Today we released a major update to our ChatGPT chatbot, the ability to hold 100% perfectly fluent conversations.
We've always had conversations, but they were clumsy, inconvenient, and didn't feel "natural". Today we can confidently claim we've got the same, if not better, conversational style chatbots than ChatGPT itself.
The way it works is that we've implemented "rolling context". This means that the context "folds" upwards as we spend all available tokens from OpenAI, eliminating messages from the beginning of the conversation, instead of the way we did previously where we would start an entirely new context if training data was found.
Then when we query our training data, we concatenate any newly found training data into the context. Not only does this perfectly implement fluent conversations, but it also allows the chatbot to associate and add additional training data as the user continues the conversation. Imagine the following conversation.
The point with the above is that the first question sets "the subject" of the conversation, which of course will trigger training data about our chatbot, perfectly answering the first question. The second question refers to the subject from the first question, which of course is the chatbot, and asks how the subject compares to the AI Search product.
Both of the above questions will trigger training data, and instead of throwing the previous training data away as we're asking follow up questions, the backend will keep the training data for both the chatbot question and the AI search question.
This creates a perfectly "fluent" conversation which is 100% "natural" and feels like a human being
Changes for you
You don't need to do anything to take advantage of this feature. All partner and client cloudlets have been automatically updated by our technology team. However, in our testing we have discovered that conversations typically flows better if you reduce your "Max context tokens" on your models. Previously for GPT4 we'd encourage most chatbot owners to use 2,000. The conversation will flow better if you reduce this to 1,000 for GPT4, and maybe 20% less for GPT3.5.
This gives you room for multiple "context snippets" making the chatbot better at conversing, since it will not evict previous training snippets before you've got a fairly long conversation with multiple training snippets. Below you can see an example of how such a conversation might proceed.
Then as I ask a follow up question, notice how it remembers my previous question, even though my follow up question triggers new training data.
Conversations on pair with OpenAI
All in all this provides equally good conversational quality as OpenAI delivers with ChatGPT - The difference of course being that our chatbot only answers questions from your training data, allowing you to create a completely custom chatbot, answering questions the way you want it to answer questions.
Yet again, if you've got your own private cloudlet, and you can't get good enough quality conversations, then try to experiment with a lower "Max context token" value.
In general this increases quality 10x on our chatbot, making it 100% perfectly impossible to determine if you're speaking to a human being or an AI chatbot