langchain_chat_with_your_data.md

LangChain - Chat with your data

Notes on the short course from DeepLearning.AI

General

• Document loaders transfer raw data into a standardized format
• You can use a speech-to-text model (e.g. OpenAI's whisper model) to load data from audi/video files
• Documents need to be split in order to be processed efficiently. Select a splitter and use create_documents() to create documents from a list of text and split_documents() to split - split_documents() accepts custom regex expressions to split
• Store splits converted to embeddings in vector store (i.e. Chroma to get started) for efficient retrieval
• Start with semantic similarity search and try maximum marginal relevance (MMR), SelfQuery (a.k.a. LLM aided retrieval) or compression
• The general retrieval workflow looks as follows:
  • User asks a question
  • Question is transformed to embedding, send to vector store and used to retrieve relevant splits
  • Relevant splits (system prompt) and original question (human prompt) and send to LLM
• Use RetrievalQA chain to ask questions about your documents and switch to ConversationalRetrievalChain and ConversationBufferMemory if you need to introduce memory