# Currently working on... ## UI implementation for viewing sources in chat - [x] Extracting sources from RAG chain - [x] Adding sources to the UI (references drop down menu inspiration: https://github.com/shashankdeshpande/langchain-chatbot/tree/master) - [x] Add references in the chat with nested expanders - [x] Deploy to Ploomber (done: https://bot-de-continuonus.ploomberapp.io/) ## LLM with memory (Bot de Continuonus) ### Embeddings - [x] Research on how the embedding works in detail - [x] Make decision: one or two indexes - [x] Figure code to upsert new vector ### Summarize and upsert - [x] Look into summarization techniques (Abstractive vs. Extractive Summarization) - [x] Design and implement the chat summarization pipeline - [x] Extract and process chat data for summarization - [x] Populate and update the vectorstore with summarized chat data - [x] Deploy the summarization pipeline to be included in the RAG chain #### Questions to ask - How does the embeddings / vector DB work? - Do the past queries and the original data live in two different DBs? - Or do they live in same one with different metadata tag? If latter, how does the DB/ embeddings get updated after a new query comes in? - Another question: Does the app start to ask the user to elaborate on their thoughts / reactions to the answer? - Asking for clarification is a bit false because LLMs aren't capable of real understanding. - But asking the human user to elaborate on some aspects of their interpretation. - Maybe we should set the focus on whatever is interesting ― i.e. emotional reaction, or something else?
