Customer Service AI Assistant with LangChain, Hugging Face, and Vertex AI

1. Listing the Most Popular Hugging Face Models for Specific Tasks

from huggingface_hub import list_models

# Define a function to list the top 5 most downloaded models for a specified task
def list_most_popular(task: str):
    for rank, model in enumerate(
        list_models(filter=task, sort="downloads", direction=-1)
):
        if rank == 5:
            break
        print(f"{model.id}, {model.downloads}\n")

# List popular models for text classification
list_most_popular("text-classification")

# List popular models for summarization
list_most_popular("summarization")

• Explanation:

  • Purpose: This code lists the top 5 models for each task (text-classification and summarization), sorted by the highest number of downloads on Hugging Face.

  • Why it’s useful: Knowing the popular models can help you quickly choose reliable models for customer service applications without trial and error.

• Sample Output:

    facebook/bart-large-mnli, 500000
    distilbert-base-uncased, 300000
    ...
  

---

2. Setting Up a Hugging Face Summarization Endpoint

from langchain_community.llms.huggingface_endpoint import HuggingFaceEndpoint

# Sample customer email text
customer_email = """..."""  # Customer’s detailed complaint (see full text above).

# Set up the summarization model, requiring an API token (set in environment as HUGGINGFACEHUB_API_TOKEN)
summarizer = HuggingFaceEndpoint(
    repo_id="facebook/bart-large-cnn",
    model_kwargs={"temperature":0, "max_length":180}
)

# Function to summarize a text
def summarize(llm, text) -> str:
    return llm(f"Summarize this: {text}!")

# Summarize the customer email
print(summarize(summarizer, customer_email))

• Explanation:

  • HuggingFaceEndpoint: Wraps the facebook/bart-large-cnn summarization model for easy use within LangChain. It requires an API token from Hugging Face.

  • Summarize Function: Sends the text to the model with the Summarize this: prompt to get a concise summary.

  • Why it’s useful: Summarizing customer emails helps quickly capture the essence of lengthy complaints, improving response efficiency.

• Sample Output:

    "Customer received a damaged coffee machine and expresses disappointment in the product and brand quality."
  

---

3. Analyzing Sentiment with Hugging Face’s Sentiment Model

from transformers import pipeline

# Set up a sentiment analysis model for evaluating customer emotions
sentiment_model = pipeline(
    task="sentiment-analysis",
    model="cardiffnlp/twitter-roberta-base-sentiment"
)

# Run sentiment analysis on the customer email
print(sentiment_model(customer_email))

# Test the model on different example texts
print(sentiment_model("I am so angry and sad, I want to kill myself!"))
print(sentiment_model("I am elated, I am so happy, this is the best thing that ever happened to me!"))
print(sentiment_model("I don't care. I guess it's ok, or not, I couldn't care one way or the other"))

• Explanation:

  • Sentiment Analysis Pipeline: Loads a pre-trained sentiment model to analyze the tone and emotion in customer messages.

  • Purpose: Provides insights into the customer’s emotional state, helping the support team handle emotionally charged situations effectively.

  • Different Inputs: Tests include strongly negative, positive, and neutral statements to assess the model’s accuracy.

• Sample Output:

    [{'label': 'NEGATIVE', 'score': 0.95}]
    [{'label': 'NEGATIVE', 'score': 0.99}]
    [{'label': 'POSITIVE', 'score': 0.98}]
    [{'label': 'NEUTRAL', 'score': 0.92}]
  

---

4. Categorizing Customer Concerns with Vertex AI and LangChain

from langchain.llms import VertexAI
from langchain import PromptTemplate, LLMChain

# Define a template to classify the customer's issue based on the email content
template = """Given this text, decide what is the issue the customer is concerned about. Valid categories are these:
* product issues
* delivery problems
* missing or late orders
* wrong product
* cancellation request
* refund or exchange
* bad support experience
* no clear reason to be upset

Text: {email}
Category:
"""
# Create the prompt template
prompt = PromptTemplate(template=template, input_variables=["email"])

# Initialize VertexAI to handle text generation
llm = VertexAI()

# Create a chain that runs the prompt through VertexAI
llm_chain = LLMChain(prompt=prompt, llm=llm, verbose=True)

# Run the customer email through the chain to classify the issue
print(llm_chain.run(customer_email))

• Explanation:

  • PromptTemplate: Provides a structured prompt with predefined categories, guiding the model to classify the complaint correctly.

  • LLMChain: Uses LangChain to pass the email content through the VertexAI model, leveraging its classification capabilities.

  • Purpose: Helps identify the specific category of the complaint (e.g., product issue, delivery problem), enabling better response routing and prioritization.

• Sample Output:

    "product issues"
  

---

Summary

This notebook sets up a Customer Service Helper that can:

1. Identify popular models to quickly access reliable tools.

2. Summarize customer emails for efficient review.

3. Analyze emotional sentiment to gauge the intensity of the complaint.

4. Classify the issue type for streamlined support and response.