Tip: tune candidate_multiplier before changing anything else

Set candidate_multiplier to 3 (the default) or higher when you initialize LLMReranker — this controls how many keyword-retrieved candidates Claude Haiku evaluates before returning the final ranked list.

Why: The reranker can only surface a relevant document that keyword retrieval already found. A multiplier that is too low starves the LLM of candidates; raising it is the cheapest way to improve recall before touching queries or indexes.

Tradeoff: Each additional candidate adds latency within the 60-second timeout window. If you triple the multiplier on a large corpus, verify that your p95 response time stays within timeout (default 60.0 seconds), or pass a larger value explicitly:

reranker = LLMReranker(
    candidate_multiplier=5,
    timeout=90.0,
)

The rerank(query, hits) method returns the same RetrievalHit objects in a new order — it does not filter them — so every index from the keyword stage is preserved in the output.

Source files

• src/attune_rag/reranker.py

Tags: reranker, hybrid-retrieval, precision, claude, haiku