[deep1blue]

Description

In RankGPTRerank (llama-index-core/llama_index/core/postprocessor/rankGPT_rerank.py):

1. _remove_duplicate checks if c not in new_response: against a list on every element ($O(K^2)$). Using list(dict.fromkeys(response)) preserves first-seen order in $O(K)$.
2. _receive_permutation appends unranked candidates via [tt for tt in range(rank_end) if tt not in response_list], performing an $O(K)$ list membership scan per candidate ($O(K^2)$ total). Precomputing seen = set(response_list) makes this $O(K)$.
3. _clean_response builds new_response via repeated += character concatenation in a loop instead of "".join(c if c.isdigit() else " " for c in response).strip().

On a 2,000-hit permutation (20 calls), _receive_permutation drops from 610.49 ms to 38.13 ms (16.0x faster).

Type of Change

• Bug fix / performance improvement (non-breaking change)

How Has This Been Tested?

• Added unit test test_rankgpt_rerank_receive_permutation_duplicates_and_missing in llama-index-core/tests/postprocessor/test_rankgpt_rerank.py
• Ran pytest llama-index-core/tests/postprocessor/test_rankgpt_rerank.py (3 passed)