VerAIzon

• 

  Image_1

• 

  Image_2

Inspiration

As explained in the demo, Verizon Chatbot is not that good in answering some context based queries, and neither is ChatGPT because it lacks context which is only present inside the Verizon documentation.

What it does

It implements RAG (Retrieval Augmented Generation) accompanied with Mistral-7B for chatbot creation, specifically tailored for Verizon customer services.

How we built it

1. Dataset Data was gathered through the iterative extraction using the data_extraction.py file. The code intially begins with just 2 links - ['https://www.verizon.com/home/internet/','https://community.verizon.com/'] and then at run-time it scrapes all the links in these two web-pages and this process goes on, creating about a dataset of 1000 pages. Another source of data are the user-guides available at https://www.verizon.com/about/terms-conditions/user-guides.

2. Splitting LLMs cannot process the whole data and thus we need to split the data into small portions of texts. The Dataset was splitted into small chunks using recursive character splitting from langchain. Once we have the small chunks we can create its embeddings using all-MiniLM-L12-v2 model via huggingface embeddings.

3. Storage Once we have the embeddings corresponding to the textual chunks, we can store it in a vector database. For our purpose, we used the FAISS (Facebook AI similarity Search) vectorstore as it is highly effecient when calculating similarity over huge amount of data.

4. Retrieval For the retrieval part we take the user question as an input and find the most matching chunks stored in the vectorstore using the FAISS search that could potentially answer the user query.

5. Output Once we have our most matching chunks we can label them as context and pass them onto the custom prompt for our model. For out purpose we used the following prompt -

   [INST] You are a Verizon company's chatbot, Only use the following pieces of context to answer the user's question. If the answer is not present in context, just say that you don't know and display the following link "https://www.verizon.com/support/residential/contact-us/contactuslanding.htm", don't try to make up an answer.[/INST]
   
       Context: {context}
       Question: {question}
       answer: 
   

   Now this prompt is passed to the LLM in order to get the output. The LLM that we used is Mistral-7B (Current Open-Source State of the Art).

Challenges we ran into

The most challenging thing was to integrate Mistral-7B as very less information is available on the internet. Moreover, we are using model from scratch without any api-calls.

Accomplishments that we're proud of

We were able to create a chatbot which is powerful enough to answer user queries and performs better than Verizon current chatbot and ChatGPT (without context).

What we learned

We learned about RAG and LLMs and how to build products using it.

What's next for Untitled

Improve the chatbot by providing it with memory.

Built With

• faiss
• langchain
• llm
• mistral-7b
• python
• rag
• similarity
• transformers
• vectordb