Handle empty input inference

docs/changelog/123763.yaml

@@ -1,5 +1,5 @@
 pr: 123763
-summary: Handle empty input inference
+summary: Skip semantic_text embedding generation when no content is provided.
 area: Relevance
 type: enhancement
 issues: []

x-pack/plugin/inference/src/main/java/org/elasticsearch/xpack/inference/action/filter/ShardBulkInferenceActionFilter.java

@@ -582,7 +582,7 @@ private Map<String, List<FieldInferenceRequest>> createFieldInferenceRequests(Bu
                         for (String v : values) {
                             if (v.isBlank()) {
                                 slot.addOrUpdateResponse(
-                                    new FieldInferenceResponse(field, sourceField, null, order++, 0, null, EMPTY_CHUNKED_INFERENCE)
+                                    new FieldInferenceResponse(field, sourceField, v, order++, 0, null, EMPTY_CHUNKED_INFERENCE)
                                 );
                             } else {
                                 fieldRequests.add(new FieldInferenceRequest(itemIndex, field, sourceField, v, order++, offsetAdjustment));

x-pack/plugin/inference/src/test/java/org/elasticsearch/xpack/inference/action/filter/ShardBulkInferenceActionFilterTests.java

@@ -333,7 +333,7 @@ public void testExplicitNull() throws Exception {
                 // item 3
                 assertNull(bulkShardRequest.items()[3].getPrimaryResponse());
                 actualRequest = getIndexRequestOrNull(bulkShardRequest.items()[3].request());
-                assertInferenceResults(useLegacyFormat, actualRequest, "obj.field1", EXPLICIT_NULL, 0);
+                assertInferenceResults(useLegacyFormat, actualRequest, "obj.field1", EXPLICIT_NULL, useLegacyFormat ? null : 0);
 
                 // item 4
                 assertNull(bulkShardRequest.items()[4].getPrimaryResponse());
@@ -381,18 +381,23 @@ public void testHandleEmptyInput() throws Exception {
         ActionFilterChain actionFilterChain = (task, action, request, listener) -> {
             try {
                 BulkShardRequest bulkShardRequest = (BulkShardRequest) request;
-                IndexRequest actualRequest = getIndexRequestOrNull(bulkShardRequest.items()[0].request());
+                assertNull(bulkShardRequest.getInferenceFieldMap());
+                assertThat(bulkShardRequest.items().length, equalTo(3));
 
                 // Create with Empty string
-                assertInferenceResults(useLegacyFormat, actualRequest, "semantic_text_field", useLegacyFormat ? EXPLICIT_NULL : "", 0);
+                assertNull(bulkShardRequest.items()[0].getPrimaryResponse());
+                IndexRequest actualRequest = getIndexRequestOrNull(bulkShardRequest.items()[0].request());
+                assertInferenceResults(useLegacyFormat, actualRequest, "semantic_text_field", "", 0);
 
                 // Create with whitespace only
+                assertNull(bulkShardRequest.items()[1].getPrimaryResponse());
                 actualRequest = getIndexRequestOrNull(bulkShardRequest.items()[1].request());
-                assertInferenceResults(useLegacyFormat, actualRequest, "semantic_text_field", useLegacyFormat ? EXPLICIT_NULL : " ", 0);
+                assertInferenceResults(useLegacyFormat, actualRequest, "semantic_text_field", " ", 0);
 
                 // Update with multiple Whitespaces
+                assertNull(bulkShardRequest.items()[2].getPrimaryResponse());
                 actualRequest = getIndexRequestOrNull(bulkShardRequest.items()[2].request());
-                assertInferenceResults(useLegacyFormat, actualRequest, "semantic_text_field", useLegacyFormat ? EXPLICIT_NULL : "  ", 0);
+                assertInferenceResults(useLegacyFormat, actualRequest, "semantic_text_field", "  ", 0);
             } finally {
                 chainExecuted.countDown();
             }
@@ -637,7 +642,7 @@ private static void assertInferenceResults(
         IndexRequest request,
         String fieldName,
         Object expectedOriginalValue,
-        int expectedChunkCount
+        Integer expectedChunkCount
     ) {
         final Map<String, Object> requestMap = request.sourceAsMap();
         if (useLegacyFormat) {
@@ -647,12 +652,11 @@ private static void assertInferenceResults(
             );
 
             List<Object> chunks = (List<Object>) XContentMapValues.extractValue(getChunksFieldName(fieldName), requestMap);
-            if (expectedChunkCount > 0) {
+            if (expectedChunkCount == null) {
+                assertNull(chunks);
+            } else {
                 assertNotNull(chunks);
                 assertThat(chunks.size(), equalTo(expectedChunkCount));
-            } else {
-                // If the expected chunk count is 0, we expect that no inference has been performed.
-                assertTrue(chunks == null || chunks.isEmpty());
             }
         } else {
             assertThat(XContentMapValues.extractValue(fieldName, requestMap, EXPLICIT_NULL), equalTo(expectedOriginalValue));

x-pack/plugin/inference/src/yamlRestTest/resources/rest-api-spec/test/inference/30_semantic_text_inference.yml

@@ -1010,7 +1010,7 @@ setup:
 "Empty semantic_text field skips embedding generation":
   - requires:
       cluster_features: "semantic_text.handle_empty_input"
-      reason: skips generating embeddings when semantic_text field is contains empty or whitespace only input
+      reason: Skips embedding generation when semantic_text is empty or contains only whitespace, effective from 8.19 and 9.1.0.
 
   - do:
       index:
@@ -1020,28 +1020,10 @@ setup:
           sparse_field: ""
         refresh: true
 
-  - do:
-      search:
-        index: test-index
-        body:
-          fields: [ _inference_fields ]
-          query:
-            match_all: { }
-
-  - match: { hits.total.value: 1 }
-  - match: { hits.hits.0._source.sparse_field: "" }
-  - not_exists: hits.hits.0._source._inference_fields
-
----
-"Whitespace-Only semantic_text field skips embedding generation":
-  - requires:
-      cluster_features: "semantic_text.handle_empty_input"
-      reason: skips generating embeddings when semantic_text field is contains empty or whitespace only input
-
   - do:
       index:
         index: test-index
-        id: doc_1
+        id: doc_2
         body:
           sparse_field: "   "
         refresh: true
@@ -1054,15 +1036,17 @@ setup:
           query:
             match_all: { }
 
-  - match: { hits.total.value: 1 }
-  - match: { hits.hits.0._source.sparse_field: "   " }
+  - match: { hits.total.value: 2 }
+  - match: { hits.hits.0._source.sparse_field: "" }
+  - match: { hits.hits.1._source.sparse_field: "   " }
   - not_exists: hits.hits.0._source._inference_fields
+  - not_exists: hits.hits.1._source._inference_fields
 
 ---
 "Reindexing with empty or whitespace semantic_text skips embedding generation":
   - requires:
       cluster_features: "semantic_text.handle_empty_input"
-      reason: skips generating embeddings when semantic_text field is contains empty or whitespace only input
+      reason: Skips embedding generation when semantic_text is empty or contains only whitespace, effective from 8.19 and 9.1.0.
 
   - do:
       index:
@@ -1118,7 +1102,7 @@ setup:
 "Empty Multi-Field skips embedding generation":
   - requires:
       cluster_features: "semantic_text.handle_empty_input"
-      reason: skips generating embeddings when semantic_text field is contains empty or whitespace only input
+      reason: Skips embedding generation when semantic_text is empty or contains only whitespace, effective from 8.19 and 9.1.0.
 
   - do:
       indices.create:
@@ -1164,7 +1148,7 @@ setup:
 "Multi chunks skips empty input embedding generation":
   - requires:
       cluster_features: "semantic_text.handle_empty_input"
-      reason: skips generating embeddings when semantic_text field is contains empty or whitespace only input
+      reason: Skips embedding generation when semantic_text is empty or contains only whitespace, effective from 8.19 and 9.1.0.
 
   - do:
       index:

x-pack/plugin/inference/src/yamlRestTest/resources/rest-api-spec/test/inference/30_semantic_text_inference_bwc.yml

@@ -675,3 +675,67 @@ setup:
 
   - match: { hits.total.value: 1 }
   - not_exists: hits.hits.0._source._inference_fields
+
+---
+"Empty semantic_text field skips embedding generation":
+  - requires:
+      cluster_features: "semantic_text.handle_empty_input"
+      reason: Skips embedding generation when semantic_text is empty or contains only whitespace, effective from 8.19 and 9.1.0.
+
+  - do:
+      index:
+        index: test-index
+        id: doc_1
+        body:
+          sparse_field: ""
+        refresh: true
+
+  - do:
+      index:
+        index: test-index
+        id: doc_2
+        body:
+          sparse_field: "  "
+        refresh: true
+
+  - do:
+      search:
+        index: test-index
+        body:
+          query:
+            match_all: { }
+
+  - match: { hits.total.value: 2 }
+  - match: { hits.hits.0._source.sparse_field.text: "" }
+  - length: { hits.hits.0._source.sparse_field.inference.chunks: 0 }
+  - match: { hits.hits.1._source.sparse_field.text: "  " }
+  - length: { hits.hits.0._source.sparse_field.inference.chunks: 0 }
+
+---
+"Multi chunks skips empty input embedding generation":
+  - requires:
+      cluster_features: "semantic_text.handle_empty_input"
+      reason: Skips embedding generation when semantic_text is empty or contains only whitespace, effective from 8.19 and 9.1.0.
+
+  - do:
+      index:
+        index: test-index
+        id: doc_1
+        body:
+          sparse_field: ["some test data", "    ", "now with chunks"]
+        refresh: true
+
+  - do:
+      search:
+        index: test-index
+        body:
+          query:
+            match_all: { }
+
+  - match: { hits.total.value: 1 }
+
+  - length: { hits.hits.0._source.sparse_field.inference.chunks: 2 }
+  - match: { hits.hits.0._source.sparse_field.inference.chunks.0.text: "some test data" }
+  - exists: hits.hits.0._source.sparse_field.inference.chunks.0.embeddings
+  - match: { hits.hits.0._source.sparse_field.inference.chunks.1.text: "now with chunks" }
+  - exists: hits.hits.0._source.sparse_field.inference.chunks.1.embeddings

x-pack/plugin/inference/src/yamlRestTest/resources/rest-api-spec/test/inference/90_semantic_text_highlighter.yml

@@ -295,7 +295,7 @@ setup:
 "Highlighting and multi chunks with empty input":
   - requires:
       cluster_features: "semantic_text.handle_empty_input"
-      reason: skips generating embeddings when semantic_text field is contains empty or whitespace only input
+      reason: Skips embedding generation when semantic_text is empty or contains only whitespace, effective from 8.19 and 9.1.0.
 
   - do:
       indices.create:
@@ -329,70 +329,10 @@ setup:
             fields:
               semantic_text_field:
                 type: "semantic"
-                number_of_fragments: 1
+                number_of_fragments: 3
 
   - match: { hits.total.value: 1 }
   - match: { hits.hits.0._id: "doc_1" }
-  - length: { hits.hits.0.highlight.semantic_text_field: 1 }
-  - match: { hits.hits.0.highlight.semantic_text_field.0: "now with chunks" }
-
-  - do:
-      search:
-        index: test-multi-chunk-index
-        body:
-          query:
-            semantic:
-              field: "semantic_text_field"
-              query: "test"
-          highlight:
-            fields:
-              semantic_text_field:
-                type: "semantic"
-                number_of_fragments: 2
-
-  - match: { hits.total.value: 1 }
-  - match: { hits.hits.0._id: "doc_1" }
-  - length: { hits.hits.0.highlight.semantic_text_field: 2}
+  - length: { hits.hits.0.highlight.semantic_text_field: 2 }
   - match: { hits.hits.0.highlight.semantic_text_field.0: "some test data" }
   - match: { hits.hits.0.highlight.semantic_text_field.1: "now with chunks" }
-
-  - do:
-      search:
-        index: test-multi-chunk-index
-        body:
-          query:
-            semantic:
-              field: "semantic_text_field"
-              query: "test"
-          highlight:
-            fields:
-              semantic_text_field:
-                type: "semantic"
-                order: "score"
-                number_of_fragments: 1
-
-  - match: { hits.total.value: 1 }
-  - match: { hits.hits.0._id: "doc_1" }
-  - length: { hits.hits.0.highlight.semantic_text_field: 1 }
-  - match: { hits.hits.0.highlight.semantic_text_field.0: "now with chunks" }
-
-  - do:
-      search:
-        index: test-multi-chunk-index
-        body:
-          query:
-            semantic:
-              field: "semantic_text_field"
-              query: "test"
-          highlight:
-            fields:
-              semantic_text_field:
-                type: "semantic"
-                order: "score"
-                number_of_fragments: 2
-
-  - match: { hits.total.value: 1 }
-  - match: { hits.hits.0._id: "doc_1" }
-  - length: { hits.hits.0.highlight.semantic_text_field: 2}
-  - match: { hits.hits.0.highlight.semantic_text_field.0: "now with chunks" }
-  - match: { hits.hits.0.highlight.semantic_text_field.1: "some test data" }

x-pack/plugin/inference/src/yamlRestTest/resources/rest-api-spec/test/inference/90_semantic_text_highlighter_bwc.yml

@@ -243,4 +243,44 @@ setup:
   - match:  { hits.hits.0.highlight.body.0: "You Know, for Search!" }
   - match:  { hits.hits.0.highlight.body.1: "ElasticSearch is an open source, distributed, RESTful, search engine which is built on top of Lucene internally and enjoys all the features it provides." }
 
+---
+"Highlighting and multi chunks with empty input":
+  - do:
+      indices.create:
+        index: test-multi-chunk-index
+        body:
+          settings:
+            index.mapping.semantic_text.use_legacy_format: true
+          mappings:
+            properties:
+              semantic_text_field:
+                type: semantic_text
+                inference_id: sparse-inference-id
 
+  - do:
+      index:
+        index: test-multi-chunk-index
+        id: doc_1
+        body:
+          semantic_text_field: ["some test data", "    ", "now with chunks"]
+        refresh: true
+
+  - do:
+      search:
+        index: test-multi-chunk-index
+        body:
+          query:
+            semantic:
+              field: "semantic_text_field"
+              query: "test"
+          highlight:
+            fields:
+              semantic_text_field:
+                type: "semantic"
+                number_of_fragments: 3
+
+  - match: { hits.total.value: 1 }
+  - match: { hits.hits.0._id: "doc_1" }
+  - length: { hits.hits.0.highlight.semantic_text_field: 2 }
+  - match: { hits.hits.0.highlight.semantic_text_field.0: "some test data" }
+  - match: { hits.hits.0.highlight.semantic_text_field.1: "now with chunks" }