Extract with AI
The Extract with AI step evaluates data sent to it, then uses the GPT AI to extract whatever pieces of information you need. By naming the new columns that you’d like to populate, you can tell this step which pieces of information you’d like to extract.
At launch, you can use Parabola AI steps at no extra charge to your team. After a beta period, we’ll move to a usage-based model for Parabola AI. (You will not be charged retroactively for usage during the beta period.)
Examples of extracting data with AI
- Processing a list of invoices and extracting the invoice amount, due date, sender, and more
- Taking a list of email addresses and extracting the domain (e.g., gmail.com)
- Taking inbound emails and extracting the sender, company, and request type
As you can tell from some of these examples, the AI can do some lightweight interpretation (e.g., naming a company from an email domain URL) as well as simple data extraction.
How to use this step
Selecting what to evaluate
You start by selecting which columns you want the AI to evaluate to produce a result.
- All columns: the AI looks at every data column to find and extract the item it’s looking for
- These columns: choose which column(s) the AI should try to extract data from
- All columns except: the AI looks at all columns except the ones you define
Note that even when the AI is looking at multiple (or all) columns, it’s still only evaluating and generating a result per row.
Identifying what to extract
The next part of this step serves two purposes simultaneously:
- Telling the AI what items you’d like to extract from the input data (e.g., 'full name')
- Naming the new column(s) that the extracted data will go into (e.g., a column named 'full name')
The step starts out with three blank fields; you can fill those and even add additional columns to extract data to. Don’t need three? The step will automatically remove any blank ones, or you can remove them yourself.
(You can always rename or trim these columns later using other Parabola steps.)
Fine tuning
Open the 'Fine tuning' drawer to see extra configuration options. Using this field, you can provide additional context or explanation to help the AI deliver the result you want.
For example, if the AI was having trouble pulling 'Invoice number' from imported invoice data, you might explain to it:
“Our invoice numbers tend to begin in 96 and are 12-15 digits long.”
The AI would then better understand what you want to extract.
Helpful tips
- Currently, the AI can only run a few thousand rows at once. Choose and trim your data accordingly
- Sometimes you’ll see a response or error back instead of a result. Those responses are often generated by the AI, and can help you modify the prompt to get what you need.
- Still having trouble getting the response you expect? Often, adding more context in the 'Fine tuning' section fixes the problem.
Working with AI in Parabola
With our Artificial Intelligence (AI) steps, Parabola lets you process your data with OpenAI’s GPT AI in specific, useful, and reliable ways. But working with AI comes with important considerations.
AI has natural limitations
AI is a new field in technology, and while the results are sometimes exciting, they’re often less dependable than traditional human-built data processes. Consider reading OpenAI’s breakdown of their AI’s limitations, and keep the following in mind when using an AI to process data:
- Model limitations: Understand GPT's knowledge cutoff. GPT can “hallucinate” or be confidently incorrect, so do not expect results to be perfectly accurate all of the time.
- Data sharing: When data is processed using Parabola’s AI steps, that data is sent to OpenAI, a 3rd party. Review their policies and practices to understand how they handle your data.
- Monitoring: Continuously assess GPT performance; take corrective actions as needed.
- Responsible use: Adhere to regulations; inform stakeholders of limitations and risks.
AI processing is … different
We’ve made data processing with an AI easier than ever before! But when you use AI as part of a Parabola Flow, those steps can be less transparent and reliable than the rules-based transform steps that your Parabola Flows normally use.
Keep this in mind especially when automating processes where exact precision is critical, like financial data. Consider using AI for steps that require “interpretation” — which AI can be quite good at! — rather than precise calculation.
Feedback feeds us!
If you have feedback about the usefulness of these steps, or the AI-generated responses you’re getting from them, please tell us!