coreason-refinery

coreason-refinery is the industrial processing plant for data ingestion at CoReason-AI. It transforms raw documents (PDF, Excel, PPTX) into semantically structured, machine-readable Markdown, ready for RAG (Retrieval-Augmented Generation) pipelines.

Unlike traditional RAG tools that blindly chunk text, coreason-refinery is Structure-Aware. It preserves the logical hierarchy of documents (headers, sections) and ensures complex artifacts like tables and equations are kept intact.

Features

• Multi-Modal Parsing:
  • PDF: Vision-first parsing to preserve table grids and layout.
  • Excel/CSV: Treated as relational data, converted to Markdown tables with header preservation.
  • PowerPoint: Flattens slides into linear narratives (planned support for speaker notes).
• Semantic Segmentation:
  • Splits documents by logical headers (#, ##) rather than character counts.
  • Table Rescue: Never splits tables mid-row; merges tables spanning multiple pages.
• Context Injection ("Breadcrumbs"):
  • Enriches every chunk with its hierarchical path (e.g., Context: Protocol > Section 4 > Toxicity).
  • Ensures LLMs understand the specific scope of any given text snippet.
• GxP Compliance Ready:
  • Designed for lineage tracking and metadata enrichment.

Installation

pip install coreason-refinery

Note: This package requires Python 3.12+.

Usage

Here is how to initialize and run a refinery job:

import uuid
from coreason_refinery.models import IngestionJob, IngestionConfig
from coreason_refinery.pipeline import RefineryPipeline

# 1. Configure the Job
config = IngestionConfig(
    chunk_strategy="HEADER",
    segment_len=500
)

job = IngestionJob(
    id=uuid.uuid4(),
    source_file_path="path/to/document.pdf",
    file_type="auto",  # Infers PDF/Excel/CSV
    config=config,
    status="PROCESSING"
)

# 2. Run the Pipeline
pipeline = RefineryPipeline()
chunks = pipeline.process(job)

# 3. Inspect Results
for chunk in chunks:
    print(f"--- Chunk ID: {chunk.id} ---")
    print(chunk.text)
    print(f"Metadata: {chunk.metadata}")

Architecture

The pipeline consists of three main stages:

1. The Cracker (Parsing): Routes files to specialized parsers (UnstructuredPdfParser, ExcelParser) to extract atomic elements.
2. The Cutter (Segmentation): Reassembles elements into RefinedChunks based on document structure, applying "Rolling Context" to preserve hierarchy.
3. The Enricher (Metadata): (Planned) Adds lineage and semantic tags.

License

This software is proprietary and dual-licensed under the Prosperity Public License 3.0. Commercial use beyond a 30-day trial requires a separate license. See LICENSE for details.