OnDemandLoaderTool Tutorial

Our OnDemandLoaderTool is a powerful agent tool that allows for "on-demand" data querying from any data source on LlamaHub.

This tool takes in a BaseReader data loader, and when called will 1) load data, 2) index data, and 3) query the data.

In this walkthrough, we show how to use the OnDemandLoaderTool to convert our Wikipedia data loader into an accessible search tool for a LangChain agent.

If you're opening this Notebook on colab, you will probably need to install LlamaIndex 🦙.

python

%pip install llama-index-readers-wikipedia

python

!pip install llama-index

python

from llama_index.core.tools.ondemand_loader_tool import OnDemandLoaderTool
from llama_index.readers.wikipedia import WikipediaReader
from typing import List

from pydantic import BaseModel

Define Tool

We first define the WikipediaReader. Note that the load_data interface to WikipediaReader takes in a list of pages. By default, this queries the Wikipedia search endpoint which will autosuggest the relevant pages.

We then wrap it into our OnDemandLoaderTool.

By default since we don't specify the index_cls, a simple vector store index is initialized.

python

reader = WikipediaReader()

python

tool = OnDemandLoaderTool.from_defaults(
    reader,
    name="Wikipedia Tool",
    description="A tool for loading and querying articles from Wikipedia",
)

Testing

We can try running the tool by itself (or as a LangChain tool), just to showcase what the interface is like!

Note that besides the arguments required for the data loader, the tool also takes in a query_str which will be the query against the index.

python

# run tool by itself
tool(["Berlin"], query_str="What's the arts and culture scene in Berlin?")

python

# run tool as langchain structured tool
lc_tool = tool.to_langchain_structured_tool(verbose=True)

python

lc_tool.run(
    tool_input={
        "pages": ["Berlin"],
        "query_str": "What's the arts and culture scene in Berlin?",
    }
)

Initialize LangChain Agent

For tutorial purposes, the agent just has access to one tool - the Wikipedia Reader

Note that we need to use Structured Tools from LangChain.

python

from langchain.agents import initialize_agent
from langchain.chat_models import ChatOpenAI

python

llm = ChatOpenAI(temperature=0, model_name="gpt-3.5-turbo", streaming=True)

python

agent = initialize_agent(
    [lc_tool],
    llm=llm,
    agent="structured-chat-zero-shot-react-description",
    verbose=True,
)

Now let's run some queries!

The OnDemandLoaderTool allows the agent to simultaneously 1) load the data from Wikipedia, 2) query that data.

python

agent.run("Tell me about the arts and culture of Berlin")

python

agent.run("Tell me about the critical reception to The Departed")