vlmparse 0.1.5__py3-none-any.whl → 0.1.7__py3-none-any.whl

vlmparse/cli.py

@@ -4,44 +4,42 @@ from loguru import logger
 
 
 class DParseCLI:
-    def serve(self, model: str, port: int | None = None, gpus: str | None = None):
+    """Parsing of pdf to text using VLMs: typ in vlmparse to get the command lists, then `vlmparse <command> --help` to get help on a specific command."""
+
+    def serve(
+        self,
+        model: str,
+        port: int | None = None,
+        gpus: str | None = None,
+        vllm_args: list[str] | None = None,
+        forget_predefined_vllm_args: bool = False,
+    ):
         """Deploy a VLLM server in a Docker container.
 
         Args:
             model: Model name
             port: VLLM server port (default: 8056)
             gpus: Comma-separated GPU device IDs (e.g., "0" or "0,1,2"). If not specified, all GPUs will be used.
+            vllm_args: Additional keyword arguments to pass to the VLLM server.
+            forget_predefined_vllm_args: If True, the predefined VLLM kwargs from the docker config will be replaced by vllm_args otherwise the predefined kwargs will be updated with vllm_args with a risk of collision of argument names.
         """
-        if port is None:
-            port = 8056
-
-        from vlmparse.registries import docker_config_registry
-
-        docker_config = docker_config_registry.get(model)
-        if docker_config is None:
-            logger.warning(
-                f"No Docker configuration found for model: {model}, using default configuration"
-            )
-            return
-
-        docker_config.docker_port = port
 
-        # Only override GPU configuration if explicitly specified
-        # This preserves CPU-only settings from the config
-        if gpus is not None:
-            docker_config.gpu_device_ids = [g.strip() for g in str(gpus).split(",")]
-        server = docker_config.get_server(auto_stop=False)
+        from vlmparse.converter_with_server import start_server
 
-        # Deploy server and leave it running (cleanup=False)
-        logger.info(
-            f"Deploying VLLM server for {docker_config.model_name} on port {port}..."
+        base_url, container, _, _ = start_server(
+            model=model,
+            gpus=gpus,
+            port=port,
+            with_vllm_server=True,
+            vllm_args=vllm_args,
+            forget_predefined_vllm_args=forget_predefined_vllm_args,
+            auto_stop=False,
         )
 
-        base_url, container = server.start()
-
         logger.info(f"✓ VLLM server ready at {base_url}")
-        logger.info(f"✓ Container ID: {container.id}")
-        logger.info(f"✓ Container name: {container.name}")
+        if container is not None:
+            logger.info(f"✓ Container ID: {container.id}")
+            logger.info(f"✓ Container name: {container.name}")
 
     def convert(
         self,
@@ -54,6 +52,7 @@ class DParseCLI:
         with_vllm_server: bool = False,
         concurrency: int = 10,
         dpi: int | None = None,
+        debug: bool = False,
     ):
         """Parse PDF documents and save results.
 
@@ -67,109 +66,20 @@ class DParseCLI:
             mode: Output mode - "document" (save as JSON zip), "md" (save as markdown file), "md_page" (save as folder of markdown pages)
             with_vllm_server: If True, a local VLLM server will be deployed if the model is not found in the registry. Note that if the model is in the registry and the uri is None, the server will be anyway deployed.
             dpi: DPI to use for the conversion. If not specified, the default DPI will be used.
+            debug: If True, run in debug mode (single-threaded, no concurrency)
         """
         from vlmparse.converter_with_server import ConverterWithServer
 
-        converter_with_server = ConverterWithServer(
+        with ConverterWithServer(
             model=model,
             uri=uri,
             gpus=gpus,
             with_vllm_server=with_vllm_server,
             concurrency=concurrency,
-        )
-
-        return converter_with_server.parse(
-            inputs=inputs, out_folder=out_folder, mode=mode, dpi=dpi
-        )
-        # from vlmparse.registries import converter_config_registry
-
-        # # Infer model from URI if provided
-        # if uri is not None and model is None:
-        #     import docker
-
-        #     try:
-        #         docker_client = docker.from_env()
-        #         containers = docker_client.containers.list()
-        #         for container in containers:
-        #             # Check both exact match and match with/without trailing slash
-        #             container_uri = container.labels.get("vlmparse_uri", "")
-        #             if container_uri and (
-        #                 container_uri == uri
-        #                 or container_uri.rstrip("/") == uri.rstrip("/")
-        #             ):
-        #                 inferred_model = container.labels.get("vlmparse_model_name")
-        #                 if inferred_model:
-        #                     logger.info(
-        #                         f"Inferred model {inferred_model} from URI {uri}"
-        #                     )
-        #                     model = inferred_model
-        #                     break
-        #     except Exception:
-        #         # If Docker is not available or fails, just proceed with provided arguments
-        #         pass
-
-        # if mode not in ["document", "md", "md_page"]:
-        #     logger.error(f"Invalid mode: {mode}. Must be one of: document, md, md_page")
-        #     return
-
-        # # Expand file paths from glob patterns
-        # file_paths = []
-        # if isinstance(inputs, str):
-        #     inputs = [inputs]
-        # for pattern in inputs:
-        #     if "*" in pattern or "?" in pattern:
-        #         file_paths.extend(glob(pattern, recursive=True))
-        #     elif os.path.isdir(pattern):
-        #         file_paths.extend(glob(os.path.join(pattern, "*.pdf"), recursive=True))
-        #     elif os.path.isfile(pattern):
-        #         file_paths.append(pattern)
-        #     else:
-        #         logger.error(f"Invalid input: {pattern}")
-
-        # # Filter to only existing PDF files
-        # file_paths = [f for f in file_paths if os.path.exists(f) and f.endswith(".pdf")]
-
-        # if not file_paths:
-        #     logger.error("No PDF files found matching the inputs patterns")
-        #     return
-
-        # logger.info(f"Processing {len(file_paths)} files with {model} converter")
-
-        # gpu_device_ids = None
-        # if gpus is not None:
-        #     gpu_device_ids = [g.strip() for g in gpus.split(",")]
-
-        # if uri is None:
-        #     from vlmparse.registries import docker_config_registry
-
-        #     docker_config = docker_config_registry.get(model, default=with_vllm_server)
-
-        #     if docker_config is not None:
-        #         docker_config.gpu_device_ids = gpu_device_ids
-        #         server = docker_config.get_server(auto_stop=True)
-        #         server.start()
-
-        #         client = docker_config.get_client(
-        #             save_folder=out_folder, save_mode=mode
-        #         )
-        #     else:
-        #         client = converter_config_registry.get(model).get_client(
-        #             save_folder=out_folder, save_mode=mode
-        #         )
-
-        # else:
-        #     client_config = converter_config_registry.get(model, uri=uri)
-        #     client = client_config.get_client(save_folder=out_folder, save_mode=mode)
-        # client.num_concurrent_files = concurrency
-        # client.num_concurrent_pages = concurrency
-        # if dpi is not None:
-        #     client.config.dpi = int(dpi)
-        # documents = client.batch(file_paths)
-
-        # if documents is not None:
-        #     logger.info(f"Processed {len(documents)} documents to {out_folder}")
-        # else:
-        #     logger.info(f"Processed {len(file_paths)} documents to {out_folder}")
+        ) as converter_with_server:
+            return converter_with_server.parse(
+                inputs=inputs, out_folder=out_folder, mode=mode, dpi=dpi, debug=debug
+            )
 
     def list(self):
         """List all containers whose name begins with vlmparse."""

vlmparse/clients/chandra.py

@@ -194,7 +194,7 @@ class ChandraConverterConfig(OpenAIConverterConfig):
     model_name: str = "datalab-to/chandra"
     prompt_type: str = "ocr"  # Default prompt type
     bbox_scale: int = 1024
-    max_retries: int = 6
+    max_retries: int = 0
     max_failure_retries: int = None
     completion_kwargs: dict = Field(
         default_factory=lambda: {

vlmparse/clients/deepseekocr.py

@@ -15,6 +15,57 @@ from vlmparse.servers.docker_server import VLLMDockerServerConfig
 from vlmparse.utils import to_base64
 
 
+class DeepSeekOCRDockerServerConfig(VLLMDockerServerConfig):
+    """Configuration for DeepSeekOCR model."""
+
+    model_name: str = "deepseek-ai/DeepSeek-OCR"
+    command_args: list[str] = Field(
+        default_factory=lambda: [
+            "--limit-mm-per-prompt",
+            '{"image": 1}',
+            "--async-scheduling",
+            "--logits_processors",
+            "vllm.model_executor.models.deepseek_ocr:NGramPerReqLogitsProcessor",
+            "--no-enable-prefix-caching",
+            "--mm-processor-cache-gb",
+            "0",
+        ]
+    )
+    aliases: list[str] = Field(default_factory=lambda: ["deepseekocr"])
+
+    @property
+    def client_config(self):
+        return DeepSeekOCRConverterConfig(llm_params=self.llm_params)
+
+
+class DeepSeekOCRConverterConfig(OpenAIConverterConfig):
+    """DeepSeekOCR converter - backward compatibility alias."""
+
+    model_name: str = "deepseek-ai/DeepSeek-OCR"
+    aliases: list[str] = Field(default_factory=lambda: ["deepseekocr"])
+
+    prompt_mode: Literal["layout", "ocr"] = "ocr"
+    completion_kwargs: dict | None = {
+        "temperature": 0.0,
+        "max_tokens": 8181,
+        "extra_body": {
+            "skip_special_tokens": False,
+            # args used to control custom logits processor
+            "vllm_xargs": {
+                "ngram_size": 30,
+                "window_size": 90,
+                # whitelist: <td>, </td>
+                "whitelist_token_ids": [128821, 128822],
+            },
+        },
+    }
+    dpi: int = 200
+    aliases: list[str] = Field(default_factory=lambda: ["deepseekocr"])
+
+    def get_client(self, **kwargs) -> "DeepSeekOCRConverterClient":
+        return DeepSeekOCRConverterClient(config=self, **kwargs)
+
+
 def re_match(text):
     pattern = r"(<\|ref\|>(.*?)<\|/ref\|><\|det\|>(.*?)<\|/det\|>)"
     matches = re.findall(pattern, text, re.DOTALL)
@@ -150,54 +201,3 @@ class DeepSeekOCRConverterClient(OpenAIConverterClient):
         logger.debug(page.text)
 
         return page
-
-
-class DeepSeekOCRDockerServerConfig(VLLMDockerServerConfig):
-    """Configuration for DeepSeekOCR model."""
-
-    model_name: str = "deepseek-ai/DeepSeek-OCR"
-    command_args: list[str] = Field(
-        default_factory=lambda: [
-            "--limit-mm-per-prompt",
-            '{"image": 1}',
-            "--async-scheduling",
-            "--logits_processors",
-            "vllm.model_executor.models.deepseek_ocr:NGramPerReqLogitsProcessor",
-            "--no-enable-prefix-caching",
-            "--mm-processor-cache-gb",
-            "0",
-        ]
-    )
-    aliases: list[str] = Field(default_factory=lambda: ["deepseekocr"])
-
-    @property
-    def client_config(self):
-        return DeepSeekOCRConverterConfig(llm_params=self.llm_params)
-
-
-class DeepSeekOCRConverterConfig(OpenAIConverterConfig):
-    """DeepSeekOCR converter - backward compatibility alias."""
-
-    model_name: str = "deepseek-ai/DeepSeek-OCR"
-    aliases: list[str] = Field(default_factory=lambda: ["deepseekocr"])
-
-    prompt_mode: Literal["layout", "ocr"] = "ocr"
-    completion_kwargs: dict | None = {
-        "temperature": 0.0,
-        "max_tokens": 8181,
-        "extra_body": {
-            "skip_special_tokens": False,
-            # args used to control custom logits processor
-            "vllm_xargs": {
-                "ngram_size": 30,
-                "window_size": 90,
-                # whitelist: <td>, </td>
-                "whitelist_token_ids": [128821, 128822],
-            },
-        },
-    }
-    dpi: int = 200
-    aliases: list[str] = Field(default_factory=lambda: ["deepseekocr"])
-
-    def get_client(self, **kwargs) -> "DeepSeekOCRConverterClient":
-        return DeepSeekOCRConverterClient(config=self, **kwargs)

vlmparse/clients/dotsocr.py

@@ -237,7 +237,7 @@ class DotsOCRConverter(OpenAIConverterClient):
         )
         prompt = self.PROMPTS[prompt_mode]
 
-        response = await self._async_inference_with_vllm(image, prompt)
+        response, usage = await self._async_inference_with_vllm(image, prompt)
 
         if prompt_mode in ["prompt_layout_all_en"]:
             try:
@@ -248,17 +248,17 @@ class DotsOCRConverter(OpenAIConverterClient):
                     image.width,
                     image.height,
                 )
-                return {}, cells, False
+                return {}, cells, False, usage
             except Exception as e:
                 logger.warning(f"cells post process error: {e}, returning raw response")
-                return {}, response, True
+                return {}, response, True, usage
         else:
-            return {}, response, None
+            return {}, response, None, usage
 
     async def async_call_inside_page(self, page: Page) -> Page:
         image = page.image
 
-        _, response, _ = await self._parse_image_vllm(
+        _, response, _, usage = await self._parse_image_vllm(
             image, prompt_mode=self.config.prompt_mode
         )
         logger.info("Response: " + str(response))
@@ -283,4 +283,8 @@ class DotsOCRConverter(OpenAIConverterClient):
         text = clean_response(response)
         text = html_to_md_keep_tables(text)
         page.text = text
+
+        page.completion_tokens = usage.completion_tokens
+        page.prompt_tokens = usage.prompt_tokens
+        page.reasoning_tokens = usage.reasoning_tokens
         return page

vlmparse/clients/granite_docling.py

@@ -34,6 +34,7 @@ class GraniteDoclingDockerServerConfig(VLLMDockerServerConfig):
 class GraniteDoclingConverterConfig(OpenAIConverterConfig):
     """Granite Docling converter configuration."""
 
+    model_name: str = "ibm-granite/granite-docling-258M"
     preprompt: str | None = None
     postprompt: str | None = "Convert this page to docling."
     completion_kwargs: dict | None = {

vlmparse/clients/hunyuanocr.py

@@ -39,7 +39,8 @@ class HunyuanOCRConverterConfig(OpenAIConverterConfig):
     completion_kwargs: dict | None = {
         "temperature": 0.0,
         "extra_body": {"top_k": 1, "repetition_penalty": 1.0},
+        "max_completion_tokens": 16384,  # max token len used in training according to the technical report is 32000, but in practice the model breaks earlier
     }
-    max_image_size: int | None = 1540
     dpi: int = 200
     aliases: list[str] = Field(default_factory=lambda: ["hunyuanocr"])
+    stream: bool = True

vlmparse/clients/mineru.py

@@ -52,24 +52,22 @@ class MinerUConverter(BaseConverter):
 
     config: MinerUConverterConfig
 
-    def __init__(self, config: MinerUConverterConfig, **kwargs):
-        super().__init__(config=config, **kwargs)
-        from httpx import AsyncClient
-
-        self.client = AsyncClient(base_url=config.base_url, timeout=config.timeout)
-
     async def _async_inference_with_api(self, image) -> list:
         """Run async inference with MinerU API."""
+        from httpx import AsyncClient
 
-        img_byte_arr = await asyncio.to_thread(to_bytes_io, image)
-        response = await self.client.post(
-            "process-image",
-            files={"image": ("image.png", img_byte_arr, "image/png")},
-        )
+        async with AsyncClient(
+            base_url=self.config.base_url, timeout=self.config.timeout
+        ) as client:
+            img_byte_arr = await asyncio.to_thread(to_bytes_io, image)
+            response = await client.post(
+                "process-image",
+                files={"image": ("image.png", img_byte_arr, "image/png")},
+            )
 
-        response.raise_for_status()
+            response.raise_for_status()
 
-        res = orjson.loads(response.content)
+            res = orjson.loads(response.content)
 
         return res
 

vlmparse/clients/olmocr.py

@@ -41,6 +41,6 @@ class OlmOCRConverterConfig(OpenAIConverterConfig):
         "temperature": 0.1,
         "max_tokens": 8000,
     }
-    max_image_size: int | None = 1288
+    # max_image_size: int | None = 1288
     dpi: int = 200
     aliases: list[str] = Field(default_factory=lambda: ["olmocr-2-fp8"])

vlmparse/clients/openai_converter.py

@@ -1,5 +1,5 @@
 import os
-from typing import Literal
+from typing import Literal, Optional
 
 from loguru import logger
 from pydantic import Field
@@ -40,6 +40,8 @@ def get_llm_params(model_name: str, uri: str | None = None):
     ]:
         base_url = None
         api_key = os.getenv("OPENAI_API_KEY")
+        if api_key is None:
+            raise ValueError("OPENAI_API_KEY environment variable not set")
     else:
         if model_name in [
             "gemini-2.5-flash-lite",
@@ -48,6 +50,8 @@ def get_llm_params(model_name: str, uri: str | None = None):
         ]:
             base_url = GOOGLE_API_BASE_URL
             api_key = os.getenv("GOOGLE_API_KEY")
+            if api_key is None:
+                raise ValueError("GOOGLE_API_KEY environment variable not set")
         else:
             return None
     return LLMParams(base_url=base_url, model_name=model_name, api_key=api_key)
@@ -97,7 +101,7 @@ class OpenAIConverterClient(BaseConverter):
 
     async def _get_chat_completion(
         self, messages: list[dict], completion_kwargs: dict | None = None
-    ) -> str:
+    ) -> tuple[str, Optional["CompletionUsage"]]:  # noqa: F821
         """Helper to handle chat completion with optional streaming."""
         if completion_kwargs is None:
             completion_kwargs = self.config.completion_kwargs
@@ -113,7 +117,8 @@ class OpenAIConverterClient(BaseConverter):
             async for chunk in response_stream:
                 if chunk.choices and chunk.choices[0].delta.content:
                     response_parts.append(chunk.choices[0].delta.content)
-            return "".join(response_parts)
+
+            return "".join(response_parts), None
         else:
             response_obj = await self.model.chat.completions.create(
                 model=self.config.llm_params.model_name,
@@ -126,7 +131,8 @@ class OpenAIConverterClient(BaseConverter):
                     "Response is None, finish reason: "
                     + response_obj.choices[0].finish_reason
                 )
-            return response_obj.choices[0].message.content
+
+            return response_obj.choices[0].message.content, response_obj.usage
 
     async def async_call_inside_page(self, page: Page) -> Page:
         """Process a single page using OpenAI-compatible API."""
@@ -163,12 +169,17 @@ class OpenAIConverterClient(BaseConverter):
             },
         ]
 
-        response = await self._get_chat_completion(messages)
-        logger.info("Response: " + str(response))
+        response, usage = await self._get_chat_completion(messages)
+        logger.debug("Response: " + str(response))
         page.raw_response = response
         text = clean_response(response)
 
         text = html_to_md_keep_tables(text)
         page.text = text
+        if usage is not None:
+            page.prompt_tokens = usage.prompt_tokens
+            page.completion_tokens = usage.completion_tokens
+            if hasattr(usage, "reasoning_tokens"):
+                page.reasoning_tokens = usage.reasoning_tokens
 
         return page

vlmparse/clients/paddleocrvl.py

@@ -42,7 +42,8 @@ class PaddleOCRVLConverterConfig(OpenAIConverterConfig):
     postprompt: str | None = TASKS["ocr"]
     completion_kwargs: dict | None = {
         "temperature": 0.0,
+        "max_completion_tokens": 16384,
     }
-    max_image_size: int | None = 1540
     dpi: int = 200
     aliases: list[str] = Field(default_factory=lambda: ["paddleocrvl"])
+    stream: bool = True

vlmparse/constants.py

@@ -1,2 +1,5 @@
+import os
+
 IMAGE_EXTENSIONS = [".jpg", ".jpeg", ".png", ".tiff", ".tif", ".bmp", ".gif", ".webp"]
 PDF_EXTENSION = ".pdf"
+DEFAULT_SERVER_PORT = os.getenv("VLMPARSE_DEFAULT_PORT", 8056)

vlmparse/converter_with_server.py

@@ -5,51 +5,102 @@ from typing import Literal
 
 from loguru import logger
 
+from vlmparse.constants import DEFAULT_SERVER_PORT
 from vlmparse.servers.utils import get_model_from_uri
 from vlmparse.utils import get_file_paths
 
 
+def start_server(
+    model: str,
+    gpus: str,
+    port: None | int = None,
+    with_vllm_server: bool = True,
+    vllm_args: list[str] = {},
+    forget_predefined_vllm_args: bool = False,
+    auto_stop: bool = False,
+):
+    from vlmparse.registries import docker_config_registry
+
+    base_url = ""
+    container = None
+    docker_config = docker_config_registry.get(model, default=with_vllm_server)
+
+    if port is None:
+        port = DEFAULT_SERVER_PORT
+
+    if docker_config is None:
+        logger.warning(
+            f"No Docker configuration found for model: {model}, using default configuration"
+        )
+        return "", container, None, docker_config
+
+    gpu_device_ids = None
+    if gpus is not None:
+        gpu_device_ids = [g.strip() for g in str(gpus).split(",")]
+
+    if docker_config is not None:
+        if port is not None:
+            docker_config.docker_port = port
+        docker_config.gpu_device_ids = gpu_device_ids
+        docker_config.update_command_args(
+            vllm_args,
+            forget_predefined_vllm_args=forget_predefined_vllm_args,
+        )
+
+        logger.info(
+            f"Deploying VLLM server for {docker_config.model_name} on port {port}..."
+        )
+        server = docker_config.get_server(auto_stop=auto_stop)
+        if server is None:
+            logger.error(f"Model server not found for model: {model}")
+            return "", container, None, docker_config
+
+        base_url, container = server.start()
+
+    return base_url, container, server, docker_config
+
+
 class ConverterWithServer:
     def __init__(
         self,
-        model: str,
+        model: str | None = None,
         uri: str | None = None,
         gpus: str | None = None,
         port: int | None = None,
         with_vllm_server: bool = False,
         concurrency: int = 10,
+        vllm_args: dict | None = None,
+        forget_predefined_vllm_args: bool = False,
     ):
-        from vlmparse.registries import (
-            converter_config_registry,
-            docker_config_registry,
-        )
-
         self.model = model
         self.uri = uri
         self.port = port
         self.gpus = gpus
         self.with_vllm_server = with_vllm_server
         self.concurrency = concurrency
+        self.vllm_args = vllm_args
+        self.forget_predefined_vllm_args = forget_predefined_vllm_args
+        self.server = None
+        self.client = None
 
         if self.uri is not None and self.model is None:
             self.model = get_model_from_uri(self.uri)
 
-        gpu_device_ids = None
-        if self.gpus is not None:
-            gpu_device_ids = [g.strip() for g in self.gpus.split(",")]
+    def start_server_and_client(self):
+        from vlmparse.registries import converter_config_registry
 
         if self.uri is None:
-            docker_config = docker_config_registry.get(
-                self.model, default=self.with_vllm_server
+            _, _, self.server, docker_config = start_server(
+                model=self.model,
+                gpus=self.gpus,
+                port=self.port,
+                with_vllm_server=self.with_vllm_server,
+                vllm_args=self.vllm_args,
+                forget_predefined_vllm_args=self.forget_predefined_vllm_args,
+                auto_stop=True,
             )
 
             if docker_config is not None:
-                if self.port is not None:
-                    docker_config.docker_port = self.port
-                docker_config.gpu_device_ids = gpu_device_ids
-                self.server = docker_config.get_server(auto_stop=True)
-                self.server.start()
-
                 self.client = docker_config.get_client()
             else:
                 self.client = converter_config_registry.get(self.model).get_client()
@@ -59,6 +110,17 @@ class ConverterWithServer:
 
             self.client = client_config.get_client()
 
+    def stop_server(self):
+        if self.server is not None and self.server.auto_stop:
+            self.server.stop()
+
+    def __enter__(self):
+        self.start_server_and_client()
+        return self
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        self.stop_server()
+
     def parse(
         self,
         inputs: str | list[str],
@@ -68,6 +130,9 @@ class ConverterWithServer:
         debug: bool = False,
         retrylast: bool = False,
     ):
+        assert (
+            self.client is not None
+        ), "Client not initialized. Call start_server_and_client() first."
         file_paths = get_file_paths(inputs)
         assert (
             out_folder is not None
@@ -119,5 +184,5 @@ class ConverterWithServer:
 
         return documents
 
-    def get_out_folder(self) -> Path:
+    def get_out_folder(self) -> str | None:
         return self.client.save_folder

vlmparse/data_model/document.py

@@ -41,6 +41,10 @@ class Page(VLMParseBaseModel):
     buffer_image: Optional[Image.Image | str | dict] = None
     latency: Optional[float] = None
     """Time taken to process the page in seconds."""
+    prompt_tokens: Optional[int] = None
+    completion_tokens: Optional[int] = None
+    """Include reasoning tokens"""
+    reasoning_tokens: Optional[int] = None
 
     @property
     def image(self):
@@ -66,7 +70,7 @@ class Page(VLMParseBaseModel):
 
         image = self.image
 
-        if layout:
+        if layout and image is not None:
             if self.items is None:
                 return image
             items = self.items
@@ -85,6 +89,9 @@ class Page(VLMParseBaseModel):
                 )
         return image
 
+    def to_markdown(self, **kwargs):
+        return self.text if self.text is not None else ""
+
 
 class Document(VLMParseBaseModel):
     file_path: str
@@ -104,6 +111,9 @@ class Document(VLMParseBaseModel):
             page.error is not None for page in self.pages
         )
 
+    def to_markdown(self, **kwargs):
+        return "\n\n".join([page.to_markdown(**kwargs) for page in self.pages])
+
     def to_zip(
         self,
         file_path,

vlmparse/registries.py

@@ -77,9 +77,7 @@ class ConverterConfigRegistry:
         """Register a config factory for a model name."""
         self._registry[model_name] = config_factory
 
-    def get(
-        self, model_name: str, uri: str | None = None
-    ) -> OpenAIConverterConfig | None:
+    def get(self, model_name: str, uri: str | None = None) -> OpenAIConverterConfig:
         """Get config for a model name. Returns default if not registered."""
         if model_name in self._registry:
             return self._registry[model_name](uri=uri)

vlmparse/servers/docker_server.py

@@ -47,6 +47,19 @@ class DockerServerConfig(BaseModel):
         """Build command for container. Override in subclasses for specific logic."""
         return self.command_args if self.command_args else None
 
+    def update_command_args(
+        self,
+        vllm_args: dict | None = None,
+        forget_predefined_vllm_args: bool = False,
+    ) -> list[str]:
+        if vllm_args is not None:
+            if forget_predefined_vllm_args:
+                self.command_args = vllm_args
+            else:
+                self.command_args.extend(vllm_args)
+
+        return self.command_args
+
     def get_volumes(self) -> dict | None:
         """Setup volumes for container. Override in subclasses for specific logic."""
         return self.volumes
@@ -144,7 +157,7 @@ class ConverterServer:
         """Start the Docker server."""
         if self._server_context is not None:
             logger.warning("Server already started")
-            return self.base_url
+            return self.base_url, self._container
 
         # Use the generic docker_server for all server types
         self._server_context = docker_server(config=self.config, cleanup=self.auto_stop)

vlmparse/servers/utils.py

@@ -2,6 +2,7 @@ import getpass
 import time
 from contextlib import contextmanager
 from pathlib import Path
+from urllib.parse import parse_qsl, urlparse
 
 import docker
 from loguru import logger
@@ -222,25 +223,52 @@ def docker_server(
             logger.info("Container stopped")
 
 
+def normalize_uri(uri: str) -> tuple:
+    u = urlparse(uri)
+
+    # --- Normalize scheme ---
+    scheme = (u.scheme or "http").lower()
+
+    # --- Normalize host ---
+    host = (u.hostname or "").lower()
+    if host in ("localhost", "0.0.0.0"):
+        host = "localhost"
+
+    # --- Normalize port (apply defaults) ---
+    if u.port:
+        port = u.port
+    else:
+        port = 443 if scheme == "https" else 80
+
+    # --- Normalize path ---
+    # Treat empty path as "/" and remove trailing slash (except root)
+    path = u.path or "/"
+    if path != "/" and path.endswith("/"):
+        path = path.rstrip("/")
+
+    # Collapse duplicate slashes
+    while "//" in path:
+        path = path.replace("//", "/")
+
+    # --- Normalize query parameters (sorted) ---
+    query_pairs = parse_qsl(u.query, keep_blank_values=True)
+    query = "&".join(f"{k}={v}" for k, v in sorted(query_pairs))
+
+    return (scheme, host, port, path, query)
+
+
 def get_model_from_uri(uri: str) -> str:
     model = None
     client = docker.from_env()
     containers = client.containers.list()
+
+    uri = normalize_uri(uri)
+
     for container in containers:
         c_uri = container.labels.get("vlmparse_uri")
         c_model = container.labels.get("vlmparse_model_name")
-        if c_uri is not None:
-            c_uri = c_uri.replace("localhost", "0.0.0.0")
-
-        # Check if user URI matches container URI (ignoring /v1 suffix if missing)
-        if c_uri and (
-            c_uri == uri or c_uri.startswith(uri.rstrip("/")) or uri.startswith(c_uri)
-        ):
-            # Update URI to the correct one from container (likely has /v1)
-            if len(c_uri) > len(uri.rstrip("/")):
-                logger.info(f"Updating URI from {uri} to {c_uri}")
-                uri = c_uri
 
+        if c_uri and uri == normalize_uri(c_uri):
             # Infer model if not provided
             if model is None and c_model:
                 logger.info(f"Inferred model {c_model} from container")

{vlmparse-0.1.5.dist-info → vlmparse-0.1.7.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: vlmparse
-Version: 0.1.5
+Version: 0.1.7
 Requires-Python: >=3.11.0
 Description-Content-Type: text/markdown
 License-File: LICENSE
@@ -199,11 +199,13 @@ server.stop()
 ```
 
 
-Converter with automatic server deployment:
+Converter with automatic server management:
 
 ```python
 from vlmparse.converter_with_server import ConverterWithServer
 
-converter_with_server = ConverterWithServer(model="mineru2.5")
-documents = converter_with_server.parse(inputs=["file1.pdf", "file2.pdf"], out_folder="./output")
+with ConverterWithServer(model="mineru2.5") as converter_with_server:
+    documents = converter_with_server.parse(inputs=["file1.pdf", "file2.pdf"], out_folder="./output")
 ```
+
+Note that if you pass an uri of a vllm server to `ConverterWithServer`, the model name is inferred automatically and no server is started.

vlmparse-0.1.7.dist-info/RECORD

@@ -0,0 +1,36 @@
+vlmparse/base_model.py,sha256=4U4UPe8SNArliKnUf8pp8zQugWYsnhg9okylt7mrW1U,381
+vlmparse/build_doc.py,sha256=LAWrnFrqamN5PwJo57AUtQOPrMFGnCGw4gBjEKZ6pYo,2127
+vlmparse/cli.py,sha256=JfR6gk0pdYAavJgFTVx4OcgWdiLktGoKJ8TcVcD_IHw,12235
+vlmparse/constants.py,sha256=DYaK7KtTW8p9MPb3iPvoP5H1r7ICRuIFo89P01q4uCI,184
+vlmparse/converter.py,sha256=F0JSY9sFYUggCvaUCb27kKGJJpnZKW2FStMDVJoIOeQ,7383
+vlmparse/converter_with_server.py,sha256=zpUHDpHbDBs4Cj7dcVjvUQw0-U_InRNDC5Ekb_gehRM,6022
+vlmparse/registries.py,sha256=yBVrrhy61rSoLwdNV-z0C4lqIpTbLoWab3V6u7aSyNM,5797
+vlmparse/utils.py,sha256=rcVrtPiQVj_8HAmFQOu___72uYIapp_X89yxrMNCBow,1236
+vlmparse/clients/chandra.py,sha256=EulsCZdwOtm0pQ6CDm320U96k8aWFN4wKqCm1Xo7VCE,9775
+vlmparse/clients/deepseekocr.py,sha256=Uw6tPvP2KVsPDlz1ZUgYdbgQSjmFPuYeFDrGMMOTBAo,6501
+vlmparse/clients/docling.py,sha256=SAkLsqseuWfkuiel8FWR1G0Z5s-SZU3dE2JbsOvF4SA,5328
+vlmparse/clients/dotsocr.py,sha256=uGJoYEiDkP3-rmfdkAnMeAX-T4RZyEPoh6jmow5_-J8,10336
+vlmparse/clients/granite_docling.py,sha256=LMJAFjpSxcgLhsVxknSqrCC35MUTmklsE9PJZvMK2O8,4691
+vlmparse/clients/hunyuanocr.py,sha256=UFqaS4b8UM9EtizyrZIxlqcYlESmxm8xrQZP7lL6tkE,1857
+vlmparse/clients/lightonocr.py,sha256=wx1Im8Z3wlRWwYbPqnSd3LqTtdAU8CnX5mzu1BuCUY8,1314
+vlmparse/clients/mineru.py,sha256=6jZ1sKn2kGwUvD8gVs4PqEDH7uUXYK8pAB5Fr1JeqnY,3617
+vlmparse/clients/nanonetocr.py,sha256=BT5vaeerCsK5agvOaHK3NvLUqWd1FfDmrMmDYbp646I,1543
+vlmparse/clients/olmocr.py,sha256=A4Vl0meYpU5QPTML_OxyyRM07xCxtfrMZedgGMYEcuU,1851
+vlmparse/clients/openai_converter.py,sha256=bIDpR7Yn70eEp0pmzFoG2dDwY-mxCj3kH1IZS9BvXVQ,6266
+vlmparse/clients/paddleocrvl.py,sha256=qFBDj_UQocyq3WCh24tUOx9Ud7S9DfSm-1n3ztikY2s,1402
+vlmparse/clients/prompts.py,sha256=-J60lqxgRzlkQ9VsQLxmWsIMaDt-gNqWqWoqHIw9CLc,4228
+vlmparse/clients/pipe_utils/cleaner.py,sha256=oxBkBTOkluN1lmeNbzajRIe0_D__ZGwUOBaI_Ph0uxE,2396
+vlmparse/clients/pipe_utils/html_to_md_conversion.py,sha256=cFFqzD2jCNw_968_eu3Wt--Ox7iJj2Rn5UoP_DZWosU,4112
+vlmparse/clients/pipe_utils/utils.py,sha256=935ecIO446I0pstszE_1nrIPHn1Ffrxunq7fVd0dsd8,315
+vlmparse/data_model/box.py,sha256=lJsh4qhjgYXZF5vTSJ1qMXD5GVlBi2_SBedBMlfJikU,16868
+vlmparse/data_model/document.py,sha256=xheaMeStOj2c9GZKmdtxcEl_Dj44V5JyVp6JnTrSpH0,4615
+vlmparse/servers/docker_server.py,sha256=UVU7VDloJ8Yfqj-WUv3Trti9AODcdC9JyTzW3sCM-l4,7032
+vlmparse/servers/utils.py,sha256=tIXhgbF9EVOJy2nYEguVq69gn9ATxtya_1F4wZSt68o,9454
+vlmparse/st_viewer/fs_nav.py,sha256=7GNH68h2Loh5pQ64Pe72-D2cs2BLhqRXevEmKdFmPX0,1616
+vlmparse/st_viewer/st_viewer.py,sha256=m2rQTtk5rlwErNmivNAg-4rkHkvNkvLhoJZxFQi7Dwk,2105
+vlmparse-0.1.7.dist-info/licenses/LICENSE,sha256=3TKJHk8hPBR5dbLWZ3IpfCftl-_m-iyBwpYQGZYxj14,1080
+vlmparse-0.1.7.dist-info/METADATA,sha256=DP--8aCeLxAgvo6vvaDog7xzzMzvZywVvCrMiAKhDbo,5597
+vlmparse-0.1.7.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+vlmparse-0.1.7.dist-info/entry_points.txt,sha256=gD5berP6HwE2wNIkls-Lw5goiceA8uMgPEd7ifnFJXs,47
+vlmparse-0.1.7.dist-info/top_level.txt,sha256=k4ni-GNH_iAX7liQEsk_KY_c3xgZgt8k9fsSs9IXLXs,9
+vlmparse-0.1.7.dist-info/RECORD,,

vlmparse-0.1.5.dist-info/RECORD

@@ -1,36 +0,0 @@
-vlmparse/base_model.py,sha256=4U4UPe8SNArliKnUf8pp8zQugWYsnhg9okylt7mrW1U,381
-vlmparse/build_doc.py,sha256=LAWrnFrqamN5PwJo57AUtQOPrMFGnCGw4gBjEKZ6pYo,2127
-vlmparse/cli.py,sha256=tQma1IkOsFnqPKqqHVO1PJh18n1w82gp4ewA7oraJkE,15855
-vlmparse/constants.py,sha256=7-47S01n4MI2ebR09bpdOo3_P16d-z-NVGsm6KJP8ls,110
-vlmparse/converter.py,sha256=F0JSY9sFYUggCvaUCb27kKGJJpnZKW2FStMDVJoIOeQ,7383
-vlmparse/converter_with_server.py,sha256=G393O7vU_lJz6Vz-qYVkrjFhf0Vmpjjl8OjPKQe2blU,3928
-vlmparse/registries.py,sha256=6bEUKTkTjc8C7c1R1ZvAHSF5NCXmAuhNpw0qNnuQ7-A,5818
-vlmparse/utils.py,sha256=rcVrtPiQVj_8HAmFQOu___72uYIapp_X89yxrMNCBow,1236
-vlmparse/clients/chandra.py,sha256=zfu-A6Slh-fIAyrtrlVoCb6QHLBimnimefap_K9YwYw,9775
-vlmparse/clients/deepseekocr.py,sha256=rQvaOaPPoDiZ0MzXqfqqH9BgUBfjmlfHu3NlMjSDgiQ,6501
-vlmparse/clients/docling.py,sha256=SAkLsqseuWfkuiel8FWR1G0Z5s-SZU3dE2JbsOvF4SA,5328
-vlmparse/clients/dotsocr.py,sha256=w2T-xkhlw1AfT-CUYoF0ectr2jDYHe9239B24XKB1UQ,10139
-vlmparse/clients/granite_docling.py,sha256=EQpsv5qSJG0HtMSacmJStER2sq4TGf1EMU5_NmJsl4g,4634
-vlmparse/clients/hunyuanocr.py,sha256=Xw0Q1l-3pQzaEgFngnfM8vrSWpnT3I99QvDaGZ8XooM,1712
-vlmparse/clients/lightonocr.py,sha256=wx1Im8Z3wlRWwYbPqnSd3LqTtdAU8CnX5mzu1BuCUY8,1314
-vlmparse/clients/mineru.py,sha256=bilDPcUoLk2rcFVqMk4q2Hx2txilc3GDUbjAEoMM_BI,3671
-vlmparse/clients/nanonetocr.py,sha256=BT5vaeerCsK5agvOaHK3NvLUqWd1FfDmrMmDYbp646I,1543
-vlmparse/clients/olmocr.py,sha256=mQEDpfyLY8a80Zlps5mG0QaWytIgnNQZVEVWKWjPIjk,1849
-vlmparse/clients/openai_converter.py,sha256=j2H0iAQTADRRpu1Zy1b-1OFfWyXuqCvrQKy2UcwggTA,5696
-vlmparse/clients/paddleocrvl.py,sha256=tmaqg3boV4edywiiiNiNiI3dBHi111wz4dFb52OISXw,1376
-vlmparse/clients/prompts.py,sha256=-J60lqxgRzlkQ9VsQLxmWsIMaDt-gNqWqWoqHIw9CLc,4228
-vlmparse/clients/pipe_utils/cleaner.py,sha256=oxBkBTOkluN1lmeNbzajRIe0_D__ZGwUOBaI_Ph0uxE,2396
-vlmparse/clients/pipe_utils/html_to_md_conversion.py,sha256=cFFqzD2jCNw_968_eu3Wt--Ox7iJj2Rn5UoP_DZWosU,4112
-vlmparse/clients/pipe_utils/utils.py,sha256=935ecIO446I0pstszE_1nrIPHn1Ffrxunq7fVd0dsd8,315
-vlmparse/data_model/box.py,sha256=lJsh4qhjgYXZF5vTSJ1qMXD5GVlBi2_SBedBMlfJikU,16868
-vlmparse/data_model/document.py,sha256=pdCZvWzRFkez53ZJpNaB4ezUW-OVUlbR3_SBmmgVzGQ,4217
-vlmparse/servers/docker_server.py,sha256=qOoZcWSHrK7kK7tAL61RJSW-Jmee93It2SEfWG3jGrc,6633
-vlmparse/servers/utils.py,sha256=qy2-rnQTCQKt6CeTV5H74tvRTXyzBV2KswQiYW8Tf-k,8908
-vlmparse/st_viewer/fs_nav.py,sha256=7GNH68h2Loh5pQ64Pe72-D2cs2BLhqRXevEmKdFmPX0,1616
-vlmparse/st_viewer/st_viewer.py,sha256=m2rQTtk5rlwErNmivNAg-4rkHkvNkvLhoJZxFQi7Dwk,2105
-vlmparse-0.1.5.dist-info/licenses/LICENSE,sha256=3TKJHk8hPBR5dbLWZ3IpfCftl-_m-iyBwpYQGZYxj14,1080
-vlmparse-0.1.5.dist-info/METADATA,sha256=LN4W1cvXJvL22hwLAgeSwd3PGTmlrt6lgqNi-tL9pes,5446
-vlmparse-0.1.5.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-vlmparse-0.1.5.dist-info/entry_points.txt,sha256=gD5berP6HwE2wNIkls-Lw5goiceA8uMgPEd7ifnFJXs,47
-vlmparse-0.1.5.dist-info/top_level.txt,sha256=k4ni-GNH_iAX7liQEsk_KY_c3xgZgt8k9fsSs9IXLXs,9
-vlmparse-0.1.5.dist-info/RECORD,,