- Input: Query
- Output: LLM chain of relevant documents
- extract "topic" from query to figure out which db to use
- extract "query-able" question from query to search for relevant texts within db
- this is because a question may be multi-layered and complex
- some parts might be inteded for the AI
- by extracting/ generating a string which is easily querable to the db, we can improve result accuracy
- using topic and query-able text, extract
k
documents from db - query relevant information into LLM
- sort incoming data into "bucket"/ new dbs of topics (to reduce cluter within topics)
- AI fn to chose which bucket to query
- UI
- Benchmark lightweight embedding models & test against OpenAI embedding
- Benchmark lightweight LLMs