A CLI chatbot that uses RAG architecture for improving and adapting LLM to specific context. It allows users to ask questions and get response directly from open-source LLMs(OpenAI, MistralAI etc.) or from the information on a website which is provided as context using the RAG architecture. This chatbot empowers users to retrieve information from a website that may not have been in the training dataset of open-source LLMs or get current information from a website that may have been updated after the open-source LLM was trained.

Asking the OpenAI GPT LLM about the Nigerian Immigration Service(Question: Who is the current comptroller general of Nigeria Immigration Service as at 2023?) by running the script chatbot simple will give some outdated result as shown below:

By using the CLI website chatbot which extends the OpenAI GPT LLM with RAG we get a more useful result. We can use this chatbot by running chatbot rag https://immigration.gov.ng/current-and-past-leaders-of-the-nis/ and asking the same question - Who is the current comptroller general of Nigeria Immigration Service as at 2023?. This gives a more current result as shown below:

The chatbot crawls the given URL to the provided depth(default = 0) and indexes the content of the website which it uses as context to respond to users' questions. Hence, the chatbot tries to generate responses from the provided website context. The more information within the context(pages crawled by the chatbot), the more likely the model will generate a response. However, the quality and accuracy of these responses are subject to the quality and accuracy of information on the website which serves as context for the chatbot under the RAG architecture.

• OpenAI - API key

• Link the CLI app to the OS by running npm link
• Install dependencies by running npm install
• create a copy of the .env.example file by running cp .env.example .env.local in terminal
• update the env variables with the required keys in the .env.local file created in the previous step.
• Run the simple version by running npm run chatbot
• See the available commands for the app by running chatbot --help
  • Start the chatbot by running chatbot simple
  • Start the chatbot with a website context by running chatbot rag <url>
  • For running the chatbot with context, you can configure the crawl depth by using the --depth option by running chatbot rag <url> --depth <number>. The default crawl depth is 0

NB:

• Be careful when changing the default crawl depth to protect the memory because the crawled website contents are indexed In-Memory. Hence, the higher the crawl depth the more the memory that will be needed to index the website data.
• I will recommend a max crawl depth of 2 to be on the safe side. However, you will need to take caution and take into consideration the memory of your machine.
• This project is intended for learning purposes where the crawl depth(by extension data volume) is expected to be small. If you wish to use in a production environment where the data volume is expected to be high, kindly use a third-party VectorDB service provider(e.g qdrant, pinecone,mongodb)

References Langchain documentation OpenAI API Embeddings RAG Article VectorDB Service

This project started out from RAG tutorial I facilitated. Watch the tutorial recording here