[azopenai] Fixing some issues with incorrect/incomplete types in generation (#22119)

Fixes:
- ToolChoice was unmodeled.
- ResponseFormat for ChatCompletions wasn't settable using the swagger as we had it (it's an object, not a string)

Author: richardpark-msft

sdk/ai/azopenai/CHANGELOG.md

@@ -1,6 +1,6 @@
 # Release History
 
-## 0.4.0 (2023-12-07)
+## 0.4.0 (2023-12-11)
 
 Support for many of the features mentioned in OpenAI's November Dev Day and Microsoft's 2023 Ignite conference
 

sdk/ai/azopenai/assets.json

@@ -2,5 +2,5 @@
   "AssetsRepo": "Azure/azure-sdk-assets",
   "AssetsRepoPrefixPath": "go",
   "TagPrefix": "go/ai/azopenai",
-  "Tag": "go/ai/azopenai_9ed7d01267"
+  "Tag": "go/ai/azopenai_d4fd4783ec"
 }

sdk/ai/azopenai/autorest.md

@@ -4,8 +4,7 @@ These settings apply only when `--go` is specified on the command line.
 
 ``` yaml
 input-file:
-- https://github.com/Azure/azure-rest-api-specs/blob/d402f685809d6d08be9c0b45065cadd7d78ab870/specification/cognitiveservices/data-plane/AzureOpenAI/inference/preview/2023-12-01-preview/generated.json
-
+- https://github.com/Azure/azure-rest-api-specs/blob/3e0e2a93ddb3c9c44ff1baf4952baa24ca98e9db/specification/cognitiveservices/data-plane/AzureOpenAI/inference/preview/2023-12-01-preview/generated.json
 output-folder: ../azopenai
 clear-output-folder: false
 module: github.com/Azure/azure-sdk-for-go/sdk/ai/azopenai
@@ -98,6 +97,20 @@ directive:
     transform: return $.replace(/InternalOYDAuthTypeRename/g, "configType")
 ```
 
+`ChatCompletionsResponseFormat.Type`
+
+```yaml
+directive:
+  - from: swagger-document
+    where: $.definitions.ChatCompletionsResponseFormat
+    transform: $.properties.type["x-ms-client-name"] = "InternalChatCompletionsResponseFormat"
+  - from:
+    - models.go
+    - models_serde.go
+    where: $
+    transform: return $.replace(/InternalChatCompletionsResponseFormat/g, "respType")
+```
+
 ## Model -> DeploymentName
 
 ```yaml
@@ -571,3 +584,21 @@ directive:
 
       return $.replace(/(func \(c ChatCompletionsOptions\) MarshalJSON\(\).+?populate\(objectMap, "frequency_penalty", c.FrequencyPenalty\))/s, "$1\n" + populateLines)
 ```
+
+Fix ToolChoice discriminated union
+
+```yaml
+directive:
+  - from: swagger-document
+    where: $.definitions.ChatCompletionsOptions.properties
+    transform: $["tool_choice"]["x-ms-client-name"] = "ToolChoiceRenameMe"
+  - from: 
+    - models.go
+    - models_serde.go
+    where: $
+    transform: |
+      return $
+        .replace(/^\s+ToolChoiceRenameMe.+$/m, "ToolChoice *ChatCompletionsToolChoice")   // update the name _and_ type for the field
+        .replace(/ToolChoiceRenameMe/g, "ToolChoice")    // rename all other references
+        .replace(/populateAny\(objectMap, "tool_choice", c\.ToolChoice\)/, 'populate(objectMap, "tool_choice", c.ToolChoice)');   // treat field as typed so nil means omit.
+```

sdk/ai/azopenai/client_chat_completions_test.go

@@ -8,6 +8,7 @@ package azopenai_test
 
 import (
 	"context"
+	"encoding/json"
 	"errors"
 	"io"
 	"net/http"
@@ -262,3 +263,40 @@ func TestClient_OpenAI_GetChatCompletions_Vision(t *testing.T) {
 
 	t.Logf(*resp.Choices[0].Message.Content)
 }
+
+func TestGetChatCompletions_usingResponseFormatForJSON(t *testing.T) {
+	testFn := func(t *testing.T, chatClient *azopenai.Client, deploymentName string) {
+		body := azopenai.ChatCompletionsOptions{
+			DeploymentName: &deploymentName,
+			Messages: []azopenai.ChatRequestMessageClassification{
+				&azopenai.ChatRequestSystemMessage{Content: to.Ptr("You are a helpful assistant designed to output JSON.")},
+				&azopenai.ChatRequestUserMessage{
+					Content: azopenai.NewChatRequestUserMessageContent("List capital cities and their states"),
+				},
+			},
+			// Without this format directive you end up getting JSON, but with a non-JSON preamble, like this:
+			// "I'm happy to help! Here are some examples of capital cities and their corresponding states:\n\n```json\n{\n" (etc)
+			ResponseFormat: &azopenai.ChatCompletionsJSONResponseFormat{},
+			Temperature:    to.Ptr[float32](0.0),
+		}
+
+		resp, err := chatClient.GetChatCompletions(context.Background(), body, nil)
+		require.NoError(t, err)
+
+		// validate that it came back as JSON data
+		var v any
+		err = json.Unmarshal([]byte(*resp.Choices[0].Message.Content), &v)
+		require.NoError(t, err)
+		require.NotEmpty(t, v)
+	}
+
+	t.Run("OpenAI", func(t *testing.T) {
+		chatClient := newOpenAIClientForTest(t)
+		testFn(t, chatClient, "gpt-3.5-turbo-1106")
+	})
+
+	t.Run("AzureOpenAI", func(t *testing.T) {
+		chatClient := newTestClient(t, azureOpenAI.DallE.Endpoint)
+		testFn(t, chatClient, "gpt-4-1106-preview")
+	})
+}

sdk/ai/azopenai/client_functions_test.go

@@ -28,15 +28,46 @@ type ParamProperty struct {
 func TestGetChatCompletions_usingFunctions(t *testing.T) {
 	// https://platform.openai.com/docs/guides/gpt/function-calling
 
+	useSpecificTool := azopenai.NewChatCompletionsToolChoice(
+		azopenai.ChatCompletionsToolChoiceFunction{Name: "get_current_weather"},
+	)
+
 	t.Run("OpenAI", func(t *testing.T) {
 		chatClient := newOpenAIClientForTest(t)
-		testChatCompletionsFunctions(t, chatClient, openAI.ChatCompletions)
-		testChatCompletionsFunctions(t, chatClient, openAI.ChatCompletionsLegacyFunctions)
+
+		testData := []struct {
+			Model      string
+			ToolChoice *azopenai.ChatCompletionsToolChoice
+		}{
+			// all of these variants use the tool provided - auto just also works since we did provide
+			// a tool reference and ask a question to use it.
+			{Model: openAI.ChatCompletions, ToolChoice: nil},
+			{Model: openAI.ChatCompletions, ToolChoice: azopenai.ChatCompletionsToolChoiceAuto},
+			{Model: openAI.ChatCompletionsLegacyFunctions, ToolChoice: useSpecificTool},
+		}
+
+		for _, td := range testData {
+			testChatCompletionsFunctions(t, chatClient, td.Model, td.ToolChoice)
+		}
 	})
 
 	t.Run("AzureOpenAI", func(t *testing.T) {
 		chatClient := newAzureOpenAIClientForTest(t, azureOpenAI)
-		testChatCompletionsFunctions(t, chatClient, azureOpenAI.ChatCompletions)
+
+		testData := []struct {
+			Model      string
+			ToolChoice *azopenai.ChatCompletionsToolChoice
+		}{
+			// all of these variants use the tool provided - auto just also works since we did provide
+			// a tool reference and ask a question to use it.
+			{Model: azureOpenAI.ChatCompletions, ToolChoice: nil},
+			{Model: azureOpenAI.ChatCompletions, ToolChoice: azopenai.ChatCompletionsToolChoiceAuto},
+			{Model: azureOpenAI.ChatCompletions, ToolChoice: useSpecificTool},
+		}
+
+		for _, td := range testData {
+			testChatCompletionsFunctions(t, chatClient, td.Model, td.ToolChoice)
+		}
 	})
 }
 
@@ -120,7 +151,7 @@ func testChatCompletionsFunctionsOlderStyle(t *testing.T, client *azopenai.Clien
 	require.Equal(t, location{Location: "Boston, MA", Unit: "celsius"}, *funcParams)
 }
 
-func testChatCompletionsFunctions(t *testing.T, chatClient *azopenai.Client, deploymentName string) {
+func testChatCompletionsFunctions(t *testing.T, chatClient *azopenai.Client, deploymentName string, toolChoice *azopenai.ChatCompletionsToolChoice) {
 	body := azopenai.ChatCompletionsOptions{
 		DeploymentName: &deploymentName,
 		Messages: []azopenai.ChatRequestMessageClassification{
@@ -150,6 +181,7 @@ func testChatCompletionsFunctions(t *testing.T, chatClient *azopenai.Client, dep
 				},
 			},
 		},
+		ToolChoice:  toolChoice,
 		Temperature: to.Ptr[float32](0.0),
 	}
 

sdk/ai/azopenai/constants.go

@@ -132,3 +132,56 @@ func (e *Error) Error() string {
 
 	return *e.message
 }
+
+// ChatCompletionsToolChoice controls which tool is used for this ChatCompletions call.
+// You can choose between:
+// - [ChatCompletionsToolChoiceAuto] means the model can pick between generating a message or calling a function.
+// - [ChatCompletionsToolChoiceNone] means the model will not call a function and instead generates a message
+// - Use the [NewChatCompletionsToolChoice] function to specify a specific tool.
+type ChatCompletionsToolChoice struct {
+	value any
+}
+
+var (
+	// ChatCompletionsToolChoiceAuto means the model can pick between generating a message or calling a function.
+	ChatCompletionsToolChoiceAuto *ChatCompletionsToolChoice = &ChatCompletionsToolChoice{value: "auto"}
+
+	// ChatCompletionsToolChoiceNone means the model will not call a function and instead generates a message.
+	ChatCompletionsToolChoiceNone *ChatCompletionsToolChoice = &ChatCompletionsToolChoice{value: "none"}
+)
+
+// NewChatCompletionsToolChoice creates a ChatCompletionsToolChoice for a specific tool.
+func NewChatCompletionsToolChoice[T ChatCompletionsToolChoiceFunction](v T) *ChatCompletionsToolChoice {
+	return &ChatCompletionsToolChoice{value: v}
+}
+
+// ChatCompletionsToolChoiceFunction can be used to force the model to call a particular function.
+type ChatCompletionsToolChoiceFunction struct {
+	// Name is the name of the function to call.
+	Name string
+}
+
+// MarshalJSON implements the json.Marshaller interface for type ChatCompletionsToolChoiceFunction.
+func (tf ChatCompletionsToolChoiceFunction) MarshalJSON() ([]byte, error) {
+	type jsonInnerFunc struct {
+		Name string `json:"name"`
+	}
+
+	type jsonFormat struct {
+		Type     string        `json:"type"`
+		Function jsonInnerFunc `json:"function"`
+	}
+
+	return json.Marshal(jsonFormat{
+		Type: "function",
+		//nolint:gosimple,can't use the ChatCompletionsToolChoiceFunction here or marshalling will be circular!
+		Function: jsonInnerFunc{
+			Name: tf.Name,
+		},
+	})
+}
+
+// MarshalJSON implements the json.Marshaller interface for type ChatCompletionsToolChoice.
+func (tc ChatCompletionsToolChoice) MarshalJSON() ([]byte, error) {
+	return json.Marshal(tc.value)
+}