Optimizing extraction performance
To improve extraction performance, you can optimize:
- document performance
- document type performance
Note that the number of documents you submit for extraction has no noticeable effect on performance. Each document gets its own worker in parallel.
Document performance
In an ideal performance scenario, you extract data from digitally generated PDFs using text-based or coordinate-based Sensible methods, such as Label, Row, Region, Text Table, and Document Range.
In order of largest to smallest impact, these factors add seconds to the ideal document processing time:
Over 10 seconds per document
| Factor | Notes |
|---|---|
| whole-document OCR for document images | Sensible takes 10 seconds or more to OCR an entire document. You can speed OCR up for documents that are 25 pages and shorter by choosing Sensible’s Google OCR option for a document type. |
| whole-document table recognition | Avoid configuring Sensible to search a whole document for tables. Instead, configure a table stop. For examples, see any of the table methods. |
Under 5 seconds per document
| Factor | Notes |
|---|---|
| OCR preprocessor | Some documents mix digital text with text images, for example by embedding scanned pages in a digital PDF. Speed this up by OCRing select pages, not the whole document. For more information, see the OCR preprocessor. |
| Fixed Tables with stops,Nearest Checkbox method | Sensible processes Fixed Tables that include a Stop parameter in less than 5 seconds. Or, convert to the faster Text Table method, which skips table recognition. |
| LLM-based methods | For methods based on large language models (LLMs), see Sensible Instruct |
Under 1 second per document
Some Sensible methods use pixels, for example to recognize borders. Pixel recognition for PDF documents requires rendering a PDF page, which can take hundreds of milliseconds. To improve processing time, use coordinate-based alternatives to these methods.
| Factor | Notes |
|---|---|
| Box method | To improve processing speed, convert the more flexible Box method to the strictly coordinate-based Region method. |
| Signature method, Checkbox method,image coordinate extraction | These methods have no alternatives. See the following section for ways to avoid running these methods except when necessary. |
Document type performance
By default, Sensible runs all the configs in a document type before choosing the best one for a given document. If your document type contains configs with computationally expensive methods, you can improve performance by selectively running and skipping configs. Use fingerprints to test whether documents contain matching text before skipping or running configs. For more information, see fingerprint.