big overall, got rid of Reasoned

Author: Zulko

README.md

@@ -24,10 +24,10 @@ format_date("12/31/2008.") # returns '2008-12-31'
 Here is another example with a more structured output:
 
 ```python
-from gpt_function_decorator import gpt_function, Reasoned
+from gpt_function_decorator import gpt_function
 
 @gpt_function
-def deduplicate_celebrities(names) -> Reasoned[list[str]]:
+def deduplicate_celebrities(names) -> list[str]:
     """Return the deduplicated version of the celebrities list."""
 
 celebrities = [
@@ -37,7 +37,7 @@ celebrities = [
     "Lionel Messi",
     "Leo diCaprio"
 ]
-answer = deduplicate_celebrities(celebrities)
+answer = deduplicate_celebrities(celebrities, gpt_reasoning=true)
 
 print (answer)
 # ['Leo Messi', 'Mozart', 'Leo diCaprio']
@@ -311,23 +311,25 @@ could_have_met("Chopin", celebrities=[
 
 It turns out that this function struggles to find the right answer. Its output (generated by `gpt-4o-mini`) varies a lot and typically includes Peggy Lee, who lived in a different century. This is because this short prompt actually requires real reasoning from the GPT: first listing everyone's birth and death years, then checking who overlapped with Chopin. 
 
-To get a smarter answer, we provide a `Reasoned` constructor for the output schema. Concretely, it requests the GPT answer to have two fields, `reasoning` and `result`. The GPT answers is more verbose, and therefore slower and more costly, but also more helpful:
+To get a smarter answer, we provide a `gpt_reasoning` parameter to gpt-decorated functions. Concretely, it requests the GPT answer to have two fields, `reasoning` and `result` (the final result will have the reasoning under the `.__reasoning__` attribute). The GPT answers is more verbose, and therefore slower and more costly, but also more helpful:
 - The `reasoning` field gives the GPT room to "think through" the problem and produce better answers.
 - It is now possible for the user to see what the GPT's "reasoning" was, and whether a wrong answer was caused by a lack of knowledge, or logics, etc.
 - It reduces the risk that some of GPT's reasoning and formatting ends up polluting the result's schema.
 
-So let's just change our function's output type hint to `Reasoned[list[str]]` and observe the improvement:
+So let's ask for a reasoning and observe the improvement:
 
 ```python
-from gpt_function_decorator import gpt_function, Reasoned
+from gpt_function_decorator import gpt_function
 
 @gpt_function
-def could_have_met(person, celebrities) -> Reasoned[list[str]]:
+def could_have_met(person, celebrities) -> list[str]:
     """List the names in {celebrities} that {person} could have met."""
 
-answer = could_have_met("Chopin", celebrities=[
+celebrities = [
     "Napoleon", "Jefferson", "Mozart", "Julius Cesar", "Peggy Lee", "Beethoven"
-])
+]
+answer = could_have_met("Chopin", celebrities, gpt_reasoning=True)
+
 print (answer)
 # ['Napoleon', 'Jefferson', 'Beethoven']
 
@@ -352,6 +354,7 @@ print (answer.__reasoning__)
 The `gpt_function` decorator adds two parameters to the function it decorates:
 - `gpt_model`: this allows the function's user to switch between `gpt-4o-mini` (the default, fast and cheap but less capable) and `gpt-4o` (any compatible version).
 - `gpt_system_prompt`: this enables the user to tweak the answer as they would like by asking the GPT to focus on some aspects, or to roleplay.
+- `gpt_reasoning` as described in the previous section.
 - `gpt_debug`: this will cause the function to print the full prompt that it sends to the GPT (useful for troubleshooting or just getting a sense of what's going on).
 
 As an example, let's start from this function:
@@ -419,8 +422,8 @@ GPTs are not yet fast and cheap enough to be used anywhere, but when they are it
 For instance instead of having developers write zillions of messages to help users troubleshoot errors, we'll use a function like this:
 
 ```python
-@gpt_function
-def help_troubleshoot(error_traceback: str) -> Reasoned[str]:
+@gpt_function(reasoning=True)
+def help_troubleshoot(error_traceback: str) -> str:
     """Return a short analysis of the Python error: {error_traceback}"""
 ```
 

examples/async_function.py

@@ -1,3 +1,10 @@
+"""This example demonstrates how to use the `gpt_function` decorator with
+async functions. We define an async function `write_story` that generates a
+short story about a given subject. We then run that story on three different
+subjects asynchronously using `asyncio.as_completed`, so that OpenAI's API
+generates the stories in parallel.
+"""
+
 from gpt_function_decorator import gpt_function
 import asyncio
 

examples/games/cheater_guess_the_word.py

@@ -3,7 +3,7 @@
 anymore without contradicting previous questions/answers.
 """
 
-from gpt_function_decorator import gpt_function, Reasoned
+from gpt_function_decorator import gpt_function
 from pydantic import BaseModel, Field
 
 
@@ -12,10 +12,8 @@ class RandomWordSelection(BaseModel):
     could_find_a_word: bool
 
 
-@gpt_function
-def select_random_word(
-    avoided_words=None, answers=None
-) -> Reasoned[RandomWordSelection]:
+@gpt_function(reasoning=True)
+def select_random_word(avoided_words=None, answers=None) -> RandomWordSelection:
     """If it is possible, pick a word not in {avoided_words} for which you  ALL
     of these answers are strictly true: {answers}. If not possible, set
     could_find_a_word=False"""
@@ -32,12 +30,11 @@ class InputEvaluation(BaseModel):
     )
     user_asked_to_quit: bool = Field(description="Does the user want to quit?")
 
-
-@gpt_function
-def evaluate_guess(text: str, secret_word: str) -> Reasoned[InputEvaluation]:
-    """In trying to guess secret word "{secret_word}", the user said "{text}".
-    Evaluate this input
-    """
+    @gpt_function(reasoning=True)
+    def evaluate_guess(text: str, secret_word: str) -> "InputEvaluation":
+        """In trying to guess secret word "{secret_word}", the user said "{text}".
+        Evaluate this input
+        """
 
 
 def play_cheater_guess_the_word(subject):

examples/games/guess_the_word.py

@@ -11,7 +11,7 @@
 you will get a more random word.
 """
 
-from gpt_function_decorator import gpt_function, Reasoned
+from gpt_function_decorator import gpt_function
 from pydantic import BaseModel, Field
 
 
@@ -22,7 +22,7 @@ class RandomWordSelection(BaseModel):
 
     @staticmethod
     @gpt_function
-    def pick() -> Reasoned["RandomWordSelection"]:
+    def pick() -> "RandomWordSelection":
         """Pick a 3 random themes, then write a very random 100-word story about
         these subjects with any strange, unrelated events happening. List the noons
         in that story, and finally select the 7th noun in the list (make it singular
@@ -45,8 +45,8 @@ class InputEvaluation(BaseModel):
     user_tried_to_cheat: bool = Field(description="Did the user try to cheat?")
 
     @staticmethod
-    @gpt_function
-    def from_user_text(text: str, secret_word: str) -> Reasoned["InputEvaluation"]:
+    @gpt_function(reasoning=True)
+    def from_user_text(text: str, secret_word: str) -> "InputEvaluation":
         """In trying to guess secret word "{secret_word}", the user said
         "{text}". Evaluate this input.
         """

examples/games/role_playing_game.py

@@ -1,3 +1,8 @@
+"""In this example, we create a role-playing game where the player makes decisions
+that affect the story. The game has a "state" (the character's health and inventory)
+that is carried over from one turn to the next, and updated after each turn.
+"""
+
 from gpt_function_decorator import gpt_function
 from pydantic import BaseModel, Field