Extract Tables to CSV

Convert bordered and borderless tables in PDFs and scans into clean CSV files with pandas.

Enable bbox=true to get cell positions for every block
Sort table blocks into rows and columns by coordinates
Export a validated CSV with pandas

Time to complete: ~15 minutes

Prerequisites

A SAGEA_API_KEY exported as an environment variable
A document with tables in PDF, PNG, JPG, TIFF, or WEBP format, max 100MB, ideally 300 DPI
Python 3.9+ with requests and pandas installed
Set up document processing
ARVA OCR model card

Step 1: Request OCR with bounding boxes

Set bbox=true and output_format="json" so each table block includes coordinates.

curl -X POST https://api.sagea.space/v1/ocr/process \
  -H "Authorization: Bearer $SAGEA_API_KEY" \
  -F model="arva-ocr" \
  -F [email protected] \
  -F output_format="json" \
  -F bbox=true \
  -o tables.json

Response shape:

{"text": "...", "pages": 2, "confidence": 0.992, "blocks": [{"type": "table", "page": 1, "bbox": [x0,y0,x1,y1], "confidence": 0.99, "text": "..."}], "model_used": "arva-ocr"}

Complex pages with tables cost $2.50 per 1000 pages. Accuracy is 99.2% on clean 300 DPI scans.

Step 2: Reconstruct rows and columns

Group table blocks by vertical position, then sort by horizontal position.

curl -X POST https://api.sagea.space/v1/ocr/process \
  -H "Authorization: Bearer $SAGEA_API_KEY" \
  -F model="arva-ocr" \
  -F [email protected] \
  -F output_format="text"

Step 3: Export to CSV with pandas

Normalize column counts and write a clean CSV file.

import pandas as pd
 
grid = [[c["text"].strip() for c in r] for r in rows]
width = max(len(r) for r in grid)
grid = [r + [""] * (width - len(r)) for r in grid]
 
df = pd.DataFrame(grid[1:], columns=grid[0])
print(df.head())
df.to_csv("extracted_table.csv", index=False)
print("wrote extracted_table.csv with shape", df.shape)

Verify

Error	Cause	Fix
`401 Unauthorized`	Missing or invalid API key	Export a valid key as `SAGEA_API_KEY` and retry
`400 unsupported_format`	File type not in PDF, PNG, JPG, TIFF, WEBP	Convert the file to PDF or PNG and retry
`413 Payload Too Large`	File is over 100MB	Compress or split the file, then retry
`429 rate_limit_exceeded`	Too many requests at once	Back off with retries and reduce concurrency