Reranker reference

Use LLMReranker to re-order keyword-retrieved candidates by relevance using Claude as a judge. Any API error falls back to the original keyword-only order, so the reranker is safe to use in production without a circuit breaker.

Classes

Class	Description
`LLMReranker`	Re-ranks keyword retrieval candidates by relevance using Claude Haiku.

`LLMReranker`

from attune_rag.reranker import LLMReranker

Constructor

Parameter	Type	Default	Description
`model`	`str`	`'claude-haiku-4-5'`	Claude model used as the relevance judge.
`api_key`	`str	None`	`None`
`candidate_multiplier`	`int`	`3`	Multiplier applied to the requested result count to determine how many candidates to retrieve before re-ranking.
`timeout`	`float`	`60.0`	Request timeout in seconds for each Claude call.

Methods

Method	Parameters	Returns	Description
`rerank`	`query: str`, `hits: list[RetrievalHit]`	`list[RetrievalHit]`	Re-ranks `hits` against `query` and returns them in relevance order. Falls back to the original order if the API call fails.

Module constants

`_SYSTEM`

System prompt passed to Claude on every rerank call. Instructs the model to act as a relevance judge and return a JSON array of 0-based indices ranked from most to least relevant.

Constant	Type	Value (stem)
`_SYSTEM`	`str`	`'You are a relevance judge for the attune-ai developer workflow documentation system…'`

Full value:

You are a relevance judge for the attune-ai developer workflow documentation system.
Given a user query and a numbered list of candidate documents (path + summary),
return a JSON array of 0-based indices ranking candidates from most to least relevant.

Ranking guidance:
- Paths with "tool-" (e.g. concepts/tool-release-prep.md) are canonical attune workflow
  references. Paths with "skill-", "task-", or "use-" are quickstarts and task guides.
  For workflow-goal queries, rank tool-* docs above skill-*, task-*, and use-* docs.
- "Version bump", "changelog", "release", "publish", or "ship" → prefer tool-release-prep.md
  (full release workflow with health, security, and changelog checks).
- "CI pipeline failing", "tests failing", "fix tests" → prefer tool-fix-test.md or
  skill-fix-test.md over task-ci-cd-pipeline.md (a setup guide, not a fix tool).
- "Orchestrate", "coordinate", or "manage" documentation → prefer tool-doc-orchestrator.md.
- "Publish to PyPI" → prefer tool-release-prep.md over task-package-publishing.md.

Include every index exactly once. Return ONLY the JSON array — no explanation.

Source files

src/attune_rag/reranker.py