We tried a self hosted OCR model a few years ago, but the quality and speed wasn’t great. From experience, it’s usually better to reverse engineer the APIs. The more complicated they are, the less they change. So it can sometimes be painful to set up the scrapers, but once they work, they tend to be more stable than other methods.
Data pollution is real. Also location specific results, personalized results, A/B testing, and my favorite, badly implemented websites are real as well.
When you encounter this, you can try scraping the data from different locations, with various tokens, cookies, referrers etc. and often you can find a pattern to make the data consistent. Websites hate scraping, but they hate showing wrong data to human users even more. So if you resemble a legit user, you’ll most likely get correct data. But of course, there are exceptions.
Comments
We tried a self hosted OCR model a few years ago, but the quality and speed wasn’t great. From experience, it’s usually better to reverse engineer the APIs. The more complicated they are, the less they change. So it can sometimes be painful to set up the scrapers, but once they work, they tend to be more stable than other methods.
Data pollution is real. Also location specific results, personalized results, A/B testing, and my favorite, badly implemented websites are real as well.
When you encounter this, you can try scraping the data from different locations, with various tokens, cookies, referrers etc. and often you can find a pattern to make the data consistent. Websites hate scraping, but they hate showing wrong data to human users even more. So if you resemble a legit user, you’ll most likely get correct data. But of course, there are exceptions.