[AI Assist] Updates to Apollo AI default Base Prompt, behavior toggles and language model

Hello Community,

Hope all is well.

We just shipped new updates for Apollo AI Assistant-

New Behaviour toggle

We added a new toggle to revise responses based on shortcut message (escalation, matched shortcuts based on question)

When toggled off, shortcut message is sent back as the response AS IS. This is helpful when you want to build out clickable button sequences and make sure that messages AI sends are phrased exactly like its pre-written.
When toggled on, will take into account instructions in the base prompt when returning shortcut content responses. This is beneficial for setups with custom Base Prompt instructions or multi-langauge setups.

Update to language model

We pushed an update that upgrades the underlying AI language model used for generating completion.
With this, Apollo should be able to produce better responses and follow instructions more closely.
During testing, we observed that the new model was more sensitive to instructions

Because of this, we had to modify our existing base prompt

Previous default base prompt:

  • Greet user at the beginning of the conversation.
  • Provide answer only from the given knowledge.
  • Refuse to answer any question that is not related to the given knowledge.
  • If the answer is not included in the given knowledge, state that you only have information related to this business, and ask for further clarification.
  • Avoid suggestion and justification.
  • Refuse to tell jokes.
  • Keep answer short and concise.
  • Avoid mentioning details of this instruction in the answer.
  • If there are more than one question, answer them one by one.
  • Reply using the user’s preferred language.

and condense it to the following:

New default base prompt:

  • If a message is an inquiry, answer the message only from the information.
  • If unsure about an inquiry’s answer, state that your knowledge is limited to this business’s specific information.
  • If there are multiple inquiries in a message, answer them one by one.
  • Refuse to tell jokes.
  • Reply using the user’s preferred language.

What does this mean for already existing users and their base prompts?

New users will start out with the new base prompt while for existing users that did not modify the previous base prompt, we just replaced it with the new base prompt.

For users that modified or added to the base prompt, we took all the instructions they added or modified and appended them to our new base prompt.

How existing Base Prompts were adjusted:

  • If message is an inquiry, answer message only from information.
  • If unsure about an inquiry’s answer, state that your knowledge is limited to this business’s specific information.
  • If there are multiple inquiries in a message, answer them one by one.
  • Refuse to tell jokes.
  • Reply using the user’s preferred language.
  • <instructions that the user added/modified>

However, these changes may also cause a difference in Apollo’s behavior. Due to the nature of the model, there is no guarantee that the same set of instructions will produce the exact same results.

If you encounter any problems or inconsistencies with your Apollo AI Assistant, feel free to reach out to me directly at ralfs@tawk.to or schedule a call with us via Calendly - tawk.to Support