Resources can be added to Cogniti agents to ground the agent’s responses in information that you provide.
It’s important to understand fundamentally how resources work to make the most of them, regardless of the guidance found on this page.
If configured for your organisation within Cogniti, you may be able to set the resource usage strictness, and information similarity cutoff, for resources for your agent.
These settings can be found under the advanced settings section in agent resources.
Resource usage strictness #
When resources are used by an agent, the agent first searches through all its provided resources to find small chunks that may relate to the user’s most recent prompt to the agent. These chunks may or may not be actually relevant to the user’s needs, depending on how specific the user’s most recent prompt was. Only a few chunks are returned each time.
The resource usage strictness setting asks Cogniti to evaluate the relevance of found chunks.
At the default setting, your agent will take in chunks of information and consider these chunks alongside the AI model’s own massive knowledgebase to generate a response for the user.
At the very strict setting, Cogniti instructs the AI in the background to try to focus only on the information present in the found chunks. If no relevant information is found, Cogniti instructs the AI to not respond at all, instead of attempting to generate a best-efforts response.
The strict setting is between these.
Information similarity scoring and cutoff #
Note that this needs to be separately enabled by an administrator. Using the similarity cutoff incurs extra cost on processing time, energy use, and finances.
When chunks of information are found from the agent’s available collection of resources, they are forwarded to the AI model for it to use in generating its response. Usually, the chunks are not necessarily sorted in order of relevance, and no filtering takes place. This can sometimes mean that less relevant chunks of information from the resources are provided to the AI.
If enabled for your organisation, the information similarity scoring and cutoff can be used to control this. You can toggle the switch to enable this, and specify a cutoff threshold.
Enabling similarity scoring will allow Cogniti to use an additional AI step to evaluate (score) the relevance of the found chunks of information.
Setting the information similarity cutoff allows you to tell Cogniti how relevant the chunks of information need to be so that they are included or discarded. A lower number is more permissive, and a higher number is more restrictive. You will need to experiment and iterate to determine the right threshold for your agent’s needs.
