A conversational tool selection uses the full power of the LLM model to either choose an action or send a reply to the user, depending on the situation and prompts provided. This step type will:
Use the conversation history
The content template becomes the system prompt fed to the LLM model
The tools available to the model will be based on the combination of the built-in actions and any custom actions defined in the options
If the model chooses to send a regular response completion (rather then to use a tool), that will be translated into the
user_interactor.send_messageaction.
Options
| The name of the model that will be used to process the prompt. |
| Temperature to use with the model. The exact mathematical definition of temperature can vary depending on the model provider. |
| The maximum number of tokens to be included in the completion from the LLM provider. |
| The number of different variations to keep in the cache for this prompt. When the input data to the LLM is exactly the same, the prompt can be ‘cached’. By default only |
| The maximum number of conversation events to include in the conversation history data fed to the bot. |
| Custom actions to include in the tool selection. These must be provided with a smart_chain_binding_name that indicates which smart chain to execute for the action. The 'text' field on the output from the smart-chain will be used as the action result. |
| Actions to exclude from the tool selection, referenced by their action_ids. Can include actions that are defined in the custom_actions section. |
Output
| The text containing the LLM’s response to be said back to the user |
| The raw text of the prompt that was sent up to the LLM provider. |