExtractionConfig

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

# Structural elements available for downstream node parsers
elements = documents[0].metadata["_xberg_elements"]

Batch Processing

Pass a list of file paths to extract multiple files in one call.

reader = XbergReader()
documents = reader.load_data(["report.pdf", "slides.pptx", "data.xlsx"])

Raw Bytes

Use data= and mime_type= keyword arguments to extract from in-memory bytes.

reader = XbergReader()

# Single bytes input
documents = reader.load_data(data=pdf_bytes, mime_type="application/pdf")

# Batch bytes input -- parallel lists of data and MIME types
documents = reader.load_data(
    data=[pdf_bytes, docx_bytes],
    mime_type=["application/pdf", "application/vnd.openxmlformats-officedocument.wordprocessingml.document"],
)

Async

aload_data provides native async extraction backed by xberg's Rust runtime.

documents = await reader.aload_data(["file1.pdf", "file2.pdf"])

SimpleDirectoryReader Integration

Register XbergReader as a file extractor for any supported extension.

from llama_index.core import SimpleDirectoryReader

reader = XbergReader()
sdr = SimpleDirectoryReader(
    input_dir="./documents",
    file_extractor={".pdf": reader, ".docx": reader, ".html": reader},
)
documents = sdr.load_data()

# Async variant works too
documents = await sdr.aload_data()

Behavior Notes

• Error tolerance: By default, raise_on_error=False -- the reader logs warnings and skips files that fail extraction.
• Strict mode: Set raise_on_error=True to propagate extraction exceptions immediately.
• Deterministic IDs: Document IDs are SHA-256 hashes of the file path (or byte content) and page number, enabling stable deduplication across pipeline runs.
• Metadata exclusion: Large metadata fields (_xberg_chunks, _xberg_elements, images) are automatically excluded from LLM and embedding metadata keys to keep prompt sizes manageable.
• Table handling: Tables extracted by xberg are appended as markdown to the document text when they are not already present in the content.
• Serialization: The reader fully supports to_dict()/from_dict() round-tripping, including ExtractionConfig with nested OcrConfig and PageConfig. This enables pipeline caching and persistence with IngestionPipeline.

Metadata Reference

Each Document produced by XbergReader includes these metadata fields (when available from the source document):