@doclo/providers-llm 0.1.7 → 0.1.9

package/dist/chunk-7YPJIWRM.js

@@ -0,0 +1,291 @@
+// src/schema-prompt-formatter.ts
+function formatSchemaForPrompt(schema, indent = 0) {
+  if (!schema || typeof schema !== "object") {
+    return "";
+  }
+  const indentStr = "  ".repeat(indent);
+  let result = "";
+  if (schema.type === "object" && schema.properties) {
+    const properties = schema.properties;
+    const required = schema.required || [];
+    for (const [fieldName, fieldSchema] of Object.entries(properties)) {
+      const isRequired = required.includes(fieldName);
+      const requiredMarker = isRequired ? " (REQUIRED)" : " (optional)";
+      result += `${indentStr}- \`${fieldName}\`${requiredMarker}`;
+      const type = getTypeDescription(fieldSchema);
+      if (type) {
+        result += `: ${type}`;
+      }
+      if (fieldSchema.description) {
+        result += `
+${indentStr}  ${fieldSchema.description}`;
+      }
+      if (fieldSchema.enum) {
+        result += `
+${indentStr}  Allowed values: ${fieldSchema.enum.map((v) => JSON.stringify(v)).join(", ")}`;
+      }
+      result += "\n";
+      if (fieldSchema.type === "object" && fieldSchema.properties) {
+        result += formatSchemaForPrompt(fieldSchema, indent + 1);
+      }
+      if (fieldSchema.type === "array" && fieldSchema.items) {
+        result += `${indentStr}  Array items:
+`;
+        const itemSchema = Array.isArray(fieldSchema.items) ? fieldSchema.items[0] : fieldSchema.items;
+        if (itemSchema && itemSchema.type === "object" && itemSchema.properties) {
+          result += formatSchemaForPrompt(itemSchema, indent + 2);
+        } else if (itemSchema) {
+          const itemType = getTypeDescription(itemSchema);
+          result += `${indentStr}    ${itemType}
+`;
+        }
+      }
+    }
+  }
+  return result;
+}
+function getTypeDescription(schema) {
+  if (!schema) return "any";
+  if (schema.type) {
+    const typeStr = Array.isArray(schema.type) ? schema.type.join(" | ") : schema.type;
+    if (typeStr === "array" || Array.isArray(schema.type) && schema.type.includes("array")) {
+      if (schema.items && !Array.isArray(schema.items) && schema.items.type) {
+        const itemType = Array.isArray(schema.items.type) ? schema.items.type.join(" | ") : schema.items.type;
+        return `array of ${itemType}`;
+      }
+      return "array";
+    }
+    if ((typeStr === "string" || Array.isArray(schema.type) && schema.type.includes("string")) && schema.format) {
+      const formatHints = {
+        "date": "YYYY-MM-DD",
+        "time": "HH:MM or HH:MM:SS",
+        "date-time": "YYYY-MM-DDTHH:MM:SS (ISO 8601)"
+      };
+      const hint = formatHints[schema.format];
+      if (hint) {
+        return `string (format: ${schema.format}, use ${hint})`;
+      }
+      return `string (format: ${schema.format})`;
+    }
+    return typeStr;
+  }
+  if (schema.anyOf) {
+    return schema.anyOf.map((s) => getTypeDescription(s)).join(" OR ");
+  }
+  if (schema.oneOf) {
+    return schema.oneOf.map((s) => getTypeDescription(s)).join(" OR ");
+  }
+  return "any";
+}
+function buildSchemaPromptSection(schema) {
+  const schemaFields = formatSchemaForPrompt(schema);
+  return `
+==================================================
+CRITICAL: OUTPUT STRUCTURE REQUIREMENTS
+==================================================
+
+YOU MUST RETURN JSON MATCHING THIS EXACT STRUCTURE:
+
+${schemaFields}
+
+CRITICAL FIELD NAME REQUIREMENTS:
+\u2713 Use EXACTLY the field names shown above (character-for-character match)
+\u2713 Preserve the exact casing (e.g., "fullName", not "full_name" or "FullName")
+\u2713 Do NOT abbreviate field names (e.g., "dob" instead of "dateOfBirth")
+\u2713 Do NOT invent alternative names (e.g., "directorName" instead of "fullName")
+\u2713 Do NOT use snake_case if the schema uses camelCase
+\u2713 Do NOT flatten nested structures or rename nested fields
+\u2713 The schema above is the SINGLE SOURCE OF TRUTH for field naming
+
+MISSING DATA:
+- If a required field has no data in the document, use null
+- If an optional field has no data, you may omit it or use null
+- Do NOT invent data that isn't in the document
+
+==================================================
+`.trim();
+}
+function combineSchemaAndUserPrompt(schema, userPrompt) {
+  const schemaSection = buildSchemaPromptSection(schema);
+  if (!userPrompt || userPrompt.trim() === "") {
+    return schemaSection + "\n\nTASK: Extract structured data from the provided document.";
+  }
+  return schemaSection + "\n\n" + userPrompt;
+}
+function buildOutputFormatPrompt(options) {
+  const parts = [];
+  if (options.outputFormat) {
+    switch (options.outputFormat) {
+      case "markdown":
+        parts.push("Format all text content using markdown syntax. Use proper headings (#, ##, ###), lists (-, *), bold (**text**), and other markdown formatting where appropriate.");
+        break;
+      case "html":
+        parts.push("Format all text content as valid HTML. Use semantic tags like <p>, <h1>-<h6>, <ul>, <ol>, <strong>, <em> where appropriate.");
+        break;
+      case "json":
+        parts.push("For text fields that contain structured data, format them as embedded JSON strings where appropriate.");
+        break;
+      case "text":
+        parts.push("Return plain text without any formatting. No markdown, HTML, or other markup.");
+        break;
+    }
+  }
+  if (options.tableFormat) {
+    switch (options.tableFormat) {
+      case "markdown":
+        parts.push("Format all tables using markdown table syntax with | column separators and header row with ---.");
+        break;
+      case "html":
+        parts.push("Format all tables as HTML <table> elements with <thead>, <tbody>, <tr>, <th>, and <td> tags.");
+        break;
+      case "csv":
+        parts.push("Format all tables as CSV with headers in the first row and comma-separated values.");
+        break;
+    }
+  }
+  if (options.pageMarkers) {
+    parts.push('Insert "---" page break markers between content from different pages of the document.');
+  }
+  return parts.join("\n");
+}
+function buildLanguageHintsPrompt(languages) {
+  if (!languages || languages.length === 0) {
+    return "";
+  }
+  return `The document is written in ${languages.join(", ")}. Extract and preserve text in the original language(s).`;
+}
+function buildConfidencePrompt() {
+  return `
+For each extracted field, assess your confidence level and include it in the "_confidence" object:
+- Use a number from 0.0 to 1.0 where:
+  - 0.9-1.0: Very high confidence - text is clear and unambiguous
+  - 0.7-0.9: High confidence - minor ambiguity but likely correct
+  - 0.5-0.7: Medium confidence - some uncertainty or partial visibility
+  - 0.3-0.5: Low confidence - significant uncertainty
+  - 0.0-0.3: Very low confidence - guessing or text was unclear
+
+Include "_confidence" as a sibling object mapping field paths to their scores.
+Example: "_confidence": { "invoiceNumber": 0.95, "amount": 0.82 }
+`.trim();
+}
+function buildSourcesPrompt() {
+  return `
+For each extracted field, identify the source location in the document and include it in the "_sources" array:
+Each source entry should contain:
+- "field": The field name/path that was extracted
+- "text": The exact text from the document used for extraction
+- "bbox": Bounding box as [y_min, x_min, y_max, x_max] normalized to 0-1000 scale
+- "page": Page number (0-indexed) where the text appears
+
+Include "_sources" as a sibling array to your extracted data.
+Example: "_sources": [{ "field": "invoiceNumber", "text": "INV-001", "bbox": [100, 50, 120, 150], "page": 0 }]
+`.trim();
+}
+function buildBlockClassificationPrompt() {
+  return `
+For each extracted element or text block, classify its type in a "_blockTypes" object:
+- "title": Main document title or major section headers
+- "heading": Section headings and subheadings
+- "paragraph": Body text paragraphs
+- "table": Tabular data
+- "list": Bulleted or numbered lists
+- "header": Page headers (repeated at top of pages)
+- "footer": Page footers (repeated at bottom of pages)
+- "caption": Image or figure captions
+- "code": Code blocks or preformatted text
+
+Include "_blockTypes" mapping field paths to their block type.
+Example: "_blockTypes": { "summary": "paragraph", "items": "list" }
+`.trim();
+}
+function buildHeaderFooterPrompt(options) {
+  const parts = [];
+  if (options.extractHeaders) {
+    parts.push('Identify and extract document headers (repeated content at the top of pages) into a "_headers" array.');
+  }
+  if (options.extractFooters) {
+    parts.push('Identify and extract document footers (repeated content at the bottom of pages, like page numbers) into a "_footers" array.');
+  }
+  if (parts.length > 0) {
+    parts.push('Each header/footer entry should include: { "text": "...", "pages": [0, 1, 2] } listing which pages contain it.');
+  }
+  return parts.join("\n");
+}
+function buildChunkingPrompt(strategy, maxChunkSize) {
+  const sizeNote = maxChunkSize ? ` Keep chunks under ${maxChunkSize} characters when possible.` : "";
+  switch (strategy) {
+    case "page":
+      return `Organize the extracted content by page. Include page number for each chunk.${sizeNote}`;
+    case "section":
+      return `Divide the document into logical sections based on headings and structure. Each section should be a coherent unit.${sizeNote}`;
+    case "paragraph":
+      return `Divide the content into individual paragraphs, preserving the natural paragraph breaks from the document.${sizeNote}`;
+    case "semantic":
+      return `Divide the document into semantically coherent chunks. Each chunk should be a self-contained unit of meaning that could stand alone.${sizeNote}`;
+    default:
+      return "";
+  }
+}
+function buildLLMDerivedFeaturesPrompt(options) {
+  const parts = [];
+  const formatPrompt = buildOutputFormatPrompt(options);
+  if (formatPrompt) {
+    parts.push(formatPrompt);
+  }
+  if (options.languageHints && options.languageHints.length > 0) {
+    parts.push(buildLanguageHintsPrompt(options.languageHints));
+  }
+  if (options.includeConfidence) {
+    parts.push(buildConfidencePrompt());
+  }
+  if (options.includeSources) {
+    parts.push(buildSourcesPrompt());
+  }
+  if (options.includeBlockTypes) {
+    parts.push(buildBlockClassificationPrompt());
+  }
+  if (options.extractHeaders || options.extractFooters) {
+    parts.push(buildHeaderFooterPrompt(options));
+  }
+  if (options.chunkingStrategy) {
+    parts.push(buildChunkingPrompt(options.chunkingStrategy, options.maxChunkSize));
+  }
+  if (parts.length === 0) {
+    return "";
+  }
+  return `
+==================================================
+ADDITIONAL OUTPUT REQUIREMENTS
+==================================================
+
+${parts.join("\n\n")}
+
+==================================================
+`.trim();
+}
+function combineSchemaUserAndDerivedPrompts(schema, userPrompt, derivedOptions) {
+  let result = combineSchemaAndUserPrompt(schema, userPrompt);
+  if (derivedOptions) {
+    const derivedPrompt = buildLLMDerivedFeaturesPrompt(derivedOptions);
+    if (derivedPrompt) {
+      result = result + "\n\n" + derivedPrompt;
+    }
+  }
+  return result;
+}
+
+export {
+  formatSchemaForPrompt,
+  buildSchemaPromptSection,
+  combineSchemaAndUserPrompt,
+  buildOutputFormatPrompt,
+  buildLanguageHintsPrompt,
+  buildConfidencePrompt,
+  buildSourcesPrompt,
+  buildBlockClassificationPrompt,
+  buildHeaderFooterPrompt,
+  buildChunkingPrompt,
+  buildLLMDerivedFeaturesPrompt,
+  combineSchemaUserAndDerivedPrompts
+};
+//# sourceMappingURL=chunk-7YPJIWRM.js.map

package/dist/chunk-7YPJIWRM.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/schema-prompt-formatter.ts"],"sourcesContent":["/**\n * Utility for converting JSON Schema to human-readable prompt text\n * that emphasizes exact field name requirements for structured extraction.\n */\n\n/**\n * JSON Schema type used for prompt formatting.\n * Uses a recursive structure to support nested schemas.\n */\nexport interface JSONSchema {\n  type?: string | string[];  // Can be array for union types (e.g., [\"string\", \"null\"])\n  properties?: Record<string, JSONSchema>;\n  items?: JSONSchema | JSONSchema[];  // Can be array for tuple validation\n  description?: string;\n  required?: string[];\n  enum?: (string | number | boolean | null)[];\n  anyOf?: JSONSchema[];\n  oneOf?: JSONSchema[];\n  allOf?: JSONSchema[];\n  format?: string;\n  [key: string]: unknown;  // Allow additional properties\n}\n\n/**\n * Formats a JSON Schema into prompt text that emphasizes exact field names.\n * This helps LLMs understand they must use the exact field names specified\n * in the schema, not invent their own based on document content.\n */\nexport function formatSchemaForPrompt(schema: JSONSchema, indent: number = 0): string {\n  if (!schema || typeof schema !== 'object') {\n    return '';\n  }\n\n  const indentStr = '  '.repeat(indent);\n  let result = '';\n\n  // Handle object type with properties\n  if (schema.type === 'object' && schema.properties) {\n    const properties = schema.properties;\n    const required = schema.required || [];\n\n    for (const [fieldName, fieldSchema] of Object.entries(properties)) {\n      const isRequired = required.includes(fieldName);\n      const requiredMarker = isRequired ? ' (REQUIRED)' : ' (optional)';\n\n      // Field name in backticks to emphasize exactness\n      result += `${indentStr}- \\`${fieldName}\\`${requiredMarker}`;\n\n      // Type information\n      const type = getTypeDescription(fieldSchema);\n      if (type) {\n        result += `: ${type}`;\n      }\n\n      // Description if available\n      if (fieldSchema.description) {\n        result += `\\n${indentStr}  ${fieldSchema.description}`;\n      }\n\n      // Enum values if specified\n      if (fieldSchema.enum) {\n        result += `\\n${indentStr}  Allowed values: ${fieldSchema.enum.map((v) => JSON.stringify(v)).join(', ')}`;\n      }\n\n      result += '\\n';\n\n      // Nested object properties\n      if (fieldSchema.type === 'object' && fieldSchema.properties) {\n        result += formatSchemaForPrompt(fieldSchema, indent + 1);\n      }\n\n      // Array item schema\n      if (fieldSchema.type === 'array' && fieldSchema.items) {\n        result += `${indentStr}  Array items:\\n`;\n        // Handle both single schema and tuple schemas (array of schemas)\n        const itemSchema = Array.isArray(fieldSchema.items)\n          ? fieldSchema.items[0]  // For tuple validation, describe first item type\n          : fieldSchema.items;\n        if (itemSchema && itemSchema.type === 'object' && itemSchema.properties) {\n          result += formatSchemaForPrompt(itemSchema, indent + 2);\n        } else if (itemSchema) {\n          const itemType = getTypeDescription(itemSchema);\n          result += `${indentStr}    ${itemType}\\n`;\n        }\n      }\n    }\n  }\n\n  return result;\n}\n\n/**\n * Gets a human-readable type description from a schema property\n */\nfunction getTypeDescription(schema: JSONSchema): string {\n  if (!schema) return 'any';\n\n  if (schema.type) {\n    // Handle array of types (e.g., [\"string\", \"null\"])\n    const typeStr = Array.isArray(schema.type) ? schema.type.join(' | ') : schema.type;\n\n    if (typeStr === 'array' || (Array.isArray(schema.type) && schema.type.includes('array'))) {\n      if (schema.items && !Array.isArray(schema.items) && schema.items.type) {\n        const itemType = Array.isArray(schema.items.type)\n          ? schema.items.type.join(' | ')\n          : schema.items.type;\n        return `array of ${itemType}`;\n      }\n      return 'array';\n    }\n    // Include format information for strings (e.g., date, time, date-time, email, uri)\n    if ((typeStr === 'string' || (Array.isArray(schema.type) && schema.type.includes('string'))) && schema.format) {\n      const formatHints: Record<string, string> = {\n        'date': 'YYYY-MM-DD',\n        'time': 'HH:MM or HH:MM:SS',\n        'date-time': 'YYYY-MM-DDTHH:MM:SS (ISO 8601)',\n      };\n      const hint = formatHints[schema.format];\n      if (hint) {\n        return `string (format: ${schema.format}, use ${hint})`;\n      }\n      return `string (format: ${schema.format})`;\n    }\n    return typeStr;\n  }\n\n  // Handle anyOf, oneOf, allOf\n  if (schema.anyOf) {\n    return schema.anyOf.map((s) => getTypeDescription(s)).join(' OR ');\n  }\n  if (schema.oneOf) {\n    return schema.oneOf.map((s) => getTypeDescription(s)).join(' OR ');\n  }\n\n  return 'any';\n}\n\n/**\n * Generates a complete prompt section with schema information and\n * strict field name instructions.\n */\nexport function buildSchemaPromptSection(schema: JSONSchema): string {\n  const schemaFields = formatSchemaForPrompt(schema);\n\n  return `\n==================================================\nCRITICAL: OUTPUT STRUCTURE REQUIREMENTS\n==================================================\n\nYOU MUST RETURN JSON MATCHING THIS EXACT STRUCTURE:\n\n${schemaFields}\n\nCRITICAL FIELD NAME REQUIREMENTS:\n✓ Use EXACTLY the field names shown above (character-for-character match)\n✓ Preserve the exact casing (e.g., \"fullName\", not \"full_name\" or \"FullName\")\n✓ Do NOT abbreviate field names (e.g., \"dob\" instead of \"dateOfBirth\")\n✓ Do NOT invent alternative names (e.g., \"directorName\" instead of \"fullName\")\n✓ Do NOT use snake_case if the schema uses camelCase\n✓ Do NOT flatten nested structures or rename nested fields\n✓ The schema above is the SINGLE SOURCE OF TRUTH for field naming\n\nMISSING DATA:\n- If a required field has no data in the document, use null\n- If an optional field has no data, you may omit it or use null\n- Do NOT invent data that isn't in the document\n\n==================================================\n`.trim();\n}\n\n/**\n * Combines schema prompt section with user's custom prompt\n */\nexport function combineSchemaAndUserPrompt(\n  schema: JSONSchema,\n  userPrompt: string\n): string {\n  const schemaSection = buildSchemaPromptSection(schema);\n\n  if (!userPrompt || userPrompt.trim() === '') {\n    return schemaSection + '\\n\\nTASK: Extract structured data from the provided document.';\n  }\n\n  return schemaSection + '\\n\\n' + userPrompt;\n}\n\n// ============================================================================\n// LLM-Derived Feature Prompts\n// ============================================================================\n\n/**\n * Output format types for LLM text generation\n */\nexport type OutputFormat = 'markdown' | 'html' | 'json' | 'text';\nexport type TableFormat = 'markdown' | 'html' | 'csv';\nexport type ChunkingStrategy = 'page' | 'section' | 'paragraph' | 'semantic';\n\n/**\n * Options for LLM-derived features that are implemented via prompting\n */\nexport interface LLMDerivedPromptOptions {\n  outputFormat?: OutputFormat;\n  tableFormat?: TableFormat;\n  pageMarkers?: boolean;\n  includeConfidence?: boolean;\n  includeSources?: boolean;\n  includeBlockTypes?: boolean;\n  extractHeaders?: boolean;\n  extractFooters?: boolean;\n  chunkingStrategy?: ChunkingStrategy;\n  maxChunkSize?: number;\n  languageHints?: string[];\n}\n\n/**\n * Builds prompt additions for output format options\n */\nexport function buildOutputFormatPrompt(options: LLMDerivedPromptOptions): string {\n  const parts: string[] = [];\n\n  // Output format\n  if (options.outputFormat) {\n    switch (options.outputFormat) {\n      case 'markdown':\n        parts.push('Format all text content using markdown syntax. Use proper headings (#, ##, ###), lists (-, *), bold (**text**), and other markdown formatting where appropriate.');\n        break;\n      case 'html':\n        parts.push('Format all text content as valid HTML. Use semantic tags like <p>, <h1>-<h6>, <ul>, <ol>, <strong>, <em> where appropriate.');\n        break;\n      case 'json':\n        parts.push('For text fields that contain structured data, format them as embedded JSON strings where appropriate.');\n        break;\n      case 'text':\n        parts.push('Return plain text without any formatting. No markdown, HTML, or other markup.');\n        break;\n    }\n  }\n\n  // Table format\n  if (options.tableFormat) {\n    switch (options.tableFormat) {\n      case 'markdown':\n        parts.push('Format all tables using markdown table syntax with | column separators and header row with ---.');\n        break;\n      case 'html':\n        parts.push('Format all tables as HTML <table> elements with <thead>, <tbody>, <tr>, <th>, and <td> tags.');\n        break;\n      case 'csv':\n        parts.push('Format all tables as CSV with headers in the first row and comma-separated values.');\n        break;\n    }\n  }\n\n  // Page markers\n  if (options.pageMarkers) {\n    parts.push('Insert \"---\" page break markers between content from different pages of the document.');\n  }\n\n  return parts.join('\\n');\n}\n\n/**\n * Builds prompt additions for language hints\n */\nexport function buildLanguageHintsPrompt(languages: string[]): string {\n  if (!languages || languages.length === 0) {\n    return '';\n  }\n  return `The document is written in ${languages.join(', ')}. Extract and preserve text in the original language(s).`;\n}\n\n/**\n * Builds prompt additions for confidence scoring\n */\nexport function buildConfidencePrompt(): string {\n  return `\nFor each extracted field, assess your confidence level and include it in the \"_confidence\" object:\n- Use a number from 0.0 to 1.0 where:\n  - 0.9-1.0: Very high confidence - text is clear and unambiguous\n  - 0.7-0.9: High confidence - minor ambiguity but likely correct\n  - 0.5-0.7: Medium confidence - some uncertainty or partial visibility\n  - 0.3-0.5: Low confidence - significant uncertainty\n  - 0.0-0.3: Very low confidence - guessing or text was unclear\n\nInclude \"_confidence\" as a sibling object mapping field paths to their scores.\nExample: \"_confidence\": { \"invoiceNumber\": 0.95, \"amount\": 0.82 }\n`.trim();\n}\n\n/**\n * Builds prompt additions for source citations with bounding boxes\n */\nexport function buildSourcesPrompt(): string {\n  return `\nFor each extracted field, identify the source location in the document and include it in the \"_sources\" array:\nEach source entry should contain:\n- \"field\": The field name/path that was extracted\n- \"text\": The exact text from the document used for extraction\n- \"bbox\": Bounding box as [y_min, x_min, y_max, x_max] normalized to 0-1000 scale\n- \"page\": Page number (0-indexed) where the text appears\n\nInclude \"_sources\" as a sibling array to your extracted data.\nExample: \"_sources\": [{ \"field\": \"invoiceNumber\", \"text\": \"INV-001\", \"bbox\": [100, 50, 120, 150], \"page\": 0 }]\n`.trim();\n}\n\n/**\n * Builds prompt additions for block type classification\n */\nexport function buildBlockClassificationPrompt(): string {\n  return `\nFor each extracted element or text block, classify its type in a \"_blockTypes\" object:\n- \"title\": Main document title or major section headers\n- \"heading\": Section headings and subheadings\n- \"paragraph\": Body text paragraphs\n- \"table\": Tabular data\n- \"list\": Bulleted or numbered lists\n- \"header\": Page headers (repeated at top of pages)\n- \"footer\": Page footers (repeated at bottom of pages)\n- \"caption\": Image or figure captions\n- \"code\": Code blocks or preformatted text\n\nInclude \"_blockTypes\" mapping field paths to their block type.\nExample: \"_blockTypes\": { \"summary\": \"paragraph\", \"items\": \"list\" }\n`.trim();\n}\n\n/**\n * Builds prompt additions for header/footer extraction\n */\nexport function buildHeaderFooterPrompt(options: { extractHeaders?: boolean; extractFooters?: boolean }): string {\n  const parts: string[] = [];\n\n  if (options.extractHeaders) {\n    parts.push('Identify and extract document headers (repeated content at the top of pages) into a \"_headers\" array.');\n  }\n\n  if (options.extractFooters) {\n    parts.push('Identify and extract document footers (repeated content at the bottom of pages, like page numbers) into a \"_footers\" array.');\n  }\n\n  if (parts.length > 0) {\n    parts.push('Each header/footer entry should include: { \"text\": \"...\", \"pages\": [0, 1, 2] } listing which pages contain it.');\n  }\n\n  return parts.join('\\n');\n}\n\n/**\n * Builds prompt additions for semantic chunking\n */\nexport function buildChunkingPrompt(strategy: ChunkingStrategy, maxChunkSize?: number): string {\n  const sizeNote = maxChunkSize\n    ? ` Keep chunks under ${maxChunkSize} characters when possible.`\n    : '';\n\n  switch (strategy) {\n    case 'page':\n      return `Organize the extracted content by page. Include page number for each chunk.${sizeNote}`;\n    case 'section':\n      return `Divide the document into logical sections based on headings and structure. Each section should be a coherent unit.${sizeNote}`;\n    case 'paragraph':\n      return `Divide the content into individual paragraphs, preserving the natural paragraph breaks from the document.${sizeNote}`;\n    case 'semantic':\n      return `Divide the document into semantically coherent chunks. Each chunk should be a self-contained unit of meaning that could stand alone.${sizeNote}`;\n    default:\n      return '';\n  }\n}\n\n/**\n * Combines all LLM-derived feature prompts into a single prompt section\n */\nexport function buildLLMDerivedFeaturesPrompt(options: LLMDerivedPromptOptions): string {\n  const parts: string[] = [];\n\n  // Output format options\n  const formatPrompt = buildOutputFormatPrompt(options);\n  if (formatPrompt) {\n    parts.push(formatPrompt);\n  }\n\n  // Language hints\n  if (options.languageHints && options.languageHints.length > 0) {\n    parts.push(buildLanguageHintsPrompt(options.languageHints));\n  }\n\n  // Metadata features (confidence, sources, block types)\n  if (options.includeConfidence) {\n    parts.push(buildConfidencePrompt());\n  }\n\n  if (options.includeSources) {\n    parts.push(buildSourcesPrompt());\n  }\n\n  if (options.includeBlockTypes) {\n    parts.push(buildBlockClassificationPrompt());\n  }\n\n  // Header/footer extraction\n  if (options.extractHeaders || options.extractFooters) {\n    parts.push(buildHeaderFooterPrompt(options));\n  }\n\n  // Chunking strategy\n  if (options.chunkingStrategy) {\n    parts.push(buildChunkingPrompt(options.chunkingStrategy, options.maxChunkSize));\n  }\n\n  if (parts.length === 0) {\n    return '';\n  }\n\n  return `\n==================================================\nADDITIONAL OUTPUT REQUIREMENTS\n==================================================\n\n${parts.join('\\n\\n')}\n\n==================================================\n`.trim();\n}\n\n/**\n * Combines schema prompt with user prompt and LLM-derived features\n */\nexport function combineSchemaUserAndDerivedPrompts(\n  schema: JSONSchema,\n  userPrompt: string,\n  derivedOptions?: LLMDerivedPromptOptions\n): string {\n  let result = combineSchemaAndUserPrompt(schema, userPrompt);\n\n  if (derivedOptions) {\n    const derivedPrompt = buildLLMDerivedFeaturesPrompt(derivedOptions);\n    if (derivedPrompt) {\n      result = result + '\\n\\n' + derivedPrompt;\n    }\n  }\n\n  return result;\n}\n"],"mappings":";AA4BO,SAAS,sBAAsB,QAAoB,SAAiB,GAAW;AACpF,MAAI,CAAC,UAAU,OAAO,WAAW,UAAU;AACzC,WAAO;AAAA,EACT;AAEA,QAAM,YAAY,KAAK,OAAO,MAAM;AACpC,MAAI,SAAS;AAGb,MAAI,OAAO,SAAS,YAAY,OAAO,YAAY;AACjD,UAAM,aAAa,OAAO;AAC1B,UAAM,WAAW,OAAO,YAAY,CAAC;AAErC,eAAW,CAAC,WAAW,WAAW,KAAK,OAAO,QAAQ,UAAU,GAAG;AACjE,YAAM,aAAa,SAAS,SAAS,SAAS;AAC9C,YAAM,iBAAiB,aAAa,gBAAgB;AAGpD,gBAAU,GAAG,SAAS,OAAO,SAAS,KAAK,cAAc;AAGzD,YAAM,OAAO,mBAAmB,WAAW;AAC3C,UAAI,MAAM;AACR,kBAAU,KAAK,IAAI;AAAA,MACrB;AAGA,UAAI,YAAY,aAAa;AAC3B,kBAAU;AAAA,EAAK,SAAS,KAAK,YAAY,WAAW;AAAA,MACtD;AAGA,UAAI,YAAY,MAAM;AACpB,kBAAU;AAAA,EAAK,SAAS,qBAAqB,YAAY,KAAK,IAAI,CAAC,MAAM,KAAK,UAAU,CAAC,CAAC,EAAE,KAAK,IAAI,CAAC;AAAA,MACxG;AAEA,gBAAU;AAGV,UAAI,YAAY,SAAS,YAAY,YAAY,YAAY;AAC3D,kBAAU,sBAAsB,aAAa,SAAS,CAAC;AAAA,MACzD;AAGA,UAAI,YAAY,SAAS,WAAW,YAAY,OAAO;AACrD,kBAAU,GAAG,SAAS;AAAA;AAEtB,cAAM,aAAa,MAAM,QAAQ,YAAY,KAAK,IAC9C,YAAY,MAAM,CAAC,IACnB,YAAY;AAChB,YAAI,cAAc,WAAW,SAAS,YAAY,WAAW,YAAY;AACvE,oBAAU,sBAAsB,YAAY,SAAS,CAAC;AAAA,QACxD,WAAW,YAAY;AACrB,gBAAM,WAAW,mBAAmB,UAAU;AAC9C,oBAAU,GAAG,SAAS,OAAO,QAAQ;AAAA;AAAA,QACvC;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,SAAO;AACT;AAKA,SAAS,mBAAmB,QAA4B;AACtD,MAAI,CAAC,OAAQ,QAAO;AAEpB,MAAI,OAAO,MAAM;AAEf,UAAM,UAAU,MAAM,QAAQ,OAAO,IAAI,IAAI,OAAO,KAAK,KAAK,KAAK,IAAI,OAAO;AAE9E,QAAI,YAAY,WAAY,MAAM,QAAQ,OAAO,IAAI,KAAK,OAAO,KAAK,SAAS,OAAO,GAAI;AACxF,UAAI,OAAO,SAAS,CAAC,MAAM,QAAQ,OAAO,KAAK,KAAK,OAAO,MAAM,MAAM;AACrE,cAAM,WAAW,MAAM,QAAQ,OAAO,MAAM,IAAI,IAC5C,OAAO,MAAM,KAAK,KAAK,KAAK,IAC5B,OAAO,MAAM;AACjB,eAAO,YAAY,QAAQ;AAAA,MAC7B;AACA,aAAO;AAAA,IACT;AAEA,SAAK,YAAY,YAAa,MAAM,QAAQ,OAAO,IAAI,KAAK,OAAO,KAAK,SAAS,QAAQ,MAAO,OAAO,QAAQ;AAC7G,YAAM,cAAsC;AAAA,QAC1C,QAAQ;AAAA,QACR,QAAQ;AAAA,QACR,aAAa;AAAA,MACf;AACA,YAAM,OAAO,YAAY,OAAO,MAAM;AACtC,UAAI,MAAM;AACR,eAAO,mBAAmB,OAAO,MAAM,SAAS,IAAI;AAAA,MACtD;AACA,aAAO,mBAAmB,OAAO,MAAM;AAAA,IACzC;AACA,WAAO;AAAA,EACT;AAGA,MAAI,OAAO,OAAO;AAChB,WAAO,OAAO,MAAM,IAAI,CAAC,MAAM,mBAAmB,CAAC,CAAC,EAAE,KAAK,MAAM;AAAA,EACnE;AACA,MAAI,OAAO,OAAO;AAChB,WAAO,OAAO,MAAM,IAAI,CAAC,MAAM,mBAAmB,CAAC,CAAC,EAAE,KAAK,MAAM;AAAA,EACnE;AAEA,SAAO;AACT;AAMO,SAAS,yBAAyB,QAA4B;AACnE,QAAM,eAAe,sBAAsB,MAAM;AAEjD,SAAO;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOP,YAAY;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAiBZ,KAAK;AACP;AAKO,SAAS,2BACd,QACA,YACQ;AACR,QAAM,gBAAgB,yBAAyB,MAAM;AAErD,MAAI,CAAC,cAAc,WAAW,KAAK,MAAM,IAAI;AAC3C,WAAO,gBAAgB;AAAA,EACzB;AAEA,SAAO,gBAAgB,SAAS;AAClC;AAiCO,SAAS,wBAAwB,SAA0C;AAChF,QAAM,QAAkB,CAAC;AAGzB,MAAI,QAAQ,cAAc;AACxB,YAAQ,QAAQ,cAAc;AAAA,MAC5B,KAAK;AACH,cAAM,KAAK,kKAAkK;AAC7K;AAAA,MACF,KAAK;AACH,cAAM,KAAK,6HAA6H;AACxI;AAAA,MACF,KAAK;AACH,cAAM,KAAK,uGAAuG;AAClH;AAAA,MACF,KAAK;AACH,cAAM,KAAK,+EAA+E;AAC1F;AAAA,IACJ;AAAA,EACF;AAGA,MAAI,QAAQ,aAAa;AACvB,YAAQ,QAAQ,aAAa;AAAA,MAC3B,KAAK;AACH,cAAM,KAAK,iGAAiG;AAC5G;AAAA,MACF,KAAK;AACH,cAAM,KAAK,8FAA8F;AACzG;AAAA,MACF,KAAK;AACH,cAAM,KAAK,oFAAoF;AAC/F;AAAA,IACJ;AAAA,EACF;AAGA,MAAI,QAAQ,aAAa;AACvB,UAAM,KAAK,uFAAuF;AAAA,EACpG;AAEA,SAAO,MAAM,KAAK,IAAI;AACxB;AAKO,SAAS,yBAAyB,WAA6B;AACpE,MAAI,CAAC,aAAa,UAAU,WAAW,GAAG;AACxC,WAAO;AAAA,EACT;AACA,SAAO,8BAA8B,UAAU,KAAK,IAAI,CAAC;AAC3D;AAKO,SAAS,wBAAgC;AAC9C,SAAO;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAWP,KAAK;AACP;AAKO,SAAS,qBAA6B;AAC3C,SAAO;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAUP,KAAK;AACP;AAKO,SAAS,iCAAyC;AACvD,SAAO;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAcP,KAAK;AACP;AAKO,SAAS,wBAAwB,SAAyE;AAC/G,QAAM,QAAkB,CAAC;AAEzB,MAAI,QAAQ,gBAAgB;AAC1B,UAAM,KAAK,uGAAuG;AAAA,EACpH;AAEA,MAAI,QAAQ,gBAAgB;AAC1B,UAAM,KAAK,6HAA6H;AAAA,EAC1I;AAEA,MAAI,MAAM,SAAS,GAAG;AACpB,UAAM,KAAK,gHAAgH;AAAA,EAC7H;AAEA,SAAO,MAAM,KAAK,IAAI;AACxB;AAKO,SAAS,oBAAoB,UAA4B,cAA+B;AAC7F,QAAM,WAAW,eACb,sBAAsB,YAAY,+BAClC;AAEJ,UAAQ,UAAU;AAAA,IAChB,KAAK;AACH,aAAO,8EAA8E,QAAQ;AAAA,IAC/F,KAAK;AACH,aAAO,qHAAqH,QAAQ;AAAA,IACtI,KAAK;AACH,aAAO,4GAA4G,QAAQ;AAAA,IAC7H,KAAK;AACH,aAAO,uIAAuI,QAAQ;AAAA,IACxJ;AACE,aAAO;AAAA,EACX;AACF;AAKO,SAAS,8BAA8B,SAA0C;AACtF,QAAM,QAAkB,CAAC;AAGzB,QAAM,eAAe,wBAAwB,OAAO;AACpD,MAAI,cAAc;AAChB,UAAM,KAAK,YAAY;AAAA,EACzB;AAGA,MAAI,QAAQ,iBAAiB,QAAQ,cAAc,SAAS,GAAG;AAC7D,UAAM,KAAK,yBAAyB,QAAQ,aAAa,CAAC;AAAA,EAC5D;AAGA,MAAI,QAAQ,mBAAmB;AAC7B,UAAM,KAAK,sBAAsB,CAAC;AAAA,EACpC;AAEA,MAAI,QAAQ,gBAAgB;AAC1B,UAAM,KAAK,mBAAmB,CAAC;AAAA,EACjC;AAEA,MAAI,QAAQ,mBAAmB;AAC7B,UAAM,KAAK,+BAA+B,CAAC;AAAA,EAC7C;AAGA,MAAI,QAAQ,kBAAkB,QAAQ,gBAAgB;AACpD,UAAM,KAAK,wBAAwB,OAAO,CAAC;AAAA,EAC7C;AAGA,MAAI,QAAQ,kBAAkB;AAC5B,UAAM,KAAK,oBAAoB,QAAQ,kBAAkB,QAAQ,YAAY,CAAC;AAAA,EAChF;AAEA,MAAI,MAAM,WAAW,GAAG;AACtB,WAAO;AAAA,EACT;AAEA,SAAO;AAAA;AAAA;AAAA;AAAA;AAAA,EAKP,MAAM,KAAK,MAAM,CAAC;AAAA;AAAA;AAAA,EAGlB,KAAK;AACP;AAKO,SAAS,mCACd,QACA,YACA,gBACQ;AACR,MAAI,SAAS,2BAA2B,QAAQ,UAAU;AAE1D,MAAI,gBAAgB;AAClB,UAAM,gBAAgB,8BAA8B,cAAc;AAClE,QAAI,eAAe;AACjB,eAAS,SAAS,SAAS;AAAA,IAC7B;AAAA,EACF;AAEA,SAAO;AACT;","names":[]}