openaivec 0.14.2__py3-none-any.whl → 0.14.4__py3-none-any.whl

openaivec/pandas_ext.py

@@ -74,6 +74,21 @@ __all__ = [
 _LOGGER = logging.getLogger(__name__)
 
 
+# ---------------------------------------------------------------------------
+# Internal helpers (not exported)
+# ---------------------------------------------------------------------------
+def _df_rows_to_json_series(df: pd.DataFrame) -> pd.Series:
+    """Return a Series of JSON strings (UTF-8, no ASCII escaping) representing DataFrame rows.
+
+    Each element is the JSON serialisation of the corresponding row as a dict. Index and
+    name are preserved so downstream operations retain alignment. This consolidates the
+    previously duplicated inline pipeline used by responses*/task* DataFrame helpers.
+    """
+    return pd.Series(df.to_dict(orient="records"), index=df.index, name="record").map(
+        lambda x: json.dumps(x, ensure_ascii=False)
+    )
+
+
 T = TypeVar("T")  # For pipe function return type
 
 
@@ -165,7 +180,29 @@ class OpenAIVecSeriesAccessor:
         response_format: Type[ResponseFormat] = str,
         temperature: float | None = 0.0,
         top_p: float = 1.0,
+        **api_kwargs,
     ) -> pd.Series:
+        """Call an LLM once for every Series element using a provided cache.
+
+        This is a lower-level method that allows explicit cache management for advanced
+        use cases. Most users should use the standard ``responses`` method instead.
+
+        Args:
+            instructions (str): System prompt prepended to every user message.
+            cache (BatchingMapProxy[str, ResponseFormat]): Explicit cache instance for
+                batching and deduplication control.
+            response_format (Type[ResponseFormat], optional): Pydantic model or built-in
+                type the assistant should return. Defaults to ``str``.
+            temperature (float | None, optional): Sampling temperature. Defaults to ``0.0``.
+            top_p (float, optional): Nucleus sampling parameter. Defaults to ``1.0``.
+
+        Additional Keyword Args:
+            Arbitrary OpenAI Responses API parameters (e.g. ``frequency_penalty``, ``presence_penalty``,
+            ``seed``, etc.) are forwarded verbatim to the underlying client.
+
+        Returns:
+            pandas.Series: Series whose values are instances of ``response_format``.
+        """
         client: BatchResponses = BatchResponses(
             client=CONTAINER.resolve(OpenAI),
             model_name=CONTAINER.resolve(ResponsesModelName).value,
@@ -176,7 +213,58 @@ class OpenAIVecSeriesAccessor:
             top_p=top_p,
         )
 
-        return pd.Series(client.parse(self._obj.tolist()), index=self._obj.index, name=self._obj.name)
+        # Forward any extra kwargs to the underlying Responses API.
+        return pd.Series(client.parse(self._obj.tolist(), **api_kwargs), index=self._obj.index, name=self._obj.name)
+
+    def responses(
+        self,
+        instructions: str,
+        response_format: Type[ResponseFormat] = str,
+        batch_size: int | None = None,
+        temperature: float | None = 0.0,
+        top_p: float = 1.0,
+        show_progress: bool = False,
+        **api_kwargs,
+    ) -> pd.Series:
+        """Call an LLM once for every Series element.
+
+        Example:
+            ```python
+            animals = pd.Series(["cat", "dog", "elephant"])
+            # Basic usage
+            animals.ai.responses("translate to French")
+
+            # With progress bar in Jupyter notebooks
+            large_series = pd.Series(["data"] * 1000)
+            large_series.ai.responses(
+                "analyze this data",
+                batch_size=32,
+                show_progress=True
+            )
+            ```
+
+        Args:
+            instructions (str): System prompt prepended to every user message.
+            response_format (Type[ResponseFormat], optional): Pydantic model or built‑in
+                type the assistant should return. Defaults to ``str``.
+            batch_size (int | None, optional): Number of prompts grouped into a single
+                request. Defaults to ``None`` (automatic batch size optimization
+                based on execution time). Set to a positive integer for fixed batch size.
+            temperature (float | None, optional): Sampling temperature. Defaults to ``0.0``.
+            top_p (float, optional): Nucleus sampling parameter. Defaults to ``1.0``.
+            show_progress (bool, optional): Show progress bar in Jupyter notebooks. Defaults to ``False``.
+
+        Returns:
+            pandas.Series: Series whose values are instances of ``response_format``.
+        """
+        return self.responses_with_cache(
+            instructions=instructions,
+            cache=BatchingMapProxy(batch_size=batch_size, show_progress=show_progress),
+            response_format=response_format,
+            temperature=temperature,
+            top_p=top_p,
+            **api_kwargs,
+        )
 
     def embeddings_with_cache(
         self,
@@ -188,15 +276,6 @@ class OpenAIVecSeriesAccessor:
         a pre-configured BatchingMapProxy instance, enabling cache sharing
         across multiple operations or custom batch size management.
 
-        Args:
-            cache (BatchingMapProxy[str, np.ndarray]): Pre-configured cache
-                instance for managing API call batching and deduplication.
-                Set cache.batch_size=None to enable automatic batch size optimization.
-
-        Returns:
-            pandas.Series: Series whose values are ``np.ndarray`` objects
-                (dtype ``float32``).
-
         Example:
             ```python
             from openaivec._proxy import BatchingMapProxy
@@ -208,6 +287,15 @@ class OpenAIVecSeriesAccessor:
             animals = pd.Series(["cat", "dog", "elephant"])
             embeddings = animals.ai.embeddings_with_cache(cache=shared_cache)
             ```
+
+        Args:
+            cache (BatchingMapProxy[str, np.ndarray]): Pre-configured cache
+                instance for managing API call batching and deduplication.
+                Set cache.batch_size=None to enable automatic batch size optimization.
+
+        Returns:
+            pandas.Series: Series whose values are ``np.ndarray`` objects
+                (dtype ``float32``).
         """
         client: BatchEmbeddings = BatchEmbeddings(
             client=CONTAINER.resolve(OpenAI),
@@ -221,96 +309,69 @@ class OpenAIVecSeriesAccessor:
             name=self._obj.name,
         )
 
-    def responses(
-        self,
-        instructions: str,
-        response_format: Type[ResponseFormat] = str,
-        batch_size: int | None = None,
-        temperature: float | None = 0.0,
-        top_p: float = 1.0,
-        show_progress: bool = False,
-    ) -> pd.Series:
-        """Call an LLM once for every Series element.
+    def embeddings(self, batch_size: int | None = None, show_progress: bool = False) -> pd.Series:
+        """Compute OpenAI embeddings for every Series element.
 
         Example:
             ```python
             animals = pd.Series(["cat", "dog", "elephant"])
             # Basic usage
-            animals.ai.responses("translate to French")
+            animals.ai.embeddings()
 
-            # With progress bar in Jupyter notebooks
-            large_series = pd.Series(["data"] * 1000)
-            large_series.ai.responses(
-                "analyze this data",
-                batch_size=32,
+            # With progress bar for large datasets
+            large_texts = pd.Series(["text"] * 5000)
+            embeddings = large_texts.ai.embeddings(
+                batch_size=100,
                 show_progress=True
             )
             ```
-            This method returns a Series of strings, each containing the
-            assistant's response to the corresponding input.
-            The model used is set by the `responses_model` function.
-            The default model is `gpt-4.1-mini`.
 
         Args:
-            instructions (str): System prompt prepended to every user message.
-            response_format (Type[ResponseFormat], optional): Pydantic model or built‑in
-                type the assistant should return. Defaults to ``str``.
-            batch_size (int | None, optional): Number of prompts grouped into a single
-                request. Defaults to ``None`` (automatic batch size optimization
+            batch_size (int | None, optional): Number of inputs grouped into a
+                single request. Defaults to ``None`` (automatic batch size optimization
                 based on execution time). Set to a positive integer for fixed batch size.
-            temperature (float, optional): Sampling temperature. Defaults to ``0.0``.
-            top_p (float, optional): Nucleus sampling parameter. Defaults to ``1.0``.
             show_progress (bool, optional): Show progress bar in Jupyter notebooks. Defaults to ``False``.
 
         Returns:
-            pandas.Series: Series whose values are instances of ``response_format``.
+            pandas.Series: Series whose values are ``np.ndarray`` objects
+                (dtype ``float32``).
         """
-        return self.responses_with_cache(
-            instructions=instructions,
+        return self.embeddings_with_cache(
             cache=BatchingMapProxy(batch_size=batch_size, show_progress=show_progress),
-            response_format=response_format,
-            temperature=temperature,
-            top_p=top_p,
         )
 
     def task_with_cache(
         self,
         task: PreparedTask[ResponseFormat],
         cache: BatchingMapProxy[str, ResponseFormat],
+        **api_kwargs,
     ) -> pd.Series:
         """Execute a prepared task on every Series element using a provided cache.
 
-        This method allows external control over caching behavior by accepting
-        a pre-configured BatchingMapProxy instance, enabling cache sharing
-        across multiple operations or custom batch size management.
-
-        Args:
-            task (PreparedTask): A pre-configured task containing instructions,
-                response format, and other parameters for processing the inputs.
-            cache (BatchingMapProxy[str, ResponseFormat]): Pre-configured cache
-                instance for managing API call batching and deduplication.
-                Set cache.batch_size=None to enable automatic batch size optimization.
-
-        Returns:
-            pandas.Series: Series whose values are instances of the task's
-                response format, aligned with the original Series index.
+        This mirrors ``responses_with_cache`` but uses the task's stored instructions,
+        response format, temperature and top_p. A supplied ``BatchingMapProxy`` enables
+        cross‑operation deduplicated reuse and external batch size / progress control.
 
         Example:
             ```python
-            from openaivec._model import PreparedTask
             from openaivec._proxy import BatchingMapProxy
-
-            # Create a shared cache with custom batch size
             shared_cache = BatchingMapProxy(batch_size=64)
+            reviews.ai.task_with_cache(sentiment_task, cache=shared_cache)
+            ```
 
-            # Assume you have a prepared task for sentiment analysis
-            sentiment_task = PreparedTask(...)
+        Args:
+            task (PreparedTask): Prepared task (instructions + response_format + sampling params).
+            cache (BatchingMapProxy[str, ResponseFormat]): Pre‑configured cache instance.
 
-            reviews = pd.Series(["Great product!", "Not satisfied", "Amazing quality"])
-            results = reviews.ai.task_with_cache(sentiment_task, cache=shared_cache)
-            ```
+        Additional Keyword Args:
+            Arbitrary OpenAI Responses API parameters (e.g. ``frequency_penalty``, ``presence_penalty``,
+            ``seed``, etc.) forwarded verbatim to the underlying client. Core routing keys
+            (``model``, system instructions, user input) are managed internally and cannot be overridden.
+
+        Returns:
+            pandas.Series: Task results aligned with the original Series index.
         """
-        client = BatchResponses(
+        client: BatchResponses = BatchResponses(
             client=CONTAINER.resolve(OpenAI),
             model_name=CONTAINER.resolve(ResponsesModelName).value,
             system_message=task.instructions,
@@ -319,15 +380,17 @@ class OpenAIVecSeriesAccessor:
             temperature=task.temperature,
             top_p=task.top_p,
         )
-        return pd.Series(client.parse(self._obj.tolist()), index=self._obj.index, name=self._obj.name)
+        return pd.Series(client.parse(self._obj.tolist(), **api_kwargs), index=self._obj.index, name=self._obj.name)
 
-    def task(self, task: PreparedTask, batch_size: int | None = None, show_progress: bool = False) -> pd.Series:
+    def task(
+        self,
+        task: PreparedTask,
+        batch_size: int | None = None,
+        show_progress: bool = False,
+        **api_kwargs,
+    ) -> pd.Series:
         """Execute a prepared task on every Series element.
 
-        This method applies a pre-configured task to each element in the Series,
-        using the task's instructions and response format to generate structured
-        responses from the language model.
-
         Example:
             ```python
             from openaivec._model import PreparedTask
@@ -347,8 +410,6 @@ class OpenAIVecSeriesAccessor:
                 show_progress=True
             )
             ```
-            This method returns a Series containing the task results for each
-            corresponding input element, following the task's defined structure.
 
         Args:
             task (PreparedTask): A pre-configured task containing instructions,
@@ -358,48 +419,19 @@ class OpenAIVecSeriesAccessor:
                 optimization based on execution time). Set to a positive integer for fixed batch size.
             show_progress (bool, optional): Show progress bar in Jupyter notebooks. Defaults to ``False``.
 
+        Additional Keyword Args:
+            Arbitrary OpenAI Responses API parameters (e.g. ``frequency_penalty``, ``presence_penalty``,
+            ``seed``, etc.) are forwarded verbatim to the underlying client. Core batching / routing
+            keys (``model``, ``instructions`` / system message, user ``input``) are managed by the
+            library and cannot be overridden.
+
         Returns:
-            pandas.Series: Series whose values are instances of the task's
-                response format, aligned with the original Series index.
+            pandas.Series: Series whose values are instances of the task's response format.
         """
         return self.task_with_cache(
             task=task,
             cache=BatchingMapProxy(batch_size=batch_size, show_progress=show_progress),
-        )
-
-    def embeddings(self, batch_size: int | None = None, show_progress: bool = False) -> pd.Series:
-        """Compute OpenAI embeddings for every Series element.
-
-        Example:
-            ```python
-            animals = pd.Series(["cat", "dog", "elephant"])
-            # Basic usage
-            animals.ai.embeddings()
-
-            # With progress bar for large datasets
-            large_texts = pd.Series(["text"] * 5000)
-            embeddings = large_texts.ai.embeddings(
-                batch_size=100,
-                show_progress=True
-            )
-            ```
-            This method returns a Series of numpy arrays, each containing the
-            embedding vector for the corresponding input.
-            The embedding model is set by the `embeddings_model` function.
-            The default embedding model is `text-embedding-3-small`.
-
-        Args:
-            batch_size (int | None, optional): Number of inputs grouped into a
-                single request. Defaults to ``None`` (automatic batch size optimization
-                based on execution time). Set to a positive integer for fixed batch size.
-            show_progress (bool, optional): Show progress bar in Jupyter notebooks. Defaults to ``False``.
-
-        Returns:
-            pandas.Series: Series whose values are ``np.ndarray`` objects
-                (dtype ``float32``).
-        """
-        return self.embeddings_with_cache(
-            cache=BatchingMapProxy(batch_size=batch_size, show_progress=show_progress),
+            **api_kwargs,
         )
 
     def count_tokens(self) -> pd.Series:
@@ -456,38 +488,6 @@ class OpenAIVecDataFrameAccessor:
     def __init__(self, df_obj: pd.DataFrame):
         self._obj = df_obj
 
-    def extract(self, column: str) -> pd.DataFrame:
-        """Flatten one column of Pydantic models/dicts into top‑level columns.
-
-        Example:
-            ```python
-            df = pd.DataFrame([
-                {"animal": {"name": "cat", "legs": 4}},
-                {"animal": {"name": "dog", "legs": 4}},
-                {"animal": {"name": "elephant", "legs": 4}},
-            ])
-            df.ai.extract("animal")
-            ```
-            This method returns a DataFrame with the same index as the original,
-            where each column corresponds to a key in the dictionaries.
-            The source column is dropped.
-
-        Args:
-            column (str): Column to expand.
-
-        Returns:
-            pandas.DataFrame: Original DataFrame with the extracted columns; the source column is dropped.
-        """
-        if column not in self._obj.columns:
-            raise ValueError(f"Column '{column}' does not exist in the DataFrame.")
-
-        return (
-            self._obj.pipe(lambda df: df.reset_index(drop=True))
-            .pipe(lambda df: df.join(df[column].ai.extract()))
-            .pipe(lambda df: df.set_index(self._obj.index))
-            .pipe(lambda df: df.drop(columns=[column], axis=1))
-        )
-
     def responses_with_cache(
         self,
         instructions: str,
@@ -495,26 +495,14 @@ class OpenAIVecDataFrameAccessor:
         response_format: Type[ResponseFormat] = str,
         temperature: float | None = 0.0,
         top_p: float = 1.0,
+        **api_kwargs,
     ) -> pd.Series:
-        """Generate a response for each row after serialising it to JSON using a provided cache.
+        """Generate a response for each row after serializing it to JSON using a provided cache.
 
         This method allows external control over caching behavior by accepting
         a pre-configured BatchingMapProxy instance, enabling cache sharing
         across multiple operations or custom batch size management.
 
-        Args:
-            instructions (str): System prompt for the assistant.
-            cache (BatchingMapProxy[str, ResponseFormat]): Pre-configured cache
-                instance for managing API call batching and deduplication.
-                Set cache.batch_size=None to enable automatic batch size optimization.
-            response_format (Type[ResponseFormat], optional): Desired Python type of the
-                responses. Defaults to ``str``.
-            temperature (float, optional): Sampling temperature. Defaults to ``0.0``.
-            top_p (float, optional): Nucleus sampling parameter. Defaults to ``1.0``.
-
-        Returns:
-            pandas.Series: Responses aligned with the DataFrame's original index.
-
         Example:
             ```python
             from openaivec._proxy import BatchingMapProxy
@@ -532,19 +520,27 @@ class OpenAIVecDataFrameAccessor:
                 cache=shared_cache
             )
             ```
+
+        Args:
+            instructions (str): System prompt for the assistant.
+            cache (BatchingMapProxy[str, ResponseFormat]): Pre-configured cache
+                instance for managing API call batching and deduplication.
+                Set cache.batch_size=None to enable automatic batch size optimization.
+            response_format (Type[ResponseFormat], optional): Desired Python type of the
+                responses. Defaults to ``str``.
+            temperature (float | None, optional): Sampling temperature. Defaults to ``0.0``.
+            top_p (float, optional): Nucleus sampling parameter. Defaults to ``1.0``.
+
+        Returns:
+            pandas.Series: Responses aligned with the DataFrame's original index.
         """
-        return self._obj.pipe(
-            lambda df: (
-                df.pipe(lambda df: pd.Series(df.to_dict(orient="records"), index=df.index, name="record"))
-                .map(lambda x: json.dumps(x, ensure_ascii=False))
-                .ai.responses_with_cache(
-                    instructions=instructions,
-                    cache=cache,
-                    response_format=response_format,
-                    temperature=temperature,
-                    top_p=top_p,
-                )
-            )
+        return _df_rows_to_json_series(self._obj).ai.responses_with_cache(
+            instructions=instructions,
+            cache=cache,
+            response_format=response_format,
+            temperature=temperature,
+            top_p=top_p,
+            **api_kwargs,
         )
 
     def responses(
@@ -555,8 +551,9 @@ class OpenAIVecDataFrameAccessor:
         temperature: float | None = 0.0,
         top_p: float = 1.0,
         show_progress: bool = False,
+        **api_kwargs,
     ) -> pd.Series:
-        """Generate a response for each row after serialising it to JSON.
+        """Generate a response for each row after serializing it to JSON.
 
         Example:
             ```python
@@ -576,11 +573,6 @@ class OpenAIVecDataFrameAccessor:
                 show_progress=True
             )
             ```
-            This method returns a Series of strings, each containing the
-            assistant's response to the corresponding input.
-            Each row is serialised to JSON before being sent to the assistant.
-            The model used is set by the `responses_model` function.
-            The default model is `gpt-4.1-mini`.
 
         Args:
             instructions (str): System prompt for the assistant.
@@ -589,7 +581,7 @@ class OpenAIVecDataFrameAccessor:
             batch_size (int | None, optional): Number of requests sent in one batch.
                 Defaults to ``None`` (automatic batch size optimization
                 based on execution time). Set to a positive integer for fixed batch size.
-            temperature (float, optional): Sampling temperature. Defaults to ``0.0``.
+            temperature (float | None, optional): Sampling temperature. Defaults to ``0.0``.
             top_p (float, optional): Nucleus sampling parameter. Defaults to ``1.0``.
             show_progress (bool, optional): Show progress bar in Jupyter notebooks. Defaults to ``False``.
 
@@ -602,15 +594,42 @@ class OpenAIVecDataFrameAccessor:
             response_format=response_format,
             temperature=temperature,
             top_p=top_p,
+            **api_kwargs,
         )
 
-    def task(self, task: PreparedTask, batch_size: int | None = None, show_progress: bool = False) -> pd.Series:
-        """Execute a prepared task on each DataFrame row after serialising it to JSON.
+    def task_with_cache(
+        self,
+        task: PreparedTask[ResponseFormat],
+        cache: BatchingMapProxy[str, ResponseFormat],
+        **api_kwargs,
+    ) -> pd.Series:
+        """Execute a prepared task on each DataFrame row after serializing it to JSON using a provided cache.
+
+        Args:
+            task (PreparedTask): Prepared task (instructions + response_format + sampling params).
+            cache (BatchingMapProxy[str, ResponseFormat]): Pre‑configured cache instance.
+
+        Additional Keyword Args:
+            Arbitrary OpenAI Responses API parameters (e.g. ``frequency_penalty``, ``presence_penalty``,
+            ``seed``) forwarded verbatim. Core routing keys are managed internally.
+
+        Returns:
+            pandas.Series: Task results aligned with the DataFrame's original index.
+        """
+        return _df_rows_to_json_series(self._obj).ai.task_with_cache(
+            task=task,
+            cache=cache,
+            **api_kwargs,
+        )
 
-        This method applies a pre-configured task to each row in the DataFrame,
-        using the task's instructions and response format to generate structured
-        responses from the language model. Each row is serialised to JSON before
-        being processed by the task.
+    def task(
+        self,
+        task: PreparedTask,
+        batch_size: int | None = None,
+        show_progress: bool = False,
+        **api_kwargs,
+    ) -> pd.Series:
+        """Execute a prepared task on each DataFrame row after serializing it to JSON.
 
         Example:
             ```python
@@ -624,10 +643,17 @@ class OpenAIVecDataFrameAccessor:
                 {"name": "dog", "legs": 4},
                 {"name": "elephant", "legs": 4},
             ])
+            # Basic usage
             results = df.ai.task(analysis_task)
+
+            # With progress bar for large datasets
+            large_df = pd.DataFrame({"id": list(range(1000))})
+            results = large_df.ai.task(
+                analysis_task,
+                batch_size=50,
+                show_progress=True
+            )
             ```
-            This method returns a Series containing the task results for each
-            corresponding row, following the task's defined structure.
 
         Args:
             task (PreparedTask): A pre-configured task containing instructions,
@@ -637,19 +663,63 @@ class OpenAIVecDataFrameAccessor:
                 optimization based on execution time). Set to a positive integer for fixed batch size.
             show_progress (bool, optional): Show progress bar in Jupyter notebooks. Defaults to ``False``.
 
+        Additional Keyword Args:
+            Arbitrary OpenAI Responses API parameters (e.g. ``frequency_penalty``, ``presence_penalty``,
+            ``seed``, etc.) are forwarded verbatim to the underlying client. Core batching / routing
+            keys (``model``, ``instructions`` / system message, user ``input``) are managed by the
+            library and cannot be overridden.
+
         Returns:
             pandas.Series: Series whose values are instances of the task's
                 response format, aligned with the DataFrame's original index.
         """
-        return self._obj.pipe(
-            lambda df: (
-                df.pipe(lambda df: pd.Series(df.to_dict(orient="records"), index=df.index, name="record"))
-                .map(lambda x: json.dumps(x, ensure_ascii=False))
-                .ai.task(task=task, batch_size=batch_size, show_progress=show_progress)
-            )
+        return _df_rows_to_json_series(self._obj).ai.task(
+            task=task,
+            batch_size=batch_size,
+            show_progress=show_progress,
+            **api_kwargs,
+        )
+
+    def extract(self, column: str) -> pd.DataFrame:
+        """Flatten one column of Pydantic models/dicts into top‑level columns.
+
+        Example:
+            ```python
+            df = pd.DataFrame([
+                {"animal": {"name": "cat", "legs": 4}},
+                {"animal": {"name": "dog", "legs": 4}},
+                {"animal": {"name": "elephant", "legs": 4}},
+            ])
+            df.ai.extract("animal")
+            ```
+            This method returns a DataFrame with the same index as the original,
+            where each column corresponds to a key in the dictionaries.
+            The source column is dropped.
+
+        Args:
+            column (str): Column to expand.
+
+        Returns:
+            pandas.DataFrame: Original DataFrame with the extracted columns; the source column is dropped.
+        """
+        if column not in self._obj.columns:
+            raise ValueError(f"Column '{column}' does not exist in the DataFrame.")
+
+        return (
+            self._obj.pipe(lambda df: df.reset_index(drop=True))
+            .pipe(lambda df: df.join(df[column].ai.extract()))
+            .pipe(lambda df: df.set_index(self._obj.index))
+            .pipe(lambda df: df.drop(columns=[column], axis=1))
         )
 
-    def fillna(self, target_column_name: str, max_examples: int = 500, batch_size: int | None = None) -> pd.DataFrame:
+    def fillna(
+        self,
+        target_column_name: str,
+        max_examples: int = 500,
+        batch_size: int | None = None,
+        show_progress: bool = False,
+        **api_kwargs,
+    ) -> pd.DataFrame:
         """Fill missing values in a DataFrame column using AI-powered inference.
 
         This method uses machine learning to intelligently fill missing (NaN) values
@@ -666,6 +736,11 @@ class OpenAIVecDataFrameAccessor:
             batch_size (int | None, optional): Number of requests sent in one batch
                 to optimize API usage. Defaults to ``None`` (automatic batch size
                 optimization based on execution time). Set to a positive integer for fixed batch size.
+            show_progress (bool, optional): Show progress bar in Jupyter notebooks. Defaults to ``False``.
+
+        Additional Keyword Args:
+            Arbitrary OpenAI Responses API parameters (e.g. ``frequency_penalty``, ``presence_penalty``,
+            ``seed``, etc.) are forwarded verbatim to the underlying task execution.
 
         Returns:
             pandas.DataFrame: A new DataFrame with missing values filled in the target
@@ -681,6 +756,10 @@ class OpenAIVecDataFrameAccessor:
 
             # Fill missing values in the 'name' column
             filled_df = df.ai.fillna('name')
+
+            # With progress bar for large datasets
+            large_df = pd.DataFrame({'name': [None] * 1000, 'age': list(range(1000))})
+            filled_df = large_df.ai.fillna('name', batch_size=32, show_progress=True)
             ```
 
         Note:
@@ -693,7 +772,9 @@ class OpenAIVecDataFrameAccessor:
         if missing_rows.empty:
             return self._obj
 
-        filled_values: List[FillNaResponse] = missing_rows.ai.task(task=task, batch_size=batch_size)
+        filled_values: List[FillNaResponse] = missing_rows.ai.task(
+            task=task, batch_size=batch_size, show_progress=show_progress, **api_kwargs
+        )
 
         # get deep copy of the DataFrame to avoid modifying the original
         df = self._obj.copy()
@@ -716,15 +797,6 @@ class OpenAIVecDataFrameAccessor:
         two columns of the DataFrame. The vectors should be numpy arrays or
         array-like objects that support dot product operations.
 
-        Args:
-            col1 (str): Name of the first column containing embedding vectors.
-            col2 (str): Name of the second column containing embedding vectors.
-
-        Returns:
-            pandas.Series: Series containing cosine similarity scores between
-                corresponding vectors in col1 and col2, with values ranging
-                from -1 to 1, where 1 indicates identical direction.
-
         Example:
             ```python
             df = pd.DataFrame({
@@ -733,188 +805,86 @@ class OpenAIVecDataFrameAccessor:
             })
             similarities = df.ai.similarity('vec1', 'vec2')
             ```
-        """
-        return self._obj.apply(
-            lambda row: np.dot(row[col1], row[col2]) / (np.linalg.norm(row[col1]) * np.linalg.norm(row[col2])),
-            axis=1,
-        ).rename("similarity")  # type: ignore[arg-type]
-
-
-@pd.api.extensions.register_series_accessor("aio")
-class AsyncOpenAIVecSeriesAccessor:
-    """pandas Series accessor (``.aio``) that adds OpenAI helpers."""
-
-    def __init__(self, series_obj: pd.Series):
-        self._obj = series_obj
-
-    async def responses_with_cache(
-        self,
-        instructions: str,
-        cache: AsyncBatchingMapProxy[str, ResponseFormat],
-        response_format: Type[ResponseFormat] = str,
-        temperature: float | None = 0.0,
-        top_p: float = 1.0,
-    ) -> pd.Series:
-        """Call an LLM once for every Series element using a provided cache (asynchronously).
-
-        This method allows external control over caching behavior by accepting
-        a pre-configured AsyncBatchingMapProxy instance, enabling cache sharing
-        across multiple operations or custom batch size management. The concurrency
-        is controlled by the cache instance itself.
 
         Args:
-            instructions (str): System prompt prepended to every user message.
-            cache (AsyncBatchingMapProxy[str, ResponseFormat]): Pre-configured cache
-                instance for managing API call batching and deduplication.
-                Set cache.batch_size=None to enable automatic batch size optimization.
-            response_format (Type[ResponseFormat], optional): Pydantic model or built‑in
-                type the assistant should return. Defaults to ``str``.
-            temperature (float, optional): Sampling temperature. Defaults to ``0.0``.
-            top_p (float, optional): Nucleus sampling parameter. Defaults to ``1.0``.
+            col1 (str): Name of the first column containing embedding vectors.
+            col2 (str): Name of the second column containing embedding vectors.
 
         Returns:
-            pandas.Series: Series whose values are instances of ``response_format``.
-
-        Example:
-            ```python
-            from openaivec._proxy import AsyncBatchingMapProxy
-
-            # Create a shared cache with custom batch size and concurrency
-            shared_cache = AsyncBatchingMapProxy(batch_size=64, max_concurrency=4)
-
-            animals = pd.Series(["cat", "dog", "elephant"])
-            # Must be awaited
-            result = await animals.aio.responses_with_cache(
-                "translate to French",
-                cache=shared_cache
-            )
-            ```
-
-        Note:
-            This is an asynchronous method and must be awaited.
+            pandas.Series: Series containing cosine similarity scores between
+                corresponding vectors in col1 and col2, with values ranging
+                from -1 to 1, where 1 indicates identical direction.
         """
-        client: AsyncBatchResponses = AsyncBatchResponses(
-            client=CONTAINER.resolve(AsyncOpenAI),
-            model_name=CONTAINER.resolve(ResponsesModelName).value,
-            system_message=instructions,
-            response_format=response_format,
-            cache=cache,
-            temperature=temperature,
-            top_p=top_p,
-        )
-        # Await the async operation
-        results = await client.parse(self._obj.tolist())
-
-        return pd.Series(results, index=self._obj.index, name=self._obj.name)
-
-    async def embeddings_with_cache(
-        self,
-        cache: AsyncBatchingMapProxy[str, np.ndarray],
-    ) -> pd.Series:
-        """Compute OpenAI embeddings for every Series element using a provided cache (asynchronously).
-
-        This method allows external control over caching behavior by accepting
-        a pre-configured AsyncBatchingMapProxy instance, enabling cache sharing
-        across multiple operations or custom batch size management. The concurrency
-        is controlled by the cache instance itself.
-
-        Args:
-            cache (AsyncBatchingMapProxy[str, np.ndarray]): Pre-configured cache
-                instance for managing API call batching and deduplication.
-                Set cache.batch_size=None to enable automatic batch size optimization.
-
-        Returns:
-            pandas.Series: Series whose values are ``np.ndarray`` objects
-                (dtype ``float32``).
-
-        Example:
-            ```python
-            from openaivec._proxy import AsyncBatchingMapProxy
-            import numpy as np
-
-            # Create a shared cache with custom batch size and concurrency
-            shared_cache = AsyncBatchingMapProxy[str, np.ndarray](
-                batch_size=64, max_concurrency=4
-            )
-
-            animals = pd.Series(["cat", "dog", "elephant"])
-            # Must be awaited
-            embeddings = await animals.aio.embeddings_with_cache(cache=shared_cache)
-            ```
+        return self._obj.apply(
+            lambda row: np.dot(row[col1], row[col2]) / (np.linalg.norm(row[col1]) * np.linalg.norm(row[col2])),
+            axis=1,
+        ).rename("similarity")  # type: ignore[arg-type]
 
-        Note:
-            This is an asynchronous method and must be awaited.
-        """
-        client: AsyncBatchEmbeddings = AsyncBatchEmbeddings(
-            client=CONTAINER.resolve(AsyncOpenAI),
-            model_name=CONTAINER.resolve(EmbeddingsModelName).value,
-            cache=cache,
-        )
 
-        # Await the async operation
-        results = await client.create(self._obj.tolist())
+@pd.api.extensions.register_series_accessor("aio")
+class AsyncOpenAIVecSeriesAccessor:
+    """pandas Series accessor (``.aio``) that adds OpenAI helpers."""
 
-        return pd.Series(
-            results,
-            index=self._obj.index,
-            name=self._obj.name,
-        )
+    def __init__(self, series_obj: pd.Series):
+        self._obj = series_obj
 
-    async def task_with_cache(
+    async def responses_with_cache(
         self,
-        task: PreparedTask[ResponseFormat],
+        instructions: str,
         cache: AsyncBatchingMapProxy[str, ResponseFormat],
+        response_format: Type[ResponseFormat] = str,
+        temperature: float | None = 0.0,
+        top_p: float = 1.0,
+        **api_kwargs,
     ) -> pd.Series:
-        """Execute a prepared task on every Series element using a provided cache (asynchronously).
+        """Call an LLM once for every Series element using a provided cache (asynchronously).
 
         This method allows external control over caching behavior by accepting
         a pre-configured AsyncBatchingMapProxy instance, enabling cache sharing
         across multiple operations or custom batch size management. The concurrency
         is controlled by the cache instance itself.
 
+        Example:
+            ```python
+            result = await series.aio.responses_with_cache(
+                "classify",
+                cache=shared,
+                max_output_tokens=256,
+                frequency_penalty=0.2,
+            )
+            ```
+
         Args:
-            task (PreparedTask): A pre-configured task containing instructions,
-                response format, and other parameters for processing the inputs.
+            instructions (str): System prompt prepended to every user message.
             cache (AsyncBatchingMapProxy[str, ResponseFormat]): Pre-configured cache
                 instance for managing API call batching and deduplication.
                 Set cache.batch_size=None to enable automatic batch size optimization.
+            response_format (Type[ResponseFormat], optional): Pydantic model or built‑in
+                type the assistant should return. Defaults to ``str``.
+            temperature (float | None, optional): Sampling temperature. ``None`` omits the
+                parameter (recommended for reasoning models). Defaults to ``0.0``.
+            top_p (float, optional): Nucleus sampling parameter. Defaults to ``1.0``.
+            **api_kwargs: Additional keyword arguments forwarded verbatim to
+                ``AsyncOpenAI.responses.parse`` (e.g. ``max_output_tokens``, penalties,
+                future parameters). Core batching keys (model, instructions, input,
+                text_format) are protected and silently ignored if provided.
 
         Returns:
-            pandas.Series: Series whose values are instances of the task's
-                response format, aligned with the original Series index.
-
-        Example:
-            ```python
-            from openaivec._model import PreparedTask
-            from openaivec._proxy import AsyncBatchingMapProxy
-
-            # Create a shared cache with custom batch size and concurrency
-            shared_cache = AsyncBatchingMapProxy(batch_size=64, max_concurrency=4)
-
-            # Assume you have a prepared task for sentiment analysis
-            sentiment_task = PreparedTask(...)
-
-            reviews = pd.Series(["Great product!", "Not satisfied", "Amazing quality"])
-            # Must be awaited
-            results = await reviews.aio.task_with_cache(sentiment_task, cache=shared_cache)
-            ```
+            pandas.Series: Series whose values are instances of ``response_format``.
 
         Note:
             This is an asynchronous method and must be awaited.
         """
-        client = AsyncBatchResponses(
+        client: AsyncBatchResponses = AsyncBatchResponses(
             client=CONTAINER.resolve(AsyncOpenAI),
             model_name=CONTAINER.resolve(ResponsesModelName).value,
-            system_message=task.instructions,
-            response_format=task.response_format,
+            system_message=instructions,
+            response_format=response_format,
             cache=cache,
-            temperature=task.temperature,
-            top_p=task.top_p,
+            temperature=temperature,
+            top_p=top_p,
         )
-
-        # Await the async operation
-        results = await client.parse(self._obj.tolist())
-
+        results = await client.parse(self._obj.tolist(), **api_kwargs)
         return pd.Series(results, index=self._obj.index, name=self._obj.name)
 
     async def responses(
@@ -926,6 +896,7 @@ class AsyncOpenAIVecSeriesAccessor:
         top_p: float = 1.0,
         max_concurrency: int = 8,
         show_progress: bool = False,
+        **api_kwargs,
     ) -> pd.Series:
         """Call an LLM once for every Series element (asynchronously).
 
@@ -944,10 +915,6 @@ class AsyncOpenAIVecSeriesAccessor:
                 show_progress=True
             )
             ```
-            This method returns a Series of strings, each containing the
-            assistant's response to the corresponding input.
-            The model used is set by the `responses_model` function.
-            The default model is `gpt-4.1-mini`.
 
         Args:
             instructions (str): System prompt prepended to every user message.
@@ -956,7 +923,7 @@ class AsyncOpenAIVecSeriesAccessor:
             batch_size (int | None, optional): Number of prompts grouped into a single
                 request. Defaults to ``None`` (automatic batch size optimization
                 based on execution time). Set to a positive integer for fixed batch size.
-            temperature (float, optional): Sampling temperature. Defaults to ``0.0``.
+            temperature (float | None, optional): Sampling temperature. Defaults to ``0.0``.
             top_p (float, optional): Nucleus sampling parameter. Defaults to ``1.0``.
             max_concurrency (int, optional): Maximum number of concurrent
                 requests. Defaults to ``8``.
@@ -976,6 +943,60 @@ class AsyncOpenAIVecSeriesAccessor:
             response_format=response_format,
             temperature=temperature,
             top_p=top_p,
+            **api_kwargs,
+        )
+
+    async def embeddings_with_cache(
+        self,
+        cache: AsyncBatchingMapProxy[str, np.ndarray],
+    ) -> pd.Series:
+        """Compute OpenAI embeddings for every Series element using a provided cache (asynchronously).
+
+        This method allows external control over caching behavior by accepting
+        a pre-configured AsyncBatchingMapProxy instance, enabling cache sharing
+        across multiple operations or custom batch size management. The concurrency
+        is controlled by the cache instance itself.
+
+        Example:
+            ```python
+            from openaivec._proxy import AsyncBatchingMapProxy
+            import numpy as np
+
+            # Create a shared cache with custom batch size and concurrency
+            shared_cache = AsyncBatchingMapProxy[str, np.ndarray](
+                batch_size=64, max_concurrency=4
+            )
+
+            animals = pd.Series(["cat", "dog", "elephant"])
+            # Must be awaited
+            embeddings = await animals.aio.embeddings_with_cache(cache=shared_cache)
+            ```
+
+        Args:
+            cache (AsyncBatchingMapProxy[str, np.ndarray]): Pre-configured cache
+                instance for managing API call batching and deduplication.
+                Set cache.batch_size=None to enable automatic batch size optimization.
+
+        Returns:
+            pandas.Series: Series whose values are ``np.ndarray`` objects
+                (dtype ``float32``).
+
+        Note:
+            This is an asynchronous method and must be awaited.
+        """
+        client: AsyncBatchEmbeddings = AsyncBatchEmbeddings(
+            client=CONTAINER.resolve(AsyncOpenAI),
+            model_name=CONTAINER.resolve(EmbeddingsModelName).value,
+            cache=cache,
+        )
+
+        # Await the async operation
+        results = await client.create(self._obj.tolist())
+
+        return pd.Series(
+            results,
+            index=self._obj.index,
+            name=self._obj.name,
         )
 
     async def embeddings(
@@ -997,10 +1018,6 @@ class AsyncOpenAIVecSeriesAccessor:
                 show_progress=True
             )
             ```
-            This method returns a Series of numpy arrays, each containing the
-            embedding vector for the corresponding input.
-            The embedding model is set by the `embeddings_model` function.
-            The default embedding model is `text-embedding-3-small`.
 
         Args:
             batch_size (int | None, optional): Number of inputs grouped into a
@@ -1023,15 +1040,79 @@ class AsyncOpenAIVecSeriesAccessor:
             ),
         )
 
+    async def task_with_cache(
+        self,
+        task: PreparedTask[ResponseFormat],
+        cache: AsyncBatchingMapProxy[str, ResponseFormat],
+        **api_kwargs,
+    ) -> pd.Series:
+        """Execute a prepared task on every Series element using a provided cache (asynchronously).
+
+        This method allows external control over caching behavior by accepting
+        a pre-configured AsyncBatchingMapProxy instance, enabling cache sharing
+        across multiple operations or custom batch size management. The concurrency
+        is controlled by the cache instance itself.
+
+        Args:
+            task (PreparedTask): A pre-configured task containing instructions,
+                response format, and other parameters for processing the inputs.
+            cache (AsyncBatchingMapProxy[str, ResponseFormat]): Pre-configured cache
+                instance for managing API call batching and deduplication.
+                Set cache.batch_size=None to enable automatic batch size optimization.
+
+        Example:
+            ```python
+            from openaivec._model import PreparedTask
+            from openaivec._proxy import AsyncBatchingMapProxy
+
+            # Create a shared cache with custom batch size and concurrency
+            shared_cache = AsyncBatchingMapProxy(batch_size=64, max_concurrency=4)
+
+            # Assume you have a prepared task for sentiment analysis
+            sentiment_task = PreparedTask(...)
+
+            reviews = pd.Series(["Great product!", "Not satisfied", "Amazing quality"])
+            # Must be awaited
+            results = await reviews.aio.task_with_cache(sentiment_task, cache=shared_cache)
+            ```
+
+        Additional Keyword Args:
+            Arbitrary OpenAI Responses API parameters (e.g. ``frequency_penalty``, ``presence_penalty``,
+            ``seed``, etc.) are forwarded verbatim to the underlying client. Core batching / routing
+            keys (``model``, ``instructions`` / system message, user ``input``) are managed by the
+            library and cannot be overridden.
+
+        Returns:
+            pandas.Series: Series whose values are instances of the task's
+                response format, aligned with the original Series index.
+
+        Note:
+            This is an asynchronous method and must be awaited.
+        """
+        client = AsyncBatchResponses(
+            client=CONTAINER.resolve(AsyncOpenAI),
+            model_name=CONTAINER.resolve(ResponsesModelName).value,
+            system_message=task.instructions,
+            response_format=task.response_format,
+            cache=cache,
+            temperature=task.temperature,
+            top_p=task.top_p,
+        )
+        # Await the async operation
+        results = await client.parse(self._obj.tolist(), **api_kwargs)
+
+        return pd.Series(results, index=self._obj.index, name=self._obj.name)
+
     async def task(
-        self, task: PreparedTask, batch_size: int | None = None, max_concurrency: int = 8, show_progress: bool = False
+        self,
+        task: PreparedTask,
+        batch_size: int | None = None,
+        max_concurrency: int = 8,
+        show_progress: bool = False,
+        **api_kwargs,
     ) -> pd.Series:
         """Execute a prepared task on every Series element (asynchronously).
 
-        This method applies a pre-configured task to each element in the Series,
-        using the task's instructions and response format to generate structured
-        responses from the language model.
-
         Example:
             ```python
             from openaivec._model import PreparedTask
@@ -1052,8 +1133,6 @@ class AsyncOpenAIVecSeriesAccessor:
                 show_progress=True
             )
             ```
-            This method returns a Series containing the task results for each
-            corresponding input element, following the task's defined structure.
 
         Args:
             task (PreparedTask): A pre-configured task containing instructions,
@@ -1065,6 +1144,12 @@ class AsyncOpenAIVecSeriesAccessor:
                 requests. Defaults to 8.
             show_progress (bool, optional): Show progress bar in Jupyter notebooks. Defaults to ``False``.
 
+        Additional Keyword Args:
+            Arbitrary OpenAI Responses API parameters (e.g. ``frequency_penalty``, ``presence_penalty``,
+            ``seed``, etc.) are forwarded verbatim to the underlying client. Core batching / routing
+            keys (``model``, ``instructions`` / system message, user ``input``) are managed by the
+            library and cannot be overridden.
+
         Returns:
             pandas.Series: Series whose values are instances of the task's
                 response format, aligned with the original Series index.
@@ -1077,6 +1162,7 @@ class AsyncOpenAIVecSeriesAccessor:
             cache=AsyncBatchingMapProxy(
                 batch_size=batch_size, max_concurrency=max_concurrency, show_progress=show_progress
             ),
+            **api_kwargs,
         )
 
 
@@ -1094,27 +1180,15 @@ class AsyncOpenAIVecDataFrameAccessor:
         response_format: Type[ResponseFormat] = str,
         temperature: float | None = 0.0,
         top_p: float = 1.0,
+        **api_kwargs,
     ) -> pd.Series:
-        """Generate a response for each row after serialising it to JSON using a provided cache (asynchronously).
+        """Generate a response for each row after serializing it to JSON using a provided cache (asynchronously).
 
         This method allows external control over caching behavior by accepting
         a pre-configured AsyncBatchingMapProxy instance, enabling cache sharing
         across multiple operations or custom batch size management. The concurrency
         is controlled by the cache instance itself.
 
-        Args:
-            instructions (str): System prompt for the assistant.
-            cache (AsyncBatchingMapProxy[str, ResponseFormat]): Pre-configured cache
-                instance for managing API call batching and deduplication.
-                Set cache.batch_size=None to enable automatic batch size optimization.
-            response_format (Type[ResponseFormat], optional): Desired Python type of the
-                responses. Defaults to ``str``.
-            temperature (float, optional): Sampling temperature. Defaults to ``0.0``.
-            top_p (float, optional): Nucleus sampling parameter. Defaults to ``1.0``.
-
-        Returns:
-            pandas.Series: Responses aligned with the DataFrame's original index.
-
         Example:
             ```python
             from openaivec._proxy import AsyncBatchingMapProxy
@@ -1134,23 +1208,30 @@ class AsyncOpenAIVecDataFrameAccessor:
             )
             ```
 
+        Args:
+            instructions (str): System prompt for the assistant.
+            cache (AsyncBatchingMapProxy[str, ResponseFormat]): Pre-configured cache
+                instance for managing API call batching and deduplication.
+                Set cache.batch_size=None to enable automatic batch size optimization.
+            response_format (Type[ResponseFormat], optional): Desired Python type of the
+                responses. Defaults to ``str``.
+            temperature (float | None, optional): Sampling temperature. Defaults to ``0.0``.
+            top_p (float, optional): Nucleus sampling parameter. Defaults to ``1.0``.
+
+        Returns:
+            pandas.Series: Responses aligned with the DataFrame's original index.
+
         Note:
             This is an asynchronous method and must be awaited.
         """
-        series_of_json = self._obj.pipe(
-            lambda df: (
-                pd.Series(df.to_dict(orient="records"), index=df.index, name="record").map(
-                    lambda x: json.dumps(x, ensure_ascii=False)
-                )
-            )
-        )
         # Await the call to the async Series method using .aio
-        return await series_of_json.aio.responses_with_cache(
+        return await _df_rows_to_json_series(self._obj).aio.responses_with_cache(
             instructions=instructions,
             cache=cache,
             response_format=response_format,
             temperature=temperature,
             top_p=top_p,
+            **api_kwargs,
         )
 
     async def responses(
@@ -1162,33 +1243,29 @@ class AsyncOpenAIVecDataFrameAccessor:
         top_p: float = 1.0,
         max_concurrency: int = 8,
         show_progress: bool = False,
+        **api_kwargs,
     ) -> pd.Series:
-        """Generate a response for each row after serialising it to JSON (asynchronously).
+        """Generate a response for each row after serializing it to JSON (asynchronously).
 
         Example:
             ```python
             df = pd.DataFrame([
-                {\"name\": \"cat\", \"legs\": 4},
-                {\"name\": \"dog\", \"legs\": 4},
-                {\"name\": \"elephant\", \"legs\": 4},
+                {"name": "cat", "legs": 4},
+                {"name": "dog", "legs": 4},
+                {"name": "elephant", "legs": 4},
             ])
             # Must be awaited
-            results = await df.aio.responses(\"what is the animal\'s name?\")
+            results = await df.aio.responses("what is the animal's name?")
 
             # With progress bar for large datasets
-            large_df = pd.DataFrame({\"id\": list(range(1000))})
+            large_df = pd.DataFrame({"id": list(range(1000))})
             results = await large_df.aio.responses(
-                \"generate a name for this ID\",
+                "generate a name for this ID",
                 batch_size=20,
                 max_concurrency=4,
                 show_progress=True
             )
             ```
-            This method returns a Series of strings, each containing the
-            assistant's response to the corresponding input.
-            Each row is serialised to JSON before being sent to the assistant.
-            The model used is set by the `responses_model` function.
-            The default model is `gpt-4.1-mini`.
 
         Args:
             instructions (str): System prompt for the assistant.
@@ -1197,7 +1274,7 @@ class AsyncOpenAIVecDataFrameAccessor:
             batch_size (int | None, optional): Number of requests sent in one batch.
                 Defaults to ``None`` (automatic batch size optimization
                 based on execution time). Set to a positive integer for fixed batch size.
-            temperature (float, optional): Sampling temperature. Defaults to ``0.0``.
+            temperature (float | None, optional): Sampling temperature. Defaults to ``0.0``.
             top_p (float, optional): Nucleus sampling parameter. Defaults to ``1.0``.
             max_concurrency (int, optional): Maximum number of concurrent
                 requests. Defaults to ``8``.
@@ -1217,17 +1294,47 @@ class AsyncOpenAIVecDataFrameAccessor:
             response_format=response_format,
             temperature=temperature,
             top_p=top_p,
+            **api_kwargs,
         )
 
-    async def task(
-        self, task: PreparedTask, batch_size: int | None = None, max_concurrency: int = 8, show_progress: bool = False
+    async def task_with_cache(
+        self,
+        task: PreparedTask[ResponseFormat],
+        cache: AsyncBatchingMapProxy[str, ResponseFormat],
+        **api_kwargs,
     ) -> pd.Series:
-        """Execute a prepared task on each DataFrame row after serialising it to JSON (asynchronously).
+        """Execute a prepared task on each DataFrame row using a provided cache (asynchronously).
+
+        After serializing each row to JSON, this method executes the prepared task.
+
+        Args:
+            task (PreparedTask): Prepared task (instructions + response_format + sampling params).
+            cache (AsyncBatchingMapProxy[str, ResponseFormat]): Pre‑configured async cache instance.
+
+        Additional Keyword Args:
+            Arbitrary OpenAI Responses API parameters forwarded verbatim. Core routing keys are protected.
+
+        Returns:
+            pandas.Series: Task results aligned with the DataFrame's original index.
+
+        Note:
+            This is an asynchronous method and must be awaited.
+        """
+        return await _df_rows_to_json_series(self._obj).aio.task_with_cache(
+            task=task,
+            cache=cache,
+            **api_kwargs,
+        )
 
-        This method applies a pre-configured task to each row in the DataFrame,
-        using the task's instructions and response format to generate structured
-        responses from the language model. Each row is serialised to JSON before
-        being processed by the task.
+    async def task(
+        self,
+        task: PreparedTask,
+        batch_size: int | None = None,
+        max_concurrency: int = 8,
+        show_progress: bool = False,
+        **api_kwargs,
+    ) -> pd.Series:
+        """Execute a prepared task on each DataFrame row after serializing it to JSON (asynchronously).
 
         Example:
             ```python
@@ -1253,8 +1360,6 @@ class AsyncOpenAIVecDataFrameAccessor:
                 show_progress=True
             )
             ```
-            This method returns a Series containing the task results for each
-            corresponding row, following the task's defined structure.
 
         Args:
             task (PreparedTask): A pre-configured task containing instructions,
@@ -1266,6 +1371,12 @@ class AsyncOpenAIVecDataFrameAccessor:
                 requests. Defaults to 8.
             show_progress (bool, optional): Show progress bar in Jupyter notebooks. Defaults to ``False``.
 
+        Additional Keyword Args:
+            Arbitrary OpenAI Responses API parameters (e.g. ``frequency_penalty``, ``presence_penalty``,
+            ``seed``, etc.) are forwarded verbatim to the underlying client. Core batching / routing
+            keys (``model``, ``instructions`` / system message, user ``input``) are managed by the
+            library and cannot be overridden.
+
         Returns:
             pandas.Series: Series whose values are instances of the task's
                 response format, aligned with the DataFrame's original index.
@@ -1273,28 +1384,33 @@ class AsyncOpenAIVecDataFrameAccessor:
         Note:
             This is an asynchronous method and must be awaited.
         """
-        series_of_json = self._obj.pipe(
-            lambda df: (
-                pd.Series(df.to_dict(orient="records"), index=df.index, name="record").map(
-                    lambda x: json.dumps(x, ensure_ascii=False)
-                )
-            )
-        )
         # Await the call to the async Series method using .aio
-        return await series_of_json.aio.task(
+        return await _df_rows_to_json_series(self._obj).aio.task(
             task=task,
             batch_size=batch_size,
             max_concurrency=max_concurrency,
             show_progress=show_progress,
+            **api_kwargs,
         )
 
     async def pipe(self, func: Callable[[pd.DataFrame], Awaitable[T] | T]) -> T:
-        """
-        Apply a function to the DataFrame, supporting both synchronous and asynchronous functions.
+        """Apply a function to the DataFrame, supporting both synchronous and asynchronous functions.
 
         This method allows chaining operations on the DataFrame, similar to pandas' `pipe` method,
         but with support for asynchronous functions.
 
+        Example:
+            ```python
+            async def process_data(df):
+                # Simulate an asynchronous computation
+                await asyncio.sleep(1)
+                return df.dropna()
+
+            df = pd.DataFrame({"col": [1, 2, None, 4]})
+            # Must be awaited
+            result = await df.aio.pipe(process_data)
+            ```
+
         Args:
             func (Callable[[pd.DataFrame], Awaitable[T] | T]): A function that takes a DataFrame
                 as input and returns either a result or an awaitable result.
@@ -1371,7 +1487,13 @@ class AsyncOpenAIVecDataFrameAccessor:
         return df_current
 
     async def fillna(
-        self, target_column_name: str, max_examples: int = 500, batch_size: int | None = None, max_concurrency: int = 8
+        self,
+        target_column_name: str,
+        max_examples: int = 500,
+        batch_size: int | None = None,
+        max_concurrency: int = 8,
+        show_progress: bool = False,
+        **api_kwargs,
     ) -> pd.DataFrame:
         """Fill missing values in a DataFrame column using AI-powered inference (asynchronously).
 
@@ -1391,6 +1513,11 @@ class AsyncOpenAIVecDataFrameAccessor:
                 optimization based on execution time). Set to a positive integer for fixed batch size.
             max_concurrency (int, optional): Maximum number of concurrent
                 requests. Defaults to 8.
+            show_progress (bool, optional): Show progress bar in Jupyter notebooks. Defaults to ``False``.
+
+        Additional Keyword Args:
+            Arbitrary OpenAI Responses API parameters (e.g. ``frequency_penalty``, ``presence_penalty``,
+            ``seed``, etc.) are forwarded verbatim to the underlying task execution.
 
         Returns:
             pandas.DataFrame: A new DataFrame with missing values filled in the target
@@ -1406,6 +1533,15 @@ class AsyncOpenAIVecDataFrameAccessor:
 
             # Fill missing values in the 'name' column (must be awaited)
             filled_df = await df.aio.fillna('name')
+
+            # With progress bar for large datasets
+            large_df = pd.DataFrame({'name': [None] * 1000, 'age': list(range(1000))})
+            filled_df = await large_df.aio.fillna(
+                'name',
+                batch_size=32,
+                max_concurrency=4,
+                show_progress=True
+            )
             ```
 
         Note:
@@ -1420,7 +1556,7 @@ class AsyncOpenAIVecDataFrameAccessor:
             return self._obj
 
         filled_values: List[FillNaResponse] = await missing_rows.aio.task(
-            task=task, batch_size=batch_size, max_concurrency=max_concurrency
+            task=task, batch_size=batch_size, max_concurrency=max_concurrency, show_progress=show_progress, **api_kwargs
         )
 
         # get deep copy of the DataFrame to avoid modifying the original