Azure Document Intelligence Table Extraction (Python)

Extract tables from PDF documents using Azure Document Intelligence (formerly Form Recognizer) prebuilt-layout model.

Features

• Uses prebuilt-layout model to detect tables
• Exports tables to CSV and/or JSON
• Pretty console rendering of tables
• Environment-based configuration

Prerequisites

1. Azure subscription
2. Azure Document Intelligence resource (Cognitive Services / Document Intelligence) with an endpoint + key.
3. Python 3.10+

Provision (Azure CLI)

If you have not created a resource yet:

az group create -n my-docint-rg -l eastus
az cognitiveservices account create `
  -n my-docint-resource `
  -g my-docint-rg `
  -l eastus `
  --kind FormRecognizer `
  --sku S0 `
  --yes
az cognitiveservices account keys list -n my-docint-resource -g my-docint-rg

Take the endpoint from:

az cognitiveservices account show -n my-docint-resource -g my-docint-rg --query properties.endpoint -o tsv

Setup

Clone / copy this folder then:

python -m venv .venv
. .venv/Scripts/Activate.ps1
pip install -r requirements.txt

Create .env from template:

Copy-Item .env.example .env

Fill in values:

AZURE_DOCUMENTINTELLIGENCE_ENDPOINT=https://<your-resource>.cognitiveservices.azure.com/
AZURE_DOCUMENTINTELLIGENCE_KEY=<key>

Usage

python .\src\di_extract_tables.py .\sample.pdf --out output --format both

Options:

• --format csv|json|both (default both)
• --no-pretty skip console table rendering

Outputs:

• output/<pdfstem>_table{n}.csv for each table
• output/<pdfstem>_tables.json aggregated structure (if json enabled)

JSON Structure

[
  {
    "table_index": 0,
    "row_count": 5,
    "column_count": 4,
    "cells": [["A1", "B1", ...], ["A2", ...]]
  }
]

Notes / Best Practices

• For semantic elements (headings, paragraphs) or styles, still use prebuilt-layout.
• For invoices, receipts, IDs: switch to specific prebuilt model (e.g. prebuilt-invoice).
• Large PDFs: consider paging or splitting before processing to reduce latency.
• Handle rate limiting: catch HttpResponseError with status 429 and backoff.

Next Steps

• Add unit tests (mock client) for table normalization
• Add option to merge multi-page tables (requires analyzing cell bounding regions & spans)
• Integrate into a larger pipeline (e.g., Azure Functions or batch job)

Troubleshooting

Issue	Cause	Fix
Auth error 401	Wrong key	Regenerate keys in portal/CLI
Empty tables	Document has no detectable grid	Verify visually / adjust source PDF
Mixed languages	Layout model returns raw text only	Use OCR languages options if/when exposed

License

Internal / sample usage. Adapt as needed.