kreuzberg-txtai 0.1.0

Kreuzberg document extraction pipeline for txtai and other callable-based frameworks

kreuzberg-txtai

A Kreuzberg-backed document extraction pipeline for txtai and any Python framework built around the __call__ convention.

KreuzbergPipeline replaces txtai's built-in Textractor (Apache Tika-based) with Kreuzberg's Rust-powered extraction stack, turning document paths into a list[dict] with content and metadata fields — surfacing title, MIME type, and page count that Tika flattens away.

Installation

pip install kreuzberg-txtai

For the txtai integration examples below:

pip install "kreuzberg-txtai[txtai]"

Requires Python 3.10+.

Quick Start

from kreuzberg_txtai import KreuzbergPipeline

pipeline = KreuzbergPipeline()
docs = pipeline(["doc1.pdf", "doc2.docx", "doc3.html"])

for doc in docs:
    print(doc["metadata"]["source"], "->", len(doc["content"]), "chars")

Each element in docs looks like:

{
    "content": "# Sample Document\n\nExtracted text...",
    "metadata": {
        "source": "doc1.pdf",
        "mime_type": "application/pdf",
        "title": "Sample Document",
        "page_count": 5,
    },
}

Features

• 88+ file formats — PDF, DOCX, PPTX, XLSX, images, HTML, Markdown, plain text, and more via Kreuzberg
• Stable dict contract — every extraction returns content + metadata with the same four keys, regardless of source format
• Rich metadata — source path, MIME type, title, and page count surface directly
• Batch support — pass a single path or a list[str]; output is always list[dict] in input order
• Full Kreuzberg control — pass an ExtractionConfig to drive output format, OCR backend/language, force_ocr, and every other Kreuzberg knob
• Framework-agnostic — txtai is an optional extra, not a hard dependency; the pipeline works in any framework that accepts a callable
• Typed — ships with a py.typed marker; full mypy strict compatibility

Usage Examples

RAG ingestion with txtai.Embeddings

The dominant real-world pattern — extract, index, search:

from kreuzberg_txtai import KreuzbergPipeline
from txtai import Embeddings

pipeline = KreuzbergPipeline()
docs = pipeline(["doc1.pdf", "doc2.docx", "doc3.html"])

embeddings = Embeddings({
    "path": "sentence-transformers/all-MiniLM-L6-v2",
    "content": True,
})
embeddings.index([(i, doc["content"], None) for i, doc in enumerate(docs)])

results = embeddings.search("query", limit=5)

Inside a txtai.workflow.Task

Task accepts any callable, so KreuzbergPipeline drops in without wrappers. Because the pipeline returns list[dict], downstream tasks that expect strings need a one-line adapter:

from txtai.workflow import Task, Workflow
from kreuzberg_txtai import KreuzbergPipeline

extract = KreuzbergPipeline()

wf = Workflow([
    Task(extract),
    Task(lambda docs: [d["content"] for d in docs]),  # flatten dicts -> strings
])

list(wf(["doc1.pdf", "doc2.pdf"]))

Framework-free loop

from kreuzberg import ExtractionConfig
from kreuzberg_txtai import KreuzbergPipeline

pipeline = KreuzbergPipeline(config=ExtractionConfig(output_format="plain"))
for doc in pipeline(["scan1.pdf", "scan2.pdf"]):
    print(doc["metadata"]["source"], "->", len(doc["content"]), "chars")

No txtai needed — the class works on just the core kreuzberg dependency.

Tuning extraction with ExtractionConfig

Every Kreuzberg knob — output format, OCR backend and language, force_ocr, chunking, custom mime handling — lives on ExtractionConfig. Build one and hand it to the pipeline:

from kreuzberg import ExtractionConfig, OcrConfig
from kreuzberg_txtai import KreuzbergPipeline

custom = ExtractionConfig(
    output_format="markdown",
    ocr=OcrConfig(backend="tesseract", language="eng+deu"),
    force_ocr=True,
)

pipeline = KreuzbergPipeline(config=custom)
docs = pipeline("scanned_report.pdf")

See the Kreuzberg docs for the full list of ExtractionConfig and OcrConfig fields.

Constructor

Parameter	Type	Default	Notes
config	ExtractionConfig | None	None	Drives output format, OCR settings, force_ocr, and every other Kreuzberg option. None falls back to Kreuzberg's defaults.

Return Shape

__call__ always returns list[dict] — a single-path input still returns a length-1 list. Each dict has exactly two top-level keys:

• content — the extracted text in the format set by config.output_format (Kreuzberg's default when no config is passed)
• metadata — a dict with exactly four keys: source, mime_type, title, page_count

Missing metadata fields are None (rather than omitted) to keep the dict shape stable across document types.

Related Projects

• kreuzberg — the extraction engine powering this package
• langchain-kreuzberg — Kreuzberg document loader for LangChain
• llama-index-kreuzberg — LlamaIndex reader and node parser
• kreuzberg-crewai — CrewAI agent tool
• kreuzberg-surrealdb — SurrealDB ingestion connector

License

MIT — see LICENSE.