langchain/cookbook/custom_agent_with_plugin_retrieval.ipynb

Custom Agent with PlugIn Retrieval

This notebook combines two concepts in order to build a custom agent that can interact with AI Plugins:

1. Custom Agent with Tool Retrieval: This introduces the concept of retrieving many tools, which is useful when trying to work with arbitrarily many plugins.
2. Natural Language API Chains: This creates Natural Language wrappers around OpenAPI endpoints. This is useful because (1) plugins use OpenAPI endpoints under the hood, (2) wrapping them in an NLAChain allows the router agent to call it more easily.

The novel idea introduced in this notebook is the idea of using retrieval to select not the tools explicitly, but the set of OpenAPI specs to use. We can then generate tools from those OpenAPI specs. The use case for this is when trying to get agents to use plugins. It may be more efficient to choose plugins first, then the endpoints, rather than the endpoints directly. This is because the plugins may contain more useful information for selection.

Set up environment

Do necessary imports, etc.

In [1]:

import re
from typing import Union

from langchain.agents import (
    AgentExecutor,
    AgentOutputParser,
    LLMSingleActionAgent,
)
from langchain.chains import LLMChain
from langchain.prompts import StringPromptTemplate
from langchain_community.agent_toolkits import NLAToolkit
from langchain_community.tools.plugin import AIPlugin
from langchain_core.agents import AgentAction, AgentFinish
from langchain_openai import OpenAI

Setup LLM

In [2]:

llm = OpenAI(temperature=0)

Set up plugins

Load and index plugins

In [3]:

urls = [
    "https://datasette.io/.well-known/ai-plugin.json",
    "https://api.speak.com/.well-known/ai-plugin.json",
    "https://www.wolframalpha.com/.well-known/ai-plugin.json",
    "https://www.zapier.com/.well-known/ai-plugin.json",
    "https://www.klarna.com/.well-known/ai-plugin.json",
    "https://www.joinmilo.com/.well-known/ai-plugin.json",
    "https://slack.com/.well-known/ai-plugin.json",
    "https://schooldigger.com/.well-known/ai-plugin.json",
]

AI_PLUGINS = [AIPlugin.from_url(url) for url in urls]

Tool Retriever

We will use a vectorstore to create embeddings for each tool description. Then, for an incoming query we can create embeddings for that query and do a similarity search for relevant tools.

In [4]:

from langchain_community.vectorstores import FAISS
from langchain_core.documents import Document
from langchain_openai import OpenAIEmbeddings

In [5]:

embeddings = OpenAIEmbeddings()
docs = [
    Document(
        page_content=plugin.description_for_model,
        metadata={"plugin_name": plugin.name_for_model},
    )
    for plugin in AI_PLUGINS
]
vector_store = FAISS.from_documents(docs, embeddings)
toolkits_dict = {
    plugin.name_for_model: NLAToolkit.from_llm_and_ai_plugin(llm, plugin)
    for plugin in AI_PLUGINS
}

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.
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.
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.
Attempting to load an OpenAPI 3.0.2 spec.  This may result in degraded performance. Convert your OpenAPI spec to 3.1.* spec for better support.
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.
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.
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.
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.
Attempting to load a Swagger 2.0 spec.  This may result in degraded performance. Convert your OpenAPI spec to 3.1.* spec for better support.

In [6]:

retriever = vector_store.as_retriever()


def get_tools(query):
    # Get documents, which contain the Plugins to use
    docs = retriever.invoke(query)
    # Get the toolkits, one for each plugin
    tool_kits = [toolkits_dict[d.metadata["plugin_name"]] for d in docs]
    # Get the tools: a separate NLAChain for each endpoint
    tools = []
    for tk in tool_kits:
        tools.extend(tk.nla_tools)
    return tools

We can now test this retriever to see if it seems to work.

In [7]:

tools = get_tools("What could I do today with my kiddo")
[t.name for t in tools]

Out [7]:

['Milo.askMilo',
 'Zapier_Natural_Language_Actions_(NLA)_API_(Dynamic)_-_Beta.search_all_actions',
 'Zapier_Natural_Language_Actions_(NLA)_API_(Dynamic)_-_Beta.preview_a_zap',
 'Zapier_Natural_Language_Actions_(NLA)_API_(Dynamic)_-_Beta.get_configuration_link',
 'Zapier_Natural_Language_Actions_(NLA)_API_(Dynamic)_-_Beta.list_exposed_actions',
 'SchoolDigger_API_V2.0.Autocomplete_GetSchools',
 'SchoolDigger_API_V2.0.Districts_GetAllDistricts2',
 'SchoolDigger_API_V2.0.Districts_GetDistrict2',
 'SchoolDigger_API_V2.0.Rankings_GetSchoolRank2',
 'SchoolDigger_API_V2.0.Rankings_GetRank_District',
 'SchoolDigger_API_V2.0.Schools_GetAllSchools20',
 'SchoolDigger_API_V2.0.Schools_GetSchool20',
 'Speak.translate',
 'Speak.explainPhrase',
 'Speak.explainTask']

In [8]:

tools = get_tools("what shirts can i buy?")
[t.name for t in tools]

Out [8]:

['Open_AI_Klarna_product_Api.productsUsingGET',
 'Milo.askMilo',
 'Zapier_Natural_Language_Actions_(NLA)_API_(Dynamic)_-_Beta.search_all_actions',
 'Zapier_Natural_Language_Actions_(NLA)_API_(Dynamic)_-_Beta.preview_a_zap',
 'Zapier_Natural_Language_Actions_(NLA)_API_(Dynamic)_-_Beta.get_configuration_link',
 'Zapier_Natural_Language_Actions_(NLA)_API_(Dynamic)_-_Beta.list_exposed_actions',
 'SchoolDigger_API_V2.0.Autocomplete_GetSchools',
 'SchoolDigger_API_V2.0.Districts_GetAllDistricts2',
 'SchoolDigger_API_V2.0.Districts_GetDistrict2',
 'SchoolDigger_API_V2.0.Rankings_GetSchoolRank2',
 'SchoolDigger_API_V2.0.Rankings_GetRank_District',
 'SchoolDigger_API_V2.0.Schools_GetAllSchools20',
 'SchoolDigger_API_V2.0.Schools_GetSchool20']

Prompt Template

The prompt template is pretty standard, because we're not actually changing that much logic in the actual prompt template, but rather we are just changing how retrieval is done.

In [9]:

# Set up the base template
template = """Answer the following questions as best you can, but speaking as a pirate might speak. You have access to the following tools:

{tools}

Use the following format:

Question: the input question you must answer
Thought: you should always think about what to do
Action: the action to take, should be one of [{tool_names}]
Action Input: the input to the action
Observation: the result of the action
... (this Thought/Action/Action Input/Observation can repeat N times)
Thought: I now know the final answer
Final Answer: the final answer to the original input question

Begin! Remember to speak as a pirate when giving your final answer. Use lots of "Arg"s

Question: {input}
{agent_scratchpad}"""

The custom prompt template now has the concept of a tools_getter, which we call on the input to select the tools to use

In [10]:

from typing import Callable


# Set up a prompt template
class CustomPromptTemplate(StringPromptTemplate):
    # The template to use
    template: str
    ############## NEW ######################
    # The list of tools available
    tools_getter: Callable

    def format(self, **kwargs) -> str:
        # Get the intermediate steps (AgentAction, Observation tuples)
        # Format them in a particular way
        intermediate_steps = kwargs.pop("intermediate_steps")
        thoughts = ""
        for action, observation in intermediate_steps:
            thoughts += action.log
            thoughts += f"\nObservation: {observation}\nThought: "
        # Set the agent_scratchpad variable to that value
        kwargs["agent_scratchpad"] = thoughts
        ############## NEW ######################
        tools = self.tools_getter(kwargs["input"])
        # Create a tools variable from the list of tools provided
        kwargs["tools"] = "\n".join(
            [f"{tool.name}: {tool.description}" for tool in tools]
        )
        # Create a list of tool names for the tools provided
        kwargs["tool_names"] = ", ".join([tool.name for tool in tools])
        return self.template.format(**kwargs)

In [11]:

prompt = CustomPromptTemplate(
    template=template,
    tools_getter=get_tools,
    # This omits the `agent_scratchpad`, `tools`, and `tool_names` variables because those are generated dynamically
    # This includes the `intermediate_steps` variable because that is needed
    input_variables=["input", "intermediate_steps"],
)

Output Parser

The output parser is unchanged from the previous notebook, since we are not changing anything about the output format.

In [12]:

class CustomOutputParser(AgentOutputParser):
    def parse(self, llm_output: str) -> Union[AgentAction, AgentFinish]:
        # Check if agent should finish
        if "Final Answer:" in llm_output:
            return AgentFinish(
                # Return values is generally always a dictionary with a single `output` key
                # It is not recommended to try anything else at the moment :)
                return_values={"output": llm_output.split("Final Answer:")[-1].strip()},
                log=llm_output,
            )
        # Parse out the action and action input
        regex = r"Action\s*\d*\s*:(.*?)\nAction\s*\d*\s*Input\s*\d*\s*:[\s]*(.*)"
        match = re.search(regex, llm_output, re.DOTALL)
        if not match:
            raise ValueError(f"Could not parse LLM output: `{llm_output}`")
        action = match.group(1).strip()
        action_input = match.group(2)
        # Return the action and action input
        return AgentAction(
            tool=action, tool_input=action_input.strip(" ").strip('"'), log=llm_output
        )

In [13]:

output_parser = CustomOutputParser()

Set up LLM, stop sequence, and the agent

Also the same as the previous notebook

In [14]:

llm = OpenAI(temperature=0)

In [15]:

# LLM chain consisting of the LLM and a prompt
llm_chain = LLMChain(llm=llm, prompt=prompt)

In [16]:

tool_names = [tool.name for tool in tools]
agent = LLMSingleActionAgent(
    llm_chain=llm_chain,
    output_parser=output_parser,
    stop=["\nObservation:"],
    allowed_tools=tool_names,
)

Use the Agent

Now we can use it!

In [17]:

agent_executor = AgentExecutor.from_agent_and_tools(
    agent=agent, tools=tools, verbose=True
)

In [18]:

agent_executor.run("what shirts can i buy?")

Out [18]:

> Entering new AgentExecutor chain...
Thought: I need to find a product API
Action: Open_AI_Klarna_product_Api.productsUsingGET
Action Input: shirts

Observation:I found 10 shirts from the API response. They range in price from $9.99 to $450.00 and come in a variety of materials, colors, and patterns. I now know what shirts I can buy
Final Answer: Arg, I found 10 shirts from the API response. They range in price from $9.99 to $450.00 and come in a variety of materials, colors, and patterns.

> Finished chain.

'Arg, I found 10 shirts from the API response. They range in price from $9.99 to $450.00 and come in a variety of materials, colors, and patterns.'

In [ ]: