> ## Documentation Index
> Fetch the complete documentation index at: https://retrievalcontextprotocol.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Rerank

> Rescore a candidate set against the query with a cross-encoder, ColBERT late interaction, or an LLM judge.

`rerank` is gated by **`rerank`**. It takes a query and a list of candidate
passages and returns a relevance score for each — the precise, expensive second
stage of a recall → rerank funnel.

<CodeGroup>
  ```json Request theme={null}
  {
    "jsonrpc": "2.0", "id": 1, "method": "rerank",
    "params": {
      "query": "efficient attention mechanisms",
      "passages": [
        "Performers use FAVOR+ to approximate softmax attention.",
        "The Eiffel Tower is in Paris.",
        "Linformer projects keys and values to a low-rank space."
      ],
      "method": "cross-encoder",
      "topN": 2
    }
  }
  ```

  ```json Response theme={null}
  { "jsonrpc": "2.0", "id": 1, "result": { "scores": [0.91, 0.02, 0.88], "order": [0, 2, 1] } }
  ```
</CodeGroup>

## Params

<ParamField path="query" type="string" required>
  The query to score passages against.
</ParamField>

<ParamField path="passages" type="string[]" required>
  The candidate passages to rescore.
</ParamField>

<ParamField path="method" type="&#x22;cross-encoder&#x22; | &#x22;colbert&#x22; | &#x22;llm&#x22;">
  The reranking strategy. **MUST** be one of the server's advertised
  `rerank.methods`.
</ParamField>

<ParamField path="topN" type="integer">
  Truncate to the best `topN` results (reflected in `order`).
</ParamField>

## Result

<ResponseField name="scores" type="float[]">
  `scores[i]` corresponds to `passages[i]`; higher is more relevant. Scores are
  server-defined and only comparable within this one response.
</ResponseField>

<ResponseField name="order" type="integer[]">
  Optional. The passage indices sorted best-first — convenient when `topN`
  truncates.
</ResponseField>

## Reranking methods

<CardGroup cols={3}>
  <Card title="cross-encoder" icon="link">
    Jointly encodes each (query, passage) pair for one relevance score. Most
    accurate, one forward pass per candidate.
  </Card>

  <Card title="colbert" icon="table-cells">
    Late interaction (MaxSim) over multi-vector representations — near
    cross-encoder accuracy at bi-encoder speed.
  </Card>

  <Card title="llm" icon="robot">
    An LLM judge scores or orders passages — flexible, slowest, best for small N.
  </Card>
</CardGroup>

<Tip>
  You rarely need `rerank` directly: [`retrieve`](/methods/retrieve) can rerank
  **inside** a single call via its `rerank` option when the server advertises
  `retrieve.rerankBuiltin`. Reach for standalone `rerank` when you run your own
  recall stage or rerank a fused federation candidate set.
</Tip>