opensearch-project
diff --git a/‎_query-dsl/specialized/script-score.md‎
Lines changed: 90 additions & 0 deletions b/‎_query-dsl/specialized/script-score.md‎
Lines changed: 90 additions & 0 deletions
@@ -373,5 +373,95 @@ GET /demo_index_v1/_search
 ```
 {% include copy-curl.html %}
 
+### Late interaction score
+
+The `lateInteractionScore` function is a Painless script scoring function that calculates document relevance using token-level vector matching. It compares each query vector against all document vectors, finds the maximum similarity for each query vector, and sums these maximum scores to produce the final document score.
+
+**Example score calculation**:
+- Query vectors: `[[0.8, 0.1], [0.2, 0.9]]`
+- Document vectors: `[[0.7, 0.2], [0.1, 0.8], [0.3, 0.4]]`
+- Query vector 1 → finds best match among document vectors → score A
+- Query vector 2 → finds best match among document vectors → score B
+- Final score = A + B
+
+This approach enables fine-grained semantic matching between queries and documents, making it particularly effective for reranking search results.
+
+#### Index mapping requirements
+
+The vector field must be mapped as either an `object` (recommended) or `float` type.
+
+We recommend mapping the vector field as an `object` with `"enabled": false` because it stores raw vectors without parsing, improving performance:
+
+```json
+{
+  "mappings": {
+    "properties": {
+      "my_vector": {
+        "type": "object",
+        "enabled": false
+      }
+    }
+  }
+}
+```
+
+Alternatively, you can map the vector field as a `float`:
+
+```json
+{
+  "mappings": {
+    "properties": {
+      "my_vector": {
+        "type": "float"
+      }
+    }
+  }
+}
+```
+
+#### Example
+
+The following example demonstrates using the `lateInteractionScore` function with cosine similarity to measure vector similarity based on direction rather than distance:
+
+```json
+GET my_index/_search
+{
+  "query": {
+    "script_score": {
+      "query": { "match_all": {} },
+      "script": {
+        "source": "lateInteractionScore(params.query_vectors, 'my_vector', params._source, params.space_type)",
+        "params": {
+          "query_vectors": [[[1.0, 0.0]], [[0.0, 1.0]]],
+          "space_type": "cosinesimil"
+        }
+      }
+    }
+  }
+}
+```
+{% include copy-curl.html %}
+
+#### Parameters
+
+The `lateInteractionFunction` supports the following parameters.
+
+| Parameter | Data type | Required | Description |
+| :--- | :--- | :--- | :--- |
+| `query_vectors` | Array of arrays | Yes | Query vectors for similarity matching |
+| `vector_field` | String | Yes | Name of the document field containing vectors |
+| `doc` | Map | Yes | Document source (use `params._source`) |
+| `space_type` | String | No | Similarity metric. Default: `"l2"` |
+
+The `space_type` parameter determines how similarity is calculated and accepts the following valid values.
+
+| Space type | Description | Higher score means |
+| :--- | :--- | :--- |
+| `innerproduct` | Dot product | More similar vectors |
+| `cosinesimil` | Cosine similarity | More similar direction |
+| `l2` (default) | Euclidean distance | Closer vectors (inverted) |
+
+For a complete example, see [Reranking by a field using an externally hosted late interaction model]({{site.url}}{{site.baseurl}}/search-plugins/search-relevance/rerank-by-field-late-interaction/).
+
 If [`search.allow_expensive_queries`]({{site.url}}{{site.baseurl}}/query-dsl/index/#expensive-queries) is set to `false`, `script_score` queries are not executed.
 {: .important}