The Knowledge Assistant is a web application designed to provide answers to user queries by leveraging a Retrieval-Augmented Generation (RAG) approach. This application utilizes various libraries and models to perform web searches, retrieve relevant documents, and generate responses based on the context provided.
Here are some examples of the application in action:
In-domain (without web-search):
Out-of-fomain (with web-search):
The following libraries are utilized in this project:
- Streamlit: For creating the web interface.
- Torch: For deep learning model operations.
- Transformers: For loading and using pre-trained models.
- Ollama: for inference LLM.
- DuckDuckGo Search: For performing web searches.
- LangGraph: For managing the workflow of the RAG process.
- LlamaIndex: For handling document embeddings and retrieval.
- Tiktoken: For tokenization of text inputs.
- SQLite: For storing chat history in a local database.
The application implements several key entities in the RAG framework:
- MyEmbeddings: A custom embedding class that generates embeddings for queries and documents using a pre-trained model.
- VectorStoreManager: Manages the vector store, loading data, and creating embeddings for documents.
- GraphState: Represents the state of the RAG workflow, including the user's question, retrieved documents, and generated responses.
- Workflow Graph: A directed graph that defines the sequence of operations (nodes) and their connections (edges) in the RAG process.
The application implements the CRAG (Conditional Retrieval-Augmented Generation) method, which enhances the response generation process by conditionally deciding whether to perform a web search or generate a response based on the retrieved documents' relevance. The workflow consists of the following steps:
- Retrieve: Fetch relevant documents from the vector store based on the user's query.
- Grade: Evaluate the relevance of the retrieved documents using a reranker.
- Web Search: If necessary, perform a web search to gather additional context.
- Generate: Synthesize a final response using the available context from both retrieved documents and web search results.