pdfluent-extract

PDF and Office files to Markdown, JSON and RAG chunks, in three calls. It is pure Python: the engine underneath is pdfluent, which pip pulls in automatically. There is nothing to compile and no system dependency to install — on Linux arm64 and Alpine/musl as much as anywhere else, because pdfluent 1.1.0 ships a wheel for every platform in the matrix below.

Install

pip install pdfluent-extract

It needs Python 3.8 or later and pdfluent 1.1.0 or later.

Use

from pdfluent_extract import to_markdown

print(to_markdown("report.pdf"))

to_markdown takes a path, bytes or a file-like object, and routes PDF and Office (.docx/.pptx/.xlsx) by magic bytes, not by the extension. pages lists 0-based PDF page indices (None for the whole file).

from pdfluent_extract import to_json

doc = to_json("report.pdf")
print(doc["name"], len(doc["texts"]))

to_json returns the file as a Docling DoclingDocument, as a dict — the same JSON DoclingDocument.model_validate_json reads.

Command line

$ pdfluent-extract report.pdf -o report.md
$ pdfluent-extract report.pdf --format json --pages 1-3 > doc.json
$ cat report.pdf | pdfluent-extract - --format text

--format is one of md, json, chunks or text; --pages counts from 1 (1-3,5). Without -o (or with -o -) the output goes to stdout.

Chunks for RAG

from pdfluent_extract import to_chunks

for chunk in to_chunks("report.pdf", max_tokens=512):
    print(chunk.metadata["page"], chunk.metadata["section_path"], chunk.text[:80])

One chunk per paragraph, heading and list item, and one per table. Every chunk carries page, section_path (the enclosing headings) and bbox in its metadata, plus source for path inputs. Tables longer than the budget split by rows with the header repeated. PDF only in v1.

With LangChain and LlamaIndex

pip install "pdfluent-extract[langchain]"

from langchain_pdfluent import PDFluentLoader

docs = PDFluentLoader("report.pdf").load()
print(docs[0].page_content, docs[0].metadata["bbox"])

pip install "pdfluent-extract[llamaindex]"

from llama_index.readers.pdfluent import PDFluentReader

docs = PDFluentReader().load_data("report.pdf")
print(docs[0].text[:200])

Supported platforms

Every platform pdfluent 1.1.0 ships a wheel for: macOS (universal2), Windows x64, Linux x86_64 and arm64 on glibc (manylinux2014) and musl (musllinux_1_2) — Python 3.8 through 3.13 from one abi3 wheel per platform.

Quality

quality="fast" is the only level in v1; anything else is refused by name. Further levels (vision/OCR) are designed later.

Licence

AGPL-3.0 or a commercial licence, at your option; see LICENSE and LICENSE-COMMERCIAL. The commercial licence is sold yearly and self-service at https://pdfluent.com/sdk/pricing/.