Skip to content
Nappai Documentation

Entities extraction

Entities extraction reads the text of a record with the help of an AI model and looks for the values you ask for: customer names, amounts, dates, products, clauses or any other concept you can describe. What it finds is added to the same record, so the next steps of the flow can use it.

If the text is long, the component splits it into pieces, analyses all of them and joins the results without repeating values.

When to use it

  • You want to get the parties, the signing date and the term out of a contract.
  • You receive customer emails and need to identify the product and order number they mention.
  • You review CVs and want to extract technologies, languages and years of experience.
  • You read news or reports and want to list the companies and people that appear.

Configuration

FieldWhat it does
Data (required)The record or list of records whose text you want to analyse.
Extract keys (required)List of what you want to extract, one concept per line. For example: Customer name, Total amount, Delivery date. You can describe each concept in more detail if needed.
Additional instructionsExtra guidance for the extraction, for example “Dates are written day/month/year” or “Ignore the email signature”.
Output instructionsHow you want to receive each value found. If you leave it empty, each value comes with its name, a score from 1 to 100 and a short justification.
Model (required)The language model that does the extraction. Connect a model component here, for example OpenAI.

Advanced options

FieldWhat it does
chunk sizeSize, in characters, of each piece a long text is split into. Default: 1500.
chunk overlapCharacters shared by two consecutive pieces, so values that fall right on the cut are not lost. Default: 150.

Output

The Extracted Data output returns the same records you sent, each with a new field, entities, holding the list of values found. You can connect it to Parse Data to turn it into text, to a filter or to a component that stores the information.

Example

A legal team wants a quick summary of every contract that comes in.

  1. Connect the contract text to Data.
  2. In Extract keys, type Contract parties, Signing date, Term and Penalties.
  3. In Additional instructions, type “If a value does not appear, do not make it up”.
  4. Connect a language model to Model and the Extracted Data output to a component that saves the result in a spreadsheet.

Good practices

  • Name concepts precisely: “Total amount including VAT” works better than “money”.
  • If you write your own Output instructions, ask for each value to include an entity field with its name: the component uses it to avoid repeated values.
  • Each piece is one request to the model. With very long texts, a larger chunk size means fewer requests.