"""Methods for creating chains that use OpenAI function-calling APIs."""
import inspect
import re
from typing import Any, Callable, Dict, List, Optional, Sequence, Tuple, Type, Union
from pydantic import BaseModel
from langchain.base_language import BaseLanguageModel
from langchain.chains import LLMChain
from langchain.output_parsers.openai_functions import (
JsonOutputFunctionsParser,
PydanticAttrOutputFunctionsParser,
PydanticOutputFunctionsParser,
)
from langchain.prompts import BasePromptTemplate
from langchain.schema import BaseLLMOutputParser
PYTHON_TO_JSON_TYPES = {
"str": "string",
"int": "number",
"float": "number",
"bool": "boolean",
}
def _get_python_function_name(function: Callable) -> str:
"""Get the name of a Python function."""
source = inspect.getsource(function)
return re.search(r"^def (.*)\(", source).groups()[0] # type: ignore
def _parse_python_function_docstring(function: Callable) -> Tuple[str, dict]:
"""Parse the function and argument descriptions from the docstring of a function.
Assumes the function docstring follows Google Python style guide.
"""
docstring = inspect.getdoc(function)
if docstring:
docstring_blocks = docstring.split("\n\n")
descriptors = []
args_block = None
past_descriptors = False
for block in docstring_blocks:
if block.startswith("Args:"):
args_block = block
break
elif block.startswith("Returns:") or block.startswith("Example:"):
# Don't break in case Args come after
past_descriptors = True
elif not past_descriptors:
descriptors.append(block)
else:
continue
description = " ".join(descriptors)
else:
description = ""
args_block = None
arg_descriptions = {}
if args_block:
arg = None
for line in args_block.split("\n")[1:]:
if ":" in line:
arg, desc = line.split(":")
arg_descriptions[arg.strip()] = desc.strip()
elif arg:
arg_descriptions[arg.strip()] += " " + line.strip()
return description, arg_descriptions
def _get_python_function_arguments(function: Callable, arg_descriptions: dict) -> dict:
"""Get JsonSchema describing a Python functions arguments.
Assumes all function arguments are of primitive types (int, float, str, bool) or
are subclasses of pydantic.BaseModel.
"""
properties = {}
annotations = inspect.getfullargspec(function).annotations
for arg, arg_type in annotations.items():
if arg == "return":
continue
if isinstance(arg_type, type) and issubclass(arg_type, BaseModel):
properties[arg] = arg_type.schema()
elif arg_type.__name__ in PYTHON_TO_JSON_TYPES:
properties[arg] = {"type": PYTHON_TO_JSON_TYPES[arg_type.__name__]}
if arg in arg_descriptions:
if arg not in properties:
properties[arg] = {}
properties[arg]["description"] = arg_descriptions[arg]
return properties
def _get_python_function_required_args(function: Callable) -> List[str]:
"""Get the required arguments for a Python function."""
spec = inspect.getfullargspec(function)
required = spec.args[: -len(spec.defaults)] if spec.defaults else spec.args
required += [k for k in spec.kwonlyargs if k not in (spec.kwonlydefaults or {})]
return required
[docs]def convert_python_function_to_openai_function(function: Callable) -> Dict[str, Any]:
"""Convert a Python function to an OpenAI function-calling API compatible dict.
Assumes the Python function has type hints and a docstring with a description. If
the docstring has Google Python style argument descriptions, these will be
included as well.
"""
description, arg_descriptions = _parse_python_function_docstring(function)
return {
"name": _get_python_function_name(function),
"description": description,
"parameters": {
"type": "object",
"properties": _get_python_function_arguments(function, arg_descriptions),
"required": _get_python_function_required_args(function),
},
}
[docs]def convert_to_openai_function(
function: Union[Dict[str, Any], Type[BaseModel], Callable]
) -> Dict[str, Any]:
"""Convert a raw function/class to an OpenAI function.
Args:
function: Either a dictionary, a pydantic.BaseModel class, or a Python function.
If a dictionary is passed in, it is assumed to already be a valid OpenAI
function.
Returns:
A dict version of the passed in function which is compatible with the
OpenAI function-calling API.
"""
if isinstance(function, dict):
return function
elif isinstance(function, type) and issubclass(function, BaseModel):
schema = function.schema()
return {
"name": schema["title"],
"description": schema["description"],
"parameters": schema,
}
elif callable(function):
return convert_python_function_to_openai_function(function)
else:
raise ValueError(
f"Unsupported function type {type(function)}. Functions must be passed in"
f" as Dict, pydantic.BaseModel, or Callable."
)
def _get_openai_output_parser(
functions: Sequence[Union[Dict[str, Any], Type[BaseModel], Callable]],
function_names: Sequence[str],
) -> BaseLLMOutputParser:
"""Get the appropriate function output parser given the user functions."""
if isinstance(functions[0], type) and issubclass(functions[0], BaseModel):
if len(functions) > 1:
pydantic_schema: Union[Dict, Type[BaseModel]] = {
name: fn for name, fn in zip(function_names, functions)
}
else:
pydantic_schema = functions[0]
output_parser: BaseLLMOutputParser = PydanticOutputFunctionsParser(
pydantic_schema=pydantic_schema
)
else:
output_parser = JsonOutputFunctionsParser(args_only=len(functions) <= 1)
return output_parser
[docs]def create_openai_fn_chain(
functions: Sequence[Union[Dict[str, Any], Type[BaseModel], Callable]],
llm: BaseLanguageModel,
prompt: BasePromptTemplate,
*,
output_parser: Optional[BaseLLMOutputParser] = None,
**kwargs: Any,
) -> LLMChain:
"""Create an LLM chain that uses OpenAI functions.
Args:
functions: A sequence of either dictionaries, pydantic.BaseModels classes, or
Python functions. If dictionaries are passed in, they are assumed to
already be a valid OpenAI functions. If only a single
function is passed in, then it will be enforced that the model use that
function. pydantic.BaseModels and Python functions should have docstrings
describing what the function does. For best results, pydantic.BaseModels
should have descriptions of the parameters and Python functions should have
Google Python style args descriptions in the docstring. Additionally,
Python functions should only use primitive types (str, int, float, bool) or
pydantic.BaseModels for arguments.
llm: Language model to use, assumed to support the OpenAI function-calling API.
prompt: BasePromptTemplate to pass to the model.
output_parser: BaseLLMOutputParser to use for parsing model outputs. By default
will be inferred from the function types. If pydantic.BaseModels are passed
in, then the OutputParser will try to parse outputs using those. Otherwise
model outputs will simply be parsed as JSON. If multiple functions are
passed in and they are not pydantic.BaseModels, the chain output will
include both the name of the function that was returned and the arguments
to pass to the function.
Returns:
An LLMChain that will pass in the given functions to the model when run.
Example:
.. code-block:: python
from langchain.chains.openai_functions import create_openai_fn_chain
from langchain.chat_models import ChatOpenAI
from langchain.prompts import ChatPromptTemplate, HumanMessagePromptTemplate
from pydantic import BaseModel, Field
class RecordPerson(BaseModel):
\"\"\"Record some identifying information about a person.\"\"\"
name: str = Field(..., description="The person's name")
age: int = Field(..., description="The person's age")
fav_food: Optional[str] = Field(None, description="The person's favorite food")
class RecordDog(BaseModel):
\"\"\"Record some identifying information about a dog.\"\"\"
name: str = Field(..., description="The dog's name")
color: str = Field(..., description="The dog's color")
fav_food: Optional[str] = Field(None, description="The dog's favorite food")
llm = ChatOpenAI(model="gpt-3.5-turbo-0613", temperature=0)
prompt_msgs = [
SystemMessage(
content="You are a world class algorithm for recording entities"
),
HumanMessage(content="Make calls to the relevant function to record the entities in the following input:"),
HumanMessagePromptTemplate.from_template("{input}"),
HumanMessage(content="Tips: Make sure to answer in the correct format"),
]
prompt = ChatPromptTemplate(messages=prompt_msgs)
chain = create_openai_fn_chain([RecordPerson, RecordDog])
chain.run("Harry was a chubby brown beagle who loved chicken")
# -> RecordDog(name="Harry", color="brown", fav_food="chicken")
""" # noqa: E501
if not functions:
raise ValueError("Need to pass in at least one function. Received zero.")
openai_functions = [convert_to_openai_function(f) for f in functions]
fn_names = [oai_fn["name"] for oai_fn in openai_functions]
output_parser = output_parser or _get_openai_output_parser(functions, fn_names)
llm_kwargs: Dict[str, Any] = {
"functions": openai_functions,
}
if len(openai_functions) == 1:
llm_kwargs["function_call"] = {"name": openai_functions[0]["name"]}
llm_chain = LLMChain(
llm=llm,
prompt=prompt,
output_parser=output_parser,
llm_kwargs=llm_kwargs,
output_key="function",
**kwargs,
)
return llm_chain
[docs]def create_structured_output_chain(
output_schema: Union[Dict[str, Any], Type[BaseModel]],
llm: BaseLanguageModel,
prompt: BasePromptTemplate,
*,
output_parser: Optional[BaseLLMOutputParser] = None,
**kwargs: Any,
) -> LLMChain:
"""Create an LLMChain that uses an OpenAI function to get a structured output.
Args:
output_schema: Either a dictionary or pydantic.BaseModel class. If a dictionary
is passed in, it's assumed to already be a valid JsonSchema.
For best results, pydantic.BaseModels should have docstrings describing what
the schema represents and descriptions for the parameters.
llm: Language model to use, assumed to support the OpenAI function-calling API.
prompt: BasePromptTemplate to pass to the model.
output_parser: BaseLLMOutputParser to use for parsing model outputs. By default
will be inferred from the function types. If pydantic.BaseModels are passed
in, then the OutputParser will try to parse outputs using those. Otherwise
model outputs will simply be parsed as JSON.
Returns:
An LLMChain that will pass the given function to the model.
Example:
.. code-block:: python
from langchain.chains.openai_functions import create_structured_output_chain
from langchain.chat_models import ChatOpenAI
from langchain.prompts import ChatPromptTemplate, HumanMessagePromptTemplate
from pydantic import BaseModel, Field
class Dog(BaseModel):
\"\"\"Identifying information about a dog.\"\"\"
name: str = Field(..., description="The dog's name")
color: str = Field(..., description="The dog's color")
fav_food: Optional[str] = Field(None, description="The dog's favorite food")
llm = ChatOpenAI(model="gpt-3.5-turbo-0613", temperature=0)
prompt_msgs = [
SystemMessage(
content="You are a world class algorithm for extracting information in structured formats."
),
HumanMessage(content="Use the given format to extract information from the following input:"),
HumanMessagePromptTemplate.from_template("{input}"),
HumanMessage(content="Tips: Make sure to answer in the correct format"),
]
prompt = ChatPromptTemplate(messages=prompt_msgs)
chain = create_structured_output_chain(Dog, llm, prompt)
chain.run("Harry was a chubby brown beagle who loved chicken")
# -> Dog(name="Harry", color="brown", fav_food="chicken")
""" # noqa: E501
if isinstance(output_schema, dict):
function: Any = {
"name": "output_formatter",
"description": (
"Output formatter. Should always be used to format your response to the"
" user."
),
"parameters": output_schema,
}
else:
class _OutputFormatter(BaseModel):
"""Output formatter. Should always be used to format your response to the user.""" # noqa: E501
output: output_schema # type: ignore
function = _OutputFormatter
output_parser = output_parser or PydanticAttrOutputFunctionsParser(
pydantic_schema=_OutputFormatter, attr_name="output"
)
return create_openai_fn_chain(
[function], llm, prompt, output_parser=output_parser, **kwargs
)