[ML] Introduce InferenceString wrapper object

server/src/main/java/org/elasticsearch/inference/ChunkInferenceInput.java

@@ -13,13 +13,20 @@
 
 import java.util.List;
 
-public record ChunkInferenceInput(String input, @Nullable ChunkingSettings chunkingSettings) {
+import static org.elasticsearch.inference.InferenceString.DataType.TEXT;
+
+public record ChunkInferenceInput(InferenceString input, @Nullable ChunkingSettings chunkingSettings) {
 
     public ChunkInferenceInput(String input) {
-        this(input, null);
+        this(new InferenceString(input, TEXT), null);
     }
 
-    public static List<String> inputs(List<ChunkInferenceInput> chunkInferenceInputs) {
+    public static List<InferenceString> inputs(List<ChunkInferenceInput> chunkInferenceInputs) {
         return chunkInferenceInputs.stream().map(ChunkInferenceInput::input).toList();
     }
+
+    public String inputText() {
+        assert input.isText();
+        return input.value();
+    }
 }

server/src/main/java/org/elasticsearch/inference/InferenceString.java

@@ -0,0 +1,63 @@
+/*
+ * Copyright Elasticsearch B.V. and/or licensed to Elasticsearch B.V. under one
+ * or more contributor license agreements. Licensed under the "Elastic License
+ * 2.0", the "GNU Affero General Public License v3.0 only", and the "Server Side
+ * Public License v 1"; you may not use this file except in compliance with, at
+ * your election, the "Elastic License 2.0", the "GNU Affero General Public
+ * License v3.0 only", or the "Server Side Public License, v 1".
+ */
+
+package org.elasticsearch.inference;
+
+import java.util.EnumSet;
+import java.util.List;
+import java.util.Objects;
+
+/**
+ * This class represents a String which may be raw text, or the String representation of some other data such as an image in base64
+ */
+public record InferenceString(String value, DataType dataType) {
+    /**
+     * Describes the type of data represented by an {@link InferenceString}
+     */
+    public enum DataType {
+        TEXT,
+        IMAGE_BASE64
+    }
+
+    private static final EnumSet<DataType> IMAGE_TYPES = EnumSet.of(DataType.IMAGE_BASE64);
+
+    /**
+     * Constructs an {@link InferenceString} with the given value and {@link DataType}
+     * @param value the String value
+     * @param dataType the type of data that the String represents
+     */
+    public InferenceString(String value, DataType dataType) {
+        this.value = Objects.requireNonNull(value);
+        this.dataType = Objects.requireNonNull(dataType);
+    }
+
+    public boolean isImage() {
+        return IMAGE_TYPES.contains(dataType);
+    }
+
+    public boolean isText() {
+        return DataType.TEXT.equals(dataType);
+    }
+
+    /**
+     * Converts a list of {@link InferenceString} to a list of {@link String}.
+     * This method should only be called in code paths that do not deal with multimodal inputs; where all inputs are guaranteed to be
+     * raw text, since it discards the {@link org.elasticsearch.inference.InferenceString.DataType} associated with each input.
+     *
+     * @param inferenceStrings The list of {@link InferenceString} to convert to a list of {@link String}
+     * @return a list of String inference inputs that do not contain any non-text inputs
+     */
+    public static List<String> toStringList(List<InferenceString> inferenceStrings) {
+        return inferenceStrings.stream().map(i -> {
+            assert i.isText() : "Non-text input passed to InferenceString.toStringList";
+            return i.value();
+        }).toList();
+    }
+
+}

x-pack/plugin/core/src/main/java/org/elasticsearch/xpack/core/inference/chunking/EmbeddingRequestChunker.java

@@ -15,6 +15,7 @@
 import org.elasticsearch.inference.ChunkingSettings;
 import org.elasticsearch.inference.ChunkingStrategy;
 import org.elasticsearch.inference.InferenceServiceResults;
+import org.elasticsearch.inference.InferenceString;
 import org.elasticsearch.rest.RestStatus;
 import org.elasticsearch.xpack.core.inference.chunking.Chunker.ChunkOffset;
 import org.elasticsearch.xpack.core.inference.results.ChunkedInferenceEmbedding;
@@ -36,22 +37,26 @@
  * chunks. Multiple inputs may be fit into a single batch or
  * a single large input that has been chunked may spread over
  * multiple batches.
- *
+ * <p>
  * The final aspect is to gather the responses from the batch
  * processing and map the results back to the original element
  * in the input list.
  */
 public class EmbeddingRequestChunker<E extends EmbeddingResults.Embedding<E>> {
 
     // Visible for testing
-    record Request(int inputIndex, int chunkIndex, ChunkOffset chunk, String input) {
-        public String chunkText() {
-            return input.substring(chunk.start(), chunk.end());
+    record Request(int inputIndex, int chunkIndex, ChunkOffset chunk, InferenceString input) {
+        public InferenceString chunkText() {
+            if (chunk.start() == 0 && chunk.end() == input.value().length()) {
+                return input;
+            } else {
+                return new InferenceString(input.value().substring(chunk.start(), chunk.end()), input.dataType());
+            }
         }
     }
 
     public record BatchRequest(List<Request> requests) {
-        public Supplier<List<String>> inputs() {
+        public Supplier<List<InferenceString>> inputs() {
             return () -> requests.stream().map(Request::chunkText).collect(Collectors.toList());
         }
     }
@@ -107,13 +112,21 @@ public EmbeddingRequestChunker(
 
         List<Request> allRequests = new ArrayList<>();
         for (int inputIndex = 0; inputIndex < inputs.size(); inputIndex++) {
-            ChunkingSettings chunkingSettings = inputs.get(inputIndex).chunkingSettings();
+            ChunkInferenceInput chunkInferenceInput = inputs.get(inputIndex);
+            ChunkingSettings chunkingSettings = chunkInferenceInput.chunkingSettings();
             if (chunkingSettings == null) {
                 chunkingSettings = defaultChunkingSettings;
             }
-            Chunker chunker = chunkers.getOrDefault(chunkingSettings.getChunkingStrategy(), defaultChunker);
-            String inputString = inputs.get(inputIndex).input();
-            List<ChunkOffset> chunks = chunker.chunk(inputString, chunkingSettings);
+            Chunker chunker;
+            if (chunkInferenceInput.input().isText()) {
+                chunker = chunkers.getOrDefault(chunkingSettings.getChunkingStrategy(), defaultChunker);
+            } else {
+                // Do not chunk non-text inputs
+                chunker = NoopChunker.INSTANCE;
+                chunkingSettings = NoneChunkingSettings.INSTANCE;
+            }
+            InferenceString inputString = chunkInferenceInput.input();
+            List<ChunkOffset> chunks = chunker.chunk(inputString.value(), chunkingSettings);
             int resultCount = Math.min(chunks.size(), MAX_CHUNKS);
             resultEmbeddings.add(new AtomicReferenceArray<>(resultCount));
             resultOffsetStarts.add(new ArrayList<>(resultCount));

x-pack/plugin/core/src/main/java/org/elasticsearch/xpack/core/inference/results/ChunkedInferenceEmbedding.java

@@ -8,6 +8,7 @@
 package org.elasticsearch.xpack.core.inference.results;
 
 import org.elasticsearch.inference.ChunkedInference;
+import org.elasticsearch.inference.InferenceString;
 import org.elasticsearch.xcontent.XContent;
 
 import java.io.IOException;
@@ -19,15 +20,18 @@
 
 public record ChunkedInferenceEmbedding(List<EmbeddingResults.Chunk> chunks) implements ChunkedInference {
 
-    public static List<ChunkedInference> listOf(List<String> inputs, SparseEmbeddingResults sparseEmbeddingResults) {
-        validateInputSizeAgainstEmbeddings(inputs, sparseEmbeddingResults.embeddings().size());
+    public static List<ChunkedInference> listOf(List<InferenceString> inputs, SparseEmbeddingResults sparseEmbeddingResults) {
+        validateInputSizeAgainstEmbeddings(inputs.size(), sparseEmbeddingResults.embeddings().size());
 
         var results = new ArrayList<ChunkedInference>(inputs.size());
         for (int i = 0; i < inputs.size(); i++) {
             results.add(
                 new ChunkedInferenceEmbedding(
                     List.of(
-                        new EmbeddingResults.Chunk(sparseEmbeddingResults.embeddings().get(i), new TextOffset(0, inputs.get(i).length()))
+                        new EmbeddingResults.Chunk(
+                            sparseEmbeddingResults.embeddings().get(i),
+                            new TextOffset(0, inputs.get(i).value().length())
+                        )
                     )
                 )
             );

x-pack/plugin/core/src/main/java/org/elasticsearch/xpack/core/inference/results/TextEmbeddingUtils.java

@@ -9,18 +9,16 @@
 
 import org.elasticsearch.common.Strings;
 
-import java.util.List;
-
 public class TextEmbeddingUtils {
 
     /**
      * Throws an exception if the number of elements in the input text list is different than the results in text embedding
      * response.
      */
-    public static void validateInputSizeAgainstEmbeddings(List<String> inputs, int embeddingSize) {
-        if (inputs.size() != embeddingSize) {
+    public static void validateInputSizeAgainstEmbeddings(int inputsSize, int embeddingSize) {
+        if (inputsSize != embeddingSize) {
             throw new IllegalArgumentException(
-                Strings.format("The number of inputs [%s] does not match the embeddings [%s]", inputs.size(), embeddingSize)
+                Strings.format("The number of inputs [%s] does not match the embeddings [%s]", inputsSize, embeddingSize)
             );
         }
     }