openingestion 0.1.4

RAG ingestion pipeline — Chef → Chunker → Refinery → Porter

openingestion

Modular RAG ingestion pipeline — from raw documents to retrieval-ready chunks.

Fetcher → Chef → Chunker → Refinery → Porter

Version: 0.1.4 · Python: 3.10 – 3.13 · License: MIT

---

Overview

openingestion orchestrates the full journey from raw documents to enriched, retrieval-ready chunks through five composable stages:

Stage	Classes	Input → Output
Fetcher	LocalFileFetcher, WebFetcher, SharepointFetcher	Source → FetchedDocument[]
Chef	MinerUChef, DoclingChef	File/dir → ContentBlock[]
Chunker	TokenChunker, SentenceChunker, SemanticChunker, SlumberChunker, BlockChunker, PageChunker, SectionChunker, RecursiveChunker	ContentBlock[] → RagChunk[]
Refinery	RagRefinery, ContextualRagRefinery, VisionRefinery	RagChunk[] → enriched RagChunk[]
Porter	JSONPorter, to_dicts, to_langchain, to_llamaindex	RagChunk[] → target format

---

Installation

From PyPI

pip install openingestion

The base install (loguru + chonkie-core) is intentionally minimal — no heavy ML dependencies. Parsers, tokenizers, and advanced chunkers are opt-in extras.

From source (editable)

git clone https://github.com/Isopope/openIngestion.git
cd openIngestion
pip install -e .

Windows / PowerShell

openingestion requires Python 3.10 – 3.13 (3.14+ not yet supported due to MinerU).

py -3.13 -m venv .venv
. .\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
python -m pip install -e .

---

Optional extras

Extra	Installs	Use case
mineru	mineru[pipeline]==3.1.0	GPU-accelerated PDF, Office, and image parsing (full layout analysis)
docling	docling	CPU-only PDF parsing (IBM Docling, no GPU required)
semantic	sentence-transformers, numpy, scipy	SemanticChunker — embedding-based splitting
slumber	openai, pydantic, tenacity, tqdm	SlumberChunker + OpenAIGenie — LLM-guided chunking
tiktoken	tiktoken	Exact OpenAI tokenizer (cl100k_base, o200k_base …)
hf-tokenizers	tokenizers	Fast HuggingFace tokenizers (Rust, BPE/WordPiece)
transformers	transformers	HuggingFace AutoTokenizer (full model loading)
langchain	langchain-core	output_format="langchain"
llamaindex	llama-index-core	output_format="llamaindex"
web	playwright	WebFetcher — render websites to PDF/HTML
sharepoint	msal, office365-rest-python-client	SharepointFetcher — Microsoft 365 / SharePoint

Convenience bundles:

# CPU pipeline (Docling + semantic + tiktoken)
pip install -e ".[cpu]"

# GPU pipeline (MinerU + semantic + tiktoken)
pip install -e ".[mineru,gpu]"

# Everything
pip install -e ".[all]"

Individual extras:

pip install -e ".[mineru]"          # MinerU parser (GPU recommended)
pip install -e ".[docling]"         # Docling parser (CPU)
pip install -e ".[semantic]"        # SemanticChunker
pip install -e ".[slumber]"         # SlumberChunker + OpenAI
pip install -e ".[web]"             # WebFetcher (then: playwright install chromium)
pip install -e ".[sharepoint]"      # SharepointFetcher

---

Quick start

from openingestion import ingest, ingest_from_output, ingest_from_json

# Parse a raw PDF with MinerU (requires [mineru] extra)
chunks = ingest("rapport.pdf")

# Parse native Office documents with MinerU 3.1+
chunks = ingest("support.docx")
chunks = ingest("deck.pptx")
chunks = ingest("tableau.xlsx")

# Skip re-parsing — reuse an existing MinerU output directory
chunks = ingest_from_output("./output/rapport/auto/")

# Load directly from a content_list.json
chunks = ingest_from_json("./output/rapport/auto/rapport_content_list.json")

# Use Docling instead of MinerU (CPU, no GPU needed)
chunks = ingest("rapport.pdf", parser="docling", strategy="by_sentence")

# Full control
chunks = ingest(
    "rapport.pdf",
    parser="mineru",                # or "docling"
    strategy="by_token",            # by_block | by_token | by_sentence | by_semantic | by_slumber
    max_tokens=512,
    overlap_tokens=64,
    image_mode="path",              # path | base64 | skip | ignore
    infer_captions=True,
    output_format="chunks",         # chunks | dicts | langchain | llamaindex
)

# Export to LangChain Documents
docs = ingest("rapport.pdf", output_format="langchain")

# Export to LlamaIndex TextNodes
nodes = ingest("rapport.pdf", output_format="llamaindex")

---

Fetchers

from openingestion.fetcher import LocalFileFetcher, WebFetcher, SharepointFetcher

# Local filesystem
fetcher = LocalFileFetcher(ext=[".pdf"])
docs = fetcher(dir="./inputs/")

# Website → PDF (requires [web] extra + playwright install chromium)
fetcher = WebFetcher(output_dir="./downloads/", mode="pdf")
docs = fetcher.fetch(urls=["https://example.com"])

# SharePoint / Microsoft 365 (requires [sharepoint] extra)
fetcher = SharepointFetcher(
    client_id="...", client_secret="...", tenant_id="...",
    output_dir="./downloads/",
)
docs = fetcher.fetch(site_url="https://tenant.sharepoint.com/sites/MySite")

---

Refineries

from openingestion.refinery import RagRefinery, ContextualRagRefinery, VisionRefinery
from openingestion.genie import OpenAIGenie

# Standard enrichment: token count, content hash, image paths
refinery = RagRefinery(output_dir="./output/doc/auto/", image_mode="path")
chunks = refinery.enrich(chunks)

# Contextual RAG: LLM-generated doc summary + per-chunk context
genie = OpenAIGenie(model="gpt-4o-mini", api_key="sk-...")
ctx_refinery = ContextualRagRefinery(genie=genie, generate_doc_summary=True)
chunks = ctx_refinery.enrich(chunks)

# Vision: extract text from scanned tables / images via GPT-4o
vision_refinery = VisionRefinery(genie=genie, only_if_empty=True)
chunks = vision_refinery.enrich(chunks)

---

Export

from openingestion.porter import JSONPorter

# JSONL (one chunk per line)
JSONPorter(lines=True)(chunks, file="output.jsonl")

# Pretty JSON array
JSONPorter(lines=False, indent=2)(chunks, file="output.json")

---

Architecture

Each stage follows a uniform Abstract Base Class pattern:

• Abstract method: process() / chunk() / enrich() / export()
• Batch processing: process_batch() / chunk_batch() / enrich_batch() / export_batch()
• Callable shortcut: instance(input) == instance.main_method(input)
• Unified logging via loguru

The three core data models flow through the whole pipeline:

FetchedDocument  →  ContentBlock  →  RagChunk
   (Fetcher)          (Chef)       (Chunker + Refinery)

BlockKind (TEXT, TITLE, TABLE, IMAGE, LIST, EQUATION, DISCARDED) is preserved from Chef through to the final export.

---

License

MIT — see LICENSE.

Note on optional dependencies: the [mineru] extra installs MinerU which is licensed under AGPL-3.0. Its licence terms apply when that extra is installed.

See specv3.md for full technical specifications.