Archives
/
langchain
mirror of https://github.com/hwchase17/langchain
2
0
Fork 1
Code Issues Packages Projects Releases Wiki Activity

Generic OpenAI fn chain (#7270)

Browse Source 
Add loading functions for openai function chains and add docs page
pull/7335/head
Bagatur 1 year ago committed by GitHub
parent fd7145970f
commit 1c8cff32f1
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 932 additions and 7 deletions
Show Stats Download Patch File Download Diff File

578
docs/extras/modules/chains/popular/openai_functions.ipynb
Unescape Escape View File

@ -0,0 +1,578 @@
{
"cells": [
{
"cell_type": "markdown",
"id": "54ccb772",
"metadata": {},
"source": [
"# Using OpenAI functions\n",
"This walkthrough demonstrates how to incorporate OpenAI function-calling API's in a chain. We'll go over: \n",
"1. How to use functions to get structured outputs from ChatOpenAI\n",
"2. How to create a generic chain that uses (multiple) functions\n",
"3. How to create a chain that actually executes the chosen function"
]
},
{
"cell_type": "code",
"execution_count": 1,
"id": "767ac575",
"metadata": {},
"outputs": [],
"source": [
"from typing import Optional\n",
"\n",
"from langchain.chains.openai_functions import (\n",
" create_openai_fn_chain, create_structured_output_chain\n",
")\n",
"from langchain.prompts import ChatPromptTemplate"
]
},
{
"cell_type": "markdown",
"id": "976b6496",
"metadata": {},
"source": [
"## Getting structured outputs\n",
"We can take advantage of OpenAI functions to try and force the model to return a particular kind of structured output. We'll use the `create_structured_output_chain` to create our chain, which takes the desired structured output either as a Pydantic object or as JsonSchema.\n",
"\n",
"See here for relevant [reference docs](https://api.python.langchain.com/en/latest/chains/langchain.chains.openai_functions.base.create_structured_output_chain.html)."
]
},
{
"cell_type": "markdown",
"id": "e052faae",
"metadata": {},
"source": [
"### Using Pydantic objects\n",
"When passing in Pydantic objects to structure our text, we need to make sure to have a docstring description for the class. It also helps to have descriptions for each of the object attributes."
]
},
{
"cell_type": "code",
"execution_count": 2,
"id": "b459a33e",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"\n",
"\u001b[1m> Entering new chain...\u001b[0m\n",
"Prompt after formatting:\n",
"\u001b[32;1m\u001b[1;3mHuman: Sally is 13\u001b[0m\n",
"\n",
"\u001b[1m> Finished chain.\u001b[0m\n"
]
},
{
"data": {
"text/plain": [
"{'name': 'Sally', 'age': 13}"
]
},
"execution_count": 2,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"from pydantic import BaseModel, Field\n",
"\n",
"class Person(BaseModel):\n",
" \"\"\"Identifying information about a person.\"\"\"\n",
" name: str = Field(..., description=\"The person's name\")\n",
" age: int = Field(..., description=\"The person's age\")\n",
" fav_food: Optional[str] = Field(None, description=\"The person's favorite food\")\n",
" \n",
"chain = create_structured_output_chain(Person, verbose=True)\n",
"chain.run(\"Sally is 13\")"
]
},
{
"cell_type": "markdown",
"id": "e3539936",
"metadata": {},
"source": [
"To extract arbitrarily many structured outputs of a given format, we can just create a wrapper Pydantic object that takes a sequence of the original object."
]
},
{
"cell_type": "code",
"execution_count": 3,
"id": "4d8ea815",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"\n",
"\u001b[1m> Entering new chain...\u001b[0m\n",
"Prompt after formatting:\n",
"\u001b[32;1m\u001b[1;3mHuman: Sally is 13, Joey just turned 12 and loves spinach. Caroline is 10 years older than Sally, so she's 23.\u001b[0m\n",
"\n",
"\u001b[1m> Finished chain.\u001b[0m\n"
]
},
{
"data": {
"text/plain": [
"{'people': [{'name': 'Sally', 'age': 13, 'fav_food': ''},\n",
" {'name': 'Joey', 'age': 12, 'fav_food': 'spinach'},\n",
" {'name': 'Caroline', 'age': 23, 'fav_food': ''}]}"
]
},
"execution_count": 3,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"from typing import Sequence\n",
"\n",
"class People(BaseModel):\n",
" \"\"\"Identifying information about all people in a text.\"\"\"\n",
" people: Sequence[Person] = Field(..., description=\"The people in the text\")\n",
" \n",
"chain = create_structured_output_chain(People, verbose=True)\n",
"chain.run(\"Sally is 13, Joey just turned 12 and loves spinach. Caroline is 10 years older than Sally, so she's 23.\")"
]
},
{
"cell_type": "markdown",
"id": "ea66e10e",
"metadata": {},
"source": [
"### Using JsonSchema\n",
"\n",
"We can also pass in JsonSchema instead of Pydantic objects to specify the desired structure. When we do this, our chain will output json corresponding to the properties described in the JsonSchema, instead of a Pydantic object."
]
},
{
"cell_type": "code",
"execution_count": 4,
"id": "3484415e",
"metadata": {},
"outputs": [],
"source": [
"json_schema = {\n",
" \"title\": \"Person\",\n",
" \"description\": \"Identifying information about a person.\",\n",
" \"type\": \"object\",\n",
" \"properties\": {\n",
" \"name\": {\n",
" \"title\": \"Name\",\n",
" \"description\": \"The person's name\",\n",
" \"type\": \"string\"\n",
" },\n",
" \"age\": {\n",
" \"title\": \"Age\",\n",
" \"description\": \"The person's age\",\n",
" \"type\": \"integer\"\n",
" },\n",
" \"fav_food\": {\n",
" \"title\": \"Fav Food\",\n",
" \"description\": \"The person's favorite food\",\n",
" \"type\": \"string\"\n",
" }\n",
" },\n",
" \"required\": [\n",
" \"name\",\n",
" \"age\"\n",
" ]\n",
"}\n"
]
},
{
"cell_type": "code",
"execution_count": 5,
"id": "be9b76b3",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"\n",
"\u001b[1m> Entering new chain...\u001b[0m\n",
"Prompt after formatting:\n",
"\u001b[32;1m\u001b[1;3mHuman: Sally is 13\u001b[0m\n",
"\n",
"\u001b[1m> Finished chain.\u001b[0m\n"
]
},
{
"data": {
"text/plain": [
"{'name': 'Sally', 'age': 13}"
]
},
"execution_count": 5,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"chain = create_structured_output_chain(json_schema, verbose=True)\n",
"chain.run(\"Sally is 13\")"
]
},
{
"cell_type": "markdown",
"id": "12394696",
"metadata": {},
"source": [
"## Creating a generic OpenAI functions chain\n",
"To create a generic OpenAI functions chain, we can use the `create_openai_fn_chain` method. This is the same as `create_structured_output_chain` except that instead of taking a single output schema, it takes a sequence of function definitions.\n",
"\n",
"Functions can be passed in as:\n",
"- dicts conforming to OpenAI functions spec,\n",
"- Pydantic objects, in which case they should have docstring descriptions of the function they represent and descriptions for each of the parameters,\n",
"- Python functions, in which case they should have docstring descriptions of the function and args, along with type hints.\n",
"\n",
"See here for relevant [reference docs](https://api.python.langchain.com/en/latest/chains/langchain.chains.openai_functions.base.create_openai_fn_chain.html)."
]
},
{
"cell_type": "markdown",
"id": "ff19be25",
"metadata": {},
"source": [
"### Using Pydantic objects"
]
},
{
"cell_type": "code",
"execution_count": 6,
"id": "a4658ad8",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"\n",
"\u001b[1m> Entering new chain...\u001b[0m\n",
"Prompt after formatting:\n",
"\u001b[32;1m\u001b[1;3mHuman: Harry was a chubby brown beagle who loved chicken\u001b[0m\n",
"\n",
"\u001b[1m> Finished chain.\u001b[0m\n"
]
},
{
"data": {
"text/plain": [
"RecordDog(name='Harry', color='brown', fav_food='chicken')"
]
},
"execution_count": 6,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"class RecordPerson(BaseModel):\n",
" \"\"\"Record some identifying information about a pe.\"\"\"\n",
" name: str = Field(..., description=\"The person's name\")\n",
" age: int = Field(..., description=\"The person's age\")\n",
" fav_food: Optional[str] = Field(None, description=\"The person's favorite food\")\n",
"\n",
"class RecordDog(BaseModel):\n",
" \"\"\"Record some identifying information about a dog.\"\"\"\n",
" name: str = Field(..., description=\"The dog's name\")\n",
" color: str = Field(..., description=\"The dog's color\")\n",
" fav_food: Optional[str] = Field(None, description=\"The dog's favorite food\")\n",
"\n",
"chain = create_openai_fn_chain([RecordPerson, RecordDog], verbose=True)\n",
"chain.run(\"Harry was a chubby brown beagle who loved chicken\")"
]
},
{
"cell_type": "markdown",
"id": "df6d9147",
"metadata": {},
"source": [
"### Using Python functions\n",
"We can pass in functions as Pydantic objects, directly as OpenAI function dicts, or Python functions. To pass Python function in directly, we'll want to make sure our parameters have type hints, we have a docstring, and we use [Google Python style docstrings](https://google.github.io/styleguide/pyguide.html#doc-function-args) to describe the parameters.\n",
"\n",
"**NOTE**: To use Python functions, make sure the function arguments are of primitive types (str, float, int, bool) or that they are Pydantic objects."
]
},
{
"cell_type": "code",
"execution_count": 41,
"id": "95ac5825",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"\n",
"\u001b[1m> Entering new chain...\u001b[0m\n",
"Prompt after formatting:\n",
"\u001b[32;1m\u001b[1;3mHuman: The most important thing to remember about Tommy, my 12 year old, is that he'll do anything for apple pie.\u001b[0m\n",
"\n",
"\u001b[1m> Finished chain.\u001b[0m\n"
]
},
{
"data": {
"text/plain": [
"{'name': 'Tommy', 'age': 12, 'fav_food': {'food': 'apple pie'}}"
]
},
"execution_count": 41,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"class OptionalFavFood(BaseModel):\n",
" \"\"\"Either a food or null.\"\"\"\n",
" food: Optional[str] = Field(None, description=\"Either the name of a food or null. Should be null if the food isn't known.\")\n",
"\n",
"def record_person(name: str, age: int, fav_food: OptionalFavFood) -> str:\n",
" \"\"\"Record some basic identifying information about a person.\n",
" \n",
" Args:\n",
" name: The person's name.\n",
" age: The person's age in years.\n",
" fav_food: An OptionalFavFood object that either contains the person's favorite food or a null value. Food should be null if it's not known.\n",
" \"\"\"\n",
" return f\"Recording person {name} of age {age} with favorite food {fav_food.food}!\"\n",
"\n",
" \n",
"chain = create_openai_fn_chain([record_person], verbose=True)\n",
"chain.run(\"The most important thing to remember about Tommy, my 12 year old, is that he'll do anything for apple pie.\")"
]
},
{
"cell_type": "markdown",
"id": "403ea5dd",
"metadata": {},
"source": [
"If we pass in multiple Python functions or OpenAI functions, then the returned output will be of the form\n",
"```python\n",
"{\"name\": \"<<function_name>>\", \"arguments\": {<<function_arguments>>}}\n",
"```"
]
},
{
"cell_type": "code",
"execution_count": 42,
"id": "8b0d11de",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"\n",
"\u001b[1m> Entering new chain...\u001b[0m\n",
"Prompt after formatting:\n",
"\u001b[32;1m\u001b[1;3mHuman: I can't find my dog Henry anywhere, he's a small brown beagle. Could you send a message about him?\u001b[0m\n",
"\n",
"\u001b[1m> Finished chain.\u001b[0m\n"
]
},
{
"data": {
"text/plain": [
"{'name': 'report_dog',\n",
" 'arguments': {'name': 'Henry', 'color': 'brown', 'fav_food': {'food': None}}}"
]
},
"execution_count": 42,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"def record_dog(name: str, color: str, fav_food: OptionalFavFood) -> str:\n",
" \"\"\"Record some basic identifying information about a dog.\n",
" \n",
" Args:\n",
" name: The dog's name.\n",
" color: The dog's color.\n",
" fav_food: An OptionalFavFood object that either contains the dog's favorite food or a null value. Food should be null if it's not known.\n",
" \"\"\"\n",
" return f\"Recording dog {name} of color {color} with favorite food {fav_food}!\"\n",
"\n",
"\n",
"chain = create_openai_fn_chain([record_person, report_dog], verbose=True)\n",
"chain.run(\"I can't find my dog Henry anywhere, he's a small brown beagle. Could you send a message about him?\")"
]
},
{
"cell_type": "markdown",
"id": "4535ce33",
"metadata": {},
"source": [
"## Creating a Chain that runs the chosen function\n",
"We can go one step further and create a chain that actually executes the function chosen by the model."
]
},
{
"cell_type": "code",
"execution_count": 43,
"id": "43b0dfe0",
"metadata": {},
"outputs": [],
"source": [
"import json\n",
"import inspect\n",
"from typing import Any, Callable, Dict, List, Optional\n",
"\n",
"from langchain.callbacks.manager import CallbackManagerForChainRun\n",
"from langchain.chains.base import Chain\n",
"from langchain.input import get_colored_text\n",
"\n",
"\n",
"class FunctionExecutorChain(Chain):\n",
" functions: Dict[str, Callable]\n",
" output_key: str = \"output\"\n",
" input_key: str = \"function\"\n",
"\n",
" @property\n",
" def input_keys(self) -> List[str]:\n",
" return [self.input_key]\n",
"\n",
" @property\n",
" def output_keys(self) -> List[str]:\n",
" return [self.output_key]\n",
"\n",
" def _call(\n",
" self,\n",
" inputs: Dict[str, Any],\n",
" run_manager: Optional[CallbackManagerForChainRun] = None,\n",
" ) -> Dict[str, Any]:\n",
" \"\"\"Run the logic of this chain and return the output.\"\"\"\n",
" _run_manager = run_manager or CallbackManagerForChainRun.get_noop_manager()\n",
" name = inputs[\"function\"].pop(\"name\")\n",
" args = inputs[\"function\"].pop(\"arguments\")\n",
" _pretty_name = get_colored_text(name, \"green\")\n",
" _pretty_args = get_colored_text(json.dumps(args, indent=2), \"green\")\n",
" _text = f\"Calling function {_pretty_name} with arguments:\\n\" + _pretty_args\n",
" _run_manager.on_text(_text)\n",
" _args = {}\n",
" function = self.functions[name]\n",
" for arg_name, arg_type in inspect.getfullargspec(function).annotations.items():\n",
" if isinstance(arg_type, type) and issubclass(arg_type, BaseModel):\n",
" args[arg_name] = arg_type.parse_obj(args[arg_name])\n",
" output = function(**args)\n",
" return {self.output_key: output}"
]
},
{
"cell_type": "code",
"execution_count": 44,
"id": "b8391857",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"\n",
"\u001b[1m> Entering new chain...\u001b[0m\n",
"\n",
"\n",
"\u001b[1m> Entering new chain...\u001b[0m\n",
"Calling function \u001b[32;1m\u001b[1;3mrecord_person\u001b[0m with arguments:\n",
"\u001b[32;1m\u001b[1;3m{\n",
" \"name\": \"Tommy\",\n",
" \"age\": 12,\n",
" \"fav_food\": {\n",
" \"food\": \"apple pie\"\n",
" }\n",
"}\u001b[0m\n",
"\u001b[1m> Finished chain.\u001b[0m\n",
"\n",
"\u001b[1m> Finished chain.\u001b[0m\n"
]
},
{
"data": {
"text/plain": [
"'Recording person Tommy of age 12 with favorite food apple pie!'"
]
},
"execution_count": 44,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"from langchain.chains import SequentialChain\n",
"from langchain.chains.openai_functions.base import convert_to_openai_function\n",
"\n",
"functions = [record_person, record_dog]\n",
"openai_functions = [convert_to_openai_function(f) for f in functions]\n",
"fn_map = {\n",
" openai_fn[\"name\"]: fn for openai_fn, fn in zip(openai_functions, functions)\n",
"}\n",
"llm_chain = create_openai_fn_chain(functions)\n",
"exec_chain = FunctionExecutorChain(functions=fn_map, verbose=True)\n",
"chain = SequentialChain(\n",
" chains=[llm_chain, exec_chain],\n",
" input_variables=llm_chain.input_keys,\n",
" output_variables=[\"output\"],\n",
" verbose=True\n",
")\n",
"chain.run(\"The most important thing to remember about Tommy, my 12 year old, is that he'll do anything for apple pie.\")"
]
},
{
"cell_type": "markdown",
"id": "5f93686b",
"metadata": {},
"source": [
"## Other Chains using OpenAI functions\n",
"\n",
"There are a number of more specific chains that use OpenAI functions.\n",
"- [Extraction](/docs/modules/chains/additional/extraction): very similar to structured output chain, intended for information/entity extraction specifically.\n",
"- [Tagging](/docs/modules/chains/additional/tagging): tag inputs.\n",
"- [OpenAPI](/docs/modules/chains/additional/openapi_openai): take an OpenAPI spec and create + execute valid requests against the API, using OpenAI functions under the hood.\n",
"- [QA with citations](/docs/modules/chains/additional/qa_citations): use OpenAI functions ability to extract citations from text."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "93425c66",
"metadata": {},
"outputs": [],
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": "venv",
"language": "python",
"name": "venv"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.11.3"
}
},
"nbformat": 4,
"nbformat_minor": 5
}

6
langchain/chains/openai_functions/__init__.py
Unescape Escape View File

@ -1,3 +1,7 @@
from langchain.chains.openai_functions.base import (
create_openai_fn_chain,
create_structured_output_chain,
)
from langchain.chains.openai_functions.citation_fuzzy_match import (
create_citation_fuzzy_match_chain,
)
@ -22,4 +26,6 @@ __all__ = [
"create_citation_fuzzy_match_chain",
"create_qa_with_structure_chain",
"create_qa_with_sources_chain",
"create_structured_output_chain",
"create_openai_fn_chain",
]

315
langchain/chains/openai_functions/base.py
Unescape Escape View File

@ -0,0 +1,315 @@
"""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.chat_models import ChatOpenAI
from langchain.output_parsers.openai_functions import (
JsonOutputFunctionsParser,
PydanticOutputFunctionsParser,
)
from langchain.prompts import BasePromptTemplate, ChatPromptTemplate
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
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),
},
}
def convert_to_openai_function(
function: Union[Dict[str, Any], BaseModel, Callable]
) -> Dict[str, Any]:
"""Convert a raw function/class to an OpenAI function.
Args:
function: Either a dictionary, a pydantic.BaseModel, 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], 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
def create_openai_fn_chain(
functions: Sequence[Union[Dict[str, Any], BaseModel, Callable]],
llm: Optional[BaseLanguageModel] = None,
prompt: Optional[BasePromptTemplate] = None,
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, 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.
Defaults to ChatOpenAI using model gpt-3.5-turbo-0613.
prompt: BasePromptTemplate to pass to the model. Defaults to a prompt that just
passes user input directly to 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 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")
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]
llm = llm or ChatOpenAI(model="gpt-3.5-turbo-0613", temperature=0)
prompt = prompt or ChatPromptTemplate.from_template("{input}")
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
def create_structured_output_chain(
output_schema: Union[Dict[str, Any], BaseModel],
llm: Optional[BaseLanguageModel] = None,
prompt: Optional[BasePromptTemplate] = None,
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. 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.
Defaults to ChatOpenAI using model gpt-3.5-turbo-0613.
prompt: BasePromptTemplate to pass to the model. Defaults to a prompt that just
passes user input directly to 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 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")
chain = create_structured_output_chain([Dog])
chain.run("Harry was a chubby brown beagle who loved chicken")
# -> Dog(name="Harry", color="brown", fav_food="chicken")
""" # noqa: E501
function: Dict = {
"name": "output_formatter",
"description": (
"Output formatter. Should always be used to format your response to the"
" user."
),
}
parameters = (
output_schema if isinstance(output_schema, dict) else output_schema.schema()
)
function["parameters"] = parameters
return create_openai_fn_chain(
[function], llm=llm, prompt=prompt, output_parser=output_parser, **kwargs
)

40
langchain/output_parsers/openai_functions.py
Unescape Escape View File

@ -1,7 +1,14 @@
import json
from typing import Any, List
from typing import Any, Dict, List, Type, Union
from langchain.schema import BaseLLMOutputParser, ChatGeneration, Generation
from pydantic import BaseModel, root_validator
from langchain.schema import (
BaseLLMOutputParser,
ChatGeneration,
Generation,
OutputParserException,
)
class OutputFunctionsParser(BaseLLMOutputParser[Any]):
@ -10,14 +17,14 @@ class OutputFunctionsParser(BaseLLMOutputParser[Any]):
def parse_result(self, result: List[Generation]) -> Any:
generation = result[0]
if not isinstance(generation, ChatGeneration):
raise ValueError(
raise OutputParserException(
"This output parser can only be used with a chat generation."
)
message = generation.message
try:
func_call = message.additional_kwargs["function_call"]
except ValueError as exc:
raise ValueError(f"Could not parse function call: {exc}")
raise OutputParserException(f"Could not parse function call: {exc}")
if self.args_only:
return func_call["arguments"]
@ -42,11 +49,30 @@ class JsonKeyOutputFunctionsParser(JsonOutputFunctionsParser):
class PydanticOutputFunctionsParser(OutputFunctionsParser):
pydantic_schema: Any
pydantic_schema: Union[Type[BaseModel], Dict[str, Type[BaseModel]]]
@root_validator(pre=True)
def validate_schema(cls, values: Dict) -> Dict:
schema = values["pydantic_schema"]
if "args_only" not in values:
values["args_only"] = isinstance(schema, type) and issubclass(
schema, BaseModel
)
elif values["args_only"] and isinstance(schema, Dict):
raise ValueError(
"If multiple pydantic schemas are provided then args_only should be"
" False."
)
return values
def parse_result(self, result: List[Generation]) -> Any:
_args = super().parse_result(result)
pydantic_args = self.pydantic_schema.parse_raw(_args)
_result = super().parse_result(result)
if self.args_only:
pydantic_args = self.pydantic_schema.parse_raw(_result) # type: ignore
else:
fn_name = _result["name"]
_args = _result["arguments"]
pydantic_args = self.pydantic_schema[fn_name].parse_raw(_args) # type: ignore # noqa: E501
return pydantic_args

Write Preview
Loading…
Cancel
Save