Keyword alone sucks for negation. Searching a set of patient documents for “Which of my patients has COPD?” to get a set of responses that states “COPD not indicated” will not endear you to the physician using your tool. Hybrid (keyword + semantic) is much more all-encompassing.
Forwarding the users query directly to the document store seems ridiculous to me. The whole point is for the LLM to interpret the context and issue multiple targeted queries based upon the interpretation(s) arrived at.
The LLM is the semantic part. FTS is the keyword part. This is the hybrid you're looking for.
You already have a model with incredibly powerful semantic understanding. Why do we need the document store to also be a smartass? The model can project multiple OR clauses into the search term based upon its interpretation of the context.
If you are using something like Lucene, queries are extremely fast and the maximum # of supported documents in one index far exceeds what AWS says they can support here.