@heripo/pdf-parser 0.1.4 → 0.1.6

package/dist/chunk-WWNI354M.js

@@ -0,0 +1,121 @@
+// src/config/vlm-models.ts
+var VLM_MODELS = {
+  // ── DocTags models (specialized document structure output) ──────────
+  "granite-docling-258M-mlx": {
+    repo_id: "ibm-granite/granite-docling-258M-mlx",
+    inference_framework: "mlx",
+    response_format: "doctags",
+    transformers_model_type: "automodel-vision2seq",
+    description: "Granite Docling 258M (MLX, Apple Silicon optimized, ~6s/page)"
+  },
+  "granite-docling-258M": {
+    repo_id: "ibm-granite/granite-docling-258M",
+    inference_framework: "transformers",
+    response_format: "doctags",
+    transformers_model_type: "automodel-vision2seq",
+    description: "Granite Docling 258M (Transformers, cross-platform)"
+  },
+  "smoldocling-256M-mlx": {
+    repo_id: "docling-project/SmolDocling-256M-preview-mlx-bf16",
+    inference_framework: "mlx",
+    response_format: "doctags",
+    transformers_model_type: "automodel-vision2seq",
+    description: "SmolDocling 256M (MLX, fastest option)"
+  },
+  "smoldocling-256M": {
+    repo_id: "docling-project/SmolDocling-256M-preview",
+    inference_framework: "transformers",
+    response_format: "doctags",
+    transformers_model_type: "automodel-vision2seq",
+    description: "SmolDocling 256M (Transformers)"
+  },
+  // ── Markdown models (general-purpose vision LLMs) ──────────────────
+  "granite-vision-2B": {
+    repo_id: "ibm-granite/granite-vision-3.2-2b",
+    inference_framework: "transformers",
+    response_format: "markdown",
+    transformers_model_type: "automodel-vision2seq",
+    description: "Granite Vision 3.2 2B (IBM, higher accuracy)"
+  },
+  "qwen25-vl-3B-mlx": {
+    repo_id: "mlx-community/Qwen2.5-VL-3B-Instruct-bf16",
+    inference_framework: "mlx",
+    response_format: "markdown",
+    transformers_model_type: "automodel-vision2seq",
+    description: "Qwen 2.5 VL 3B (MLX, multilingual, good KCJ support)"
+  },
+  phi4: {
+    repo_id: "microsoft/Phi-4-multimodal-instruct",
+    inference_framework: "transformers",
+    response_format: "markdown",
+    transformers_model_type: "automodel",
+    description: "Phi-4 Multimodal (Microsoft, CausalLM)"
+  },
+  "pixtral-12B-mlx": {
+    repo_id: "mlx-community/pixtral-12b-bf16",
+    inference_framework: "mlx",
+    response_format: "markdown",
+    transformers_model_type: "automodel-vision2seq",
+    description: "Pixtral 12B (MLX, Mistral, high accuracy)"
+  },
+  "pixtral-12B": {
+    repo_id: "mistral-community/pixtral-12b",
+    inference_framework: "transformers",
+    response_format: "markdown",
+    transformers_model_type: "automodel-vision2seq",
+    description: "Pixtral 12B (Transformers, Mistral)"
+  },
+  got2: {
+    repo_id: "stepfun-ai/GOT-OCR-2.0-hf",
+    inference_framework: "transformers",
+    response_format: "markdown",
+    transformers_model_type: "automodel-vision2seq",
+    description: "GOT-OCR 2.0 (StepFun, OCR-specialized)"
+  },
+  "gemma3-12B-mlx": {
+    repo_id: "mlx-community/gemma-3-12b-it-bf16",
+    inference_framework: "mlx",
+    response_format: "markdown",
+    transformers_model_type: "automodel-vision2seq",
+    description: "Gemma 3 12B (MLX, Google)"
+  },
+  "gemma3-27B-mlx": {
+    repo_id: "mlx-community/gemma-3-27b-it-bf16",
+    inference_framework: "mlx",
+    response_format: "markdown",
+    transformers_model_type: "automodel-vision2seq",
+    description: "Gemma 3 27B (MLX, Google, highest accuracy)"
+  },
+  dolphin: {
+    repo_id: "ByteDance/Dolphin",
+    inference_framework: "transformers",
+    response_format: "markdown",
+    transformers_model_type: "automodel-vision2seq",
+    description: "Dolphin (ByteDance, document-oriented)"
+  }
+};
+var DEFAULT_VLM_MODEL = "granite-docling-258M-mlx";
+function resolveVlmModel(model) {
+  if (typeof model === "string") {
+    const preset = VLM_MODELS[model];
+    if (!preset) {
+      throw new Error(
+        `Unknown VLM model preset: "${model}". Available presets: ${Object.keys(VLM_MODELS).join(", ")}`
+      );
+    }
+    return {
+      repo_id: preset.repo_id,
+      inference_framework: preset.inference_framework,
+      response_format: preset.response_format,
+      transformers_model_type: preset.transformers_model_type
+    };
+  }
+  return model;
+}
+
+export {
+  VLM_MODELS,
+  DEFAULT_VLM_MODEL,
+  resolveVlmModel
+};
+//# sourceMappingURL=chunk-WWNI354M.js.map

package/dist/chunk-WWNI354M.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/config/vlm-models.ts"],"sourcesContent":["import type { VlmModelLocal } from 'docling-sdk';\n\n/**\n * VLM model preset with description\n */\nexport interface VlmModelPreset {\n  repo_id: string;\n  inference_framework: 'mlx' | 'transformers';\n  response_format: 'doctags' | 'markdown';\n  transformers_model_type: 'automodel-vision2seq' | 'automodel';\n  description: string;\n}\n\n/**\n * Available VLM model presets\n *\n * Based on Docling's official VLM model specs:\n * https://docling-project.github.io/docling/usage/vision_models/#available-local-models\n *\n * Users can select a preset key or provide a custom VlmModelLocal object.\n */\nexport const VLM_MODELS: Record<string, VlmModelPreset> = {\n  // ── DocTags models (specialized document structure output) ──────────\n\n  'granite-docling-258M-mlx': {\n    repo_id: 'ibm-granite/granite-docling-258M-mlx',\n    inference_framework: 'mlx',\n    response_format: 'doctags',\n    transformers_model_type: 'automodel-vision2seq',\n    description:\n      'Granite Docling 258M (MLX, Apple Silicon optimized, ~6s/page)',\n  },\n  'granite-docling-258M': {\n    repo_id: 'ibm-granite/granite-docling-258M',\n    inference_framework: 'transformers',\n    response_format: 'doctags',\n    transformers_model_type: 'automodel-vision2seq',\n    description: 'Granite Docling 258M (Transformers, cross-platform)',\n  },\n  'smoldocling-256M-mlx': {\n    repo_id: 'docling-project/SmolDocling-256M-preview-mlx-bf16',\n    inference_framework: 'mlx',\n    response_format: 'doctags',\n    transformers_model_type: 'automodel-vision2seq',\n    description: 'SmolDocling 256M (MLX, fastest option)',\n  },\n  'smoldocling-256M': {\n    repo_id: 'docling-project/SmolDocling-256M-preview',\n    inference_framework: 'transformers',\n    response_format: 'doctags',\n    transformers_model_type: 'automodel-vision2seq',\n    description: 'SmolDocling 256M (Transformers)',\n  },\n\n  // ── Markdown models (general-purpose vision LLMs) ──────────────────\n\n  'granite-vision-2B': {\n    repo_id: 'ibm-granite/granite-vision-3.2-2b',\n    inference_framework: 'transformers',\n    response_format: 'markdown',\n    transformers_model_type: 'automodel-vision2seq',\n    description: 'Granite Vision 3.2 2B (IBM, higher accuracy)',\n  },\n  'qwen25-vl-3B-mlx': {\n    repo_id: 'mlx-community/Qwen2.5-VL-3B-Instruct-bf16',\n    inference_framework: 'mlx',\n    response_format: 'markdown',\n    transformers_model_type: 'automodel-vision2seq',\n    description: 'Qwen 2.5 VL 3B (MLX, multilingual, good KCJ support)',\n  },\n  phi4: {\n    repo_id: 'microsoft/Phi-4-multimodal-instruct',\n    inference_framework: 'transformers',\n    response_format: 'markdown',\n    transformers_model_type: 'automodel',\n    description: 'Phi-4 Multimodal (Microsoft, CausalLM)',\n  },\n  'pixtral-12B-mlx': {\n    repo_id: 'mlx-community/pixtral-12b-bf16',\n    inference_framework: 'mlx',\n    response_format: 'markdown',\n    transformers_model_type: 'automodel-vision2seq',\n    description: 'Pixtral 12B (MLX, Mistral, high accuracy)',\n  },\n  'pixtral-12B': {\n    repo_id: 'mistral-community/pixtral-12b',\n    inference_framework: 'transformers',\n    response_format: 'markdown',\n    transformers_model_type: 'automodel-vision2seq',\n    description: 'Pixtral 12B (Transformers, Mistral)',\n  },\n  got2: {\n    repo_id: 'stepfun-ai/GOT-OCR-2.0-hf',\n    inference_framework: 'transformers',\n    response_format: 'markdown',\n    transformers_model_type: 'automodel-vision2seq',\n    description: 'GOT-OCR 2.0 (StepFun, OCR-specialized)',\n  },\n  'gemma3-12B-mlx': {\n    repo_id: 'mlx-community/gemma-3-12b-it-bf16',\n    inference_framework: 'mlx',\n    response_format: 'markdown',\n    transformers_model_type: 'automodel-vision2seq',\n    description: 'Gemma 3 12B (MLX, Google)',\n  },\n  'gemma3-27B-mlx': {\n    repo_id: 'mlx-community/gemma-3-27b-it-bf16',\n    inference_framework: 'mlx',\n    response_format: 'markdown',\n    transformers_model_type: 'automodel-vision2seq',\n    description: 'Gemma 3 27B (MLX, Google, highest accuracy)',\n  },\n  dolphin: {\n    repo_id: 'ByteDance/Dolphin',\n    inference_framework: 'transformers',\n    response_format: 'markdown',\n    transformers_model_type: 'automodel-vision2seq',\n    description: 'Dolphin (ByteDance, document-oriented)',\n  },\n} as const;\n\n/**\n * Default VLM model preset key\n */\nexport const DEFAULT_VLM_MODEL = 'granite-docling-258M-mlx';\n\n/**\n * Resolve a VLM model from a preset key or custom VlmModelLocal object.\n *\n * When using a preset key, only required fields are populated.\n * Optional fields (prompt, scale, extra_generation_config) use Docling defaults.\n */\nexport function resolveVlmModel(model: string | VlmModelLocal): VlmModelLocal {\n  if (typeof model === 'string') {\n    const preset = VLM_MODELS[model];\n    if (!preset) {\n      throw new Error(\n        `Unknown VLM model preset: \"${model}\". Available presets: ${Object.keys(VLM_MODELS).join(', ')}`,\n      );\n    }\n    return {\n      repo_id: preset.repo_id,\n      inference_framework: preset.inference_framework,\n      response_format: preset.response_format,\n      transformers_model_type: preset.transformers_model_type,\n    } as VlmModelLocal;\n  }\n  return model;\n}\n"],"mappings":";AAqBO,IAAM,aAA6C;AAAA;AAAA,EAGxD,4BAA4B;AAAA,IAC1B,SAAS;AAAA,IACT,qBAAqB;AAAA,IACrB,iBAAiB;AAAA,IACjB,yBAAyB;AAAA,IACzB,aACE;AAAA,EACJ;AAAA,EACA,wBAAwB;AAAA,IACtB,SAAS;AAAA,IACT,qBAAqB;AAAA,IACrB,iBAAiB;AAAA,IACjB,yBAAyB;AAAA,IACzB,aAAa;AAAA,EACf;AAAA,EACA,wBAAwB;AAAA,IACtB,SAAS;AAAA,IACT,qBAAqB;AAAA,IACrB,iBAAiB;AAAA,IACjB,yBAAyB;AAAA,IACzB,aAAa;AAAA,EACf;AAAA,EACA,oBAAoB;AAAA,IAClB,SAAS;AAAA,IACT,qBAAqB;AAAA,IACrB,iBAAiB;AAAA,IACjB,yBAAyB;AAAA,IACzB,aAAa;AAAA,EACf;AAAA;AAAA,EAIA,qBAAqB;AAAA,IACnB,SAAS;AAAA,IACT,qBAAqB;AAAA,IACrB,iBAAiB;AAAA,IACjB,yBAAyB;AAAA,IACzB,aAAa;AAAA,EACf;AAAA,EACA,oBAAoB;AAAA,IAClB,SAAS;AAAA,IACT,qBAAqB;AAAA,IACrB,iBAAiB;AAAA,IACjB,yBAAyB;AAAA,IACzB,aAAa;AAAA,EACf;AAAA,EACA,MAAM;AAAA,IACJ,SAAS;AAAA,IACT,qBAAqB;AAAA,IACrB,iBAAiB;AAAA,IACjB,yBAAyB;AAAA,IACzB,aAAa;AAAA,EACf;AAAA,EACA,mBAAmB;AAAA,IACjB,SAAS;AAAA,IACT,qBAAqB;AAAA,IACrB,iBAAiB;AAAA,IACjB,yBAAyB;AAAA,IACzB,aAAa;AAAA,EACf;AAAA,EACA,eAAe;AAAA,IACb,SAAS;AAAA,IACT,qBAAqB;AAAA,IACrB,iBAAiB;AAAA,IACjB,yBAAyB;AAAA,IACzB,aAAa;AAAA,EACf;AAAA,EACA,MAAM;AAAA,IACJ,SAAS;AAAA,IACT,qBAAqB;AAAA,IACrB,iBAAiB;AAAA,IACjB,yBAAyB;AAAA,IACzB,aAAa;AAAA,EACf;AAAA,EACA,kBAAkB;AAAA,IAChB,SAAS;AAAA,IACT,qBAAqB;AAAA,IACrB,iBAAiB;AAAA,IACjB,yBAAyB;AAAA,IACzB,aAAa;AAAA,EACf;AAAA,EACA,kBAAkB;AAAA,IAChB,SAAS;AAAA,IACT,qBAAqB;AAAA,IACrB,iBAAiB;AAAA,IACjB,yBAAyB;AAAA,IACzB,aAAa;AAAA,EACf;AAAA,EACA,SAAS;AAAA,IACP,SAAS;AAAA,IACT,qBAAqB;AAAA,IACrB,iBAAiB;AAAA,IACjB,yBAAyB;AAAA,IACzB,aAAa;AAAA,EACf;AACF;AAKO,IAAM,oBAAoB;AAQ1B,SAAS,gBAAgB,OAA8C;AAC5E,MAAI,OAAO,UAAU,UAAU;AAC7B,UAAM,SAAS,WAAW,KAAK;AAC/B,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR,8BAA8B,KAAK,yBAAyB,OAAO,KAAK,UAAU,EAAE,KAAK,IAAI,CAAC;AAAA,MAChG;AAAA,IACF;AACA,WAAO;AAAA,MACL,SAAS,OAAO;AAAA,MAChB,qBAAqB,OAAO;AAAA,MAC5B,iBAAiB,OAAO;AAAA,MACxB,yBAAyB,OAAO;AAAA,IAClC;AAAA,EACF;AACA,SAAO;AACT;","names":[]}

package/dist/index.cjs

@@ -30,15 +30,18 @@ var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: tru
 // src/index.ts
 var src_exports = {};
 __export(src_exports, {
+  DEFAULT_VLM_MODEL: () => DEFAULT_VLM_MODEL,
   ImagePdfFallbackError: () => ImagePdfFallbackError,
-  PDFParser: () => PDFParser
+  PDFParser: () => PDFParser,
+  VLM_MODELS: () => VLM_MODELS,
+  resolveVlmModel: () => resolveVlmModel
 });
 module.exports = __toCommonJS(src_exports);
 
 // src/core/pdf-parser.ts
-var import_docling_sdk = require("docling-sdk");
+var import_docling_sdk2 = require("docling-sdk");
 var import_node_child_process3 = require("child_process");
-var import_node_os2 = require("os");
+var import_node_os3 = require("os");
 var import_node_path6 = require("path");
 
 // src/config/constants.ts
@@ -68,7 +71,11 @@ var PDF_CONVERTER = {
   /**
    * Interval for progress polling in milliseconds
    */
-  POLL_INTERVAL_MS: 1e3
+  POLL_INTERVAL_MS: 1e3,
+  /**
+   * Default timeout for task completion in milliseconds (30 minutes)
+   */
+  DEFAULT_TIMEOUT_MS: 18e5
 };
 var DOCLING_ENVIRONMENT = {
   /**
@@ -86,6 +93,19 @@ var IMAGE_PDF_CONVERTER = {
    */
   QUALITY: 100
 };
+var VLM_ENVIRONMENT = {
+  /**
+   * Timeout for VLM dependency installation (pip install) in milliseconds (3 hours).
+   * VLM packages can be very large and may require extended download times
+   * depending on network conditions.
+   */
+  SETUP_TIMEOUT_MS: 108e5,
+  /**
+   * Timeout for VLM model download in milliseconds (3 hours).
+   * Large VLM models (e.g., multi-GB weights) need sufficient time to download.
+   */
+  MODEL_DOWNLOAD_TIMEOUT_MS: 108e5
+};
 
 // ../shared/dist/index.mjs
 var import_child_process = require("child_process");
@@ -118,6 +138,7 @@ function spawnAsync(command, args, options = {}) {
 
 // src/environment/docling-environment.ts
 var import_node_child_process = require("child_process");
+var import_node_os = require("os");
 var import_node_path = require("path");
 
 // src/utils/python-version.ts
@@ -159,6 +180,7 @@ var DoclingEnvironment = class _DoclingEnvironment {
   venvPath;
   port;
   killExistingProcess;
+  vlmDependenciesInstalled = false;
   constructor(options) {
     this.logger = options.logger;
     this.venvPath = options.venvPath;
@@ -288,6 +310,81 @@ var DoclingEnvironment = class _DoclingEnvironment {
       );
     }
   }
+  /**
+   * Install VLM-specific dependencies for the Docling VLM pipeline.
+   *
+   * Installs:
+   * 1. docling-serve[vlm] - VLM model support for docling-serve
+   * 2. mlx + mlx-lm (macOS ARM64 only) - Apple Silicon optimized inference
+   *
+   * This is idempotent - subsequent calls skip if already installed.
+   */
+  async setupVlmDependencies() {
+    if (this.vlmDependenciesInstalled) {
+      this.logger.info(
+        "[DoclingEnvironment] VLM dependencies already installed, skipping"
+      );
+      return;
+    }
+    if (await this.isVlmReady()) {
+      this.vlmDependenciesInstalled = true;
+      this.logger.info(
+        "[DoclingEnvironment] VLM dependencies already installed, skipping"
+      );
+      return;
+    }
+    this.logger.info("[DoclingEnvironment] Installing VLM dependencies...");
+    const pipPath = (0, import_node_path.join)(this.venvPath, "bin", "pip");
+    this.logger.info("[DoclingEnvironment] Installing docling[vlm]...");
+    const vlmResult = await spawnAsync(
+      pipPath,
+      ["install", "docling-serve[vlm]"],
+      { timeout: VLM_ENVIRONMENT.SETUP_TIMEOUT_MS }
+    );
+    if (vlmResult.code !== 0) {
+      this.logger.error(
+        "[DoclingEnvironment] Failed to install docling-serve[vlm]:",
+        vlmResult.stderr
+      );
+      throw new Error(
+        `Failed to install docling-serve[vlm]. Exit code: ${vlmResult.code}`
+      );
+    }
+    if ((0, import_node_os.platform)() === "darwin" && (0, import_node_os.arch)() === "arm64") {
+      this.logger.info(
+        "[DoclingEnvironment] Installing mlx + mlx-lm for Apple Silicon..."
+      );
+      const mlxResult = await spawnAsync(
+        pipPath,
+        ["install", "mlx", "mlx-lm"],
+        { timeout: VLM_ENVIRONMENT.SETUP_TIMEOUT_MS }
+      );
+      if (mlxResult.code !== 0) {
+        this.logger.error(
+          "[DoclingEnvironment] Failed to install mlx/mlx-lm:",
+          mlxResult.stderr
+        );
+        throw new Error(
+          `Failed to install mlx/mlx-lm. Exit code: ${mlxResult.code}`
+        );
+      }
+    }
+    this.vlmDependenciesInstalled = true;
+    this.logger.info(
+      "[DoclingEnvironment] VLM dependencies installed successfully"
+    );
+  }
+  /**
+   * Check if VLM dependencies are ready by verifying Python module imports
+   */
+  async isVlmReady() {
+    const pythonPath = (0, import_node_path.join)(this.venvPath, "bin", "python");
+    const result = await spawnAsync(pythonPath, [
+      "-c",
+      "import docling_core; import docling"
+    ]);
+    return result.code === 0;
+  }
   async isPortInUse(port) {
     try {
       const result = await spawnAsync("lsof", ["-ti", `:${port}`]);
@@ -374,11 +471,127 @@ var DoclingEnvironment = class _DoclingEnvironment {
 };
 
 // src/core/pdf-converter.ts
+var import_docling_sdk = require("docling-sdk");
 var import_es_toolkit = require("es-toolkit");
 var import_node_fs4 = require("fs");
 var import_node_path5 = require("path");
 var import_promises = require("stream/promises");
 
+// src/config/vlm-models.ts
+var VLM_MODELS = {
+  // ── DocTags models (specialized document structure output) ──────────
+  "granite-docling-258M-mlx": {
+    repo_id: "ibm-granite/granite-docling-258M-mlx",
+    inference_framework: "mlx",
+    response_format: "doctags",
+    transformers_model_type: "automodel-vision2seq",
+    description: "Granite Docling 258M (MLX, Apple Silicon optimized, ~6s/page)"
+  },
+  "granite-docling-258M": {
+    repo_id: "ibm-granite/granite-docling-258M",
+    inference_framework: "transformers",
+    response_format: "doctags",
+    transformers_model_type: "automodel-vision2seq",
+    description: "Granite Docling 258M (Transformers, cross-platform)"
+  },
+  "smoldocling-256M-mlx": {
+    repo_id: "docling-project/SmolDocling-256M-preview-mlx-bf16",
+    inference_framework: "mlx",
+    response_format: "doctags",
+    transformers_model_type: "automodel-vision2seq",
+    description: "SmolDocling 256M (MLX, fastest option)"
+  },
+  "smoldocling-256M": {
+    repo_id: "docling-project/SmolDocling-256M-preview",
+    inference_framework: "transformers",
+    response_format: "doctags",
+    transformers_model_type: "automodel-vision2seq",
+    description: "SmolDocling 256M (Transformers)"
+  },
+  // ── Markdown models (general-purpose vision LLMs) ──────────────────
+  "granite-vision-2B": {
+    repo_id: "ibm-granite/granite-vision-3.2-2b",
+    inference_framework: "transformers",
+    response_format: "markdown",
+    transformers_model_type: "automodel-vision2seq",
+    description: "Granite Vision 3.2 2B (IBM, higher accuracy)"
+  },
+  "qwen25-vl-3B-mlx": {
+    repo_id: "mlx-community/Qwen2.5-VL-3B-Instruct-bf16",
+    inference_framework: "mlx",
+    response_format: "markdown",
+    transformers_model_type: "automodel-vision2seq",
+    description: "Qwen 2.5 VL 3B (MLX, multilingual, good KCJ support)"
+  },
+  phi4: {
+    repo_id: "microsoft/Phi-4-multimodal-instruct",
+    inference_framework: "transformers",
+    response_format: "markdown",
+    transformers_model_type: "automodel",
+    description: "Phi-4 Multimodal (Microsoft, CausalLM)"
+  },
+  "pixtral-12B-mlx": {
+    repo_id: "mlx-community/pixtral-12b-bf16",
+    inference_framework: "mlx",
+    response_format: "markdown",
+    transformers_model_type: "automodel-vision2seq",
+    description: "Pixtral 12B (MLX, Mistral, high accuracy)"
+  },
+  "pixtral-12B": {
+    repo_id: "mistral-community/pixtral-12b",
+    inference_framework: "transformers",
+    response_format: "markdown",
+    transformers_model_type: "automodel-vision2seq",
+    description: "Pixtral 12B (Transformers, Mistral)"
+  },
+  got2: {
+    repo_id: "stepfun-ai/GOT-OCR-2.0-hf",
+    inference_framework: "transformers",
+    response_format: "markdown",
+    transformers_model_type: "automodel-vision2seq",
+    description: "GOT-OCR 2.0 (StepFun, OCR-specialized)"
+  },
+  "gemma3-12B-mlx": {
+    repo_id: "mlx-community/gemma-3-12b-it-bf16",
+    inference_framework: "mlx",
+    response_format: "markdown",
+    transformers_model_type: "automodel-vision2seq",
+    description: "Gemma 3 12B (MLX, Google)"
+  },
+  "gemma3-27B-mlx": {
+    repo_id: "mlx-community/gemma-3-27b-it-bf16",
+    inference_framework: "mlx",
+    response_format: "markdown",
+    transformers_model_type: "automodel-vision2seq",
+    description: "Gemma 3 27B (MLX, Google, highest accuracy)"
+  },
+  dolphin: {
+    repo_id: "ByteDance/Dolphin",
+    inference_framework: "transformers",
+    response_format: "markdown",
+    transformers_model_type: "automodel-vision2seq",
+    description: "Dolphin (ByteDance, document-oriented)"
+  }
+};
+var DEFAULT_VLM_MODEL = "granite-docling-258M-mlx";
+function resolveVlmModel(model) {
+  if (typeof model === "string") {
+    const preset = VLM_MODELS[model];
+    if (!preset) {
+      throw new Error(
+        `Unknown VLM model preset: "${model}". Available presets: ${Object.keys(VLM_MODELS).join(", ")}`
+      );
+    }
+    return {
+      repo_id: preset.repo_id,
+      inference_framework: preset.inference_framework,
+      response_format: preset.response_format,
+      transformers_model_type: preset.transformers_model_type
+    };
+  }
+  return model;
+}
+
 // src/errors/image-pdf-fallback-error.ts
 var ImagePdfFallbackError = class extends Error {
   constructor(originalError, fallbackError) {
@@ -730,7 +943,7 @@ var LocalFileServer = class {
 
 // src/core/image-pdf-converter.ts
 var import_node_fs3 = require("fs");
-var import_node_os = require("os");
+var import_node_os2 = require("os");
 var import_node_path4 = require("path");
 var ImagePdfConverter = class {
   constructor(logger) {
@@ -746,7 +959,7 @@ var ImagePdfConverter = class {
    */
   async convert(pdfUrl, reportId) {
     const timestamp = Date.now();
-    const tempDir = (0, import_node_os.tmpdir)();
+    const tempDir = (0, import_node_os2.tmpdir)();
     const inputPath = (0, import_node_path4.join)(tempDir, `${reportId}-${timestamp}-input.pdf`);
     const outputPath = (0, import_node_path4.join)(tempDir, `${reportId}-${timestamp}-image.pdf`);
     try {
@@ -816,11 +1029,17 @@ var ImagePdfConverter = class {
 };
 
 // src/core/pdf-converter.ts
+var _origAssertValidConversionOptions = import_docling_sdk.ValidationUtils.assertValidConversionOptions.bind(import_docling_sdk.ValidationUtils);
+import_docling_sdk.ValidationUtils.assertValidConversionOptions = (options) => {
+  const { pipeline: _pipeline, ...rest } = options;
+  _origAssertValidConversionOptions(rest);
+};
 var PDFConverter = class {
-  constructor(logger, client, enableImagePdfFallback = false) {
+  constructor(logger, client, enableImagePdfFallback = false, timeout = PDF_CONVERTER.DEFAULT_TIMEOUT_MS) {
     this.logger = logger;
     this.client = client;
     this.enableImagePdfFallback = enableImagePdfFallback;
+    this.timeout = timeout;
   }
   async convert(url, reportId, onComplete, cleanupAfterCallback, options, abortSignal) {
     this.logger.info("[PDFConverter] Converting:", url);
@@ -875,7 +1094,15 @@ var PDFConverter = class {
   }
   async performConversion(url, reportId, onComplete, cleanupAfterCallback, options, abortSignal) {
     const startTime = Date.now();
-    const conversionOptions = this.buildConversionOptions(options);
+    const pipelineType = options.pipeline ?? "standard";
+    const conversionOptions = pipelineType === "vlm" ? this.buildVlmConversionOptions(options) : this.buildConversionOptions(options);
+    if (pipelineType === "vlm") {
+      this.logger.info("[PDFConverter] Using VLM pipeline");
+    } else {
+      this.logger.info(
+        `[PDFConverter] OCR languages: ${JSON.stringify(conversionOptions.ocr_options?.lang)}`
+      );
+    }
     this.logger.info(
       "[PDFConverter] Converting document with Async Source API..."
     );
@@ -942,7 +1169,7 @@ var PDFConverter = class {
   }
   buildConversionOptions(options) {
     return {
-      ...(0, import_es_toolkit.omit)(options, ["num_threads"]),
+      ...(0, import_es_toolkit.omit)(options, ["num_threads", "pipeline", "vlm_model"]),
       to_formats: ["json", "html"],
       image_export_mode: "embedded",
       ocr_engine: "ocrmac",
@@ -968,6 +1195,31 @@ var PDFConverter = class {
       }
     };
   }
+  /**
+   * Build conversion options for VLM pipeline.
+   *
+   * VLM pipeline uses a Vision Language Model instead of traditional OCR,
+   * providing better accuracy for KCJ characters and complex layouts.
+   */
+  buildVlmConversionOptions(options) {
+    const vlmModel = resolveVlmModel(options.vlm_model ?? DEFAULT_VLM_MODEL);
+    this.logger.info(
+      `[PDFConverter] VLM model: ${vlmModel.repo_id} (framework: ${vlmModel.inference_framework}, format: ${vlmModel.response_format})`
+    );
+    return {
+      ...(0, import_es_toolkit.omit)(options, ["num_threads", "pipeline", "vlm_model", "ocr_lang"]),
+      to_formats: ["json", "html"],
+      image_export_mode: "embedded",
+      pipeline: "vlm",
+      vlm_pipeline_model_local: vlmModel,
+      generate_picture_images: true,
+      images_scale: 2,
+      accelerator_options: {
+        device: "mps",
+        num_threads: options.num_threads
+      }
+    };
+  }
   async startConversionTask(url, conversionOptions) {
     const task = await this.client.convertSourceAsync({
       sources: [
@@ -1003,38 +1255,42 @@ var PDFConverter = class {
   }
   async trackTaskProgress(task) {
     const conversionStartTime = Date.now();
-    let lastStatus = "";
-    let isCompleted = false;
-    const pollInterval = setInterval(() => {
-      if (isCompleted) return;
-      const elapsed = Math.floor((Date.now() - conversionStartTime) / 1e3);
-      process.stdout.write(
-        `\r[PDFConverter] Status: ${lastStatus || "processing"} (${elapsed}s elapsed)`
-      );
-    }, PDF_CONVERTER.POLL_INTERVAL_MS);
-    task.on("progress", (status) => {
-      lastStatus = status.task_status;
+    let lastProgressLine = "";
+    const logProgress = (status) => {
+      const parts = [`Status: ${status.task_status}`];
       if (status.task_position !== void 0) {
-        process.stdout.write(
-          `\r[PDFConverter] Status: ${status.task_status} (position: ${status.task_position})`
-        );
+        parts.push(`position: ${status.task_position}`);
       }
-    });
-    task.on("complete", () => {
-      isCompleted = true;
-      clearInterval(pollInterval);
-      this.logger.info("\n[PDFConverter] Conversion completed!");
-    });
-    task.on("error", (error) => {
-      isCompleted = true;
-      clearInterval(pollInterval);
-      this.logger.error("\n[PDFConverter] Conversion error:", error.message);
-    });
-    try {
-      await task.waitForCompletion();
-    } finally {
-      isCompleted = true;
-      clearInterval(pollInterval);
+      const meta = status.task_meta;
+      if (meta) {
+        if (meta.processed_documents !== void 0 && meta.total_documents !== void 0) {
+          parts.push(
+            `progress: ${meta.processed_documents}/${meta.total_documents}`
+          );
+        }
+      }
+      const progressLine = `\r[PDFConverter] ${parts.join(" | ")}`;
+      if (progressLine !== lastProgressLine) {
+        lastProgressLine = progressLine;
+        process.stdout.write(progressLine);
+      }
+    };
+    while (true) {
+      if (Date.now() - conversionStartTime > this.timeout) {
+        throw new Error("Task timeout");
+      }
+      const status = await task.poll();
+      logProgress(status);
+      if (status.task_status === "success") {
+        this.logger.info("\n[PDFConverter] Conversion completed!");
+        return;
+      }
+      if (status.task_status === "failure") {
+        throw new Error("Task failed with status: failure");
+      }
+      await new Promise(
+        (resolve) => setTimeout(resolve, PDF_CONVERTER.POLL_INTERVAL_MS)
+      );
     }
   }
   async downloadResult(taskId) {
@@ -1070,6 +1326,7 @@ var PDFParser = class {
   killExistingProcess;
   enableImagePdfFallback;
   client = null;
+  environment;
   constructor(options) {
     const {
       logger,
@@ -1106,7 +1363,7 @@ var PDFParser = class {
     }
     if (this.baseUrl) {
       this.logger.info("[PDFParser] Using external server:", this.baseUrl);
-      this.client = new import_docling_sdk.Docling({
+      this.client = new import_docling_sdk2.Docling({
         api: { baseUrl: this.baseUrl, timeout: this.timeout }
       });
       await this.waitForServerReady();
@@ -1114,15 +1371,15 @@ var PDFParser = class {
     }
     this.logger.info("[PDFParser] Setting up local server...");
     try {
-      const environment = new DoclingEnvironment({
+      this.environment = new DoclingEnvironment({
         logger: this.logger,
         venvPath: this.venvPath,
         port: this.port,
         killExistingProcess: this.killExistingProcess
       });
-      await environment.setup();
+      await this.environment.setup();
       const clientUrl = `http://localhost:${this.port}`;
-      this.client = new import_docling_sdk.Docling({
+      this.client = new import_docling_sdk2.Docling({
         api: {
           baseUrl: clientUrl,
           timeout: this.timeout
@@ -1136,9 +1393,9 @@ var PDFParser = class {
     }
   }
   checkOperatingSystem() {
-    if ((0, import_node_os2.platform)() !== "darwin") {
+    if ((0, import_node_os3.platform)() !== "darwin") {
       throw new Error(
-        "PDFParser is only supported on macOS. Current platform: " + (0, import_node_os2.platform)()
+        "PDFParser is only supported on macOS. Current platform: " + (0, import_node_os3.platform)()
       );
     }
   }
@@ -1222,7 +1479,7 @@ var PDFParser = class {
     });
     await environment.startServer();
     this.client?.destroy();
-    this.client = new import_docling_sdk.Docling({
+    this.client = new import_docling_sdk2.Docling({
       api: {
         baseUrl: `http://localhost:${this.port}`,
         timeout: this.timeout
@@ -1266,6 +1523,12 @@ var PDFParser = class {
         "PDFParser is not initialized. Call init() before using parse()"
       );
     }
+    if (options.pipeline === "vlm" && this.environment && !this.baseUrl) {
+      this.logger.info(
+        "[PDFParser] VLM pipeline requested, ensuring VLM dependencies..."
+      );
+      await this.environment.setupVlmDependencies();
+    }
     const canRecover = !this.baseUrl && this.port !== void 0;
     const maxAttempts = PDF_PARSER.MAX_SERVER_RECOVERY_ATTEMPTS;
     let attempt = 0;
@@ -1275,7 +1538,8 @@ var PDFParser = class {
         const converter = new PDFConverter(
           this.logger,
           this.client,
-          effectiveFallbackEnabled
+          effectiveFallbackEnabled,
+          this.timeout
         );
         return await converter.convert(
           url,
@@ -1323,7 +1587,10 @@ var PDFParser = class {
 };
 // Annotate the CommonJS export names for ESM import in node:
 0 && (module.exports = {
+  DEFAULT_VLM_MODEL,
   ImagePdfFallbackError,
-  PDFParser
+  PDFParser,
+  VLM_MODELS,
+  resolveVlmModel
 });
 //# sourceMappingURL=index.cjs.map