gmakstutis/graphiti
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Use OpenAI structured output API for response validation

Browse source 
Replace prompt-based schema injection with native json_schema response format. This improves token efficiency and reliability by having OpenAI enforce the schema directly instead of embedding it in the prompt message.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
Daniel Chalef 2025-11-09 20:51:46 -08:00
parent d4a92772ec
commit f8a17ce435
1 changed files with 14 additions and 9 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

23
graphiti_core/llm_client/openai_generic_client.py
View file

@ -99,12 +99,25 @@ class OpenAIGenericClient(LLMClient):
elif m.role == 'system': elif m.role == 'system':
openai_messages.append({'role': 'system', 'content': m.content}) openai_messages.append({'role': 'system', 'content': m.content})
try: try:
# Prepare response format
response_format = {'type': 'json_object'}
if response_model is not None:
schema_name = getattr(response_model, '__name__', 'structured_response')
json_schema = response_model.model_json_schema()
response_format = {
'type': 'json_schema',
'json_schema': {
'name': schema_name,
'schema': json_schema,
},
}
response = await self.client.chat.completions.create( response = await self.client.chat.completions.create(
model=self.model or DEFAULT_MODEL, model=self.model or DEFAULT_MODEL,
messages=openai_messages, messages=openai_messages,
temperature=self.temperature, temperature=self.temperature,
max_tokens=self.max_tokens, max_tokens=self.max_tokens,
response_format={'type': 'json_object'}, response_format=response_format,
) )
result = response.choices[0].message.content or '' result = response.choices[0].message.content or ''
return json.loads(result) return json.loads(result)
@ -126,14 +139,6 @@ class OpenAIGenericClient(LLMClient):
if max_tokens is None: if max_tokens is None:
max_tokens = self.max_tokens max_tokens = self.max_tokens
if response_model is not None:
serialized_model = json.dumps(response_model.model_json_schema())
messages[
-1
].content += (
f'\n\nRespond with a JSON object in the following format:\n\n{serialized_model}'
)
# Add multilingual extraction instructions # Add multilingual extraction instructions
messages[0].content += get_extraction_language_instruction(group_id) messages[0].content += get_extraction_language_instruction(group_id)