Since its release in November 2022, ChatGPT has been getting a lot of traction, which is great for AI and applications of Chat and QA (Question Answering) tools.
We wanted to highlight some of the similarities and differences between Cognistx’s SQUARE (Scalable Question Answering and Recommendation Engine) and ChatGPT from Open.ai. Our goal is to showcase these differences to you, to help you choose the next tool for your chat, QA, or search application.
SQUARE works on your corpus of documents versus all publicly available information. This ensures the results are specific to your field, domain and documents.
Overall the underlying technology and language models for both tools are similar. SQUARE utilizes an ensemble of models versus a good deep-learning model. SQUARE can have custom-trained models work hand in hand with state-of-the-art general-purpose models.
SQUARE links you back to the text or source document versus just coming up with an answer. It is based on extractive models that provide facts from a given set of documents. ChatGPT, on the other hand, is a generative model and sometimes comes up with answers that are made up and inaccurate.
SQUARE overall is a good fit for business domains where it's essential to point to the source document(s) versus generating the best possible answer.
Training SQUARE requires less data than training ChatGPT-like generative models, as SQUARE is based on extractive QA. Hence, less effort in the annotation process for organizations.
Generative models generate new answers for the same question every time they are executed. Even though they might generate multiple correct answers for the same question, different answers for the same questions create inconsistency.
Unlike ChatGPT, SQUARE provides evidence or facts for questions instead of generating answers. Generating answers can be harmful because often, these answers are convincing even though they are incorrect. For example, we tried generative models in the past for one of our customers. When we asked about the license cost, the model generated an answer of one million dollars with supporting statements. However, the correct answer was just $2,000. It's hard for users to validate if generated answers are correct.
Customers can choose how long the final answer should be with SQUARE - Does it need to be a short - exact answer? Does it need to be a long - complete sentence or paragraph, etc.?
We hope this clarifies the difference between SQUARE and ChatGPT. We look forward to conversing more with you.