DocBot ⛵

LLMs suffer from issues such as Knowledge cut-offs and hallucinations. They are not exposed to private data and lag the capability of being a fact engine.

Retrieval-Augmented Generation (RAG) framework can solve these issues. This chatbot is built leveraging the ChatCohere Model and Embeddings API. It processes uploaded PDF documents by breaking them down into manageable chunks, creates a searchable vector store, and generates accurate answers to user queries. Docbot cashes the embeddings to avoid recomputation, can memorize the previous conversation, and has hybrid search capabilities that ensure high-quality, contextually relevant responses for an efficient and effective user experience. (DocBot will not hallucinate an answer if it's missing context.)

Here is the app flow:

Technical Overview

• PDF Processing: Utilizes PyPDF2 for extracting text from PDF documents.
• Text Chunking: Employs the RecursiveCharacterTextSplitter from LangChain for dividing the extracted text into manageable chunks.
• Embeddings: Utilizes SOTA "embed-english-v3.0" model from CohereEmbeddings to generate embeddings of context documents and user queries.
• Vector Store Creation: Uses FAISS (or Chroma) for creating a searchable vector store from text chunks.
• Answer Generation: Leverages ChatCohere from Cohere for generating answers to user queries using the context provided by the uploaded documents.
• Deployment: The web app has been deployed on Streamlit.

System Overview:

Advanced RAG Concepts Implemented

• Caching Embeddings: Embeddings can be stored or temporarily cached to avoid needing to recompute them using CacheBackedEmbeddings
• Conversational History: Remembers and considers the context of previous questions and answers enabling the system to have a coherent and contextual conversation.
• Hybrid vector Search: It has the advantage of doing keyword search as well as the advantage of doing a semantic lookup that we get from embeddings and a vector search. I have used FAISS for semantic search and BM25 Algorithm for keyword search to implementing Hybrid Search using langchainEnsembleRetriever.
• InMemory Caching: Caching happens for every user query. The response is generated for each user query if it does not match with previously requested queries.

Below is a screenshot of the App:

To Run

1. Clone this Repository
2. Create a Python Virtual Environment
3. Install requirements: pip install -r requirements.txt
4. Run the app: streamlit run src/streamlit_app.py
5. Open the 'External URL' in your browser and start Chatting ⛵