The Source of Truth for Enterprise ChatGPT

OpenAI explains that one of the limitations of ChatGPT that is very challenging to overcome (and critical to fix so it can be used within the enterprise) relates to the underlying GPT 3.5 series of models on which ChatGPT is based on having no ‘source of truth.’ This fundamental problem gives rise to the phenomenon where ChatGPT sometimes writes plausible-sounding but incorrect or nonsensical answers.

Getting the Facts Straight

When asked, ChatGPT responds to a question about the range of a Tesla Model 3 as follows:

“The range of a Tesla Model 3 varies depending on the model and battery type. The Long Range model has an EPA-estimated range of up to 310 miles, while the Standard Range Plus model has an EPA-estimated range of up to 250 miles."

Very nice, you think, if only it were true. Per Tesla’s website, the Tesla Model 3 comes in 3 variants (not 2), and the (EPA est.) ranges are:

So, wrong all around.? We need to solve this ‘source of truth’ problem if ChatGPT is to be helpful within the Enterprise.

ChatGPT + Enterprise Search

We are starting to see articles being published about ChapGPT taking over search in its entirety. ? On Jan 3, 2023, Aaron Holmes wrote an article published in The Information? about Microsoft and OpenAI collaborating in an effort to challenge Google.

While this may be a good approach for searching the web, we need something more specialized and focused for the enterprise. At Colibri, we leverage Enterprise Search to provide the facts and data from the enterprise's own data sources (Website, Wiki, KBs, Product Specs, Community, etc.) to feed into ChatGPT.

Here’s How it Works

During a conversation between a salesperson and a prospective customer:

1. Colibri transcribes the customer’s conversation and identifies that they have asked a question
2. Colibri searches Enterprise KBs/Wikis (or articles on the web) for the top 10 results
3. Results are ranked using the classic BERT-based sentence-transformer
4. Question and search results are sent to Generative AI/ChatGPT for summarization
5. The answer is fed back to the salesperson as a cue card for them to read back to the customer while still on the call

Smooth Sounding Facts & Data

The results have been very impressive, with ChatGPT now able to regurgitate answers/facts in its usual smooth and readable output. Check out the following actual generated output using OpenAI’s ChatGPT accessed via their open source Python Package:

“The Tesla Model 3 is available in three different versions, each with its own range. The Rear-Wheel Drive version has a range of 272 miles, the Performance version has a range of 315 miles, and the Long Range AWD version has a range of 358 miles.”

The answer is nicely worded and includes all models, with accurate range information. This is the type of outpu that enterprises are going to need to drive business benefits from Generative AI/ChatGPT.

Let’s Build Together

If your company is contemplating how it can derive business benefits from Generative AI and/or ChatGPT, talk to us here at Colibri.ai, where we can deliver incredible solutions that leverage the best that ChatGPT offers by combining it with accurate enterprise/CRM data.??