Imagine if you're on your own in a foreign country and you've been bitten by a stray dog, and you need to be treated ASAP. Could you imagine trying to read through a 99-page, fine-print document to find out if you're covered whilst in immense pain? No, you couldn't. Which is why you need that friend who can search for you.
You need something you can interact with and in a short few words, receive all the information you need, rapidly. Taking this into account, we propose a chatbot that is able to find the solution to your query.
PromptMyPaul is the chatbot on a website that runs on locally hosted Llama2 chat model, that's here to answer any query about the policy you've paid for and the next steps you should take to make your claim, in the blink of an eye. The interface is user-friendly, thus making it accessible to any level of computer expertise. With any policy-related question and need for speed, we know this is a website customers of Zurich Insurance will regularly be using and thoroughly appreciate.
We used Langchain python library along with locally hosted Llama 2 model for tokenization and chat completion. As we are using pdf documents with policy information, we are using PDFLoaders from langchain and tokenizing the content.
We are also using OpenAI for fetching data from SQL databases which can help Zurich Insurance to get and idea about how their policies are performing.
As we are deploying Llama2 model in our laptops, it performs really slow. We tried to run this on GPU in cloud but the quota limitations didn't permit it.
We could perform uploading of policy documents and ask questions from it and it performed with high accuracy. For the company's perspective, employee's of Zurich Insurance can query about the performances of the policies in simple languages like "How many policies were claimed last year?" and then the answer is fetched from SQL databases.
We learned the use of Langchain to build prompt engineering and to use local models to train from PDF data and answer questions from it.
We can give an option to choose between OpenAI and Llama2 models or to plug your own LLM provider. OpenAI for better accuracy or self hosted Llama2 for better privacy. The user can choose according to his preference.