Does llama-index "remember" the last query? Why is response time faster for repeated or similar queries?

Question

I am using llama-index to retrieve information from legal documents. (model: llama3 run with Ollama)

My use case does not require maintaining a chat history, so I am using a standalone QueryEngine for each query.

Even if I don't use a chat history, I’ve observed a pattern where the model's response time is significantly shorter when a question similar to the previous one is asked. For example:

• Question 1: Response time - 40 seconds
• Question 2 (similar to Question 1): Response time - 3 seconds
• Question 3: Response time - 35 seconds
• Question 4: Response time - 42 seconds
• Question 5 (similar to Question 4): Response time - 5 seconds

By "similar," I mean the questions have the same meaning and expect the same answer, though they might be worded differently.

The pattern suggests that the model somehow "remembers" the previous query and uses this information to speed up the response for similar questions. However, this effect seems limited to the immediately preceding query (it does not apply to questions asked before that).

My Questions:

• Does llama-index cache or reuse results from the last query? If so, how is this behavior implemented?
• Is there a way to make response times consistently faster, as if the question has been asked before?

I would appreciate any insight into how llama-index handles query processing and if there are any optimizations I can apply to benefit from faster response times.