From 2337d36f7b3968cfeff06b00613f7464c8d0ca93 Mon Sep 17 00:00:00 2001 From: Andrej Milicevic Date: Thu, 13 Nov 2025 18:25:07 +0100 Subject: [PATCH] feat: add variable to control instructor mode --- cognee/infrastructure/llm/config.py | 2 ++ .../litellm_instructor/llm/anthropic/adapter.py | 8 +++++++- .../litellm_instructor/llm/gemini/adapter.py | 12 +++++++++++- .../llm/generic_llm_api/adapter.py | 12 +++++++++++- .../litellm_instructor/llm/mistral/adapter.py | 8 +++++++- .../litellm_instructor/llm/ollama/adapter.py | 12 +++++++++++- .../litellm_instructor/llm/openai/adapter.py | 11 +++++++++-- 7 files changed, 58 insertions(+), 7 deletions(-) diff --git a/cognee/infrastructure/llm/config.py b/cognee/infrastructure/llm/config.py index 8fd196eaf..c87054ff6 100644 --- a/cognee/infrastructure/llm/config.py +++ b/cognee/infrastructure/llm/config.py @@ -38,6 +38,7 @@ class LLMConfig(BaseSettings): """ structured_output_framework: str = "instructor" + llm_instructor_mode: Optional[str] = None llm_provider: str = "openai" llm_model: str = "openai/gpt-5-mini" llm_endpoint: str = "" @@ -181,6 +182,7 @@ class LLMConfig(BaseSettings): instance. """ return { + "llm_instructor_mode": self.llm_instructor_mode, "provider": self.llm_provider, "model": self.llm_model, "endpoint": self.llm_endpoint, diff --git a/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/anthropic/adapter.py b/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/anthropic/adapter.py index bf19d6e86..6fb78718e 100644 --- a/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/anthropic/adapter.py +++ b/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/anthropic/adapter.py @@ -28,13 +28,19 @@ class AnthropicAdapter(LLMInterface): name = "Anthropic" model: str + default_instructor_mode = "anthropic_tools" def __init__(self, max_completion_tokens: int, model: str = None): import anthropic + config_instructor_mode = get_llm_config().llm_instructor_mode + instructor_mode = ( + config_instructor_mode if config_instructor_mode else self.default_instructor_mode + ) + self.aclient = instructor.patch( create=anthropic.AsyncAnthropic(api_key=get_llm_config().llm_api_key).messages.create, - mode=instructor.Mode.ANTHROPIC_TOOLS, + mode=instructor.Mode(instructor_mode), ) self.model = model diff --git a/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/gemini/adapter.py b/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/gemini/adapter.py index 1187e0cad..68dddc7b7 100644 --- a/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/gemini/adapter.py +++ b/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/gemini/adapter.py @@ -41,6 +41,7 @@ class GeminiAdapter(LLMInterface): name: str model: str api_key: str + default_instructor_mode = "json_mode" def __init__( self, @@ -63,7 +64,16 @@ class GeminiAdapter(LLMInterface): self.fallback_api_key = fallback_api_key self.fallback_endpoint = fallback_endpoint - self.aclient = instructor.from_litellm(litellm.acompletion, mode=instructor.Mode.JSON) + from cognee.infrastructure.llm.config import get_llm_config + + config_instructor_mode = get_llm_config().llm_instructor_mode + instructor_mode = ( + config_instructor_mode if config_instructor_mode else self.default_instructor_mode + ) + + self.aclient = instructor.from_litellm( + litellm.acompletion, mode=instructor.Mode(instructor_mode) + ) @retry( stop=stop_after_delay(128), diff --git a/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/generic_llm_api/adapter.py b/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/generic_llm_api/adapter.py index 8bbbaa2cc..ea32dced1 100644 --- a/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/generic_llm_api/adapter.py +++ b/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/generic_llm_api/adapter.py @@ -41,6 +41,7 @@ class GenericAPIAdapter(LLMInterface): name: str model: str api_key: str + default_instructor_mode = "json_mode" def __init__( self, @@ -63,7 +64,16 @@ class GenericAPIAdapter(LLMInterface): self.fallback_api_key = fallback_api_key self.fallback_endpoint = fallback_endpoint - self.aclient = instructor.from_litellm(litellm.acompletion, mode=instructor.Mode.JSON) + from cognee.infrastructure.llm.config import get_llm_config + + config_instructor_mode = get_llm_config().llm_instructor_mode + instructor_mode = ( + config_instructor_mode if config_instructor_mode else self.default_instructor_mode + ) + + self.aclient = instructor.from_litellm( + litellm.acompletion, mode=instructor.Mode(instructor_mode) + ) @retry( stop=stop_after_delay(128), diff --git a/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/mistral/adapter.py b/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/mistral/adapter.py index 78a3cbff5..bed88ce3c 100644 --- a/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/mistral/adapter.py +++ b/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/mistral/adapter.py @@ -37,6 +37,7 @@ class MistralAdapter(LLMInterface): model: str api_key: str max_completion_tokens: int + default_instructor_mode = "mistral_tools" def __init__(self, api_key: str, model: str, max_completion_tokens: int, endpoint: str = None): from mistralai import Mistral @@ -44,9 +45,14 @@ class MistralAdapter(LLMInterface): self.model = model self.max_completion_tokens = max_completion_tokens + config_instructor_mode = get_llm_config().llm_instructor_mode + instructor_mode = ( + config_instructor_mode if config_instructor_mode else self.default_instructor_mode + ) + self.aclient = instructor.from_litellm( litellm.acompletion, - mode=instructor.Mode.MISTRAL_TOOLS, + mode=instructor.Mode(instructor_mode), api_key=get_llm_config().llm_api_key, ) diff --git a/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/ollama/adapter.py b/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/ollama/adapter.py index 9c3d185aa..aa24a7911 100644 --- a/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/ollama/adapter.py +++ b/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/ollama/adapter.py @@ -42,6 +42,8 @@ class OllamaAPIAdapter(LLMInterface): - aclient """ + default_instructor_mode = "json_mode" + def __init__( self, endpoint: str, api_key: str, model: str, name: str, max_completion_tokens: int ): @@ -51,8 +53,16 @@ class OllamaAPIAdapter(LLMInterface): self.endpoint = endpoint self.max_completion_tokens = max_completion_tokens + from cognee.infrastructure.llm.config import get_llm_config + + config_instructor_mode = get_llm_config().llm_instructor_mode + instructor_mode = ( + config_instructor_mode if config_instructor_mode else self.default_instructor_mode + ) + self.aclient = instructor.from_openai( - OpenAI(base_url=self.endpoint, api_key=self.api_key), mode=instructor.Mode.JSON + OpenAI(base_url=self.endpoint, api_key=self.api_key), + mode=instructor.Mode(instructor_mode), ) @retry( diff --git a/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/openai/adapter.py b/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/openai/adapter.py index 305b426b8..69367602d 100644 --- a/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/openai/adapter.py +++ b/cognee/infrastructure/llm/structured_output_framework/litellm_instructor/llm/openai/adapter.py @@ -56,6 +56,7 @@ class OpenAIAdapter(LLMInterface): model: str api_key: str api_version: str + default_instructor_mode = "json_schema_mode" MAX_RETRIES = 5 @@ -74,14 +75,20 @@ class OpenAIAdapter(LLMInterface): fallback_api_key: str = None, fallback_endpoint: str = None, ): + from cognee.infrastructure.llm.config import get_llm_config + + config_instructor_mode = get_llm_config().llm_instructor_mode + instructor_mode = ( + config_instructor_mode if config_instructor_mode else self.default_instructor_mode + ) # TODO: With gpt5 series models OpenAI expects JSON_SCHEMA as a mode for structured outputs. # Make sure all new gpt models will work with this mode as well. if "gpt-5" in model: self.aclient = instructor.from_litellm( - litellm.acompletion, mode=instructor.Mode.JSON_SCHEMA + litellm.acompletion, mode=instructor.Mode(instructor_mode) ) self.client = instructor.from_litellm( - litellm.completion, mode=instructor.Mode.JSON_SCHEMA + litellm.completion, mode=instructor.Mode(instructor_mode) ) else: self.aclient = instructor.from_litellm(litellm.acompletion)