GenerativeSettings

Settings for Generative AI.

JSON representation 
{
  "name": string,
  "fallbackSettings": {
    object (FallbackSettings)
  },
  "generativeSafetySettings": {
    object (SafetySettings)
  },
  "knowledgeConnectorSettings": {
    object (KnowledgeConnectorSettings)
  },
  "languageCode": string,
  "llmModelSettings": {
    object (LlmModelSettings)
  }
}
Fields
name

string

Format: projects/<ProjectID>/locations/<LocationID>/agents/<AgentID>/generativeSettings.
fallbackSettings

object (FallbackSettings)

Settings for Generative Fallback.
generativeSafetySettings

object (SafetySettings)

Settings for Generative Safety.
knowledgeConnectorSettings

object (KnowledgeConnectorSettings)

Settings for knowledge connector.
languageCode

string

Language for this settings.
llmModelSettings

object (LlmModelSettings)

LLM model settings.

FallbackSettings

Settings for Generative Fallback.

JSON representation 
{
  "selectedPrompt": string,
  "promptTemplates": [
    {
      object (PromptTemplate)
    }
  ]
}
Fields
selectedPrompt

string

Display name of the selected prompt.
promptTemplates[]

object (PromptTemplate)

Stored prompts that can be selected, for example default templates like "conservative" or "chatty", or user defined ones.

PromptTemplate

Prompt template.

JSON representation 
{
  "displayName": string,
  "promptText": string,
  "frozen": boolean
}
Fields
displayName

string

Prompt name.
promptText

string

Prompt text that is sent to a LLM on no-match default, placeholders are filled downstream. For example: "Here is a conversation $conversation, a response is: "
frozen

boolean

If the flag is true, the prompt is frozen and cannot be modified by users.

SafetySettings

Settings for Generative Safety.

JSON representation 
{
  "defaultBannedPhraseMatchStrategy": enum (PhraseMatchStrategy),
  "bannedPhrases": [
    {
      object (Phrase)
    }
  ],
  "promptSecuritySettings": {
    object (PromptSecuritySettings)
  }
}
Fields
defaultBannedPhraseMatchStrategy

enum (PhraseMatchStrategy)

Optional. Default phrase match strategy for banned phrases.
bannedPhrases[]

object (Phrase)

Banned phrases for generated text.
promptSecuritySettings

object (PromptSecuritySettings)

Optional. Settings for prompt security checks.

PhraseMatchStrategy

Strategy for matching phrases.

Enums
PHRASE_MATCH_STRATEGY_UNSPECIFIED Unspecified, defaults to PARTIAL_MATCH.
PARTIAL_MATCH Text that contains the phrase as a substring will be matched, e.g. "foo" will match "afoobar".
WORD_MATCH Text that contains the tokenized words of the phrase will be matched, e.g. "foo" will match "a foo bar" and "foo bar", but not "foobar".

Phrase

Text input which can be used for prompt or banned phrases.

JSON representation 
{
  "text": string,
  "languageCode": string
}
Fields
text

string

Required. Text input which can be used for prompt or banned phrases.
languageCode

string

Required. Language code of the phrase.

PromptSecuritySettings

Settings for prompt security checks.

JSON representation 
{
  "enablePromptSecurity": boolean
}
Fields
enablePromptSecurity

boolean

Optional. Enable prompt security checks.

KnowledgeConnectorSettings

Settings for knowledge connector. These parameters are used for LLM prompt like "You are . You are a helpful and verbose at , . Your task is to help humans on ".

JSON representation 
{
  "business": string,
  "agent": string,
  "agentIdentity": string,
  "businessDescription": string,
  "agentScope": string,
  "disableDataStoreFallback": boolean
}
Fields
business

string

Name of the company, organization or other entity that the agent represents. Used for knowledge connector LLM prompt and for knowledge search.
agent

string

Name of the virtual agent. Used for LLM prompt. Can be left empty.
agentIdentity

string

Identity of the agent, e.g. "virtual agent", "AI assistant".
businessDescription

string

Company description, used for LLM prompt, e.g. "a family company selling freshly roasted coffee beans".
agentScope

string

Agent scope, e.g. "Example company website", "internal Example company website for employees", "manual of car owner".
disableDataStoreFallback

boolean

Whether to disable fallback to Data Store search results (in case the LLM couldn't pick a proper answer). Per default the feature is enabled.