Cutting OCR Costs by 85% Without Losing Accuracy

This is a sample post so you can see how the blog renders. Replace the text in data/blog/cutting-ocr-costs.md with your own - the formatting below shows what's supported.

When you process millions of documents a month, OCR is not a line item you can ignore. Third-party OCR is billed per page, and at scale those pages add up fast. Here's the approach that let us cut OCR spend by roughly 85% while improving extraction accuracy.

1. Only OCR what you need

Most documents have a lot of dead space. Instead of sending the whole page to OCR, we crop to the regions of interest first:

• Detect the table or field block
• Crop tightly around it
• Send only that region downstream

Fewer pixels means faster, cheaper, and more accurate recognition.

2. Cache aggressively

The same templates show up again and again. A content-addressed cache keyed on a hash of the cropped image means we never pay to OCR the same thing twice.

key = hashlib.sha256(image_bytes).hexdigest()
if key in cache:
    return cache[key]

3. Let the LLM clean up, not read everything

OCR gets you raw text; an LLM is great at structuring it. Pushing the cleanup and classification step to a well-prompted model - rather than more OCR passes - is where the cost curve really bends.

The cheapest OCR call is the one you never make.

That's the whole trick: do less work, cache the rest, and reserve the expensive tools for the parts that actually need them.