FAQ Assistant

A simple Python-based FAQ Assistant leveraging DistilBERT for semantic understanding of questions. This assistant matches user queries to a set of pre-defined questions and returns the most relevant answer with confidence.

---

Features

• Semantic search of FAQ questions using DistilBERT embeddings.
• Configurable similarity threshold for determining matches.
• Handles multiple FAQ entries with ease.

---

Requirements

Ensure you have the following Python packages installed:

• Transformers: For using the DistilBERT model.
• Torch: For processing model inputs and outputs.
• Scikit-learn: For cosine similarity calculations.
• NumPy: For handling embeddings.
• Pandas: (Optional) If you plan to use data frames.

Install dependencies using pip:

pip install transformers torch scikit-learn numpy pandas

---

Setup

1. Clone or download the repository containing the code.
2. Save the script as main.py.

---

Usage

Step 1: Initialize the Assistant

assistant = Assistant()

Step 2: Add FAQ Data

Provide a list of question-answer pairs as a tuple data type:

faqs = [
    ("How do I reset my password?", "Click the 'Forgot Password' link on the login page and follow the email instructions."),
    ("Where can I find my account settings?", "Account settings are in the top-right menu under your profile icon."),
    ("What payment methods do you accept?", "We accept credit cards, PayPal, and bank transfers."),
]
assistant.add_data(faqs)

Step 3: Ask Questions

Input a user query and get a response:

answer, confidence = assistant.answer_question("Do you take PayPal?")
print(f"Answer: {answer} | Confidence: {confidence}")

If the question doesn't meet the similarity threshold, currently set to 0.7 by default:

"I am as confused as you are right now!"

---

How It Works

1. Data Loading: FAQ data (questions and answers) is added and converted into embeddings.
2. Embeddings: Questions are processed into numeric embeddings using DistilBERT.
3. Similarity Search: The cosine similarity between user query embeddings and FAQ embeddings determines the most relevant question.

---

Code Comments Explained

torch.no_grad()

This is a PyTorch context manager used to disable gradient calculations. It reduces memory usage and speeds up inference, as gradients are unnecessary for model evaluation.

outputs.last_hidden_state[:, 0, :]

• last_hidden_state: Outputs the embeddings for all tokens in the input sequence.
• [:, 0, :]: Extracts the embedding for the [CLS] token, which is often used as the representation of the entire input sequence.

cosine_similarity

Calculates the similarity between two embeddings. The result is a value between -1 (completely dissimilar) and 1 (identical).

Thresholding

The threshold parameter (default: 0.7) ensures the assistant only returns answers with a sufficient confidence level.

---

Sample Output

Answer: We accept credit cards, PayPal, and bank transfers. | Confidence: 0.87

---

Future Enhancements

• Support for dynamic data loading from files (e.g., CSV, JSON).
• Integration with web frameworks for API-based queries.
• Fine-tuning of DistilBERT on custom FAQ datasets.

---

Let us know if you need further clarifications! Happy Hacking!