Add rescore parameter documentation (#11467)

* Add rescore parameter documentation

Signed-off-by: Fanit Kolchina <[email protected]>

* Apply suggestions from code review

Signed-off-by: Nathan Bower <[email protected]>

---------

Signed-off-by: Fanit Kolchina <[email protected]>
Signed-off-by: Nathan Bower <[email protected]>
Co-authored-by: Nathan Bower <[email protected]>
(cherry picked from commit 6cfd636)
Signed-off-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

Author: github-actions[bot]

_query-dsl/rescore.md

@@ -0,0 +1,139 @@
+---
+layout: default
+title: Rescore
+nav_order: 99
+---
+
+# Rescore
+
+The `rescore` parameter improves search precision by reordering only the highest-ranked documents returned from your initial query. Rather than applying an expensive algorithm to all documents in the index, rescoring focuses computational resources on reranking a smaller window of top results using a secondary scoring method.
+
+When you include a `rescore` parameter in your search request, OpenSearch processes the results in the following sequence:
+
+1. **Initial search**: The primary query and any post-filters execute across all relevant documents.
+2. **Shard-level rescoring**: Each shard applies the rescoring algorithm to its top results.
+3. **Final coordination**: The coordinating node combines rescored results from all shards.
+
+This approach provides better relevance while maintaining acceptable performance.
+
+When using the `rescore` parameter, note the following important considerations:
+
+- You cannot use explicit sorting (other than `_score` in descending order) with rescoring. OpenSearch returns an error if you attempt to combine custom sorting with a rescore query.
+
+- When implementing pagination, maintain the same `window_size` across all pages. Changing the window size between pages can cause result inconsistencies as users navigate through search results.
+
+## Query rescoring
+
+Query rescoring applies a secondary query to refine the scores of top-ranked documents. You can control how many documents each shard examines using the `window_size` parameter (default is `10`).
+
+### Basic rescoring syntax
+
+```json
+POST /_search
+{
+  "query": {
+    "match": {
+      "content": {
+        "query": "OpenSearch query optimization",
+        "operator": "or"
+      }
+    }
+  },
+  "rescore": {
+    "window_size": 100,
+    "query": {
+      "rescore_query": {
+        "match_phrase": {
+          "content": {
+            "query": "OpenSearch query optimization",
+            "slop": 1
+          }
+        }
+      },
+      "query_weight": 0.8,
+      "rescore_query_weight": 1.3
+    }
+  }
+}
+```
+{% include copy-curl.html %}
+
+### Rescore parameters
+
+The `rescore` object supports the following parameters.
+
+Parameter | Type | Description
+--- | --- | ---
+`window_size` | Integer | Number of top documents to rescore per shard. Default is `10`.
+`query_weight` | Float | Weight applied to the original query score. Default is `1.0`.
+`rescore_query_weight` | Float | Weight applied to the rescore query score. Default is `1.0`.
+`score_mode` | String | Method for combining original and rescore query scores. Default is `total`. See [Score combination modes](#score-combination-modes).
+
+### Score combination modes
+
+The `score_mode` parameter determines how OpenSearch combines the original score with the rescore query score. This parameter accepts the following values.
+
+Mode | Description | Use case
+--- | --- | ---
+`total` | Adds original score + rescore score | General relevance improvement (default)
+`multiply` | Multiplies original score × rescore score | Effective with function queries that return values between 0 and 1
+`avg` | Averages the two scores | Balanced approach when both scores are equally important
+`max` | Uses the higher of the two scores | Ensures documents with high scores in either query rank well
+`min` | Uses the lower of the two scores | Conservative approach that requires both queries to agree
+
+## Multiple rescoring stages
+
+You can chain multiple rescoring operations to apply increasingly sophisticated ranking algorithms:
+
+```json
+POST /_search
+{
+  "query": {
+    "match": {
+      "title": {
+        "query": "search engine technology",
+        "operator": "or"
+      }
+    }
+  },
+  "rescore": [
+    {
+      "window_size": 200,
+      "query": {
+        "rescore_query": {
+          "match_phrase": {
+            "title": {
+              "query": "search engine technology",
+              "slop": 2
+            }
+          }
+        },
+        "query_weight": 0.6,
+        "rescore_query_weight": 1.4
+      }
+    },
+    {
+      "window_size": 50,
+      "query": {
+        "score_mode": "multiply",
+        "rescore_query": {
+          "function_score": {
+            "script_score": {
+              "script": {
+                "source": "Math.log10(doc['popularity'].value + 2)"
+              }
+            }
+          }
+        }
+      }
+    }
+  ]
+}
+```
+{% include copy-curl.html %}
+
+In this multi-stage example:
+1. The **first rescoring stage** examines 200 documents per shard, applying phrase matching to improve relevance.
+2. The **second rescoring stage** takes the top 50 results from the first stage and applies popularity-based scoring using a logarithmic function.
+
+Each stage processes the results from the previous stage, creating a refinement pipeline where computationally expensive operations only operate on the most promising candidates.