Class LLMUtilityAI

constructor

• new LLMUtilityAI(utilityId?): LLMUtilityAI

• Parameters

  • Optional utilityId: string

  Returns LLMUtilityAI

initialize

• initialize(config): Promise<void>

• Parameters

  • config: LLMUtilityAIConfig

  Returns Promise<void>

summarize

• summarize(textToSummarize, options?): Promise<string>

• Parameters

  • textToSummarize: string
  • Optional options: SummarizationOptions

  Returns Promise<string>

parseJsonSafe

• parseJsonSafe<T>(jsonString, options?): Promise<null | T>

• The JSON the text holds, read as StatisticalUtilityAI reads it (parseReplyJson: the text as it is, else inside the first Markdown code fence it holds, backticks inside a sentence opening none, with a language tag or none and with words before or after it, up to the first later line that starts with as many backticks as opened the fence or more), then a check against targetSchema with Ajv when given, as StatisticalUtilityAI checks it: the standard formats checked, a keyword or a format Ajv does not know passed over, and a value coerced to the type its schema names where it can (targetSchemaChecks). With attemptFixWithLLM, a text that fails either is sent for repair, up to maxRepairAttempts times (1 unless it is a whole number of 1 or more), each repair on the text the last one answered, whose answer is read the same way, on llmModelIdForFix (else jsonFixerModelId, else defaultModelId) and llmProviderIdForFix (else defaultProviderId). A repair request names targetSchema when given, and the schema's errors when the text parsed and failed it. Answers null when no repair parses and passes the schema, or a repair call fails. No text can pass a targetSchema that Ajv cannot compile, that takes the draft-07 meta-schema's $id or that is marked $async: the call answers null at once, with a warning, and asks for no repair.

  Type Parameters

  • T = any

  Parameters

  • jsonString: string
  • Optional options: ParseJsonOptions<T>

  Returns Promise<null | T>

classifyText

• classifyText(textToClassify, options): Promise<ClassificationResult>

• Parameters

  • textToClassify: string
  • options: ClassificationOptions

  Returns Promise<ClassificationResult>

extractKeywords

• extractKeywords(textToAnalyze, options?): Promise<string[]>

• Parameters

  • textToAnalyze: string
  • Optional options: KeywordExtractionOptions

  Returns Promise<string[]>

analyzeSentiment

• analyzeSentiment(text, options?): Promise<SentimentResult>

• Parameters

  • text: string
  • Optional options: SentimentAnalysisOptions

  Returns Promise<SentimentResult>

detectLanguage

• detectLanguage(text, options?): Promise<LanguageDetectionResult[]>

• Parameters

  • text: string
  • Optional options: LanguageDetectionOptions

  Returns Promise<LanguageDetectionResult[]>

tokenize

• tokenize(text, options?): Promise<string[]>

• Parameters

  • text: string
  • Optional options: TokenizationOptions

  Returns Promise<string[]>

stemTokens

• stemTokens(tokens, _options?): Promise<string[]>

• Parameters

  • tokens: string[]
  • Optional _options: StemmingOptions

  Returns Promise<string[]>

calculateSimilarity

• calculateSimilarity(text1, text2, options?): Promise<number>

• Parameters

  • text1: string
  • text2: string
  • Optional options: SimilarityOptions

  Returns Promise<number>

normalizeText

• normalizeText(text, options?): Promise<string>

• Parameters

  • text: string
  • Optional options: TextNormalizationOptions

  Returns Promise<string>

generateNGrams

• generateNGrams(tokens, options): Promise<Record<number, string[][]>>

• Parameters

  • tokens: string[]
  • options: NGramOptions

  Returns Promise<Record<number, string[][]>>

calculateReadability

• calculateReadability(text, options): Promise<ReadabilityResult>

• Parameters

  • text: string
  • options: ReadabilityOptions

  Returns Promise<ReadabilityResult>

summarizeConversationHistory

• summarizeConversationHistory(messages, targetTokenCount, modelInfo, preserveImportantMessages?): Promise<{
      summaryMessages: ConversationMessage[];
      originalTokenCount: number;
      finalTokenCount: number;
      messagesSummarized: number;
  }>

• Summarizes conversation history to fit within token constraints, attempting to preserve key information.

  Parameters

  • messages: readonly ConversationMessage[]

    The array of conversation messages to summarize.
  • targetTokenCount: number

    The desired maximum token count for the summary.
  • modelInfo: Readonly<ModelTargetInfo>

    Information about the model for which the summary is being prepared.
  • Optional preserveImportantMessages: boolean

    If true, attempt to identify and keep important messages verbatim.

  Returns Promise<{
      summaryMessages: ConversationMessage[];
      originalTokenCount: number;
      finalTokenCount: number;
      messagesSummarized: number;
  }>

  A summary (which might be a single system message or a condensed list of messages), and metadata about the summarization.

summarizeRAGContext

• summarizeRAGContext(context, targetTokenCount, modelInfo, preserveSourceAttribution?): Promise<{
      summary: string;
      originalTokenCount: number;
      finalTokenCount: number;
      preservedSources?: string[];
  }>

• Summarizes retrieved RAG context to fit token limits, ideally preserving source attribution if possible.

  Parameters

  • context: string | readonly {
        source: string;
        content: string;
        relevance?: number;
    }[]

    The RAG context to summarize.
  • targetTokenCount: number

    The desired maximum token count for the summarized context.
  • modelInfo: Readonly<ModelTargetInfo>

    Information about the model.
  • Optional preserveSourceAttribution: boolean

    If true, attempt to retain source information in the summary.

  Returns Promise<{
      summary: string;
      originalTokenCount: number;
      finalTokenCount: number;
      preservedSources?: string[];
  }>

  The summarized text and metadata.

checkHealth

• checkHealth(): Promise<{
      isHealthy: boolean;
      details?: any;
      dependencies?: {
          name: string;
          isHealthy: boolean;
          details?: any;
      }[];
  }>

• Returns Promise<{
      isHealthy: boolean;
      details?: any;
      dependencies?: {
          name: string;
          isHealthy: boolean;
          details?: any;
      }[];
  }>

trainModel

• trainModel(): Promise<{
      success: boolean;
      message?: string;
      modelId?: string;
  }>

• Returns Promise<{
      success: boolean;
      message?: string;
      modelId?: string;
  }>

saveTrainedModel

• saveTrainedModel(): Promise<{
      success: boolean;
      pathOrStoreId?: string;
      message?: string;
  }>

• Returns Promise<{
      success: boolean;
      pathOrStoreId?: string;
      message?: string;
  }>

loadTrainedModel

• loadTrainedModel(): Promise<{
      success: boolean;
      message?: string;
  }>

• Returns Promise<{
      success: boolean;
      message?: string;
  }>

shutdown

• shutdown(): Promise<void>

• Returns Promise<void>

Readonly utilityId