Add inner hits support to semantic query

docs/changelog/111834.yaml

@@ -0,0 +1,5 @@
+pr: 111834
+summary: Add inner hits support to semantic query
+area: Search
+type: enhancement
+issues: []

docs/reference/query-dsl/semantic-query.asciidoc

@@ -25,7 +25,7 @@ GET my-index-000001/_search
   }
 }
 ------------------------------------------------------------
-// TEST[skip:TBD]
+// TEST[skip: Requires inference endpoints]
 
 
 [discrete]
@@ -40,9 +40,209 @@ The `semantic_text` field to perform the query on.
 (Required, string)
 The query text to be searched for on the field.
 
+`inner_hits`::
+(Optional, object)
+Retrieves the specific passages that match the query.
+See <<semantic-query-passage-ranking, Passage ranking with the `semantic` query>> for more information.
++
+.Properties of `inner_hits`
+[%collapsible%open]
+====
+`from`::
+(Optional, integer)
+The offset from the first matching passage to fetch.
+Used to paginate through the passages.
+Defaults to `0`.
+
+`size`::
+(Optional, integer)
+The maximum number of matching passages to return.
+Defaults to `3`.
+====
 
 Refer to <<semantic-search-semantic-text,this tutorial>> to learn more about semantic search using `semantic_text` and `semantic` query.
 
+[discrete]
+[[semantic-query-passage-ranking]]
+==== Passage ranking with the `semantic` query
+The `inner_hits` parameter can be used for _passage ranking_, which allows you to determine which passages in the document best match the query.
+For example, if you have a document that covers varying topics:
+
+[source,console]
+------------------------------------------------------------
+POST my-index/_doc/lake_tahoe
+{
+  "inference_field": [
+    "Lake Tahoe is the largest alpine lake in North America",
+    "When hiking in the area, please be on alert for bears"
+  ]
+}
+------------------------------------------------------------
+// TEST[skip: Requires inference endpoints]
+
+You can use passage ranking to find the passage that best matches your query:
+
+[source,console]
+------------------------------------------------------------
+GET my-index/_search
+{
+  "query": {
+    "semantic": {
+      "field": "inference_field",
+      "query": "mountain lake",
+      "inner_hits": { }
+    }
+  }
+}
+------------------------------------------------------------
+// TEST[skip: Requires inference endpoints]
+
+[source,console-result]
+------------------------------------------------------------
+{
+    "took": 67,
+    "timed_out": false,
+    "_shards": {
+        "total": 1,
+        "successful": 1,
+        "skipped": 0,
+        "failed": 0
+    },
+    "hits": {
+        "total": {
+            "value": 1,
+            "relation": "eq"
+        },
+        "max_score": 10.844536,
+        "hits": [
+            {
+                "_index": "my-index",
+                "_id": "lake_tahoe",
+                "_score": 10.844536,
+                "_source": {
+                    ...
+                },
+                "inner_hits": { <1>
+                    "inference_field": {
+                        "hits": {
+                            "total": {
+                                "value": 2,
+                                "relation": "eq"
+                            },
+                            "max_score": 10.844536,
+                            "hits": [
+                                {
+                                    "_index": "my-index",
+                                    "_id": "lake_tahoe",
+                                    "_nested": {
+                                        "field": "inference_field.inference.chunks",
+                                        "offset": 0
+                                    },
+                                    "_score": 10.844536,
+                                    "_source": {
+                                        "text": "Lake Tahoe is the largest alpine lake in North America"
+                                    }
+                                },
+                                {
+                                    "_index": "my-index",
+                                    "_id": "lake_tahoe",
+                                    "_nested": {
+                                        "field": "inference_field.inference.chunks",
+                                        "offset": 1
+                                    },
+                                    "_score": 3.2726858,
+                                    "_source": {
+                                        "text": "When hiking in the area, please be on alert for bears"
+                                    }
+                                }
+                            ]
+                        }
+                    }
+                }
+            }
+        ]
+    }
+}
+------------------------------------------------------------
+<1> Ranked passages will be returned using the <<inner-hits,`inner_hits` response format>>, with `<inner_hits_name>` set to the `semantic_text` field name.
+
+By default, the top three matching passages will be returned.
+You can use the `size` parameter to control the number of passages returned and the `from` parameter to page through the matching passages:
+
+[source,console]
+------------------------------------------------------------
+GET my-index/_search
+{
+  "query": {
+    "semantic": {
+      "field": "inference_field",
+      "query": "mountain lake",
+      "inner_hits": {
+        "from": 1,
+        "size": 1
+      }
+    }
+  }
+}
+------------------------------------------------------------
+// TEST[skip: Requires inference endpoints]
+
+[source,console-result]
+------------------------------------------------------------
+{
+    "took": 42,
+    "timed_out": false,
+    "_shards": {
+        "total": 1,
+        "successful": 1,
+        "skipped": 0,
+        "failed": 0
+    },
+    "hits": {
+        "total": {
+            "value": 1,
+            "relation": "eq"
+        },
+        "max_score": 10.844536,
+        "hits": [
+            {
+                "_index": "my-index",
+                "_id": "lake_tahoe",
+                "_score": 10.844536,
+                "_source": {
+                    ...
+                },
+                "inner_hits": {
+                    "inference_field": {
+                        "hits": {
+                            "total": {
+                                "value": 2,
+                                "relation": "eq"
+                            },
+                            "max_score": 10.844536,
+                            "hits": [
+                                {
+                                    "_index": "my-index",
+                                    "_id": "lake_tahoe",
+                                    "_nested": {
+                                        "field": "inference_field.inference.chunks",
+                                        "offset": 1
+                                    },
+                                    "_score": 3.2726858,
+                                    "_source": {
+                                        "text": "When hiking in the area, please be on alert for bears"
+                                    }
+                                }
+                            ]
+                        }
+                    }
+                }
+            }
+        ]
+    }
+}
+------------------------------------------------------------
+
 [discrete]
 [[hybrid-search-semantic]]
 ==== Hybrid search with the `semantic` query
@@ -79,7 +279,7 @@ POST my-index/_search
   }
 }
 ------------------------------------------------------------
-// TEST[skip:TBD]
+// TEST[skip: Requires inference endpoints]
 
 You can also use semantic_text as part of <<rrf,Reciprocal Rank Fusion>> to make ranking relevant results easier:
 
@@ -116,12 +316,12 @@ GET my-index/_search
   }
 }
 ------------------------------------------------------------
-// TEST[skip:TBD]
+// TEST[skip: Requires inference endpoints]
 
 
 [discrete]
 [[advanced-search]]
-=== Advanced search on `semantic_text` fields
+==== Advanced search on `semantic_text` fields
 
 The `semantic` query uses default settings for searching on `semantic_text` fields for ease of use.
 If you want to fine-tune a search on a `semantic_text` field, you need to know the task type used by the `inference_id` configured in `semantic_text`.
@@ -135,7 +335,7 @@ on a `semantic_text` field, it is not supported to use the `semantic_query` on a
 
 [discrete]
 [[search-sparse-inference]]
-==== Search with `sparse_embedding` inference
+===== Search with `sparse_embedding` inference
 
 When the {infer} endpoint uses a `sparse_embedding` model, you can use a <<query-dsl-sparse-vector-query,`sparse_vector` query>> on a <<semantic-text,`semantic_text`>> field in the following way:
 
@@ -157,14 +357,14 @@ GET test-index/_search
   }
 }
 ------------------------------------------------------------
-// TEST[skip:TBD]
+// TEST[skip: Requires inference endpoints]
 
 You can customize the `sparse_vector` query to include specific settings, like <<sparse-vector-query-with-pruning-config-and-rescore-example,pruning configuration>>.
 
 
 [discrete]
 [[search-text-inferece]]
-==== Search with `text_embedding` inference
+===== Search with `text_embedding` inference
 
 When the {infer} endpoint uses a `text_embedding` model, you can use a <<query-dsl-knn-query,`knn` query>> on a `semantic_text` field in the following way:
 
@@ -190,6 +390,6 @@ GET test-index/_search
   }
 }
 ------------------------------------------------------------
-// TEST[skip:TBD]
+// TEST[skip: Requires inference endpoints]
 
 You can customize the `knn` query to include specific settings, like `num_candidates` and `k`.

server/src/main/java/org/elasticsearch/TransportVersions.java

@@ -224,6 +224,7 @@ static TransportVersion def(int id) {
     public static final TransportVersion ML_TELEMETRY_MEMORY_ADDED = def(8_748_00_0);
     public static final TransportVersion ILM_ADD_SEARCHABLE_SNAPSHOT_TOTAL_SHARDS_PER_NODE = def(8_749_00_0);
     public static final TransportVersion SEMANTIC_TEXT_SEARCH_INFERENCE_ID = def(8_750_00_0);
+    public static final TransportVersion SEMANTIC_QUERY_INNER_HITS = def(8_751_00_0);
 
     /*
      * STOP! READ THIS FIRST! No, really,

server/src/main/java/org/elasticsearch/index/query/InnerHitBuilder.java

@@ -50,9 +50,9 @@ public final class InnerHitBuilder implements Writeable, ToXContentObject {
     public static final ParseField COLLAPSE_FIELD = new ParseField("collapse");
     public static final ParseField FIELD_FIELD = new ParseField("field");
 
+    public static final int DEFAULT_FROM = 0;
+    public static final int DEFAULT_SIZE = 3;
     private static final boolean DEFAULT_IGNORE_UNAMPPED = false;
-    private static final int DEFAULT_FROM = 0;
-    private static final int DEFAULT_SIZE = 3;
     private static final boolean DEFAULT_VERSION = false;
     private static final boolean DEFAULT_SEQ_NO_AND_PRIMARY_TERM = false;
     private static final boolean DEFAULT_EXPLAIN = false;

x-pack/plugin/inference/src/main/java/org/elasticsearch/xpack/inference/InferenceFeatures.java

@@ -10,6 +10,7 @@
 import org.elasticsearch.features.FeatureSpecification;
 import org.elasticsearch.features.NodeFeature;
 import org.elasticsearch.xpack.inference.mapper.SemanticTextFieldMapper;
+import org.elasticsearch.xpack.inference.queries.SemanticQueryBuilder;
 import org.elasticsearch.xpack.inference.rank.random.RandomRankRetrieverBuilder;
 import org.elasticsearch.xpack.inference.rank.textsimilarity.TextSimilarityRankRetrieverBuilder;
 
@@ -25,7 +26,8 @@ public Set<NodeFeature> getFeatures() {
         return Set.of(
             TextSimilarityRankRetrieverBuilder.TEXT_SIMILARITY_RERANKER_RETRIEVER_SUPPORTED,
             RandomRankRetrieverBuilder.RANDOM_RERANKER_RETRIEVER_SUPPORTED,
-            SemanticTextFieldMapper.SEMANTIC_TEXT_SEARCH_INFERENCE_ID
+            SemanticTextFieldMapper.SEMANTIC_TEXT_SEARCH_INFERENCE_ID,
+            SemanticQueryBuilder.SEMANTIC_TEXT_INNER_HITS
         );
     }
 

x-pack/plugin/inference/src/main/java/org/elasticsearch/xpack/inference/mapper/SemanticTextFieldMapper.java

@@ -40,6 +40,7 @@
 import org.elasticsearch.index.mapper.ValueFetcher;
 import org.elasticsearch.index.mapper.vectors.DenseVectorFieldMapper;
 import org.elasticsearch.index.mapper.vectors.SparseVectorFieldMapper;
+import org.elasticsearch.index.query.InnerHitBuilder;
 import org.elasticsearch.index.query.MatchNoneQueryBuilder;
 import org.elasticsearch.index.query.NestedQueryBuilder;
 import org.elasticsearch.index.query.QueryBuilder;
@@ -54,6 +55,7 @@
 import org.elasticsearch.xcontent.XContentParserConfiguration;
 import org.elasticsearch.xpack.core.ml.inference.results.MlTextEmbeddingResults;
 import org.elasticsearch.xpack.core.ml.inference.results.TextExpansionResults;
+import org.elasticsearch.xpack.inference.queries.SemanticQueryInnerHitBuilder;
 
 import java.io.IOException;
 import java.util.ArrayList;
@@ -468,7 +470,12 @@ public boolean fieldHasValue(FieldInfos fieldInfos) {
             return fieldInfos.fieldInfo(getEmbeddingsFieldName(name())) != null;
         }
 
-        public QueryBuilder semanticQuery(InferenceResults inferenceResults, float boost, String queryName) {
+        public QueryBuilder semanticQuery(
+            InferenceResults inferenceResults,
+            float boost,
+            String queryName,
+            SemanticQueryInnerHitBuilder semanticInnerHitBuilder
+        ) {
             String nestedFieldPath = getChunksFieldName(name());
             String inferenceResultsFieldName = getEmbeddingsFieldName(name());
             QueryBuilder childQueryBuilder;
@@ -524,7 +531,10 @@ public QueryBuilder semanticQuery(InferenceResults inferenceResults, float boost
                 };
             }
 
-            return new NestedQueryBuilder(nestedFieldPath, childQueryBuilder, ScoreMode.Max).boost(boost).queryName(queryName);
+            InnerHitBuilder innerHitBuilder = semanticInnerHitBuilder != null ? semanticInnerHitBuilder.toInnerHitBuilder() : null;
+            return new NestedQueryBuilder(nestedFieldPath, childQueryBuilder, ScoreMode.Max).boost(boost)
+                .queryName(queryName)
+                .innerHit(innerHitBuilder);
         }
 
         private String generateQueryInferenceResultsTypeMismatchMessage(InferenceResults inferenceResults, String expectedResultsType) {