multixtract — Vendor-neutral document extraction, OCR, chunking, and embeddings for RAG pipelines

Pull text, tables, and images out of PDFs, Word, PowerPoint, Excel/CSV and more — let any vision model describe the images, chunk everything for retrieval, embed it, and store the result anywhere.

The core is tiny (just Pillow + ImageHash). Every format parser and every cloud SDK is an optional extra — install only what you need.

Highlights

✅ Vendor-neutral — swap OpenAI for Azure, Qwen, Llama, or your own model with one line
✅ Extract text, tables, and images from 15+ file formats
✅ Modular — use only extraction, or run the full extract → vision → chunk → embed → store pipeline
✅ Fully offline — local vision models, no API key, no cloud
✅ Tiny core install — only Pillow + ImageHash; every heavy dependency is optional
✅ Fully typed — mypy and pyright compatible out of the box

Why Multixtract?

Most libraries optimise for one part of the workflow — parse documents, run OCR, generate embeddings, or store vectors. You end up stitching together five packages with incompatible interfaces and rebuilding the same pipeline on every project.

multixtract connects all of them without locking you into a provider. Swap OpenAI for Azure or a local model, swap Azure Blob for S3, add a new file format — none of it touches the rest of the pipeline.

What multixtract is not: It is not a vector database, retrieval framework, or chat system. It focuses on document ingestion and preparation — getting clean, structured, chunked content into whatever AI system you're building.

Quick Start

pip install "multixtract[pdf,docx,pptx,xlsx]"

One call. Any document. Done.

from multixtract import Pipeline

Pipeline().process("report.pdf")                      # extract → filter → chunk
Pipeline().process("report.pdf", split_chunks=True)   # + write individual chunk files

Or stay close to the data:

from multixtract import extract_document, chunk_document

document, images = extract_document("report.pdf")
chunks = chunk_document(document, base_name="report")

Supported formats: PDF · DOCX · PPTX · XLSX · CSV · EPUB · HTML · RTF · Email · Images · Plain text · Markdown — and legacy .doc / .ppt via LibreOffice.

Vision providers: OpenAI · Azure OpenAI · Qwen2.5-VL · Llama 3.2 Vision · SmolVLM (CPU) · bring your own.

→ Full installation guide · Recipes · Provider setup

Document Schema

Every call to extract_document returns the same structure regardless of input format:

PDF / DOCX / PPTX / XLSX / …
        │
        ▼
{
  "_base_name": "report",
  "metadata": { "format": "pdf", "page_count": 12, ... },
  "pgs": [
    {
      "pg_num": 1,
      "kind":   "page",
      "title":  "Executive Summary",
      "txt":    "The quarterly results show a 12% increase...",
      "tables": [
        [["Region", "Q1", "Q2"], ["North", "1.2M", "1.4M"], ...]
      ],
      "imgs": [
        { "image_id": "report-p1-img0", "width": 800, "height": 600 }
      ],
      "hyperlinks": ["https://example.com/data"]
    },
    ...
  ]
}

→ Full data model and chunk schema

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                        Your document                            │
│          PDF · DOCX · PPTX · XLSX · EPUB · HTML · RTF …         │
└────────────────────────────┬────────────────────────────────────┘
                             │
                    ┌────────▼────────┐
                    │   Extractors    │  (registry — one per format)
                    └────────┬────────┘
                             │
              ┌──────────────┼──────────────┐
              │              │              │
         ┌────▼────┐   ┌─────▼─────┐  ┌─────▼───┐
         │  Text   │   │  Tables   │  │ Images  │
         └────┬────┘   └─────┬─────┘  └─────┬───┘
              │              │              │
              │              │    ┌─────────▼───────────┐
              │              │    │  ImageFilterPipeline│
              │              │    │  · dimension        │
              │              │    │  · solid-color      │
              │              │    │  · icon rejection   │
              │              │    │  · logo dedup (hash)│
              │              │    └─────────┬───────────┘
              │              │              │
              │              │    ┌─────────▼───────────┐
              │              │    │     VisionModel     │
              │              │    │   OpenAI · Azure    │
              │              │    │  Qwen · Llama · CPU │
              │              │    │  (or skip entirely) │
              │              │    └─────────┬───────────┘
              │              │              │
              └──────────────┴──────────────┘
                             │
                    ┌────────▼────────┐
                    │    Chunking     │  sliding-window · table-MD · image
                    └────────┬────────┘
                             │
                    ┌────────▼────────┐
                    │    Embedder     │  OpenAI · Azure · BYO · (skip)
                    └────────┬────────┘
                             │
                    ┌────────▼────────┐
                    │    BlobStore    │  LocalDisk · AzureBlob · S3 · BYO
                    └─────────────────┘

The pipeline talks only to three interfaces — it never imports a vendor directly:

Add a new format with register_extractor. Plug in S3, GCS, or any backend by implementing three methods on BlobStore.

Integrations

Multixtract is an extraction and chunking layer, not a RAG framework. It fits underneath the tools you already use:

from multixtract import extract_document, chunk_document
from langchain.schema import Document as LCDocument

document, _ = extract_document("report.pdf")
chunks = chunk_document(document, base_name="report")

# LangChain
lc_docs = [LCDocument(page_content=c["content"], metadata={"pg": c["pg_num"]}) for c in chunks]

# LlamaIndex
from llama_index.core import Document as LIDocument
li_docs = [LIDocument(text=c["content"], metadata={"chunk_id": c["chunk_id"]}) for c in chunks]

Example projects

Each example is a self-contained ingest.py with a --query flag so you can extract, store, and query in one command.

Features

• Multi-format: PDF, Word, PowerPoint, Excel/CSV, EPUB, HTML, RTF, email, images (+ legacy .doc/.ppt via LibreOffice)
• Cross-page image deduplication via xref tracking
• Image filters: solid-color / tiny-icon / dimension / reference-logo (perceptual hash)
• Sliding-window text chunking (~500 tokens, ~50 overlap) at sentence boundaries
• Tables serialized to Markdown; images embedded once and reused
• Parallel vision calls (vision_workers), batched embeddings
• Resume support — skip documents already in the store (skip_if_exists)
• Two-stage chunking: _chunks.json written automatically; pass split_chunks=True to also write flat individual chunk documents ready for Azure AI Search or any vector store
• build_index_document() — transforms a raw chunk into a flat, AI-Search-ready document (renames embedding → content_vector, flattens metadata)
• safe_index_key() — sanitizes any string to a valid Azure AI Search document key
• Fully typed — py.typed marker, compatible with mypy and pyright

Roadmap

• PDF / DOCX / PPTX / XLSX extraction
• EPUB / HTML / RTF / email extraction
• Azure OpenAI vision + embeddings integration
• Local vision models (Qwen2.5-VL, Llama 3.2, SmolVLM)
• Azure Blob Storage backend
• Sliding-window chunking with sentence-boundary awareness
• Smart image filtering (dimension, solid-color, logo dedup)
• Document-level metadata on every chunk (file_path, doc_id, last_updated, …)
• Individual chunk splitting — split_chunks=True or split_chunks_file() writes per-chunk documents for AI Search ingestion
• build_index_document() — flat AI-Search-optimized output with content_vector, flattened metadata
• Figure-caption association (link extracted images to their nearest caption)
• Table-of-contents aware chunking (respect heading hierarchy)
• multisense — companion RAG pipeline library built on multixtract

PRs and feature requests welcome via GitHub Issues.

Performance

Extracts a 50-page PDF in ~4 s and a 100-slide PPTX in ~0.14 s on a standard developer machine (no GPU, no API key). Chunking adds negligible overhead.

→ Full benchmark results and methodology

Documentation

Contributing

pip install -e ".[dev,pdf,docx,pptx,xlsx,epub,html,rtf]"
pytest
ruff check src tests
mypy src/multixtract --ignore-missing-imports --no-error-summary
python benchmarks/run_benchmarks.py

See CONTRIBUTING.md for guidelines. Bug reports and PRs are welcome.

Example Applications

• Internal RAG systems on Azure OpenAI
• Enterprise search over mixed document libraries
• Research document processing pipelines

Using multixtract in your project? Open a PR to add it here.

License

MIT — see LICENSE.