llama-index-readers-oxidize-pdf

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

0.1.1 (2026-04-22) — Requires oxidize-pdf>=0.4.3 (oxidize-pdf-core 2.5.5). Fixes a regression carried by 0.1.0: the underlying HybridChunker in oxidize-pdf-core 2.5.4 re-injected just-flushed elements through its overlap branch, so each Document produced by mode="rag" contained the previous one as a prefix (quadratic accumulation). 0.1.1 is the first release where the disjointness contract is verified end-to-end by semantic regression tests; 0.1.0 should not be used for RAG ingestion.

Install

pip install llama-index-readers-oxidize-pdf

Usage

RAG chunks (default)

from llama_index.readers.oxidize_pdf import OxidizePdfReader

reader = OxidizePdfReader()  # mode="rag" by default
documents = reader.load_data("paper.pdf")

for doc in documents:
    print(doc.metadata["chunk_index"], doc.metadata["heading_context"])
    print(doc.text[:200])

Each Document carries:

One document per page

reader = OxidizePdfReader(mode="pages")
docs = reader.load_data("paper.pdf")
for doc in docs:
    print(doc.metadata["page_number"], len(doc.text))

Whole PDF as markdown

reader = OxidizePdfReader(mode="markdown")
[doc] = reader.load_data("paper.pdf")
print(doc.text)

Why oxidize-pdf

• Rust parser: fast on large PDFs, low memory footprint.
• Native RAG primitives: element-disjoint semantic chunking, element partitioning, heading-aware context — no post-processing needed. The disjointness contract (no chunk's text is a substring of another's; each source element appears in exactly one chunk) is enforced by regression tests in both this reader and the underlying bridge.
• CJK friendly: compact output for multibyte documents (see oxidize-pdf 2.5.4 subsetter fixes).
• Pure Python install: ships as a wheel for Linux/macOS/Windows via the oxidize-pdf package; no system dependencies.

Source

Part of oxidize-pdf-integrations, the ecosystem of integrations around oxidize-pdf. The Rust core and Python bridge live in oxidize-python.

License

MIT