Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Original file line number Diff line number Diff line change
Expand Up @@ -22,11 +22,13 @@
import com.fasterxml.jackson.databind.ObjectMapper;
import com.openai.client.OpenAIClient;
import com.openai.client.okhttp.OpenAIOkHttpClient;
import com.openai.core.JsonSchemaLocalValidation;
import com.openai.core.JsonValue;
import com.openai.models.ChatModel;
import com.openai.models.FunctionDefinition;
import com.openai.models.FunctionParameters;
import com.openai.models.ReasoningEffort;
import com.openai.models.ResponseFormatJsonSchema;
import com.openai.models.chat.completions.ChatCompletion;
import com.openai.models.chat.completions.ChatCompletionCreateParams;
import com.openai.models.chat.completions.ChatCompletionFunctionTool;
Expand All @@ -43,6 +45,7 @@
import java.util.HashMap;
import java.util.List;
import java.util.Map;
import java.util.Set;

/**
* A chat model integration for the OpenAI Chat Completions service using the official Java SDK.
Expand Down Expand Up @@ -119,11 +122,53 @@ public OpenAICompletionsConnection(
this.client = builder.build();
}

// Models for which OpenAI documents json_schema strict Structured Outputs support.
// Source of truth: https://platform.openai.com/docs/guides/structured-outputs — json_schema is
// supported on the gpt-4o-mini and gpt-4o-2024-08-06 snapshots "and later"; gpt-4-turbo,
// earlier models, and gpt-3.5-turbo get JSON mode only.
//
// "and later" is temporal, not a name prefix: gpt-4o-2024-05-13 predates the cutoff and does
// NOT support Structured Outputs, so matching a bare "gpt-4o" prefix would misclassify it as
// capable and fail silently at the provider. Prefix matching is therefore used only for the
// gpt-4o-mini family, whose entire lifetime post-dates the cutoff; every other capable model is
// matched exactly. An unrecognized model reports not-capable and degrades to the prompt
// fallback rather than failing at the provider.
private static final String NATIVE_STRUCTURED_OUTPUT_FAMILY_PREFIX = "gpt-4o-mini";
private static final Set<String> NATIVE_STRUCTURED_OUTPUT_MODELS =
Set.of("gpt-4o", "gpt-4o-2024-08-06", "gpt-4o-2024-11-20");

@Override
protected boolean supportsNativeStructuredOutput(String effectiveModel) {
if (effectiveModel == null) {
return false;
}
return effectiveModel.startsWith(NATIVE_STRUCTURED_OUTPUT_FAMILY_PREFIX)
|| NATIVE_STRUCTURED_OUTPUT_MODELS.contains(effectiveModel);

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The capability predicate appears to be inaccurate in both directions.

Using startsWith("gpt-4o-mini") also classifies models such as gpt-4o-mini-audio-preview and gpt-4o-mini-realtime-preview as capable, although these variants do not support Structured Outputs. Conversely, models such as gpt-4.1, gpt-5, o3, and o4-mini do support Structured Outputs but currently return false.

Once #912 uses this predicate for AUTO, the false positives may produce provider errors, while the false negatives will unnecessarily fall back to prompting. Could we use stricter alias/snapshot matching, include the currently supported model families, and add tests covering both cases?

References:

}

@Override
public ChatMessage chat(
List<ChatMessage> messages, List<Tool> tools, Map<String, Object> modelParams) {
return doChat(messages, tools, modelParams, null);
}

@Override
public ChatMessage chat(
List<ChatMessage> messages,
List<Tool> tools,
Map<String, Object> modelParams,
Object outputSchema) {
return doChat(messages, tools, modelParams, outputSchema);
}

private ChatMessage doChat(
List<ChatMessage> messages,
List<Tool> tools,
Map<String, Object> modelParams,
Object outputSchema) {
try {
ChatCompletionCreateParams params = buildRequest(messages, tools, modelParams);
ChatCompletionCreateParams params =
buildRequest(messages, tools, modelParams, outputSchema);
ChatCompletion completion = client.chat().completions().create(params);
ChatMessage response =
OpenAIChatCompletionsUtils.convertFromOpenAIMessage(
Expand All @@ -150,8 +195,13 @@ public ChatMessage chat(
}
}

private ChatCompletionCreateParams buildRequest(
List<ChatMessage> messages, List<Tool> tools, Map<String, Object> rawModelParams) {
// Package-private so the request body (including the native response_format) can be asserted
// without issuing a live API call through the final OpenAI client.
ChatCompletionCreateParams buildRequest(
List<ChatMessage> messages,
List<Tool> tools,
Map<String, Object> rawModelParams,
Object outputSchema) {
Map<String, Object> modelParams =
rawModelParams != null ? new HashMap<>(rawModelParams) : new HashMap<>();

Expand All @@ -170,6 +220,13 @@ private ChatCompletionCreateParams buildRequest(
builder.tools(convertTools(tools, strictMode));
}

// Native structured output applies only for a POJO Class schema on a model the provider
// documents as capable; a RowTypeInfo (wrapped in OutputSchema) or an incapable model keeps
// the prompt-engineering fallback.
if (outputSchema instanceof Class && supportsNativeStructuredOutput(modelName)) {

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I agree that the final fix belongs in #912, where AUTO can be distinguished from explicit NATIVE. Could we add a clear TODO referencing #912 at this capability check, noting that NATIVE must bypass the connection-side fallback or fail explicitly? This should help prevent the dormant issue from being missed when the strategy wiring is implemented.

builder.responseFormat(toNativeResponseFormat((Class<?>) outputSchema));
}

Object temperature = modelParams.remove("temperature");
if (temperature instanceof Number) {
builder.temperature(((Number) temperature).doubleValue());
Expand Down Expand Up @@ -208,6 +265,26 @@ private ChatCompletionCreateParams buildRequest(
return builder.build();
}

// Derives the strict json_schema response format from a POJO class via the SDK's typed
// structured-output builder. The Kotlin-facade StructuredOutputsKt.responseFormatFromClass is
// not callable from Java, so the response format is extracted through the typed builder, which
// generates the same strict draft-2020-12 schema, and then reattached to the standard builder.
private static <T> ResponseFormatJsonSchema toNativeResponseFormat(Class<T> schemaClass) {
return ChatCompletionCreateParams.builder()
.model(ChatModel.of(""))
.addUserMessage("")
.responseFormat(schemaClass, JsonSchemaLocalValidation.NO)
.build()
.rawParams()
.responseFormat()
.orElseThrow(
() ->
new IllegalStateException(
"OpenAI SDK did not produce a response_format for schema "
+ schemaClass.getName()))
.asJsonSchema();
}

private List<ChatCompletionTool> convertTools(List<Tool> tools, boolean strictMode) {
List<ChatCompletionTool> openaiTools = new ArrayList<>(tools.size());
for (Tool tool : tools) {
Expand Down
Original file line number Diff line number Diff line change
@@ -0,0 +1,191 @@
/*
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you under the Apache License, Version 2.0 (the
* "License"); you may not use this file except in compliance
* with the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/

package org.apache.flink.agents.integrations.chatmodels.openai;

import com.openai.models.ResponseFormatJsonSchema;
import com.openai.models.chat.completions.ChatCompletionCreateParams;
import org.apache.flink.agents.api.chat.messages.ChatMessage;
import org.apache.flink.agents.api.chat.messages.MessageRole;
import org.apache.flink.agents.api.resource.ResourceContext;
import org.apache.flink.agents.api.resource.ResourceDescriptor;
import org.apache.flink.agents.api.tools.Tool;
import org.apache.flink.agents.api.tools.ToolMetadata;
import org.apache.flink.agents.api.tools.ToolParameters;
import org.apache.flink.agents.api.tools.ToolResponse;
import org.apache.flink.agents.api.tools.ToolType;
import org.junit.jupiter.api.DisplayName;
import org.junit.jupiter.api.Test;

import java.util.HashMap;
import java.util.List;
import java.util.Map;

import static org.assertj.core.api.Assertions.assertThat;

/**
* Unit tests for {@link OpenAICompletionsConnection}'s native structured-output behavior. These
* assert the built request body without a live API call by inspecting {@code buildRequest}, and
* exercise the model-dependent capability predicate directly.
*/
class OpenAICompletionsConnectionTest {

private static final ResourceContext NOOP = ResourceContext.fromGetResource((a, b) -> null);

/** A representative POJO output schema. */
public static class Person {
public String name;
public int age;
}

private static OpenAICompletionsConnection connection() {
ResourceDescriptor desc =
ResourceDescriptor.Builder.newBuilder(OpenAICompletionsConnection.class.getName())
.addInitialArgument("api_key", "test-key")
.addInitialArgument("model", "gpt-4o")
.build();
return new OpenAICompletionsConnection(desc, NOOP);
}

private static Map<String, Object> params(String model) {
Map<String, Object> params = new HashMap<>();
params.put("model", model);
return params;
}

private static List<ChatMessage> userMessage() {
return List.of(new ChatMessage(MessageRole.USER, "hi"));
}

@Test
@DisplayName("Native response_format json_schema strict applied for a POJO on a capable model")
void testNativeAppliedForPojoCapableModel() {
ChatCompletionCreateParams params =
connection().buildRequest(userMessage(), List.of(), params("gpt-4o"), Person.class);

assertThat(params.responseFormat()).isPresent();
ResponseFormatJsonSchema jsonSchema = params.responseFormat().get().asJsonSchema();
assertThat(jsonSchema.jsonSchema().strict()).contains(true);
}

@Test
@DisplayName("Native NOT applied for a POJO on an incapable model (prompt fallback)")
void testNativeNotAppliedForIncapableModel() {
ChatCompletionCreateParams params =
connection()
.buildRequest(
userMessage(), List.of(), params("gpt-3.5-turbo"), Person.class);

assertThat(params.responseFormat()).isEmpty();
}

@Test
@DisplayName("Native NOT applied for a pre-cutoff same-family gpt-4o snapshot")
void testNativeNotAppliedForPreCutoffSnapshot() {
// gpt-4o-2024-05-13 predates the Structured Outputs cutoff even though it shares the gpt-4o
// prefix; treating it as capable would fail silently at the provider.
ChatCompletionCreateParams params =
connection()
.buildRequest(
userMessage(),
List.of(),
params("gpt-4o-2024-05-13"),
Person.class);

assertThat(params.responseFormat()).isEmpty();
}

@Test
@DisplayName("Native NOT applied when no output schema is supplied")
void testNativeNotAppliedWhenSchemaNull() {
ChatCompletionCreateParams params =
connection().buildRequest(userMessage(), List.of(), params("gpt-4o"), null);

assertThat(params.responseFormat()).isEmpty();
}

@Test
@DisplayName("Native NOT applied for a non-POJO schema form (POJO-only scope)")
void testNativeNotAppliedForNonPojoSchema() {
// A RowTypeInfo schema arrives wrapped in OutputSchema (not a bare POJO Class), so it must
// not activate native structured output; any non-Class schema object exercises the same
// instanceof gate.
Object nonClassSchema = "row<name STRING>";

ChatCompletionCreateParams params =
connection()
.buildRequest(userMessage(), List.of(), params("gpt-4o"), nonClassSchema);

assertThat(params.responseFormat()).isEmpty();
}

@Test
@DisplayName("Native applied for a POJO even when tools are bound (no empty-tools gate)")
void testNativeAppliedEvenWhenToolsBound() {
ChatCompletionCreateParams params =
connection()
.buildRequest(
userMessage(),
List.of(new StubTool()),
params("gpt-4o"),
Person.class);

assertThat(params.responseFormat()).isPresent();
}

@Test
@DisplayName("Capability predicate accepts the documented capable models")
void testCapabilityPredicateAcceptsCapableModels() {
OpenAICompletionsConnection connection = connection();

assertThat(connection.supportsNativeStructuredOutput("gpt-4o")).isTrue();
assertThat(connection.supportsNativeStructuredOutput("gpt-4o-2024-08-06")).isTrue();
assertThat(connection.supportsNativeStructuredOutput("gpt-4o-2024-11-20")).isTrue();
assertThat(connection.supportsNativeStructuredOutput("gpt-4o-mini")).isTrue();
assertThat(connection.supportsNativeStructuredOutput("gpt-4o-mini-2024-07-18")).isTrue();
}

@Test
@DisplayName("Capability predicate rejects incapable, pre-cutoff, unknown, and null models")
void testCapabilityPredicateRejectsIncapableModels() {
OpenAICompletionsConnection connection = connection();

assertThat(connection.supportsNativeStructuredOutput("gpt-3.5-turbo")).isFalse();
assertThat(connection.supportsNativeStructuredOutput("gpt-4-turbo")).isFalse();
assertThat(connection.supportsNativeStructuredOutput("gpt-4o-2024-05-13")).isFalse();
assertThat(connection.supportsNativeStructuredOutput("some-unknown-model")).isFalse();
assertThat(connection.supportsNativeStructuredOutput(null)).isFalse();
}

/** Minimal tool stub; only its presence in the tools list matters. */
private static class StubTool extends Tool {
StubTool() {
super(new ToolMetadata("add", "adds", "{\"type\":\"object\"}"));
}

@Override
public ToolType getToolType() {
return ToolType.FUNCTION;
}

@Override
public ToolResponse call(ToolParameters parameters) {
return ToolResponse.success(null);
}
}
}
Loading
Loading