Tune Instructions
How to change Aissistant behavior by change of instructions and contexts.
Last updated
How to change Aissistant behavior by change of instructions and contexts.
Last updated
Aissistant's behavior is driven by the instructions and contexts provided. There are three levels of instructions and contexts, Global, Session, and Interaction, that can be used to efficiently fine-tune the behavior of Aissistant.
Below is a summary of three levels of contexts & instructions
Global
Workspace
Always
Throughout entire session
High level, global, and important guidance that applies to all circumstance.
Session (Contextual)
Smart Taggers
Triggered with smart tag
Once triggered, throughout entire session
Related to specific topic, and important to be sticky to the session once the topic is identified. For example, some workflow, procedure, or guidance to handle a specific topic.
Interaction (Transient)
Assets
Triggered with current user input
Only the current interaction
e.g., FAQ (Frequently Asked Questions). It can also be used to provide information on workflow, procedures, or guidance, especially when they are hard to describe in a simple format, for example, diagnosis.
In general, instruction can be used to change the response, however, can not be used to retrieve information (this can be done through smart action) or output tag / comment / action (this is done through smart tagging or smart action)
Dictate the format, length or tone of the response
Create tag (this can be done via Smart Tagger)
Direct specific response under provided circumstances
Retrieve information from URL or Database (this can be done via Smart Action)
Demand AI to pay attention to specific information when compose response
Create summary or comment in the Agent Platform (this can be done via Smart Tagger)
Ask AI to follow described logic or sequence in diagnosis
Assign the conversation / ticket to a team or user (not supported)
Provide guidelines for AI response. (specific ones work better)
Outbound actions like send email / notification or change the database. (some of this can be done via Smart Action)
Context, whether in Global, Tag or Asset, is usually used to provide contextual information for AI to generate the best informed response.
Global instructions and contexts are defined in the Workspace Configuration, which take effect in all interactions throughout the session. Those instructions and contexts are supposed to be abstract, high-level, and/or important information for AI, e.g., instructions like "be succinct and no more than 160 characters in response", "never confirm the appointment in the conversation, instead tell users that you will let the team know and someone will reach out for confirmation.", etc.
Other special type of instructions and contexts include business name, identity, purpose, and tasks, etc, which are also in the workspace configuration. Those special instructions also can substantially change the behavior of Aissistant in specific circumstance.
Session instructions and contexts are defined in tagger which often represents specific topics or contexts. Therefore the instructions and contexts are also called "contextual instruction" and "contextual context" respectively, which are only triggered when the conversation hit specific tag (equivalent to topic of context).
Comparing with Global ones, session instructions and contexts are often more effective because it reduces chance of confusion to AI instead of providing a massive amount of information, some of which might be conflicting, to AI as Global instructions do. For context specific instruction or information, we highly recommend to organize them in the tags.
Session instructions and contexts, once triggered, stay throughout the entire session. This is the major difference between the information in Session instruction & context vs. Interaction instruction & context, which only exists in the current interaction.
Interaction instruction and contexts are defined in assets which can be website, Google Doc and others. Those information are triggered by the current interaction and only exist in the lifecycle of current interaction, therefore transient.
Please refer to "Turn assets into AI" for details on how to use assets to provide information to AI.
Information retrieved from "Smart Action" is also considered as interaction contexts.
