haystack-oxidize-pdf

Haystack converter backed by oxidize-pdf, a fast Rust-powered PDF engine with first-class RAG chunking.

0.1.0 (2026-05-19) — Requires oxidize-pdf>=0.4.3 (oxidize-pdf-core 2.5.5+). Ships from day one with the full semantic regression suite (test_converter_disjoint.py, 6 tests) that guarantees the RAG-chunk disjointness contract end-to-end. Mirrors the discipline applied to langchain-oxidize-pdf 0.1.0 after the llama-index-readers-oxidize-pdf 0.1.0 → 0.1.1 incident, where shape-only tests missed a quadratic accumulation bug in the underlying chunker.

Install

pip install haystack-oxidize-pdf

Usage

The converter is a Haystack @component and can be dropped straight into a Pipeline. Sources accept paths (str / pathlib.Path) and ByteStream objects interchangeably.

Three modes

RAG chunks (default)

from haystack_oxidize_pdf import OxidizePdfConverter

converter = OxidizePdfConverter()  # mode="rag"
result = converter.run(sources=["paper.pdf"])

for doc in result["documents"]:
    print(doc.meta["chunk_index"], doc.meta["heading_context"])
    print(doc.content[:200])

Each Document.meta for mode="rag" carries:

Pipeline integration

from haystack import Pipeline
from haystack_oxidize_pdf import OxidizePdfConverter

pipeline = Pipeline()
pipeline.add_component("converter", OxidizePdfConverter(mode="rag"))
# ...add embedder, writer, etc.

result = pipeline.run({"converter": {"sources": ["paper.pdf"]}})

ByteStream input

Unlike LangChain / LlamaIndex loaders which take only file paths, the Haystack converter accepts ByteStream objects natively, leveraging PdfReader.from_bytes under the hood:

from haystack.dataclasses import ByteStream
from haystack_oxidize_pdf import OxidizePdfConverter

with open("paper.pdf", "rb") as f:
    stream = ByteStream(data=f.read(), mime_type="application/pdf",
                        meta={"upstream_origin": "s3://bucket/key"})

docs = OxidizePdfConverter().run(sources=[stream])["documents"]
# stream.meta is merged into each Document.meta (here: upstream_origin)

Batch sources with per-source metadata

docs = OxidizePdfConverter(mode="markdown").run(
    sources=["doc-a.pdf", "doc-b.pdf"],
    meta=[{"tag": "first"}, {"tag": "second"}],
)["documents"]
# docs[0].meta["tag"] == "first"
# docs[1].meta["tag"] == "second"

Or broadcast a single dict to every output document:

docs = OxidizePdfConverter().run(
    sources=["a.pdf", "b.pdf"], meta={"source_tag": "batch-A"}
)["documents"]
# all docs carry source_tag == "batch-A"

Metadata precedence

Three layers, deepest wins:

1. Base file-level fields (file_path, file_name, total_pages, pdf_version).
2. Caller-supplied meta (overrides base fields by design, lets callers re-label).
3. Per-document fields (chunk_index, page_numbers, page_number) applied last and never overwritten.

Related packages

• langchain-oxidize-pdf — same engine, LangChain BaseLoader interface.
• llama-index-readers-oxidize-pdf — same engine, LlamaIndex BaseReader interface.
• oxidize-pdf — the underlying PyO3 bridge (also ships the oxidize-mcp MCP server entry point).
• OxidizePdf.NET — .NET bindings.
• oxidize-pdf core (Rust) — the Rust engine, 99.3% parse success on 9k+ real-world PDFs.

License

MIT.