query updating logic clarification

[seohyun8825]

Hi, thanks for releasing the code and repo.

I'm trying to reproduce the behavior described in the paper, especially the parts about:

• updating the query (query update) and using the updated query for retrieval, and
• updating memory over time.

From what I can see in the current RetrieverAgent implementation:

• question is taken once from message["question"]
• if use_embedding_based_retrieval is enabled, embeddings are computed using that same question
• query_llm_for_page_retrieval is also called with the original question
• the LLM output is then used only to extract relevant_pages and a document_summary

I don't see any place where:

• the LLM rewrites/refines the query and that updated query is used for another retrieval step, or
• some kind of “memory” is stored and then updated/used in later calls.

So I wanted to check:

1. Is the query-update / memory-update logic implemented somewhere else in the repo (or another script/branch) that I might have missed?
2. If it's not in this code release, how did you implement:
   • query update (e.g., refining the query with the LLM and re-running retrieval), and
   • memory update
     in the experiments reported in the paper?
3. Do we need to add an extra loop or surrounding agent logic to reproduce the full behavior, or is there an existing component that already handles this?

Right now it looks like retrieval always uses the same original query, so I just want to make sure I'm not misunderstanding the intended pipeline.

Thanks in advance for the clarification!

[seohyun8825]

Also, I noticed you’re using the following parsing logic
pred_answer = iteration.get("answer", "")
In practice, outputs can often end up as an empty string (when there is no answer tag), and gpt sometimes mark this as correct. maybe this might slightly affect the evaluation