[Feat] support for "O" series models of OpenAI (#160)

* Added condition to check whether model is of o1 type

* Update unstract-sdk's version in __init__.py

Signed-off-by: Praveen Kumar <praveen@zipstack.com>

* Added docs to init method of openai adapter

* Updated comment in open_ai.py

* added pdm.lock

---------

Signed-off-by: Praveen Kumar <praveen@zipstack.com>

Author: pk-zipstack

pdm.lock

@@ -34,7 +34,7 @@ dependencies = [
     "llama-index-vector-stores-weaviate==1.3.1",
     "llama-index-vector-stores-pinecone==0.4.2",
     "llama-index-vector-stores-qdrant==0.4.2",
-    "llama-index-llms-openai==0.3.12",
+    "llama-index-llms-openai==0.3.17",
     "llama-index-llms-palm==0.3.0",
     "llama-index-llms-mistralai==0.3.1",
     "mistralai==1.2.5",

pyproject.toml

@@ -1,4 +1,4 @@
-__version__ = "0.57.0rc2"
+__version__ = "0.57.0rc3"
 
 
 def get_sdk_version():

src/unstract/sdk/__init__.py

@@ -4,6 +4,7 @@
 from typing import Optional
 
 from llama_index.core.llms import LLM, MockLLM
+from llama_index.llms.openai.utils import O1_MODELS
 
 from unstract.sdk.adapters.base import Adapter
 from unstract.sdk.adapters.enums import AdapterTypes
@@ -72,9 +73,14 @@ def _test_llm_instance(llm: Optional[LLM]) -> bool:
                 message="Unable to connect to LLM, please recheck the configuration",
                 status_code=400,
             )
+        # Get completion kwargs based on model capabilities
+        completion_kwargs = {}
+        if hasattr(llm, 'model') and getattr(llm, 'model') not in O1_MODELS:
+            completion_kwargs['temperature'] = 0.003
+            
         response = llm.complete(
             "The capital of Tamilnadu is ",
-            temperature=0.003,
+            **completion_kwargs
         )
         response_lower_case: str = response.text.lower()
         find_match = re.search("chennai", response_lower_case)

src/unstract/sdk/adapters/llm/llm_adapter.py

@@ -3,14 +3,14 @@
 
 from llama_index.core.llms import LLM
 from llama_index.llms.openai import OpenAI
+from llama_index.llms.openai.utils import O1_MODELS
 from openai import APIError as OpenAIAPIError
 
 from unstract.sdk.adapters.exceptions import AdapterError
 from unstract.sdk.adapters.llm.constants import LLMKeys
 from unstract.sdk.adapters.llm.llm_adapter import LLMAdapter
 from unstract.sdk.exceptions import LLMError
 
-
 class Constants:
     MODEL = "model"
     API_KEY = "api_key"
@@ -23,6 +23,7 @@ class Constants:
 
 
 class OpenAILLM(LLMAdapter):
+    
     def __init__(self, settings: dict[str, Any]):
         super().__init__("OpenAI")
         self.config = settings
@@ -53,21 +54,24 @@ def get_llm_instance(self) -> LLM:
         try:
             max_tokens = self.config.get(Constants.MAX_TOKENS)
             max_tokens = int(max_tokens) if max_tokens else None
-            llm: LLM = OpenAI(
-                model=str(self.config.get(Constants.MODEL)),
-                api_key=str(self.config.get(Constants.API_KEY)),
-                api_base=str(self.config.get(Constants.API_BASE)),
-                api_version=str(self.config.get(Constants.API_VERSION)),
-                max_retries=int(
-                    self.config.get(Constants.MAX_RETRIES, LLMKeys.DEFAULT_MAX_RETRIES)
-                ),
-                api_type="openai",
-                temperature=0,
-                timeout=float(
-                    self.config.get(Constants.TIMEOUT, LLMKeys.DEFAULT_TIMEOUT)
-                ),
-                max_tokens=max_tokens,
-            )
+            model = str(self.config.get(Constants.MODEL))
+
+            llm_kwargs = {
+                "model": model,
+                "api_key": str(self.config.get(Constants.API_KEY)),
+                "api_base": str(self.config.get(Constants.API_BASE)),
+                "api_version": str(self.config.get(Constants.API_VERSION)),
+                "max_retries": int(self.config.get(Constants.MAX_RETRIES, LLMKeys.DEFAULT_MAX_RETRIES)),
+                "api_type": "openai",
+                "timeout": float(self.config.get(Constants.TIMEOUT, LLMKeys.DEFAULT_TIMEOUT)),
+                "max_tokens": max_tokens,
+            }
+
+            # O-series models default to temperature=1, ignoring passed values, so it's not set explicitly.
+            if model not in O1_MODELS:
+                llm_kwargs["temperature"] = 0
+
+            llm = OpenAI(**llm_kwargs)
             return llm
         except Exception as e:
             raise AdapterError(str(e))