dima/langchain
1
0 Fork 0
mirror of https://github.com/langchain-ai/langchain.git synced 2026-10-16 14:49:21 +03:00
Code Issues Packages Projects Releases Wiki Activity
Files
5fda8383461225f77949c94de5dbeedef1068cfa
langchain/docs/extras/use_cases/apis.ipynb
T

20 KiB
Raw Blame History

Cell:
[Cell type raw - unsupported, skipped]

Open In Collab

Use case

Suppose you want an LLM to interact with external APIs.

This can be very useful for retrieving context for the LLM to utilize.

And, more generally, it allows us to interact with APIs using natural language!

Overview

There are two primary ways to interface LLMs with external APIs:

  • Functions: For example, OpenAI functions is one popular means of doing this.
  • LLM-generated interface: Use an LLM with access to API documentation to create an interface.

Image description

Quickstart

Many APIs are already compatible with OpenAI function calling.

For example, Klarna has a YAML file that describes its API and allows OpenAI to interact with it:

https://www.klarna.com/us/shopping/public/openai/v0/api-docs/

Other options include:

We can supply the specification to get_openapi_chain directly in order to query the API with OpenAI functions:

In [ ]:
pip install langchain openai 

# Set env var OPENAI_API_KEY or load from a .env file:
# import dotenv
# dotenv.load_dotenv()
In [2]:
from langchain.chains.openai_functions.openapi import get_openapi_chain
chain = get_openapi_chain("https://www.klarna.com/us/shopping/public/openai/v0/api-docs/")
chain("What are some options for a men's large blue button down shirt")
Out [2]:
Attempting to load an OpenAPI 3.0.1 spec.  This may result in degraded performance. Convert your OpenAPI spec to 3.1.* spec for better support.

{'query': "What are some options for a men's large blue button down shirt",
 'response': {'products': [{'name': 'Cubavera Four Pocket Guayabera Shirt',
    'url': 'https://www.klarna.com/us/shopping/pl/cl10001/3202055522/Clothing/Cubavera-Four-Pocket-Guayabera-Shirt/?utm_source=openai&ref-site=openai_plugin',
    'price': '$13.50',
    'attributes': ['Material:Polyester,Cotton',
     'Target Group:Man',
     'Color:Red,White,Blue,Black',
     'Properties:Pockets',
     'Pattern:Solid Color',
     'Size (Small-Large):S,XL,L,M,XXL']},
   {'name': 'Polo Ralph Lauren Plaid Short Sleeve Button-down Oxford Shirt',
    'url': 'https://www.klarna.com/us/shopping/pl/cl10001/3207163438/Clothing/Polo-Ralph-Lauren-Plaid-Short-Sleeve-Button-down-Oxford-Shirt/?utm_source=openai&ref-site=openai_plugin',
    'price': '$52.20',
    'attributes': ['Material:Cotton',
     'Target Group:Man',
     'Color:Red,Blue,Multicolor',
     'Size (Small-Large):S,XL,L,M,XXL']},
   {'name': 'Brixton Bowery Flannel Shirt',
    'url': 'https://www.klarna.com/us/shopping/pl/cl10001/3202331096/Clothing/Brixton-Bowery-Flannel-Shirt/?utm_source=openai&ref-site=openai_plugin',
    'price': '$27.48',
    'attributes': ['Material:Cotton',
     'Target Group:Man',
     'Color:Gray,Blue,Black,Orange',
     'Properties:Pockets',
     'Pattern:Checkered',
     'Size (Small-Large):XL,3XL,4XL,5XL,L,M,XXL']},
   {'name': 'Vineyard Vines Gingham On-The-Go brrr Classic Fit Shirt Crystal',
    'url': 'https://www.klarna.com/us/shopping/pl/cl10001/3201938510/Clothing/Vineyard-Vines-Gingham-On-The-Go-brrr-Classic-Fit-Shirt-Crystal/?utm_source=openai&ref-site=openai_plugin',
    'price': '$80.64',
    'attributes': ['Material:Cotton',
     'Target Group:Man',
     'Color:Blue',
     'Size (Small-Large):XL,XS,L,M']},
   {'name': "Carhartt Men's Loose Fit Midweight Short Sleeve Plaid Shirt",
    'url': 'https://www.klarna.com/us/shopping/pl/cl10001/3201826024/Clothing/Carhartt-Men-s-Loose-Fit-Midweight-Short-Sleeve-Plaid-Shirt/?utm_source=openai&ref-site=openai_plugin',
    'price': '$17.99',
    'attributes': ['Material:Cotton',
     'Target Group:Man',
     'Color:Red,Brown,Blue,Green',
     'Properties:Pockets',
     'Pattern:Checkered',
     'Size (Small-Large):S,XL,L,M']}]}}

Functions

We can unpack what is happening when we use the functions to call external APIs.

Let's look at the LangSmith trace:

  • See here that we call the OpenAI LLM with the provided API spec:
https://www.klarna.com/us/shopping/public/openai/v0/api-docs/
  • The prompt then tells the LLM to use the API spec with input question:
Use the provided APIs to respond to this user query:
What are some options for a men's large blue button down shirt

function_call:
  name: productsUsingGET
  arguments: |-
    {
      "params": {
        "countryCode": "US",
        "q": "men's large blue button down shirt",
        "size": 5,
        "min_price": 0,
        "max_price": 100
      }
    }

Image description

API Chain

We can also build our own interface to external APIs using the APIChain and provided API documentation.

In [8]:
from langchain.llms import OpenAI
from langchain.chains import APIChain
from langchain.chains.api import open_meteo_docs
llm = OpenAI(temperature=0)
chain = APIChain.from_llm_and_api_docs(llm, open_meteo_docs.OPEN_METEO_DOCS, verbose=True)
chain.run('What is the weather like right now in Munich, Germany in degrees Fahrenheit?')
Out [8]:

> Entering new APIChain chain...
https://api.open-meteo.com/v1/forecast?latitude=48.1351&longitude=11.5820&hourly=temperature_2m&temperature_unit=fahrenheit&current_weather=true
{"latitude":48.14,"longitude":11.58,"generationtime_ms":1.0769367218017578,"utc_offset_seconds":0,"timezone":"GMT","timezone_abbreviation":"GMT","elevation":521.0,"current_weather":{"temperature":52.9,"windspeed":12.6,"winddirection":239.0,"weathercode":3,"is_day":0,"time":"2023-08-07T22:00"},"hourly_units":{"time":"iso8601","temperature_2m":"°F"},"hourly":{"time":["2023-08-07T00:00","2023-08-07T01:00","2023-08-07T02:00","2023-08-07T03:00","2023-08-07T04:00","2023-08-07T05:00","2023-08-07T06:00","2023-08-07T07:00","2023-08-07T08:00","2023-08-07T09:00","2023-08-07T10:00","2023-08-07T11:00","2023-08-07T12:00","2023-08-07T13:00","2023-08-07T14:00","2023-08-07T15:00","2023-08-07T16:00","2023-08-07T17:00","2023-08-07T18:00","2023-08-07T19:00","2023-08-07T20:00","2023-08-07T21:00","2023-08-07T22:00","2023-08-07T23:00","2023-08-08T00:00","2023-08-08T01:00","2023-08-08T02:00","2023-08-08T03:00","2023-08-08T04:00","2023-08-08T05:00","2023-08-08T06:00","2023-08-08T07:00","2023-08-08T08:00","2023-08-08T09:00","2023-08-08T10:00","2023-08-08T11:00","2023-08-08T12:00","2023-08-08T13:00","2023-08-08T14:00","2023-08-08T15:00","2023-08-08T16:00","2023-08-08T17:00","2023-08-08T18:00","2023-08-08T19:00","2023-08-08T20:00","2023-08-08T21:00","2023-08-08T22:00","2023-08-08T23:00","2023-08-09T00:00","2023-08-09T01:00","2023-08-09T02:00","2023-08-09T03:00","2023-08-09T04:00","2023-08-09T05:00","2023-08-09T06:00","2023-08-09T07:00","2023-08-09T08:00","2023-08-09T09:00","2023-08-09T10:00","2023-08-09T11:00","2023-08-09T12:00","2023-08-09T13:00","2023-08-09T14:00","2023-08-09T15:00","2023-08-09T16:00","2023-08-09T17:00","2023-08-09T18:00","2023-08-09T19:00","2023-08-09T20:00","2023-08-09T21:00","2023-08-09T22:00","2023-08-09T23:00","2023-08-10T00:00","2023-08-10T01:00","2023-08-10T02:00","2023-08-10T03:00","2023-08-10T04:00","2023-08-10T05:00","2023-08-10T06:00","2023-08-10T07:00","2023-08-10T08:00","2023-08-10T09:00","2023-08-10T10:00","2023-08-10T11:00","2023-08-10T12:00","2023-08-10T13:00","2023-08-10T14:00","2023-08-10T15:00","2023-08-10T16:00","2023-08-10T17:00","2023-08-10T18:00","2023-08-10T19:00","2023-08-10T20:00","2023-08-10T21:00","2023-08-10T22:00","2023-08-10T23:00","2023-08-11T00:00","2023-08-11T01:00","2023-08-11T02:00","2023-08-11T03:00","2023-08-11T04:00","2023-08-11T05:00","2023-08-11T06:00","2023-08-11T07:00","2023-08-11T08:00","2023-08-11T09:00","2023-08-11T10:00","2023-08-11T11:00","2023-08-11T12:00","2023-08-11T13:00","2023-08-11T14:00","2023-08-11T15:00","2023-08-11T16:00","2023-08-11T17:00","2023-08-11T18:00","2023-08-11T19:00","2023-08-11T20:00","2023-08-11T21:00","2023-08-11T22:00","2023-08-11T23:00","2023-08-12T00:00","2023-08-12T01:00","2023-08-12T02:00","2023-08-12T03:00","2023-08-12T04:00","2023-08-12T05:00","2023-08-12T06:00","2023-08-12T07:00","2023-08-12T08:00","2023-08-12T09:00","2023-08-12T10:00","2023-08-12T11:00","2023-08-12T12:00","2023-08-12T13:00","2023-08-12T14:00","2023-08-12T15:00","2023-08-12T16:00","2023-08-12T17:00","2023-08-12T18:00","2023-08-12T19:00","2023-08-12T20:00","2023-08-12T21:00","2023-08-12T22:00","2023-08-12T23:00","2023-08-13T00:00","2023-08-13T01:00","2023-08-13T02:00","2023-08-13T03:00","2023-08-13T04:00","2023-08-13T05:00","2023-08-13T06:00","2023-08-13T07:00","2023-08-13T08:00","2023-08-13T09:00","2023-08-13T10:00","2023-08-13T11:00","2023-08-13T12:00","2023-08-13T13:00","2023-08-13T14:00","2023-08-13T15:00","2023-08-13T16:00","2023-08-13T17:00","2023-08-13T18:00","2023-08-13T19:00","2023-08-13T20:00","2023-08-13T21:00","2023-08-13T22:00","2023-08-13T23:00"],"temperature_2m":[53.0,51.2,50.9,50.4,50.7,51.3,51.7,52.9,54.3,56.1,57.4,59.3,59.1,60.7,59.7,58.8,58.8,57.8,56.6,55.3,53.9,52.7,52.9,53.2,52.0,51.8,51.3,50.7,50.8,51.5,53.9,57.7,61.2,63.2,64.7,66.6,67.5,67.0,68.7,68.7,67.9,66.2,64.4,61.4,59.8,58.9,57.9,56.3,55.7,55.3,55.5,55.4,55.7,56.5,57.6,58.8,59.7,59.1,58.9,60.6,59.9,59.8,59.9,61.7,63.2,63.6,62.3,58.9,57.3,57.1,57.0,56.5,56.2,56.0,55.3,54.7,54.4,55.2,57.8,60.7,63.0,65.3,66.9,68.2,70.1,72.1,72.6,71.4,69.7,68.6,66.2,63.6,61.8,60.6,59.6,58.9,58.0,57.1,56.3,56.2,56.7,57.9,59.9,63.7,68.4,72.4,75.0,76.8,78.0,78.7,78.9,78.4,76.9,74.8,72.5,70.1,67.6,65.6,64.4,63.9,63.4,62.7,62.2,62.1,62.5,63.4,65.1,68.0,71.7,74.8,76.8,78.2,79.1,79.6,79.7,79.2,77.6,75.3,73.7,68.6,66.8,65.3,64.2,63.4,62.6,61.7,60.9,60.6,60.9,61.6,63.2,65.9,69.3,72.2,74.4,76.2,77.6,78.8,79.6,79.6,78.4,76.4,74.3,72.3,70.4,68.7,67.6,66.8]}}

> Finished chain.

' The current temperature in Munich, Germany is 52.9°F.'

Note that we supply information about the API:

In [37]:
open_meteo_docs.OPEN_METEO_DOCS[0:500]
Out [37]:
'BASE URL: https://api.open-meteo.com/\n\nAPI Documentation\nThe API endpoint /v1/forecast accepts a geographical coordinate, a list of weather variables and responds with a JSON hourly weather forecast for 7 days. Time always starts at 0:00 today and contains 168 hours. All URL parameters are listed below:\n\nParameter\tFormat\tRequired\tDefault\tDescription\nlatitude, longitude\tFloating point\tYes\t\tGeographical WGS84 coordinate of the location\nhourly\tString array\tNo\t\tA list of weather variables which shou'

Under the hood, we do two things:

  • api_request_chain: Generate an API URL based on the input question and the api_docs
  • api_answer_chain: generate a final answer based on the API response

We can look at the LangSmith trace to inspect this:

  • The api_request_chain produces the API url from our question and the API documentation:

Image description

  • Here we make the API request with the API url.
  • The api_answer_chain takes the response from the API and provides us with a natural language response:

Image description

Going deeper

Test with other APIs

In [ ]:
import os
os.environ['TMDB_BEARER_TOKEN'] = ""
from langchain.chains.api import tmdb_docs
headers = {"Authorization": f"Bearer {os.environ['TMDB_BEARER_TOKEN']}"}
chain = APIChain.from_llm_and_api_docs(llm, tmdb_docs.TMDB_DOCS, headers=headers, verbose=True)
chain.run("Search for 'Avatar'")
In [ ]:
import os
from langchain.llms import OpenAI
from langchain.chains.api import podcast_docs
from langchain.chains import APIChain
 
listen_api_key = 'xxx' # Get api key here: https://www.listennotes.com/api/pricing/
llm = OpenAI(temperature=0)
headers = {"X-ListenAPI-Key": listen_api_key}
chain = APIChain.from_llm_and_api_docs(llm, podcast_docs.PODCAST_DOCS, headers=headers, verbose=True)
chain.run("Search for 'silicon valley bank' podcast episodes, audio length is more than 30 minutes, return only 1 results")

Web requests

URL requests are such a common use-case that we have the LLMRequestsChain, which makes an HTTP GET request.

In [39]:
from langchain.llms import OpenAI
from langchain.prompts import PromptTemplate
from langchain.chains import LLMRequestsChain, LLMChain
In [40]:
template = """Between >>> and <<< are the raw search result text from google.
Extract the answer to the question '{query}' or say "not found" if the information is not contained.
Use the format
Extracted:<answer or "not found">
>>> {requests_result} <<<
Extracted:"""

PROMPT = PromptTemplate(
    input_variables=["query", "requests_result"],
    template=template,
)
In [43]:
chain = LLMRequestsChain(llm_chain=LLMChain(llm=OpenAI(temperature=0), prompt=PROMPT))
question = "What are the Three (3) biggest countries, and their respective sizes?"
inputs = {
    "query": question,
    "url": "https://www.google.com/search?q=" + question.replace(" ", "+"),
}
chain(inputs)
Out [43]:
{'query': 'What are the Three (3) biggest countries, and their respective sizes?',
 'url': 'https://www.google.com/search?q=What+are+the+Three+(3)+biggest+countries,+and+their+respective+sizes?',
 'output': ' Russia (17,098,242 km²), Canada (9,984,670 km²), China (9,706,961 km²)'}
Reference in new issue View Git Blame Copy Permalink