Openai Structured Outputs -Schema mit Zustand ⇐ Python

[Anonymous]

Ich versuche, OpenAI -strukturierte Ausgänge mit einem JSON -Schema zu verwenden, so dass:
• Wenn Teil1 "Ja" ist, muss Part2 vorhanden sein (erforderlich).
• Wenn Teil1 ist " Nein ", dann darf Part2 überhaupt nicht enthalten sein. OpenAI akzeptiert dieses Schema jedoch aufgrund des "Allof" -Teils nicht.

Code: Select all

import openai
import json
import jsonschema
from dotenv import load_dotenv
import os

load_dotenv()
api_key = os.getenv("API_KEY")
api_version = os.getenv("API_VERSION")
base_url = os.getenv("API_BASE_URL_4o")

# 1. Configure OpenAI credentials
openai.api_key = api_key
openai.api_version = api_version
openai.api_type = "azure"
openai.base_url = base_url

# 2. Define the function schema
answer_question_function = {
"name": "answer_question",
"description": "Answers user question with conditional parts",
"parameters": {
"type": "object",
"properties": {
"part1": {
"type": "string",
"enum": ["yes", "no"],
"description": "First part answer"
},
"part2": {
"type": "string",
"enum": ["yes", "no"],
"description": "Second part answer; required only if part1 is yes"
}
},
"required": ["part1"],
"allOf": [
{
"if": {
"properties": {
"part1": {"const": "yes"}
}
},
"then": {
"required": ["part2"]
}
},
{
"if": {
"properties": {
"part1": {"const": "no"}
}
},
"then": {
"not": {"required": ["part2"]}
}
}
]
}
}

# 3. Write a helper function to validate the JSON result against our schema
def validate_function_call_arguments(args):
schema = answer_question_function["parameters"]
jsonschema.validate(instance=args, schema=schema)

# 4.  E2E test function
def run_e2e_test(user_prompt):
"""
Sends a prompt to the model, requests a function call,
and validates the JSON arguments returned.
"""

load_dotenv()
api_key = os.getenv("API_KEY")
api_version = os.getenv("API_VERSION")
base_url = os.getenv("API_BASE_URL_4o")

openai.api_key = api_key
openai.api_version = api_version
openai.api_type = "azure"
openai.base_url = base_url

try:
response = openai.chat.completions.create(
model="gpt-4o",
messages=[
{
"role": "user",
"content": user_prompt
}
],
response_format={
"type": "json_schema",
"json_schema": {
"name": "answer_question",
"schema": answer_question_function["parameters"]
}
},
)

# Check if it's a function call:
choice = response["choices"][0]

if choice["finish_reason"] == "function_call":
function_call = choice["message"]["function_call"]
function_name = function_call["name"]
arguments_json = function_call["arguments"]

print(f"Function called: {function_name}")
print(f"Raw arguments: {arguments_json}")

# Convert the JSON string into a Python dict
arguments = json.loads(arguments_json)

# Validate
validate_function_call_arguments(arguments)
print("✅ JSON arguments valid against the schema.")
print(f"Final arguments object: {arguments}")

else:
# If no function call was made, print the plain text response
print("No function call. Text response from model:")
print(choice["message"]["content"])

except jsonschema.ValidationError as e:
print("❌ JSON arguments failed validation:")
print(str(e))
except Exception as e:
print("❌ Something went wrong:")
print(str(e))

# 5. Run a few test scenarios
if __name__ == "__main__":
print("===== TEST: Expecting part1 = yes, then part2 is required =====")
run_e2e_test("Should part1 be yes? If so, give me part2 too.")

print("\n===== TEST: Expecting part1 = no, then part2 should not appear =====")
run_e2e_test("Should part1 be no? Then we don’t want part2.")
print()