cosine similarity, will it be better to pass content with HTML tags, helping to answer better for context when passed to LLM?
but the issue is, due to these tags the distance is large in vector space for user queries and chunks.
As a general rule, your stored embeddings should be a close to your search terms as possible. Typically with text this means stripping anything that might influence semantic meaning. Better models always help with this, but one should attempt to minimise all external factors where possible.