groundx 3.7.2__tar.gz → 3.7.3__tar.gz

{groundx-3.7.2 → groundx-3.7.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: groundx
-Version: 3.7.2
+Version: 3.7.3
 Summary: 
 License: MIT
 Requires-Python: >=3.10,<4.0

{groundx-3.7.2 → groundx-3.7.3}/pyproject.toml

@@ -4,7 +4,7 @@ dynamic = ["version"]
 
 [tool.poetry]
 name = "groundx"
-version = "3.7.2"
+version = "3.7.3"
 description = ""
 readme = "README.md"
 authors = []

{groundx-3.7.2 → groundx-3.7.3}/src/groundx/core/client_wrapper.py

@@ -27,12 +27,12 @@ class BaseClientWrapper:
         import platform
 
         headers: typing.Dict[str, str] = {
-            "User-Agent": "groundx/3.7.2",
+            "User-Agent": "groundx/3.7.3",
             "X-Fern-Language": "Python",
             "X-Fern-Runtime": f"python/{platform.python_version()}",
             "X-Fern-Platform": f"{platform.system().lower()}/{platform.release()}",
             "X-Fern-SDK-Name": "groundx",
-            "X-Fern-SDK-Version": "3.7.2",
+            "X-Fern-SDK-Version": "3.7.3",
             **(self.get_custom_headers() or {}),
         }
         headers["X-API-Key"] = self.api_key

{groundx-3.7.2 → groundx-3.7.3}/src/groundx/extract/agents/agent.py

@@ -1,13 +1,22 @@
 from __future__ import annotations
 
 import json, traceback, typing
+from dataclasses import dataclass, field
+from urllib.parse import urlparse
 
 from smolagents import (  # pyright: ignore[reportMissingTypeStubs]
     CodeAgent,
     Tool,
     ToolCallingAgent,
 )
+from smolagents.memory import (  # pyright: ignore[reportMissingTypeStubs]
+    FinalAnswerStep,
+    SystemPromptStep,
+    TaskStep,
+)
 from smolagents.models import (  # pyright: ignore[reportMissingTypeStubs]
+    ChatMessage,
+    MessageRole,
     OpenAIServerModel,
 )
 
@@ -27,6 +36,37 @@ Return only your response using the `final_answer` tool format:
 ```
 """
 
+SUPPORTED_IMAGE_TRANSPORTS = {"pil", "data_url", "remote_url"}
+
+
+@dataclass
+class RemoteImageTaskStep(TaskStep):
+    image_urls: typing.List[str] = field(default_factory=list)
+
+    def to_messages(self, summary_mode: bool = False) -> typing.List[ChatMessage]:
+        content: typing.List[typing.Dict[str, typing.Any]] = [
+            {"type": "text", "text": f"New task:\n{self.task}"}
+        ]
+        content.extend(
+            {"type": "image_url", "image_url": {"url": image_url}}
+            for image_url in self.image_urls
+        )
+        return [ChatMessage(role=MessageRole.USER, content=content)]
+
+    def dict(self) -> typing.Dict[str, typing.Any]:
+        return {
+            "task": self.task,
+            "task_images": None,
+            "image_urls": [_sanitize_image_url(image_url) for image_url in self.image_urls],
+        }
+
+
+def _sanitize_image_url(image_url: str) -> str:
+    parsed = urlparse(image_url)
+    if not parsed.scheme or not parsed.netloc:
+        return "<invalid-image-url>"
+    return f"{parsed.scheme}://{parsed.netloc}{parsed.path}"
+
 
 def extract_response(res: typing.Dict[str, typing.Any]) -> typing.Any:
     if "answer" in res and "type" in res["answer"]:
@@ -209,6 +249,7 @@ class AgentTool(ToolCallingAgent):
         )
 
         self.log = log
+        self.image_transport = settings.image_transport
 
     def process(
         self,
@@ -217,11 +258,25 @@ class AgentTool(ToolCallingAgent):
         expected_types: typing.Union[type, typing.Tuple[type, ...]] = dict,
         attempt: int = 0,
         type: str = "",
+        image_urls: typing.Optional[typing.Sequence[str]] = None,
+        image_transport: typing.Optional[str] = None,
     ) -> typing.Any:
-        res = super().run(  # pyright: ignore[reportUnknownMemberType]
-            conflict + prompt_suffix,
+        selected_image_transport = image_transport or self.image_transport
+        transport = self._validate_image_inputs(
             images=images,
+            image_urls=image_urls,
+            image_transport=selected_image_transport,
         )
+        if transport == "remote_url":
+            res = self._run_with_remote_image_urls(
+                conflict + prompt_suffix,
+                list(image_urls or []),
+            )
+        else:
+            res = super().run(  # pyright: ignore[reportUnknownMemberType]
+                conflict + prompt_suffix,
+                images=images,
+            )
 
         try:
             return process_response(res=res, expected_types=expected_types)
@@ -236,4 +291,73 @@ class AgentTool(ToolCallingAgent):
                 f"agent process result is not of expected type(s) {expected_types!r}: [{e}], attempting again [{attempt+1}]\n\n{res}"
             )
 
-            return self.process(conflict, images, expected_types, attempt + 1)
+            return self.process(
+                conflict,
+                images,
+                expected_types,
+                attempt + 1,
+                type,
+                image_urls=image_urls,
+                image_transport=transport,
+            )
+
+    def _run_with_remote_image_urls(
+        self,
+        task: str,
+        image_urls: typing.List[str],
+    ) -> typing.Any:
+        self.task = task
+        self.interrupt_switch = False
+        self.memory.system_prompt = SystemPromptStep(system_prompt=self.system_prompt)
+        self.memory.reset()
+        self.monitor.reset()
+        self.memory.steps.append(
+            RemoteImageTaskStep(task=self.task, task_images=None, image_urls=image_urls)
+        )
+
+        try:
+            try:
+                steps = list(self._run_stream(task=self.task, max_steps=self.max_steps, images=None))
+            except Exception as exc:
+                raise RuntimeError(f"remote_url image transport failed: {exc}") from exc
+            assert isinstance(steps[-1], FinalAnswerStep)
+            return steps[-1].output
+        finally:
+            self._sanitize_remote_image_urls_in_memory()
+
+    def _validate_image_inputs(
+        self,
+        *,
+        images: typing.Sequence[Image],
+        image_urls: typing.Optional[typing.Sequence[str]],
+        image_transport: str,
+    ) -> str:
+        if image_transport not in SUPPORTED_IMAGE_TRANSPORTS:
+            raise ValueError(
+                "unsupported image_transport "
+                f"[{image_transport}]; expected one of {sorted(SUPPORTED_IMAGE_TRANSPORTS)}"
+            )
+        if image_urls and image_transport != "remote_url":
+            raise ValueError("image_urls require image_transport='remote_url'")
+        if image_transport == "remote_url":
+            if images:
+                raise ValueError("remote_url transport does not accept PIL images")
+            if not image_urls:
+                raise ValueError("remote_url transport requires image_urls")
+        return image_transport
+
+    def _sanitize_remote_image_urls_in_memory(self) -> None:
+        for step in self.memory.steps:
+            model_input_messages = getattr(step, "model_input_messages", None)
+            if not model_input_messages:
+                continue
+            for message in model_input_messages:
+                content = getattr(message, "content", None)
+                if not isinstance(content, list):
+                    continue
+                for item in content:
+                    if not isinstance(item, dict) or item.get("type") != "image_url":
+                        continue
+                    image_url = item.get("image_url")
+                    if isinstance(image_url, dict) and isinstance(image_url.get("url"), str):
+                        image_url["url"] = _sanitize_image_url(image_url["url"])

{groundx-3.7.2 → groundx-3.7.3}/src/groundx/extract/settings/settings.py

@@ -43,6 +43,7 @@ class AgentSettings(BaseModel):
     max_steps: int = 7
     model_id: str = "gpt-5-mini"
     reasoning_effort: typing.Optional[str] = "medium"
+    image_transport: str = "pil"
     model_kwargs: typing.Optional[typing.Dict[str, typing.Any]] = None
 
     def get_api_key(self) -> str: