Apppublic

mansourkama/document-extraction-pipeline

sourceHugging Faceupdated 1mo agoView on Hugging Face

0likes

App README

Document Chat & Extraction Pipeline

Talk to your documents. Upload a PDF or image, then ask anything in plain language — summarize, extract, restructure, translate, or export in any format. All models run free from HuggingFace, no API key required.

🚀 [Live demo on HuggingFace Spaces](https://huggingface.co/spaces/mansourkama/document-extraction-pipeline)

What you can do

Instruction example	What you get
"Summarize in 3 bullet points"	Concise summary
"Extract all dates and amounts as JSON"	Structured JSON object
"List all parties and their roles"	Named entity list
"Convert to a markdown table"	Formatted table
"What are the payment terms?"	Direct answer
"Translate the key fields to English"	Translated output

Plus template-based structured extraction (Invoice, Contract, Receipt, Personal Form) that returns a validated Pydantic model.

Architecture

Document (PDF or image)
        │
        ▼
┌───────────────┐
│  OCR Engine   │   doctr — db_resnet50 (detection) + crnn_vgg16_bn (recognition)
│  (src/ocr.py) │   Pretrained weights from HuggingFace, no account needed
└───────┬───────┘
        │  raw text
        ▼
┌──────────────────────┐
│  LLM                 │   Qwen2.5-1.5B-Instruct (HuggingFace)
│  (src/extractor.py)  │   ├─ chat()    — free-form instruction → any output format
│                      │   └─ extract() — schema-guided → validated Pydantic model
└──────────────────────┘

Project Structure

document-extraction-pipeline/
│
├── src/
│   ├── ocr.py                   # OCR engine (doctr)
│   ├── extractor.py             # LLM: chat() + extract() methods
│   ├── pipeline.py              # Orchestrates OCR + LLM; exposes chat_file(), process()
│   └── templates/
│       ├── base.py              # ExtractionTemplate base class (schema_prompt + type coercion)
│       ├── invoice.py           # Invoice / billing
│       ├── contract.py          # Contract / mission order
│       ├── receipt.py           # Receipt / expense
│       └── form.py              # Personal information form
│
├── app.py                       # Gradio web interface (HuggingFace Spaces)
├── tests/
│   ├── fixtures/generate_fixtures.py   # Synthetic test images (Pillow)
│   └── test_pipeline.py               # Unit + integration tests
├── examples/demo.py             # CLI demo
└── requirements.txt

Quickstart

bash

pip install -r requirements.txt

Free-form chat:

python

from src.pipeline import DocumentPipeline

pipeline = DocumentPipeline()

# From a file (OCR + LLM)
response = pipeline.chat_file("contract.pdf", "List all parties and the total value.")
print(response)

# From text (LLM only)
response = pipeline.chat(my_text, "Convert to a markdown table.")
print(response)

Structured extraction:

python

from src.pipeline import DocumentPipeline
from src.templates import InvoiceTemplate

pipeline = DocumentPipeline()
result = pipeline.process("invoice.pdf", InvoiceTemplate)
print(result.model_dump_json(indent=2))

json

{
  "invoice_number": "INV-2024-0042",
  "date": "2024-11-15",
  "vendor_name": "TechSolutions SARL",
  "total_amount": 8160.0,
  "currency": "EUR"
}

Built-in Templates

Template	Document type	Key fields
`InvoiceTemplate`	Invoices, billing	invoicenumber, date, vendor, client, subtotal, tax, total, currency, duedate
`ContractTemplate`	Contracts, mission orders	parties, missiondescription, startdate, enddate, totalvalue, payment_terms
`ReceiptTemplate`	Receipts, expense slips	merchant, date, total, currency, payment_method, items
`PersonalFormTemplate`	Registration, KYC forms	fullname, dateof_birth, email, phone, address, nationality, occupation

Custom template:

python

from pydantic import Field
from typing import Optional
from src.templates.base import ExtractionTemplate

class MedicalReportTemplate(ExtractionTemplate):
    patient_name: Optional[str] = Field(None, description="Patient full name")
    diagnosis: Optional[str] = Field(None, description="Main diagnosis")
    physician: Optional[str] = Field(None, description="Treating physician name")
    date: Optional[str] = Field(None, description="Report date")

result = pipeline.process("report.pdf", MedicalReportTemplate)

Run Tests

bash

python tests/fixtures/generate_fixtures.py   # generate synthetic test images
pytest tests/ -v

Configuration

python

pipeline = DocumentPipeline(
    ocr_det_arch="db_resnet50",
    ocr_reco_arch="crnn_vgg16_bn",
    llm_model="Qwen/Qwen2.5-1.5B-Instruct",   # swap for any HF text-gen model
)

Stack

Component	Library
Document OCR	doctr (HuggingFace pretrained)
Language model	Qwen2.5-1.5B-Instruct
Schema validation	Pydantic v2
Web interface	Gradio · deployed on HuggingFace Spaces
Test fixtures	Pillow