@promptbook/core 0.92.0-4 → 0.92.0-6

package/esm/index.es.js

@@ -27,7 +27,7 @@ const BOOK_LANGUAGE_VERSION = '1.0.0';
  * @generated
  * @see https://github.com/webgptorg/promptbook
  */
-const PROMPTBOOK_ENGINE_VERSION = '0.92.0-4';
+const PROMPTBOOK_ENGINE_VERSION = '0.92.0-6';
 /**
  * TODO: string_promptbook_version should be constrained to the all versions of Promptbook engine
  * Note: [💞] Ignore a discrepancy between file name and entity name
@@ -1893,6 +1893,45 @@ function $randomToken(randomness) {
  * TODO: Maybe use nanoid instead https://github.com/ai/nanoid
  */
 
+/**
+ * Converts a JavaScript Object Notation (JSON) string into an object.
+ *
+ * Note: This is wrapper around `JSON.parse()` with better error and type handling
+ *
+ * @public exported from `@promptbook/utils`
+ */
+function jsonParse(value) {
+    if (value === undefined) {
+        throw new Error(`Can not parse JSON from undefined value.`);
+    }
+    else if (typeof value !== 'string') {
+        console.error('Can not parse JSON from non-string value.', { text: value });
+        throw new Error(spaceTrim(`
+                    Can not parse JSON from non-string value.
+
+                    The value type: ${typeof value}
+                    See more in console.
+                `));
+    }
+    try {
+        return JSON.parse(value);
+    }
+    catch (error) {
+        if (!(error instanceof Error)) {
+            throw error;
+        }
+        throw new Error(spaceTrim((block) => `
+                    ${block(error.message)}
+            
+                    The JSON text:
+                    ${block(value)}
+                `));
+    }
+}
+/**
+ * TODO: !!!! Use in Promptbook.studio
+ */
+
 /**
  * Recursively converts JSON strings to JSON objects
 
@@ -1911,7 +1950,7 @@ function jsonStringsToJsons(object) {
     const newObject = { ...object };
     for (const [key, value] of Object.entries(object)) {
         if (typeof value === 'string' && isValidJsonString(value)) {
-            newObject[key] = JSON.parse(value);
+            newObject[key] = jsonParse(value);
         }
         else {
             newObject[key] = jsonStringsToJsons(value);
@@ -4327,13 +4366,79 @@ async function getExamplesForTask(task) {
 /**
  * @@@
  *
+ * Here is the place where RAG (retrieval-augmented generation) happens
+ *
  * @private internal utility of `createPipelineExecutor`
  */
 async function getKnowledgeForTask(options) {
-    const { preparedPipeline, task } = options;
-    return preparedPipeline.knowledgePieces.map(({ content }) => `- ${content}`).join('\n');
+    const { tools, preparedPipeline, task } = options;
+    const firstKnowlegePiece = preparedPipeline.knowledgePieces[0];
+    const firstKnowlegeIndex = firstKnowlegePiece === null || firstKnowlegePiece === void 0 ? void 0 : firstKnowlegePiece.index[0];
+    // <- TODO: Do not use just first knowledge piece and first index to determine embedding model, use also keyword search
+    if (firstKnowlegePiece === undefined || firstKnowlegeIndex === undefined) {
+        return 'No knowledge pieces found';
+    }
+    // TODO: [🚐] Make arrayable LLMs -> single LLM DRY
+    const _llms = arrayableToArray(tools.llm);
+    const llmTools = _llms.length === 1 ? _llms[0] : joinLlmExecutionTools(..._llms);
+    const taskEmbeddingPrompt = {
+        title: 'Knowledge Search',
+        modelRequirements: {
+            modelVariant: 'EMBEDDING',
+            modelName: firstKnowlegeIndex.modelName,
+        },
+        content: task.content,
+        parameters: {
+        /* !!!!!!!! */
+        },
+    };
+    const taskEmbeddingResult = await llmTools.callEmbeddingModel(taskEmbeddingPrompt);
+    const knowledgePiecesWithRelevance = preparedPipeline.knowledgePieces.map((knowledgePiece) => {
+        const { index } = knowledgePiece;
+        const knowledgePieceIndex = index.find((i) => i.modelName === firstKnowlegeIndex.modelName);
+        // <- TODO: Do not use just first knowledge piece and first index to determine embedding model
+        if (knowledgePieceIndex === undefined) {
+            return {
+                content: knowledgePiece.content,
+                relevance: 0,
+            };
+        }
+        const relevance = computeCosineSimilarity(knowledgePieceIndex.position, taskEmbeddingResult.content);
+        return {
+            content: knowledgePiece.content,
+            relevance,
+        };
+    });
+    const knowledgePiecesSorted = knowledgePiecesWithRelevance.sort((a, b) => a.relevance - b.relevance);
+    const knowledgePiecesLimited = knowledgePiecesSorted.slice(0, 5);
+    console.log('!!! Embedding', {
+        task,
+        taskEmbeddingPrompt,
+        taskEmbeddingResult,
+        firstKnowlegePiece,
+        firstKnowlegeIndex,
+        knowledgePiecesWithRelevance,
+        knowledgePiecesSorted,
+        knowledgePiecesLimited,
+    });
+    return knowledgePiecesLimited.map(({ content }) => `- ${content}`).join('\n');
     //                                                      <- TODO: [🧠] Some smart aggregation of knowledge pieces, single-line vs multi-line vs mixed
 }
+// TODO: !!!!!! Annotate + to new file
+function computeCosineSimilarity(embeddingVector1, embeddingVector2) {
+    if (embeddingVector1.length !== embeddingVector2.length) {
+        throw new TypeError('Embedding vectors must have the same length');
+    }
+    const dotProduct = embeddingVector1.reduce((sum, value, index) => sum + value * embeddingVector2[index], 0);
+    const magnitude1 = Math.sqrt(embeddingVector1.reduce((sum, value) => sum + value * value, 0));
+    const magnitude2 = Math.sqrt(embeddingVector2.reduce((sum, value) => sum + value * value, 0));
+    return 1 - dotProduct / (magnitude1 * magnitude2);
+}
+/**
+ * TODO: !!!! Verify if this is working
+ * TODO: [♨] Implement Better - use keyword search
+ * TODO: [♨] Examples of values
+ */
 
 /**
  * @@@
@@ -4341,9 +4446,9 @@ async function getKnowledgeForTask(options) {
  * @private internal utility of `createPipelineExecutor`
  */
 async function getReservedParametersForTask(options) {
-    const { preparedPipeline, task, pipelineIdentification } = options;
+    const { tools, preparedPipeline, task, pipelineIdentification } = options;
     const context = await getContextForTask(); // <- [🏍]
-    const knowledge = await getKnowledgeForTask({ preparedPipeline, task });
+    const knowledge = await getKnowledgeForTask({ tools, preparedPipeline, task });
     const examples = await getExamplesForTask();
     const currentDate = new Date().toISOString(); // <- TODO: [🧠][💩] Better
     const modelName = RESERVED_PARAMETER_MISSING_VALUE;
@@ -4405,6 +4510,7 @@ async function executeTask(options) {
     }
     const definedParameters = Object.freeze({
         ...(await getReservedParametersForTask({
+            tools,
             preparedPipeline,
             task: currentTask,
             pipelineIdentification,
@@ -4976,7 +5082,7 @@ async function preparePersona(personaDescription, tools, options) {
     }).asPromise();
     const { outputParameters } = result;
     const { modelsRequirements: modelsRequirementsJson } = outputParameters;
-    const modelsRequirementsUnchecked = JSON.parse(modelsRequirementsJson);
+    const modelsRequirementsUnchecked = jsonParse(modelsRequirementsJson);
     if (isVerbose) {
         console.info(`PERSONA ${personaDescription}`, modelsRequirementsUnchecked);
     }
@@ -5563,7 +5669,7 @@ async function makeKnowledgeSourceHandler(knowledgeSource, tools, options) {
                 > },
             */
             async asJson() {
-                return JSON.parse(await tools.fs.readFile(filename, 'utf-8'));
+                return jsonParse(await tools.fs.readFile(filename, 'utf-8'));
             },
             async asText() {
                 return await tools.fs.readFile(filename, 'utf-8');
@@ -10209,6 +10315,48 @@ function createLlmToolsFromConfiguration(configuration, options = {}) {
  * TODO: [®] DRY Register logic
  */
 
+/**
+ * How is the model provider trusted?
+ *
+ * @public exported from `@promptbook/core`
+ */
+const MODEL_TRUST_LEVEL = {
+    FULL: `Model is running on the local machine, training data and model weights are known, data are ethically sourced`,
+    OPEN: `Model is open source, training data and model weights are known`,
+    PARTIALLY_OPEN: `Model is open source, but training data and model weights are not (fully) known`,
+    CLOSED_LOCAL: `Model can be run locally, but it is not open source`,
+    CLOSED_FREE: `Model is behind API gateway but free to use`,
+    CLOSED_BUSINESS: `Model is behind API gateway and paid but has good SLA, TOS, privacy policy and in general is a good to use in business applications`,
+    CLOSED: `Model is behind API gateway and paid`,
+    UNTRUSTED: `Model has questions about the training data and ethics, but it is not known if it is a problem or not`,
+    VURNABLE: `Model has some known serious vulnerabilities, leaks, ethical problems, etc.`,
+};
+// <- TODO: Maybe do better levels of trust
+/**
+ * How is the model provider important?
+ *
+ * @public exported from `@promptbook/core`
+ */
+const MODEL_ORDER = {
+    /**
+     * Top-tier models, e.g. OpenAI, Anthropic,...
+     */
+    TOP_TIER: 333,
+    /**
+     * Mid-tier models, e.g. Llama, Mistral, etc.
+     */
+    NORMAL: 100,
+    /**
+     * Low-tier models, e.g. Phi, Tiny, etc.
+     */
+    LOW_TIER: 0,
+};
+/**
+ * TODO: Add configuration schema and maybe some documentation link
+ * TODO: Maybe constrain LlmToolsConfiguration[number] by generic to ensure that `createConfigurationFromEnv` and `getBoilerplateConfiguration` always create same `packageName` and `className`
+ * TODO: [®] DRY Register logic
+ */
+
 /**
  * Stores data in memory (HEAP)
  *
@@ -10408,9 +10556,11 @@ const _AnthropicClaudeMetadataRegistration = $llmToolsMetadataRegister.register(
     packageName: '@promptbook/anthropic-claude',
     className: 'AnthropicClaudeExecutionTools',
     envVariables: ['ANTHROPIC_CLAUDE_API_KEY'],
+    trustLevel: 'CLOSED',
+    order: MODEL_ORDER.TOP_TIER,
     getBoilerplateConfiguration() {
         return {
-            title: 'Anthropic Claude (boilerplate)',
+            title: 'Anthropic Claude',
             packageName: '@promptbook/anthropic-claude',
             className: 'AnthropicClaudeExecutionTools',
             options: {
@@ -10453,9 +10603,11 @@ const _AzureOpenAiMetadataRegistration = $llmToolsMetadataRegister.register({
     packageName: '@promptbook/azure-openai',
     className: 'AzureOpenAiExecutionTools',
     envVariables: ['AZUREOPENAI_RESOURCE_NAME', 'AZUREOPENAI_DEPLOYMENT_NAME', 'AZUREOPENAI_API_KEY'],
+    trustLevel: 'CLOSED_BUSINESS',
+    order: MODEL_ORDER.NORMAL,
     getBoilerplateConfiguration() {
         return {
-            title: 'Azure Open AI (boilerplate)',
+            title: 'Azure Open AI',
             packageName: '@promptbook/azure-openai',
             className: 'AzureOpenAiExecutionTools',
             options: {
@@ -10539,9 +10691,11 @@ const _DeepseekMetadataRegistration = $llmToolsMetadataRegister.register({
     packageName: '@promptbook/deepseek',
     className: 'DeepseekExecutionTools',
     envVariables: ['DEEPSEEK_GENERATIVE_AI_API_KEY'],
+    trustLevel: 'UNTRUSTED',
+    order: MODEL_ORDER.NORMAL,
     getBoilerplateConfiguration() {
         return {
-            title: 'Deepseek (boilerplate)',
+            title: 'Deepseek',
             packageName: '@promptbook/deepseek',
             className: 'DeepseekExecutionTools',
             options: {
@@ -10588,9 +10742,11 @@ const _GoogleMetadataRegistration = $llmToolsMetadataRegister.register({
     packageName: '@promptbook/google',
     className: 'GoogleExecutionTools',
     envVariables: ['GOOGLE_GENERATIVE_AI_API_KEY'],
+    trustLevel: 'CLOSED',
+    order: MODEL_ORDER.NORMAL,
     getBoilerplateConfiguration() {
         return {
-            title: 'Google Gemini (boilerplate)',
+            title: 'Google Gemini',
             packageName: '@promptbook/google',
             className: 'GoogleExecutionTools',
             options: {
@@ -10637,9 +10793,11 @@ const _OpenAiMetadataRegistration = $llmToolsMetadataRegister.register({
     packageName: '@promptbook/openai',
     className: 'OpenAiExecutionTools',
     envVariables: ['OPENAI_API_KEY'],
+    trustLevel: 'CLOSED',
+    order: MODEL_ORDER.TOP_TIER,
     getBoilerplateConfiguration() {
         return {
-            title: 'Open AI (boilerplate)',
+            title: 'Open AI',
             packageName: '@promptbook/openai',
             className: 'OpenAiExecutionTools',
             options: {
@@ -10677,9 +10835,11 @@ const _OpenAiAssistantMetadataRegistration = $llmToolsMetadataRegister.register(
     className: 'OpenAiAssistantExecutionTools',
     envVariables: null,
     //            <- TODO: ['OPENAI_API_KEY', 'OPENAI_ASSISTANT_ID']
+    trustLevel: 'CLOSED',
+    order: MODEL_ORDER.NORMAL,
     getBoilerplateConfiguration() {
         return {
-            title: 'Open AI Assistant (boilerplate)',
+            title: 'Open AI Assistant',
             packageName: '@promptbook/openai',
             className: 'OpenAiAssistantExecutionTools',
             options: {
@@ -11143,5 +11303,5 @@ class PrefixStorage {
     }
 }
 
-export { $llmToolsMetadataRegister, $llmToolsRegister, $scrapersMetadataRegister, $scrapersRegister, ADMIN_EMAIL, ADMIN_GITHUB_NAME, AbstractFormatError, AuthenticationError, BOOK_LANGUAGE_VERSION, BlackholeStorage, BoilerplateError, BoilerplateFormfactorDefinition, CLAIM, CLI_APP_ID, CallbackInterfaceTools, ChatbotFormfactorDefinition, CollectionError, CsvFormatDefinition, CsvFormatError, DEFAULT_BOOKS_DIRNAME, DEFAULT_BOOK_OUTPUT_PARAMETER_NAME, DEFAULT_BOOK_TITLE, DEFAULT_CSV_SETTINGS, DEFAULT_DOWNLOAD_CACHE_DIRNAME, DEFAULT_EXECUTION_CACHE_DIRNAME, DEFAULT_GET_PIPELINE_COLLECTION_FUNCTION_NAME, DEFAULT_INTERMEDIATE_FILES_STRATEGY, DEFAULT_IS_AUTO_INSTALLED, DEFAULT_IS_VERBOSE, DEFAULT_MAX_EXECUTION_ATTEMPTS, DEFAULT_MAX_FILE_SIZE, DEFAULT_MAX_KNOWLEDGE_SOURCES_SCRAPING_DEPTH, DEFAULT_MAX_KNOWLEDGE_SOURCES_SCRAPING_TOTAL, DEFAULT_MAX_PARALLEL_COUNT, DEFAULT_PIPELINE_COLLECTION_BASE_FILENAME, DEFAULT_PROMPT_TASK_TITLE, DEFAULT_REMOTE_SERVER_URL, DEFAULT_SCRAPE_CACHE_DIRNAME, DEFAULT_TASK_TITLE, EXPECTATION_UNITS, EnvironmentMismatchError, ExecutionReportStringOptionsDefaults, ExpectError, FORMFACTOR_DEFINITIONS, GENERIC_PIPELINE_INTERFACE, GeneratorFormfactorDefinition, GenericFormfactorDefinition, ImageGeneratorFormfactorDefinition, KnowledgeScrapeError, LimitReachedError, MANDATORY_CSV_SETTINGS, MAX_FILENAME_LENGTH, MODEL_VARIANTS, MatcherFormfactorDefinition, MemoryStorage, MissingToolsError, MultipleLlmExecutionTools, NAME, NonTaskSectionTypes, NotFoundError, NotYetImplementedError, ORDER_OF_PIPELINE_JSON, PLAYGROUND_APP_ID, PROMPTBOOK_ENGINE_VERSION, PROMPTBOOK_ERRORS, ParseError, PipelineExecutionError, PipelineLogicError, PipelineUrlError, PrefixStorage, PromptbookFetchError, REMOTE_SERVER_URLS, RESERVED_PARAMETER_NAMES, SET_IS_VERBOSE, SectionTypes, SheetsFormfactorDefinition, TaskTypes, TextFormatDefinition, TranslatorFormfactorDefinition, UNCERTAIN_USAGE, UNCERTAIN_ZERO_VALUE, UnexpectedError, WrappedError, ZERO_USAGE, ZERO_VALUE, _AnthropicClaudeMetadataRegistration, _AzureOpenAiMetadataRegistration, _BoilerplateScraperMetadataRegistration, _DeepseekMetadataRegistration, _DocumentScraperMetadataRegistration, _GoogleMetadataRegistration, _LegacyDocumentScraperMetadataRegistration, _MarkdownScraperMetadataRegistration, _MarkitdownScraperMetadataRegistration, _OpenAiAssistantMetadataRegistration, _OpenAiMetadataRegistration, _PdfScraperMetadataRegistration, _WebsiteScraperMetadataRegistration, addUsage, book, cacheLlmTools, collectionToJson, compilePipeline, countUsage, createCollectionFromJson, createCollectionFromPromise, createCollectionFromUrl, createLlmToolsFromConfiguration, createPipelineExecutor, createSubcollection, embeddingVectorToString, executionReportJsonToString, extractParameterNamesFromTask, filterModels, getPipelineInterface, identificationToPromptbookToken, isPassingExpectations, isPipelineImplementingInterface, isPipelineInterfacesEqual, isPipelinePrepared, isValidPipelineString, joinLlmExecutionTools, limitTotalUsage, makeKnowledgeSourceHandler, parsePipeline, pipelineJsonToString, prepareKnowledgePieces, preparePersona, preparePipeline, prepareTasks, prettifyPipelineString, promptbookFetch, promptbookTokenToIdentification, unpreparePipeline, usageToHuman, usageToWorktime, validatePipeline, validatePipelineString };
+export { $llmToolsMetadataRegister, $llmToolsRegister, $scrapersMetadataRegister, $scrapersRegister, ADMIN_EMAIL, ADMIN_GITHUB_NAME, AbstractFormatError, AuthenticationError, BOOK_LANGUAGE_VERSION, BlackholeStorage, BoilerplateError, BoilerplateFormfactorDefinition, CLAIM, CLI_APP_ID, CallbackInterfaceTools, ChatbotFormfactorDefinition, CollectionError, CsvFormatDefinition, CsvFormatError, DEFAULT_BOOKS_DIRNAME, DEFAULT_BOOK_OUTPUT_PARAMETER_NAME, DEFAULT_BOOK_TITLE, DEFAULT_CSV_SETTINGS, DEFAULT_DOWNLOAD_CACHE_DIRNAME, DEFAULT_EXECUTION_CACHE_DIRNAME, DEFAULT_GET_PIPELINE_COLLECTION_FUNCTION_NAME, DEFAULT_INTERMEDIATE_FILES_STRATEGY, DEFAULT_IS_AUTO_INSTALLED, DEFAULT_IS_VERBOSE, DEFAULT_MAX_EXECUTION_ATTEMPTS, DEFAULT_MAX_FILE_SIZE, DEFAULT_MAX_KNOWLEDGE_SOURCES_SCRAPING_DEPTH, DEFAULT_MAX_KNOWLEDGE_SOURCES_SCRAPING_TOTAL, DEFAULT_MAX_PARALLEL_COUNT, DEFAULT_PIPELINE_COLLECTION_BASE_FILENAME, DEFAULT_PROMPT_TASK_TITLE, DEFAULT_REMOTE_SERVER_URL, DEFAULT_SCRAPE_CACHE_DIRNAME, DEFAULT_TASK_TITLE, EXPECTATION_UNITS, EnvironmentMismatchError, ExecutionReportStringOptionsDefaults, ExpectError, FORMFACTOR_DEFINITIONS, GENERIC_PIPELINE_INTERFACE, GeneratorFormfactorDefinition, GenericFormfactorDefinition, ImageGeneratorFormfactorDefinition, KnowledgeScrapeError, LimitReachedError, MANDATORY_CSV_SETTINGS, MAX_FILENAME_LENGTH, MODEL_ORDER, MODEL_TRUST_LEVEL, MODEL_VARIANTS, MatcherFormfactorDefinition, MemoryStorage, MissingToolsError, MultipleLlmExecutionTools, NAME, NonTaskSectionTypes, NotFoundError, NotYetImplementedError, ORDER_OF_PIPELINE_JSON, PLAYGROUND_APP_ID, PROMPTBOOK_ENGINE_VERSION, PROMPTBOOK_ERRORS, ParseError, PipelineExecutionError, PipelineLogicError, PipelineUrlError, PrefixStorage, PromptbookFetchError, REMOTE_SERVER_URLS, RESERVED_PARAMETER_NAMES, SET_IS_VERBOSE, SectionTypes, SheetsFormfactorDefinition, TaskTypes, TextFormatDefinition, TranslatorFormfactorDefinition, UNCERTAIN_USAGE, UNCERTAIN_ZERO_VALUE, UnexpectedError, WrappedError, ZERO_USAGE, ZERO_VALUE, _AnthropicClaudeMetadataRegistration, _AzureOpenAiMetadataRegistration, _BoilerplateScraperMetadataRegistration, _DeepseekMetadataRegistration, _DocumentScraperMetadataRegistration, _GoogleMetadataRegistration, _LegacyDocumentScraperMetadataRegistration, _MarkdownScraperMetadataRegistration, _MarkitdownScraperMetadataRegistration, _OpenAiAssistantMetadataRegistration, _OpenAiMetadataRegistration, _PdfScraperMetadataRegistration, _WebsiteScraperMetadataRegistration, addUsage, book, cacheLlmTools, collectionToJson, compilePipeline, countUsage, createCollectionFromJson, createCollectionFromPromise, createCollectionFromUrl, createLlmToolsFromConfiguration, createPipelineExecutor, createSubcollection, embeddingVectorToString, executionReportJsonToString, extractParameterNamesFromTask, filterModels, getPipelineInterface, identificationToPromptbookToken, isPassingExpectations, isPipelineImplementingInterface, isPipelineInterfacesEqual, isPipelinePrepared, isValidPipelineString, joinLlmExecutionTools, limitTotalUsage, makeKnowledgeSourceHandler, parsePipeline, pipelineJsonToString, prepareKnowledgePieces, preparePersona, preparePipeline, prepareTasks, prettifyPipelineString, promptbookFetch, promptbookTokenToIdentification, unpreparePipeline, usageToHuman, usageToWorktime, validatePipeline, validatePipelineString };
 //# sourceMappingURL=index.es.js.map