feat(client): support parsing custom response types (#1111)

src/openai/__init__.py

@@ -9,6 +9,7 @@
 from ._types import NoneType, Transport, ProxiesTypes
 from ._utils import file_from_path
 from ._client import Client, OpenAI, Stream, Timeout, Transport, AsyncClient, AsyncOpenAI, AsyncStream, RequestOptions
+from ._models import BaseModel
 from ._version import __title__, __version__
 from ._response import APIResponse as APIResponse, AsyncAPIResponse as AsyncAPIResponse
 from ._exceptions import (
@@ -59,6 +60,7 @@
     "OpenAI",
     "AsyncOpenAI",
     "file_from_path",
+    "BaseModel",
 ]
 
 from .lib import azure as _azure

src/openai/_legacy_response.py

@@ -5,25 +5,28 @@
 import logging
 import datetime
 import functools
-from typing import TYPE_CHECKING, Any, Union, Generic, TypeVar, Callable, Iterator, AsyncIterator, cast
-from typing_extensions import Awaitable, ParamSpec, get_args, override, deprecated, get_origin
+from typing import TYPE_CHECKING, Any, Union, Generic, TypeVar, Callable, Iterator, AsyncIterator, cast, overload
+from typing_extensions import Awaitable, ParamSpec, override, deprecated, get_origin
 
 import anyio
 import httpx
+import pydantic
 
 from ._types import NoneType
 from ._utils import is_given
 from ._models import BaseModel, is_basemodel
 from ._constants import RAW_RESPONSE_HEADER
+from ._streaming import Stream, AsyncStream, is_stream_class_type, extract_stream_chunk_type
 from ._exceptions import APIResponseValidationError
 
 if TYPE_CHECKING:
     from ._models import FinalRequestOptions
-    from ._base_client import Stream, BaseClient, AsyncStream
+    from ._base_client import BaseClient
 
 
 P = ParamSpec("P")
 R = TypeVar("R")
+_T = TypeVar("_T")
 
 log: logging.Logger = logging.getLogger(__name__)
 
@@ -43,7 +46,7 @@ class LegacyAPIResponse(Generic[R]):
 
     _cast_to: type[R]
     _client: BaseClient[Any, Any]
-    _parsed: R | None
+    _parsed_by_type: dict[type[Any], Any]
     _stream: bool
     _stream_cls: type[Stream[Any]] | type[AsyncStream[Any]] | None
     _options: FinalRequestOptions
@@ -62,27 +65,60 @@ def __init__(
     ) -> None:
         self._cast_to = cast_to
         self._client = client
-        self._parsed = None
+        self._parsed_by_type = {}
         self._stream = stream
         self._stream_cls = stream_cls
         self._options = options
         self.http_response = raw
 
+    @overload
+    def parse(self, *, to: type[_T]) -> _T:
+        ...
+
+    @overload
     def parse(self) -> R:
+        ...
+
+    def parse(self, *, to: type[_T] | None = None) -> R | _T:
         """Returns the rich python representation of this response's data.
 
+        NOTE: For the async client: this will become a coroutine in the next major version.
+
         For lower-level control, see `.read()`, `.json()`, `.iter_bytes()`.
 
-        NOTE: For the async client: this will become a coroutine in the next major version.
+        You can customise the type that the response is parsed into through
+        the `to` argument, e.g.
+
+        ```py
+        from openai import BaseModel
+
+
+        class MyModel(BaseModel):
+            foo: str
+
+
+        obj = response.parse(to=MyModel)
+        print(obj.foo)
+        ```
+
+        We support parsing:
+          - `BaseModel`
+          - `dict`
+          - `list`
+          - `Union`
+          - `str`
+          - `httpx.Response`
         """
-        if self._parsed is not None:
-            return self._parsed
+        cache_key = to if to is not None else self._cast_to
+        cached = self._parsed_by_type.get(cache_key)
+        if cached is not None:
+            return cached  # type: ignore[no-any-return]
 
-        parsed = self._parse()
+        parsed = self._parse(to=to)
         if is_given(self._options.post_parser):
             parsed = self._options.post_parser(parsed)
 
-        self._parsed = parsed
+        self._parsed_by_type[cache_key] = parsed
         return parsed
 
     @property
@@ -135,13 +171,29 @@ def elapsed(self) -> datetime.timedelta:
         """The time taken for the complete request/response cycle to complete."""
         return self.http_response.elapsed
 
-    def _parse(self) -> R:
+    def _parse(self, *, to: type[_T] | None = None) -> R | _T:
         if self._stream:
+            if to:
+                if not is_stream_class_type(to):
+                    raise TypeError(f"Expected custom parse type to be a subclass of {Stream} or {AsyncStream}")
+
+                return cast(
+                    _T,
+                    to(
+                        cast_to=extract_stream_chunk_type(
+                            to,
+                            failure_message="Expected custom stream type to be passed with a type argument, e.g. Stream[ChunkType]",
+                        ),
+                        response=self.http_response,
+                        client=cast(Any, self._client),
+                    ),
+                )
+
             if self._stream_cls:
                 return cast(
                     R,
                     self._stream_cls(
-                        cast_to=_extract_stream_chunk_type(self._stream_cls),
+                        cast_to=extract_stream_chunk_type(self._stream_cls),
                         response=self.http_response,
                         client=cast(Any, self._client),
                     ),
@@ -160,7 +212,7 @@ def _parse(self) -> R:
                 ),
             )
 
-        cast_to = self._cast_to
+        cast_to = to if to is not None else self._cast_to
         if cast_to is NoneType:
             return cast(R, None)
 
@@ -186,14 +238,9 @@ def _parse(self) -> R:
                 raise ValueError(f"Subclasses of httpx.Response cannot be passed to `cast_to`")
             return cast(R, response)
 
-        # The check here is necessary as we are subverting the the type system
-        # with casts as the relationship between TypeVars and Types are very strict
-        # which means we must return *exactly* what was input or transform it in a
-        # way that retains the TypeVar state. As we cannot do that in this function
-        # then we have to resort to using `cast`. At the time of writing, we know this
-        # to be safe as we have handled all the types that could be bound to the
-        # `ResponseT` TypeVar, however if that TypeVar is ever updated in the future, then
-        # this function would become unsafe but a type checker would not report an error.
+        if inspect.isclass(origin) and not issubclass(origin, BaseModel) and issubclass(origin, pydantic.BaseModel):
+            raise TypeError("Pydantic models must subclass our base model type, e.g. `from openai import BaseModel`")
+
         if (
             cast_to is not object
             and not origin is list
@@ -202,12 +249,12 @@ def _parse(self) -> R:
             and not issubclass(origin, BaseModel)
         ):
             raise RuntimeError(
-                f"Invalid state, expected {cast_to} to be a subclass type of {BaseModel}, {dict}, {list} or {Union}."
+                f"Unsupported type, expected {cast_to} to be a subclass of {BaseModel}, {dict}, {list}, {Union}, {NoneType}, {str} or {httpx.Response}."
             )
 
         # split is required to handle cases where additional information is included
         # in the response, e.g. application/json; charset=utf-8
-        content_type, *_ = response.headers.get("content-type").split(";")
+        content_type, *_ = response.headers.get("content-type", "*").split(";")
         if content_type != "application/json":
             if is_basemodel(cast_to):
                 try:
@@ -253,15 +300,6 @@ def __init__(self) -> None:
         )
 
 
-def _extract_stream_chunk_type(stream_cls: type) -> type:
-    args = get_args(stream_cls)
-    if not args:
-        raise TypeError(
-            f"Expected stream_cls to have been given a generic type argument, e.g. Stream[Foo] but received {stream_cls}",
-        )
-    return cast(type, args[0])
-
-
 def to_raw_response_wrapper(func: Callable[P, R]) -> Callable[P, LegacyAPIResponse[R]]:
     """Higher order function that takes one of our bound API methods and wraps it
     to support returning the raw `APIResponse` object directly.