v0.4.1

CHANGELOG.md

@@ -86,4 +86,9 @@ New parameters in llm decorator
 ## Version 0.4.0 (2023-11-25)
 - Input kwargs augmentations by implementing the llm_prompt function  (checkout example: [code_examples/augmenting_llm_prompt_inputs.py](https://github.com/ju-bezdek/langchain-decorators/blob/main/code_examples/augmenting_llm_prompt_inputs.py) )
 - support for automatic json fix using if `json_repair` is installed 
-(*not even OpenAI JSON format is not yet perfect*)
+(*not even OpenAI JSON format is not yet perfect*)
+
+## Version 0.4.1 (2023-12-18)
+- support for func_description passed as part of llm_function decorator
+- allowed not having func_description
+- minor fixes

src/langchain_decorators/__init__.py

@@ -6,6 +6,6 @@
 from .function_decorator import llm_function, get_function_schema
 from .chains import FunctionsProvider, FollowupHandle
 
-__version__="0.4.0"
+__version__="0.4.1"
 
 

src/langchain_decorators/common.py

@@ -13,6 +13,7 @@
 from langchain.chat_models import ChatOpenAI
 from langchain.schema import BaseMessage
 from langchain.prompts.chat import ChatMessagePromptTemplate
+from .schema import OutputWithFunctionCall
 from typing_inspect import is_generic_type, is_union_type
 
 import pydantic
@@ -186,17 +187,17 @@ def define_settings(cls,
         """
         if llm_selector is None and default_llm is None and default_streaming_llm is None:
             # only use llm_selector if no default_llm and default_streaming_llm is defined, because than we dont know what rules to set up
-            default_llm = ChatOpenAI(temperature=0.0, model="gpt-3.5-turbo-1106" if USE_PREVIEW_MODELS else "gpt-3.5-turbo", request_timeout=90) #  '-0613' - has function calling
+            default_llm = ChatOpenAI(temperature=0.0, model="gpt-3.5-turbo-1106" if USE_PREVIEW_MODELS else "gpt-3.5-turbo", request_timeout=30) #  '-0613' - has function calling
             default_streaming_llm = make_llm_streamable(default_llm)
             llm_selector = LlmSelector()\
                 .with_llm(default_llm, llm_selector_rule_key="chatGPT")\
-                .with_llm(ChatOpenAI(temperature=0.0, model="gpt-4-1106-preview"  if USE_PREVIEW_MODELS else "gpt-3.5-turbo-16k",  request_timeout=120), llm_selector_rule_key="GPT4")\
+                .with_llm(ChatOpenAI(temperature=0.0, model="gpt-4-1106-preview"  if USE_PREVIEW_MODELS else "gpt-3.5-turbo-16k",  request_timeout=60), llm_selector_rule_key="GPT4")\
                 #.with_llm(ChatOpenAI(temperature=0.0, model="gpt-3.5-turbo-1106"), llm_selector_rule_key="chatGPT")\
                 #.with_llm(ChatOpenAI(temperature=0.0, model="gpt-4-32k"), llm_selector_rule_key="GPT4") 
 
         else:
             if default_llm is None:
-                default_llm = ChatOpenAI(temperature=0.0, model="gpt-3.5-turbo-1106" if USE_PREVIEW_MODELS else "gpt-3.5-turbo", request_timeout=90)  #  '-0613' - has function calling
+                default_llm = ChatOpenAI(temperature=0.0, model="gpt-3.5-turbo-1106" if USE_PREVIEW_MODELS else "gpt-3.5-turbo", request_timeout=60)  #  '-0613' - has function calling
             if default_streaming_llm is None:
                 default_streaming_llm = make_llm_streamable(default_llm)
 
@@ -352,7 +353,10 @@ def get_func_return_type(func: callable, with_args:bool=False)->Union[Type, Tupl
             raise Exception(f"Invalid Union annotation {return_type}. Expected Union[ <return_type>, None] or just <return_type>")
     elif is_generic_type(return_type):
         # this should cover list and dict
-        return get_origin(return_type) if not with_args else (get_origin(return_type), get_args(return_type))
+        if get_origin(return_type) !=OutputWithFunctionCall and return_type!=OutputWithFunctionCall:
+            return get_origin(return_type) if not with_args else (get_origin(return_type), get_args(return_type))
+        else:
+            return get_args(return_type)[0]
     else:
         return return_type if not with_args else (return_type, None)
 

src/langchain_decorators/function_decorator.py

@@ -333,13 +333,13 @@ def pop_prop_title(schema):
 
     description = parse_function_description_from_docstrings(func_docs) if func_docs else func_description
 
-    if not description:
-        raise ValueError(f"LLM Function {get_function_full_name(func)} has no description in docstrings")
-    return {
+    result_schema = {
         "name":func_name,
-        "description":description,
         "parameters":args_schema
     }
+    if description:
+        result_schema["description"]=description
+    return result_schema
 
 
 

src/langchain_decorators/prompt_decorator.py

@@ -17,9 +17,11 @@
 from .common import *
 from .prompt_template import PromptDecoratorTemplate
 from .output_parsers import *
+from .schema import OutputWithFunctionCall
 from .streaming_context import StreamingContext
 from .function_decorator import is_dynamic_llm_func, get_dynamic_function_template_args
 
+
 SPECIAL_KWARGS=["callbacks","followup_handle","llm_selector_rule_key","memory","functions","function_call","capture_stream","llm_selector_rule_key", "stop", "output_parser", "llm_kwargs"]
 
 def llm_prompt(
@@ -115,8 +117,10 @@ def llm_prompt(
 
     def decorator(func):
         name=func.__name__
+
         full_name=f"{func.__module__}.{name}" if func.__module__!="__main__" else name
         is_async = inspect.iscoroutinefunction(func)
+
         _llm_selector_rule_key=llm_selector_rule_key
 
 
@@ -134,10 +138,6 @@ def build_chain(*args, **kwargs)->LLMDecoratorChain:
 
             capture_stream=_capture_stream
 
-
-
-
-
             if "capture_stream" in kwargs:
                 if not isinstance(capture_stream,bool):
                     raise ValueError("capture_stream is a reserved kwarg and must be of type bool")
@@ -401,6 +401,14 @@ def wrapper(*args, **kwargs):
                 llmChain = build_chain(*args, **kwargs)
                 return llmChain.execute()
             wrapper.build_chain=build_chain
+
+            if inspect.signature(func).parameters.get("functions"):
+                if not func.__annotations__.get('return') or func.__annotations__.get('return') == OutputWithFunctionCall:
+                    wrapper.__annotations__['return']= OutputWithFunctionCall
+                else:
+                    wrapper.__annotations__['return']= OutputWithFunctionCall[func.__annotations__.get('return') ]
+
+
             return wrapper
 
         else:
@@ -419,6 +427,11 @@ async def async_wrapper(*args, **kwargs):
                 return await llmChain.aexecute()
 
             async_wrapper.build_chain=build_chain
+            if inspect.signature(func).parameters.get("functions"):
+                if not func.__annotations__.get('return') or func.__annotations__.get('return') == OutputWithFunctionCall  or func.__annotations__.get('return') ==  Coroutine[Any,Any,OutputWithFunctionCall]:
+                    async_wrapper.__annotations__['return'] = Coroutine[Any,Any,OutputWithFunctionCall]
+                else:
+                    async_wrapper.__annotations__['return'] = Coroutine[Any,Any,OutputWithFunctionCall[func.__annotations__.get('return') ]]
             return async_wrapper
     if func:
         return decorator(func)

src/langchain_decorators/prompt_template.py

@@ -288,7 +288,7 @@ def from_func(cls,
             elif return_type==bool:
                 output_parser = "boolean"
             elif issubclass(return_type, OutputWithFunctionCall):
-                return_type = "str"
+                output_parser = "str"
             elif issubclass(return_type,BaseModel):
                 output_parser = PydanticOutputParser(model=return_type)
             else: