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

# Ranking and fairness

> Understand BM25 ranking, candidate selection, and multi-query interleaving.

# Ranking and fairness

Kestrel combines provider ordering with local content relevance while preserving
diversity across multiple queries.

## Candidate selection

Search results are round-robin merged before fetching. Taking a bounded prefix
therefore preserves query and provider diversity better than concatenating one
provider's complete list before the next.

The default fetch pool is:

```text theme={null}
3 × top-k
```

Override it with `--fetch-candidates` when you need a different recall and
latency tradeoff.

## BM25 ranking

Kestrel tokenizes the original query and extracted page content into lowercase
word tokens. BM25 scores each fetched result within its originating-query group.

Results with positive scores are sorted from highest to lowest. If an entire
query group has no positive BM25 scores, its original provider order is retained
instead of dropping the group.

## Multi-query interleaving

Ranked query groups are interleaved round-robin. This prevents a broad or
high-volume query from monopolizing the final results.

Because BM25 scores belong to different query-specific corpora, compare them
within a query group rather than treating them as globally calibrated values.

## Disable ranking

Keep the merged provider order with:

```bash theme={null}
kestrelsearch search "python typing" --no-rank
```

Ranking requires fetched page content. `--no-fetch` therefore bypasses both
extraction and BM25.