Which components can i best use to build a RAG-Chat that works for short inquires, after first answer was given?
I am currently using ContextChatEngine, which works fine for the first interaction, but if a user poses a follow up question, which is short and more related to the previous interaction as to the content in the vector store.

To make it understandable hopefully:

1. Vectorstore Content [".... Granny Smith Apples ... ", " ... Green Frog...",...]
2. Question: What colors can apples have?
3. Answer by ContextChat Engine: Apples can have red, green (like Granny Smith) or yellow color
4. Question: Give me some more green varieties?
5. Answer by ContextChat Engine: .... Green Frog ... <= So the similarity to green frog is closest to green frog and the answer would be something related to frogs

Hi! I want to parse longer articles from multiple sources like mail, pdf files etc., which sometimes exceed the token limit so the output gets truncated.

Details: The articles are surrounded with noise, like page information, header, footer etc. The noise is not easily removed, with e.g. BeautifulSoup or similar, because they are all different.
Status: For articles that fit into the token limit it already works fine and only the resulting article with a SEO friendly headline is produced. This is what should happen.
Issue: Longer texts... So i do not want to summarize and only pick some chunks, but the whole article without any words changed.

Question: Are there options for a multi prompt setup that e.g. each prompt takes some chunks and then the output is recombined? As far as i understand: https://python.langchain.com/docs/use_cases/question_answering/how_to/analyze_document is just summarising, which is not what i want. Has somebody a better idea on how to approach it?

What is the best retriever to be used for input that sometimes require exact matches (similarity search fails).
So the most basic and stupid idea i had was to let the llm generate keywords and search for these. Problem here would probably be, that keywords would be generated, that have a high similarity but exist in many documents