Large and in Charge
Innovations in Large Language Models (LLMs) have changed the landscape of artificial intelligence. Companies like OpenAI and their GPT models have developed LLM-based technologies, like ChatGPT, that are highly effective at natural language processing. Being fully committed to advancing our capabilities through new AI integrations, Grooper further expands LLM functionality to enable users to extract data faster and easier than ever before.
Grooper currently supports two types of LLMs, OpenAI-compatible and Azure-deployed.
This allows for thousands of different models to be utilized within Grooper. While OpenAI has a handful of models, most open-source LLMs use an OpenAI-compatible API to enable usage with minimal effort. This means Grooper can talk to all these different LLMs. Learn more about alternative LLM deployment methods with Grooper here.
Similarly, Azure allows access to more than a thousand LLM models through their Azure AI Studio. Here, models can be deployed in several ways and are accessible to Grooper via the endpoints and API keys provided by Microsoft.
What this means for you, the Grooper user, is that you can bring the power of your favorite LLM directly to Grooper and your documents and data.
LLMs for Extraction
First and foremost, in Grooper 2024, there are new and improved ways to extract data from your documents using LLMs. Because LLMs are so good at processing natural language, these new extraction techniques are set up in a fraction of the time of traditional extractors in Grooper.
Don’t Prompt Me, Bro
We have removed the need to design prompts within Grooper – we build the prompts for you – you don’t need to understand “prompt engineering” anymore. Just give your fields meaningful names and an optional description.
Grooper will do the rest.
Here’s a sample of the new LLM constructs in Grooper:
AI Extract: A “Fill Method” to extract a Grooper object with virtually no configuration. It can be used in conjunction with traditional extraction techniques and can be used at the data model, section, or table level.
LLM Table Extract: A new tabular extraction method that extracts data from tables quickly with little setup.
Clause Detection: A new Data Section extraction method designed to find clauses of a specific type in unstructured documents.
Ask AI: This extractor type replaces the deprecated “GPT Complete” extractor with new functionality that allows table extraction and JSON parsing using responses from LLMs.
The Fine (Tuned) Print
The next new LLM feature we’re most excited about around the office is the fine-tuning of LLMs with Grooper.
Without Grooper, you must prepare your fine-tuned data manually if you want to create a fine-tuned OpenAI LLM model. This means you’d have to write the prompts and proper responses, then format them in the very specific “JSONL” file format and upload them to OpenAI to create your fine-tuned model.
You do have a spare data scientist sitting around, don’t you? We didn’t think so. There’s a better way now with Grooper.
Let’s say you want to train your invoice model from a batch of invoices. Jeez, you bought Grooper to eliminate manual data entry, right?
We hear you; that’s why we created a simple way to create fine-tuned data from ANY Grooper batch. Grooper now creates the fine-tuning file OpenAI needs to create fine-tuned models with the simple click of a button.
Why is this important? Fine-tuning LLMs allows you to use a cheaper, fine-tuned model and get better extraction accuracy for significantly less cost than the best “generic” model.
As of this writing, we’re seeing a cost for the fine-tuned model that is 1/10th that of GPT-4o and 60% more accurate for extraction.
Grooper can keep the fine-tuning data files around, so as you find a better base model, you can add your training data to your new model with a simple click of the mouse.
Fine-tuning does not happen automatically. As with everything in Grooper, an administrator can choose what gets submitted and which batches are used to create fine-tuning data.
With the process in place, you can now have a system that “gets better over time.” Read more about fine-tuning here.
Grooper’s all the RAG(e)
Stay tuned – Grooper is implementing RAG (Retrieval-Augmented Generation)! Imagine searching across all your data – document repositories, file directories, AND databases with one natural language search!
