Skip to main content
PaperJet uses LLMs in the extraction process. The size and capabilities of the LLM can vary greatly. If you want to get good results, especially with smaller models, you must limit the amount of “work” that the LLM must perform. This is a set of guidelines to help you get good results with your documents. We’ve found this to work well for both small inputs (images) as well as very large documents (500+ pages)

​
Field names are important

Field names are used to match the data from the source document. This means that using descriptive, self-explaining field names is crucial for good results. You don’t have to avoid spaces in field names, but it is recommended. Don’t try to be clever.
Good example: invoice_number
Bad example: the text from the 3rd row on the left side

​
Use descriptions to clarify intent

Descriptions act as an additional layer of instructions to help match data from the source document. You can add a description to any object, field or column to help the engine with extracting the correct data.
Good example: Full name of the individual
Occasionally, the data you would like to extract can have different names/labels on the source documents. This is a good use case for descriptions to handle - list out the names under which a field can appear in a source document
Field name: “Tax ID”
Description: “Can also appear as VAT ID or EIN number”

​
Use descriptions for data manipulation and filtering

You can also use the description to perform filtering and/or data manipulations, such as “Format the date in YYYY-MM-DD” or “Skip rows containing X”
Don’t overuse data manipulations, especially with smaller models
If possible, it’s best to do filtering in postprocessing, outside of PaperJet. See the next page for some examples.