Agent VectorDB Question Answering Benchmarking#

Here we go over how to benchmark performance on a question answering task using an agent to route between multiple vectordatabases.

It is highly reccomended that you do any evaluation/benchmarking with tracing enabled. See here for an explanation of what tracing is and how to set it up.

# Comment this out if you are NOT using tracing
import os
os.environ["LANGCHAIN_HANDLER"] = "langchain"

Loading the data#

First, let’s load the data.

from langchain.evaluation.loading import load_dataset
dataset = load_dataset("agent-vectordb-qa-sota-pg")
Found cached dataset json (/Users/harrisonchase/.cache/huggingface/datasets/LangChainDatasets___json/LangChainDatasets--agent-vectordb-qa-sota-pg-d3ae24016b514f92/0.0.0/0f7e3662623656454fcd2b650f34e886a7db4b9104504885bd462096cc7a9f51)
dataset[0]
{'question': 'What is the purpose of the NATO Alliance?',
 'answer': 'The purpose of the NATO Alliance is to secure peace and stability in Europe after World War 2.',
 'steps': [{'tool': 'State of Union QA System', 'tool_input': None},
  {'tool': None, 'tool_input': 'What is the purpose of the NATO Alliance?'}]}
dataset[-1]
{'question': 'What is the purpose of YC?',
 'answer': 'The purpose of YC is to cause startups to be founded that would not otherwise have existed.',
 'steps': [{'tool': 'Paul Graham QA System', 'tool_input': None},
  {'tool': None, 'tool_input': 'What is the purpose of YC?'}]}

Setting up a chain#

Now we need to create some pipelines for doing question answering. Step one in that is creating indexes over the data in question.

from langchain.document_loaders import TextLoader
loader = TextLoader("../../modules/state_of_the_union.txt")
from langchain.indexes import VectorstoreIndexCreator
vectorstore_sota = VectorstoreIndexCreator(vectorstore_kwargs={"collection_name":"sota"}).from_loaders([loader]).vectorstore
Running Chroma using direct local API.
Using DuckDB in-memory for database. Data will be transient.

Now we can create a question answering chain.

from langchain.chains import RetrievalQA
from langchain.llms import OpenAI
chain_sota = RetrievalQA.from_chain_type(llm=OpenAI(temperature=0), chain_type="stuff", retriever=vectorstore_sota, input_key="question")

Now we do the same for the Paul Graham data.

loader = TextLoader("../../modules/paul_graham_essay.txt")
vectorstore_pg = VectorstoreIndexCreator(vectorstore_kwargs={"collection_name":"paul_graham"}).from_loaders([loader]).vectorstore
Running Chroma using direct local API.
Using DuckDB in-memory for database. Data will be transient.
chain_pg = RetrievalQA.from_chain_type(llm=OpenAI(temperature=0), chain_type="stuff", retriever=vectorstore_pg, input_key="question")

We can now set up an agent to route between them.

from langchain.agents import initialize_agent, Tool
tools = [
    Tool(
        name = "State of Union QA System",
        func=chain_sota.run,
        description="useful for when you need to answer questions about the most recent state of the union address. Input should be a fully formed question."
    ),
    Tool(
        name = "Paul Graham System",
        func=chain_pg.run,
        description="useful for when you need to answer questions about Paul Graham. Input should be a fully formed question."
    ),
]
agent = initialize_agent(tools, OpenAI(temperature=0), agent="zero-shot-react-description", max_iterations=3)

Make a prediction#

First, we can make predictions one datapoint at a time. Doing it at this level of granularity allows use to explore the outputs in detail, and also is a lot cheaper than running over multiple datapoints

agent.run(dataset[0]['question'])
'The purpose of the NATO Alliance is to promote peace and security in the North Atlantic region by providing a collective defense against potential threats.'

Make many predictions#

Now we can make predictions

predictions = []
predicted_dataset = []
error_dataset = []
for data in dataset:
    new_data = {"input": data["question"], "answer": data["answer"]}
    try:
        predictions.append(agent(new_data))
        predicted_dataset.append(new_data)
    except Exception:
        error_dataset.append(new_data)

Evaluate performance#

Now we can evaluate the predictions. The first thing we can do is look at them by eye.

predictions[0]

Next, we can use a language model to score them programatically

from langchain.evaluation.qa import QAEvalChain
llm = OpenAI(temperature=0)
eval_chain = QAEvalChain.from_llm(llm)
graded_outputs = eval_chain.evaluate(predicted_dataset, predictions, question_key="input", prediction_key="output")

We can add in the graded output to the predictions dict and then get a count of the grades.

for i, prediction in enumerate(predictions):
    prediction['grade'] = graded_outputs[i]['text']
from collections import Counter
Counter([pred['grade'] for pred in predictions])
Counter({' CORRECT': 19, ' INCORRECT': 14})

We can also filter the datapoints to the incorrect examples and look at them.

incorrect = [pred for pred in predictions if pred['grade'] == " INCORRECT"]
incorrect[0]
{'input': 'What is the purpose of the Bipartisan Innovation Act mentioned in the text?',
 'answer': 'The Bipartisan Innovation Act will make record investments in emerging technologies and American manufacturing to level the playing field with China and other competitors.',
 'output': 'The purpose of the Bipartisan Innovation Act is to promote innovation and entrepreneurship in the United States by providing tax incentives and other support for startups and small businesses.',
 'grade': ' INCORRECT'}