CrewAI

The crewai-xberg package gives CrewAI agents document intelligence. It wraps Xberg’s extraction pipeline as agent tools that pull text, metadata, keywords, entities, and summaries from 98 file formats, with OCR where needed.

Tools

The package exposes three CrewAI BaseTools.

Tool	Input	Returns
XbergExtractTool	file_path	Extracted text, plus any requested rich results.
XbergExtractBatchTool	file_paths	One section per document via extract_batch, then errors.
XbergExtractMetadataTool	file_path	Title, authors, dates, page/table/image counts, format.

The batch tool routes every input through Xberg’s extract_batch in a single native call — faster than looping single extractions, since concurrency happens Rust-side. Per-file failures land in a trailing Errors section instead of aborting the batch.

Installation

pip install crewai-xberg

Requires Python 3.10+.

Quick start

from crewai import Agent
from crewai_xberg import XbergExtractTool, XbergExtractBatchTool, XbergExtractMetadataTool

agent = Agent(
    role="Document Analyst",
    goal="Extract and analyze document content",
    backstory="You process documents of any format.",
    tools=[XbergExtractTool(), XbergExtractBatchTool(), XbergExtractMetadataTool()],
)

Extraction options

Both extraction tools accept the same options; each toggles an Xberg ExtractionConfig capability.

Option	Default	Effect
output_format	"markdown"	plain, markdown, or html.
force_ocr	False	Run OCR on every page, even with a text layer.
chunk	False	Split into semantic chunks; report the chunk count.
extract_keywords	False	Append a keyword list.
extract_entities	False	Append named entities (people, orgs, locations).
summarize	False	Append a short summary.

Detected languages and tables are surfaced automatically when present.

tool = XbergExtractTool()

# Force OCR on a scan and surface keywords, entities, and a summary
enriched = tool.run(
    file_path="scan.pdf",
    force_ocr=True,
    extract_keywords=True,
    extract_entities=True,
    summarize=True,
)

# Many files in one batched extraction
combined = XbergExtractBatchTool().run(file_paths=["report.pdf", "notes.docx"])

The tools are async at the core. Inside an event loop, await tool.arun(...); the synchronous run bridges to it and must not run inside a running loop.

For the full extraction API and configuration options, see the Xberg documentation and the format support list.