@promptbook/core 0.98.0-6 → 0.98.0-8

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 = '0.98.0-6';
+const PROMPTBOOK_ENGINE_VERSION = '0.98.0-8';
 /**
  * TODO: string_promptbook_version should be constrained to the all versions of Promptbook engine
  * Note: [💞] Ignore a discrepancy between file name and entity name
@@ -354,7 +354,7 @@ const DEFAULT_MAX_PARALLEL_COUNT = 5; // <- TODO: [🤹‍♂️]
  *
  * @public exported from `@promptbook/core`
  */
-const DEFAULT_MAX_EXECUTION_ATTEMPTS = 3; // <- TODO: [🤹‍♂️]
+const DEFAULT_MAX_EXECUTION_ATTEMPTS = 7; // <- TODO: [🤹‍♂️]
 /**
  * The maximum depth to which knowledge sources will be scraped when building a knowledge base.
  * This prevents infinite recursion and limits resource usage.
@@ -2255,7 +2255,7 @@ function jsonParse(value) {
         throw new Error(spaceTrim((block) => `
                     ${block(error.message)}
 
-                    The JSON text:
+                    The expected JSON text:
                     ${block(value)}
                 `));
     }
@@ -3387,108 +3387,6 @@ function joinLlmExecutionTools(...llmExecutionTools) {
  * TODO: [👷‍♂️] @@@ Manual about construction of llmTools
  */
 
-/**
- * Extracts all code blocks from markdown.
- *
- * Note: There are multiple similar functions:
- * - `extractBlock` just extracts the content of the code block which is also used as built-in function for postprocessing
- * - `extractJsonBlock` extracts exactly one valid JSON code block
- * - `extractOneBlockFromMarkdown` extracts exactly one code block with language of the code block
- * - `extractAllBlocksFromMarkdown` extracts all code blocks with language of the code block
- *
- * @param markdown any valid markdown
- * @returns code blocks with language and content
- * @throws {ParseError} if block is not closed properly
- * @public exported from `@promptbook/markdown-utils`
- */
-function extractAllBlocksFromMarkdown(markdown) {
-    const codeBlocks = [];
-    const lines = markdown.split('\n');
-    // Note: [0] Ensure that the last block notated by gt > will be closed
-    lines.push('');
-    let currentCodeBlock = null;
-    for (const line of lines) {
-        if (line.startsWith('> ') || line === '>') {
-            if (currentCodeBlock === null) {
-                currentCodeBlock = { blockNotation: '>', language: null, content: '' };
-            } /* not else */
-            if (currentCodeBlock.blockNotation === '>') {
-                if (currentCodeBlock.content !== '') {
-                    currentCodeBlock.content += '\n';
-                }
-                currentCodeBlock.content += line.slice(2);
-            }
-        }
-        else if (currentCodeBlock !== null && currentCodeBlock.blockNotation === '>' /* <- Note: [0] */) {
-            codeBlocks.push(currentCodeBlock);
-            currentCodeBlock = null;
-        }
-        /* not else */
-        if (line.startsWith('```')) {
-            const language = line.slice(3).trim() || null;
-            if (currentCodeBlock === null) {
-                currentCodeBlock = { blockNotation: '```', language, content: '' };
-            }
-            else {
-                if (language !== null) {
-                    throw new ParseError(`${capitalize(currentCodeBlock.language || 'the')} code block was not closed and already opening new ${language} code block`);
-                }
-                codeBlocks.push(currentCodeBlock);
-                currentCodeBlock = null;
-            }
-        }
-        else if (currentCodeBlock !== null && currentCodeBlock.blockNotation === '```') {
-            if (currentCodeBlock.content !== '') {
-                currentCodeBlock.content += '\n';
-            }
-            currentCodeBlock.content += line.split('\\`\\`\\`').join('```') /* <- TODO: Maybe make proper unescape */;
-        }
-    }
-    if (currentCodeBlock !== null) {
-        throw new ParseError(`${capitalize(currentCodeBlock.language || 'the')} code block was not closed at the end of the markdown`);
-    }
-    return codeBlocks;
-}
-/**
- * TODO: Maybe name for `blockNotation` instead of  '```' and '>'
- */
-
-/**
- * Extracts  extracts exactly one valid JSON code block
- *
- * - When given string is a valid JSON as it is, it just returns it
- * - When there is no JSON code block the function throws a `ParseError`
- * - When there are multiple JSON code blocks the function throws a `ParseError`
- *
- * Note: It is not important if marked as ```json BUT if it is VALID JSON
- * Note: There are multiple similar function:
- * - `extractBlock` just extracts the content of the code block which is also used as build-in function for postprocessing
- * - `extractJsonBlock` extracts exactly one valid JSON code block
- * - `extractOneBlockFromMarkdown` extracts exactly one code block with language of the code block
- * - `extractAllBlocksFromMarkdown` extracts all code blocks with language of the code block
- *
- * @public exported from `@promptbook/markdown-utils`
- * @throws {ParseError} if there is no valid JSON block in the markdown
- */
-function extractJsonBlock(markdown) {
-    if (isValidJsonString(markdown)) {
-        return markdown;
-    }
-    const codeBlocks = extractAllBlocksFromMarkdown(markdown);
-    const jsonBlocks = codeBlocks.filter(({ content }) => isValidJsonString(content));
-    if (jsonBlocks.length === 0) {
-        throw new Error('There is no valid JSON block in the markdown');
-    }
-    if (jsonBlocks.length > 1) {
-        throw new Error('There are multiple JSON code blocks in the markdown');
-    }
-    return jsonBlocks[0].content;
-}
-/**
- * TODO: Add some auto-healing logic + extract YAML, JSON5, TOML, etc.
- * TODO: [🏢] Make this logic part of `JsonFormatParser` or `isValidJsonString`
- */
-
 /**
  * Takes an item or an array of items and returns an array of items
  *
@@ -3596,6 +3494,108 @@ function templateParameters(template, parameters) {
     return replacedTemplates;
 }
 
+/**
+ * Extracts all code blocks from markdown.
+ *
+ * Note: There are multiple similar functions:
+ * - `extractBlock` just extracts the content of the code block which is also used as built-in function for postprocessing
+ * - `extractJsonBlock` extracts exactly one valid JSON code block
+ * - `extractOneBlockFromMarkdown` extracts exactly one code block with language of the code block
+ * - `extractAllBlocksFromMarkdown` extracts all code blocks with language of the code block
+ *
+ * @param markdown any valid markdown
+ * @returns code blocks with language and content
+ * @throws {ParseError} if block is not closed properly
+ * @public exported from `@promptbook/markdown-utils`
+ */
+function extractAllBlocksFromMarkdown(markdown) {
+    const codeBlocks = [];
+    const lines = markdown.split('\n');
+    // Note: [0] Ensure that the last block notated by gt > will be closed
+    lines.push('');
+    let currentCodeBlock = null;
+    for (const line of lines) {
+        if (line.startsWith('> ') || line === '>') {
+            if (currentCodeBlock === null) {
+                currentCodeBlock = { blockNotation: '>', language: null, content: '' };
+            } /* not else */
+            if (currentCodeBlock.blockNotation === '>') {
+                if (currentCodeBlock.content !== '') {
+                    currentCodeBlock.content += '\n';
+                }
+                currentCodeBlock.content += line.slice(2);
+            }
+        }
+        else if (currentCodeBlock !== null && currentCodeBlock.blockNotation === '>' /* <- Note: [0] */) {
+            codeBlocks.push(currentCodeBlock);
+            currentCodeBlock = null;
+        }
+        /* not else */
+        if (line.startsWith('```')) {
+            const language = line.slice(3).trim() || null;
+            if (currentCodeBlock === null) {
+                currentCodeBlock = { blockNotation: '```', language, content: '' };
+            }
+            else {
+                if (language !== null) {
+                    throw new ParseError(`${capitalize(currentCodeBlock.language || 'the')} code block was not closed and already opening new ${language} code block`);
+                }
+                codeBlocks.push(currentCodeBlock);
+                currentCodeBlock = null;
+            }
+        }
+        else if (currentCodeBlock !== null && currentCodeBlock.blockNotation === '```') {
+            if (currentCodeBlock.content !== '') {
+                currentCodeBlock.content += '\n';
+            }
+            currentCodeBlock.content += line.split('\\`\\`\\`').join('```') /* <- TODO: Maybe make proper unescape */;
+        }
+    }
+    if (currentCodeBlock !== null) {
+        throw new ParseError(`${capitalize(currentCodeBlock.language || 'the')} code block was not closed at the end of the markdown`);
+    }
+    return codeBlocks;
+}
+/**
+ * TODO: Maybe name for `blockNotation` instead of  '```' and '>'
+ */
+
+/**
+ * Extracts  extracts exactly one valid JSON code block
+ *
+ * - When given string is a valid JSON as it is, it just returns it
+ * - When there is no JSON code block the function throws a `ParseError`
+ * - When there are multiple JSON code blocks the function throws a `ParseError`
+ *
+ * Note: It is not important if marked as ```json BUT if it is VALID JSON
+ * Note: There are multiple similar function:
+ * - `extractBlock` just extracts the content of the code block which is also used as build-in function for postprocessing
+ * - `extractJsonBlock` extracts exactly one valid JSON code block
+ * - `extractOneBlockFromMarkdown` extracts exactly one code block with language of the code block
+ * - `extractAllBlocksFromMarkdown` extracts all code blocks with language of the code block
+ *
+ * @public exported from `@promptbook/markdown-utils`
+ * @throws {ParseError} if there is no valid JSON block in the markdown
+ */
+function extractJsonBlock(markdown) {
+    if (isValidJsonString(markdown)) {
+        return markdown;
+    }
+    const codeBlocks = extractAllBlocksFromMarkdown(markdown);
+    const jsonBlocks = codeBlocks.filter(({ content }) => isValidJsonString(content));
+    if (jsonBlocks.length === 0) {
+        throw new Error('There is no valid JSON block in the markdown');
+    }
+    if (jsonBlocks.length > 1) {
+        throw new Error('There are multiple JSON code blocks in the markdown');
+    }
+    return jsonBlocks[0].content;
+}
+/**
+ * TODO: Add some auto-healing logic + extract YAML, JSON5, TOML, etc.
+ * TODO: [🏢] Make this logic part of `JsonFormatParser` or `isValidJsonString`
+ */
+
 /**
  * Counts number of characters in the text
  *
@@ -4039,6 +4039,68 @@ function isPassingExpectations(expectations, value) {
  * Note: [💝] and [🤠] are interconnected together
  */
 
+/**
+ * Validates a prompt result against expectations and format requirements.
+ * This function provides a common abstraction for result validation that can be used
+ * by both execution logic and caching logic to ensure consistency.
+ *
+ * @param options - The validation options including result string, expectations, and format
+ * @returns Validation result with processed string and validity status
+ * @private internal function of `createPipelineExecutor` and `cacheLlmTools`
+ */
+function validatePromptResult(options) {
+    const { resultString, expectations, format } = options;
+    let processedResultString = resultString;
+    let validationError;
+    try {
+        // TODO: [💝] Unite object for expecting amount and format
+        if (format) {
+            if (format === 'JSON') {
+                if (!isValidJsonString(processedResultString)) {
+                    // TODO: [🏢] Do more universally via `FormatParser`
+                    try {
+                        processedResultString = extractJsonBlock(processedResultString);
+                    }
+                    catch (error) {
+                        keepUnused(error);
+                        throw new ExpectError(spaceTrim$1((block) => `
+                                    Expected valid JSON string
+
+                                    The expected JSON text:
+                                    ${block(processedResultString)}
+                                `));
+                    }
+                }
+            }
+            else {
+                throw new UnexpectedError(`Unknown format "${format}"`);
+            }
+        }
+        // TODO: [💝] Unite object for expecting amount and format
+        if (expectations) {
+            checkExpectations(expectations, processedResultString);
+        }
+        return {
+            isValid: true,
+            processedResultString,
+        };
+    }
+    catch (error) {
+        if (error instanceof ExpectError) {
+            validationError = error;
+        }
+        else {
+            // Re-throw non-ExpectError errors (like UnexpectedError)
+            throw error;
+        }
+        return {
+            isValid: false,
+            processedResultString,
+            error: validationError,
+        };
+    }
+}
+
 /**
  * Executes a pipeline task with multiple attempts, including joker and retry logic. Handles different task types
  * (prompt, script, dialog, etc.), applies postprocessing, checks expectations, and updates the execution report.
@@ -4061,13 +4123,13 @@ async function executeAttempts(options) {
     // TODO: [🚐] Make arrayable LLMs -> single LLM DRY
     const _llms = arrayableToArray(tools.llm);
     const llmTools = _llms.length === 1 ? _llms[0] : joinLlmExecutionTools(..._llms);
-    attempts: for (let attempt = -jokerParameterNames.length; attempt < maxAttempts; attempt++) {
-        const isJokerAttempt = attempt < 0;
-        const jokerParameterName = jokerParameterNames[jokerParameterNames.length + attempt];
+    attempts: for (let attemptIndex = -jokerParameterNames.length; attemptIndex < maxAttempts; attemptIndex++) {
+        const isJokerAttempt = attemptIndex < 0;
+        const jokerParameterName = jokerParameterNames[jokerParameterNames.length + attemptIndex];
         // TODO: [🧠][🍭] JOKERS, EXPECTATIONS, POSTPROCESSING and FOREACH
         if (isJokerAttempt && !jokerParameterName) {
             throw new UnexpectedError(spaceTrim$1((block) => `
-                        Joker not found in attempt ${attempt}
+                        Joker not found in attempt ${attemptIndex}
 
                         ${block(pipelineIdentification)}
                     `));
@@ -4265,35 +4327,18 @@ async function executeAttempts(options) {
                 }
             }
             // TODO: [💝] Unite object for expecting amount and format
-            if (task.format) {
-                if (task.format === 'JSON') {
-                    if (!isValidJsonString($ongoingTaskResult.$resultString || '')) {
-                        // TODO: [🏢] Do more universally via `FormatParser`
-                        try {
-                            $ongoingTaskResult.$resultString = extractJsonBlock($ongoingTaskResult.$resultString || '');
-                        }
-                        catch (error) {
-                            keepUnused(error);
-                            throw new ExpectError(spaceTrim$1((block) => `
-                                        Expected valid JSON string
-
-                                        ${block(
-                            /*<- Note: No need for `pipelineIdentification`, it will be catched and added later */ '')}
-                                    `));
-                        }
-                    }
-                }
-                else {
-                    throw new UnexpectedError(spaceTrim$1((block) => `
-                                Unknown format "${task.format}"
-
-                                ${block(pipelineIdentification)}
-                            `));
+            // Use the common validation function for both format and expectations
+            if (task.format || task.expectations) {
+                const validationResult = validatePromptResult({
+                    resultString: $ongoingTaskResult.$resultString || '',
+                    expectations: task.expectations,
+                    format: task.format,
+                });
+                if (!validationResult.isValid) {
+                    throw validationResult.error;
                 }
-            }
-            // TODO: [💝] Unite object for expecting amount and format
-            if (task.expectations) {
-                checkExpectations(task.expectations, $ongoingTaskResult.$resultString || '');
+                // Update the result string in case format processing modified it (e.g., JSON extraction)
+                $ongoingTaskResult.$resultString = validationResult.processedResultString;
             }
             break attempts;
         }
@@ -4307,6 +4352,7 @@ async function executeAttempts(options) {
                 $ongoingTaskResult.$failedResults = [];
             }
             $ongoingTaskResult.$failedResults.push({
+                attemptIndex,
                 result: $ongoingTaskResult.$resultString,
                 error: error,
             });
@@ -4331,19 +4377,13 @@ async function executeAttempts(options) {
                 });
             }
         }
-        if ($ongoingTaskResult.$expectError !== null && attempt === maxAttempts - 1) {
-            // Store the current failure before throwing
-            $ongoingTaskResult.$failedResults = $ongoingTaskResult.$failedResults || [];
-            $ongoingTaskResult.$failedResults.push({
-                result: $ongoingTaskResult.$resultString,
-                error: $ongoingTaskResult.$expectError,
-            });
-            // Create a summary of all failures
+        if ($ongoingTaskResult.$expectError !== null && attemptIndex === maxAttempts - 1) {
+            // Note: Create a summary of all failures
             const failuresSummary = $ongoingTaskResult.$failedResults
-                .map((failure, index) => spaceTrim$1((block) => {
+                .map((failure) => spaceTrim$1((block) => {
                 var _a, _b;
                 return `
-                            Attempt ${index + 1}:
+                            Attempt ${failure.attemptIndex + 1}:
                             Error ${((_a = failure.error) === null || _a === void 0 ? void 0 : _a.name) || ''}:
                             ${block((_b = failure.error) === null || _b === void 0 ? void 0 : _b.message.split('\n').map((line) => `> ${line}`).join('\n'))}
 
@@ -10780,6 +10820,7 @@ function cacheLlmTools(llmTools, options = {}) {
         },
     };
     const callCommonModel = async (prompt) => {
+        var _a;
         const { parameters, content, modelRequirements } = prompt;
         // <- Note: These are relevant things from the prompt that the cache key should depend on.
         // TODO: Maybe some standalone function for normalization of content for cache
@@ -10835,11 +10876,42 @@ function cacheLlmTools(llmTools, options = {}) {
         // 1. It has a content property that is null or undefined
         // 2. It has an error property that is truthy
         // 3. It has a success property that is explicitly false
-        const isFailedResult = promptResult.content === null ||
+        // 4. It doesn't meet the prompt's expectations or format requirements
+        const isBasicFailedResult = promptResult.content === null ||
             promptResult.content === undefined ||
             promptResult.error ||
             promptResult.success === false;
-        if (!isFailedResult) {
+        let shouldCache = !isBasicFailedResult;
+        // If the basic result is valid, check against expectations and format
+        if (shouldCache && promptResult.content) {
+            try {
+                const validationResult = validatePromptResult({
+                    resultString: promptResult.content,
+                    expectations: prompt.expectations,
+                    format: prompt.format,
+                });
+                shouldCache = validationResult.isValid;
+                if (!shouldCache && isVerbose) {
+                    console.info('Not caching result that fails expectations/format validation for key:', key, {
+                        content: promptResult.content,
+                        expectations: prompt.expectations,
+                        format: prompt.format,
+                        validationError: (_a = validationResult.error) === null || _a === void 0 ? void 0 : _a.message,
+                    });
+                }
+            }
+            catch (error) {
+                // If validation throws an unexpected error, don't cache
+                shouldCache = false;
+                if (isVerbose) {
+                    console.info('Not caching result due to validation error for key:', key, {
+                        content: promptResult.content,
+                        validationError: error instanceof Error ? error.message : String(error),
+                    });
+                }
+            }
+        }
+        if (shouldCache) {
             await storage.setItem(key, {
                 date: $getCurrentDate(),
                 promptbookVersion: PROMPTBOOK_ENGINE_VERSION,
@@ -10856,7 +10928,7 @@ function cacheLlmTools(llmTools, options = {}) {
                 promptResult,
             });
         }
-        else if (isVerbose) {
+        else if (isVerbose && isBasicFailedResult) {
             console.info('Not caching failed result for key:', key, {
                 content: promptResult.content,
                 error: promptResult.error,
@@ -11326,7 +11398,7 @@ const _OpenAiCompatibleMetadataRegistration = $llmToolsMetadataRegister.register
     title: 'Open AI Compatible',
     packageName: '@promptbook/openai',
     className: 'OpenAiCompatibleExecutionTools',
-    envVariables: ['OPENAI_API_KEY'],
+    envVariables: ['OPENAI_API_KEY', 'OPENAI_BASE_URL'],
     trustLevel: 'CLOSED',
     order: MODEL_ORDERS.TOP_TIER,
     getBoilerplateConfiguration() {
@@ -11336,11 +11408,35 @@ const _OpenAiCompatibleMetadataRegistration = $llmToolsMetadataRegister.register
             className: 'OpenAiCompatibleExecutionTools',
             options: {
                 apiKey: 'sk-',
+                baseURL: 'https://api.openai.com/v1',
+                isProxied: false,
+                remoteServerUrl: DEFAULT_REMOTE_SERVER_URL,
                 maxRequestsPerMinute: DEFAULT_MAX_REQUESTS_PER_MINUTE,
             },
         };
     },
     createConfigurationFromEnv(env) {
+        // Note: OpenAiCompatibleExecutionTools is an abstract class and cannot be instantiated directly
+        // However, we can provide configuration for users who want to manually instantiate it
+        if (typeof env.OPENAI_API_KEY === 'string') {
+            const options = {
+                apiKey: env.OPENAI_API_KEY,
+                isProxied: false,
+                remoteServerUrl: DEFAULT_REMOTE_SERVER_URL,
+                maxRequestsPerMinute: DEFAULT_MAX_REQUESTS_PER_MINUTE,
+                defaultModelName: 'gpt-4-turbo',
+            };
+            // Add baseURL if provided in environment
+            if (typeof env.OPENAI_BASE_URL === 'string') {
+                options.baseURL = env.OPENAI_BASE_URL;
+            }
+            return {
+                title: 'Open AI Compatible (from env)',
+                packageName: '@promptbook/openai',
+                className: 'OpenAiCompatibleExecutionTools',
+                options,
+            };
+        }
         return null;
     },
 });