The session-level cache is an interesting interaction with your short-query expansion. If an answer depends on the previous message, a key containing only the normalised current question can return an answer from an earlier conversational context.
I would add a paired case to the 21-case suite: ask the same short question after two different product references in one session and require the answer to follow the new reference. The key can then include the resolved standalone query and relevant corpus version. That checks whether consistency comes from repeatable grounding rather than simply replaying the first answer.
Ahmet Özel
AI Engineer. Computer Vision, RAG and LLM agents.
The session-level cache is an interesting interaction with your short-query expansion. If an answer depends on the previous message, a key containing only the normalised current question can return an answer from an earlier conversational context.
I would add a paired case to the 21-case suite: ask the same short question after two different product references in one session and require the answer to follow the new reference. The key can then include the resolved standalone query and relevant corpus version. That checks whether consistency comes from repeatable grounding rather than simply replaying the first answer.