Entity Detection

Entity Detection is a term that refers to using AI models to automatically extract certain named elements from a piece of text. It can be used to extract parameters from a command, extract important data points from an article such as the people, places and dates involved, or in many other use cases. 

As an example, say you provide a scheduling workflow, and your user types a message such as 'schedule a meeting with sales at 2pm on Friday'. Using entity detection, you can ask the AI to pick out the entities 'department' and 'date' and the task will extract 'sales' and '2pm on Friday'. 

To configure the Detect Entities, task, simply provide the text you want to extract entities from, and the name of the entities you want to extract. 

Text - the text you want to extract entities from

Entity Labels  - the entities you want to extract. These should be categories of data, like department, date, time, account number, person, etc.

AI Connection  - which AI connection to use. Paid users can configure their own Open AI account, and all users can use Flow XO Ai credits. How many credits are charged are based on the selected model.

Model - the Open AI model you want to use for the entity detection. You can choose from ChatGPT or GPT-4. ChatGPT will be the least expensive and the fastest (1 credit if using Flow XO AI Credits) and Gpt4 will be slower, more expensive, but also more accurate. It costs 30 Flow XO AI credits per request. The default is ChatGPT.

Here is how it looks:

If you aren't getting the results you expect, you may need to experiment with the labels and the models. You can see the raw response in your interaction logs:

For example, let's say in the example above, you chose 'date' and 'department' as your entity labels. This gave us the following results:

Notice that the time was not included, because the entity we asked for was date. To get the time, we had to use either "datetime" as the entity label, or have two labels, one for "Date" and one for "Time", which produced the following results in each case:

And when we split into date and time labels:

How you structure your labels will depend on how you plan to use the data, which labels produce the most accurate results, and what model (ChatGPT/GPT4) you are using. As with most things AI, experimentation and testing are required to fine tune the results to work on your actual data.

And that's it!

Happy Flowing!