should apply the top-k here

In this case query.limit <= 0 indicates the 'k' is 0/none, therefore no top-k is applied.

COMBINED_SUFFICIENCY_AND_REWRITE_PROMPT also contains mid-word hard wraps (e.g., orig\ninal, qu\nery). These look accidental and can materially affect the chain-of-query rewrite/sufficiency behavior. Suggest reflowing the prompt to remove mid-word newlines so instructions are unambiguous.

if isinstance(result.index, int): will never be true because result.index is a pandas Index object, so the type column mapping won’t run. If you want to map numeric category IDs to names, consider using pd.api.types.is_integer_dtype(result.index) (or attempt result.index.astype(int) in a try/except) before applying the mapping.

_do_rerank ignores query.limit when self._reranker is None (or when len(episodes) <= query.limit), returning all episodes sorted by created_at. This can violate the caller’s top_k/limit contract and potentially return a much larger payload than expected. Consider always truncating to query.limit when it’s > 0 (even without a reranker), and only then applying the final chronological sort if that ordering is required.

TOOL_SELECT_PROMPT appears to have accidental hard line breaks mid-word (e.g., ...{query}; d\no not..., findi\nng, for\nm). This will degrade the routing prompt quality and can lead to incorrect tool selection. Suggest reformatting the prompt to remove mid-word newlines/stray characters and keep sentences intact.

Tool selection currently checks if tool.agent_name in rsp, which is a substring match across the full model output. If the model ever returns extra text (e.g., repeats multiple tool names, includes examples, etc.), this can select the wrong tool nondeterministically. Consider strict parsing (e.g., take the first non-empty line, strip it, and require an exact match against known tool names; treat anything else as NONE/fallback).

SPLIT_QUERY_PROMPT contains mid-word line breaks (e.g., co\nntain, increase\n ,, app\nears) which look unintentional. These artifacts can confuse the model and reduce split quality. Suggest removing the hard wraps so the prompt text is contiguous and readable.

When building the final rerank query, param.query += "\n".join(sub_queries) appends without a leading newline, so the original query and the first sub-query can run together (e.g., "orig?sub1?\nsub2?"). Consider replacing the query with a clearly delimited combined string (or add an explicit leading newline) so reranking uses the intended text.