[ML] Ignore unrecognized openai sse fields (#114715)

Azure / Llama sends back fields we do not expect - rewriting the parser
to better handle unknown fields (by dropping them).

Author: prwhelan

docs/changelog/114715.yaml

@@ -0,0 +1,5 @@
+pr: 114715
+summary: Ignore unrecognized openai sse fields
+area: Machine Learning
+type: bug
+issues: []

x-pack/plugin/inference/src/main/java/org/elasticsearch/xpack/inference/external/openai/OpenAiStreamingProcessor.java

@@ -110,8 +110,6 @@ public class OpenAiStreamingProcessor extends DelegatingProcessor<Deque<ServerSe
     private static final String CHOICES_FIELD = "choices";
     private static final String DELTA_FIELD = "delta";
     private static final String CONTENT_FIELD = "content";
-    private static final String FINISH_REASON_FIELD = "finish_reason";
-    private static final String STOP_MESSAGE = "stop";
     private static final String DONE_MESSAGE = "[done]";
 
     @Override
@@ -162,21 +160,27 @@ private Iterator<StreamingChatCompletionResults.Result> parse(XContentParserConf
                 ensureExpectedToken(XContentParser.Token.START_OBJECT, currentToken, parser);
 
                 currentToken = parser.nextToken();
-                if (currentToken == XContentParser.Token.END_OBJECT) {
-                    consumeUntilObjectEnd(parser); // end choices
-                    return ""; // stopped
-                }
 
-                if (currentToken == XContentParser.Token.FIELD_NAME && parser.currentName().equals(CONTENT_FIELD)) {
-                    parser.nextToken();
-                } else {
-                    positionParserAtTokenAfterField(parser, CONTENT_FIELD, FAILED_TO_FIND_FIELD_TEMPLATE);
+                // continue until the end of delta
+                while (currentToken != null && currentToken != XContentParser.Token.END_OBJECT) {
+                    if (currentToken == XContentParser.Token.START_OBJECT || currentToken == XContentParser.Token.START_ARRAY) {
+                        parser.skipChildren();
+                    }
+
+                    if (currentToken == XContentParser.Token.FIELD_NAME && parser.currentName().equals(CONTENT_FIELD)) {
+                        parser.nextToken();
+                        ensureExpectedToken(XContentParser.Token.VALUE_STRING, parser.currentToken(), parser);
+                        var content = parser.text();
+                        consumeUntilObjectEnd(parser); // end delta
+                        consumeUntilObjectEnd(parser); // end choices
+                        return content;
+                    }
+
+                    currentToken = parser.nextToken();
                 }
-                ensureExpectedToken(XContentParser.Token.VALUE_STRING, parser.currentToken(), parser);
-                var content = parser.text();
-                consumeUntilObjectEnd(parser); // end delta
+
                 consumeUntilObjectEnd(parser); // end choices
-                return content;
+                return ""; // stopped
             }).stream()
                 .filter(Objects::nonNull)
                 .filter(Predicate.not(String::isEmpty))

x-pack/plugin/inference/src/test/java/org/elasticsearch/xpack/inference/external/openai/OpenAiStreamingProcessorTests.java

@@ -149,6 +149,42 @@ public void testDoneMessageIsIgnored() throws Exception {
         verify(downstream, times(0)).onNext(any());
     }
 
+    public void testInitialLlamaResponseIsIgnored() throws Exception {
+        var item = new ArrayDeque<ServerSentEvent>();
+        item.offer(new ServerSentEvent(ServerSentEventField.DATA, """
+            {
+                "id":"12345",
+                "object":"chat.completion.chunk",
+                "created":123456789,
+                "model":"Llama-2-7b-chat",
+                "system_fingerprint": "123456789",
+                "choices":[
+                    {
+                        "index":0,
+                        "delta":{
+                            "role":"assistant"
+                        },
+                        "logprobs":null,
+                        "finish_reason":null
+                    }
+                ]
+            }
+            """));
+
+        var processor = new OpenAiStreamingProcessor();
+
+        Flow.Subscriber<ChunkedToXContent> downstream = mock();
+        processor.subscribe(downstream);
+
+        Flow.Subscription upstream = mock();
+        processor.onSubscribe(upstream);
+
+        processor.next(item);
+
+        verify(upstream, times(1)).request(1);
+        verify(downstream, times(0)).onNext(any());
+    }
+
     private String toJsonString(ChunkedToXContent chunkedToXContent) throws IOException {
         try (var builder = XContentFactory.jsonBuilder()) {
             chunkedToXContent.toXContentChunked(EMPTY_PARAMS).forEachRemaining(xContent -> {