Update otel chat client / embedding generator for 1.29 draft of the spec

src/Libraries/Microsoft.Extensions.AI.Abstractions/ChatCompletion/ChatClientMetadata.cs

@@ -9,7 +9,10 @@ namespace Microsoft.Extensions.AI;
 public class ChatClientMetadata
 {
     /// <summary>Initializes a new instance of the <see cref="ChatClientMetadata"/> class.</summary>
-    /// <param name="providerName">The name of the chat completion provider, if applicable.</param>
+    /// <param name="providerName">
+    /// The name of the chat completion provider, if applicable. Where possible, this should map to the
+    /// appropriate name defined in the OpenTelemetry Semantic Conventions for Generative AI systems.
+    /// </param>
     /// <param name="providerUri">The URL for accessing the chat completion provider, if applicable.</param>
     /// <param name="modelId">The ID of the chat completion model used, if applicable.</param>
     public ChatClientMetadata(string? providerName = null, Uri? providerUri = null, string? modelId = null)
@@ -20,12 +23,19 @@ public ChatClientMetadata(string? providerName = null, Uri? providerUri = null,
     }
 
     /// <summary>Gets the name of the chat completion provider.</summary>
+    /// <remarks>
+    /// Where possible, this maps to the appropriate name defined in the
+    /// OpenTelemetry Semantic Conventions for Generative AI systems.
+    /// </remarks>
     public string? ProviderName { get; }
 
     /// <summary>Gets the URL for accessing the chat completion provider.</summary>
     public Uri? ProviderUri { get; }
 
     /// <summary>Gets the ID of the model used by this chat completion provider.</summary>
-    /// <remarks>This value can be null if either the name is unknown or there are multiple possible models associated with this instance.</remarks>
+    /// <remarks>
+    /// This value can be null if either the name is unknown or there are multiple possible models associated with this instance.
+    /// An individual request may override this value via <see cref="ChatOptions.ModelId"/>.
+    /// </remarks>
     public string? ModelId { get; }
 }

src/Libraries/Microsoft.Extensions.AI.Abstractions/ChatCompletion/StreamingChatCompletionUpdateExtensions.cs

@@ -136,17 +136,33 @@ private static void AddMessagesToCompletion(Dictionary<int, ChatMessage> message
     {
         if (messages.Count <= 1)
         {
+            // Add the single message if there is one.
             foreach (var entry in messages)
             {
                 AddMessage(completion, coalesceContent, entry);
             }
+
+            // In the vast majority case where there's only one choice, promote any additional properties
+            // from the single message to the chat completion, making them more discoverable and more similar
+            // to how they're typically surfaced from non-streaming services.
+            if (completion.Choices.Count == 1 &&
+                completion.Choices[0].AdditionalProperties is { } messageProps)
+            {
+                completion.Choices[0].AdditionalProperties = null;
+                completion.AdditionalProperties = messageProps;
+            }
         }
         else
         {
+            // Add all of the messages, sorted by choice index.
             foreach (var entry in messages.OrderBy(entry => entry.Key))
             {
                 AddMessage(completion, coalesceContent, entry);
             }
+
+            // If there are multiple choices, we don't promote additional properties from the individual messages.
+            // At a minimum, we'd want to know which choice the additional properties applied to, and if there were
+            // conflicting values across the choices, it would be unclear which one should be used.
         }
 
         static void AddMessage(ChatCompletion completion, bool coalesceContent, KeyValuePair<int, ChatMessage> entry)

src/Libraries/Microsoft.Extensions.AI.Abstractions/Embeddings/EmbeddingGeneratorMetadata.cs

@@ -9,7 +9,11 @@ namespace Microsoft.Extensions.AI;
 public class EmbeddingGeneratorMetadata
 {
     /// <summary>Initializes a new instance of the <see cref="EmbeddingGeneratorMetadata"/> class.</summary>
-    /// <param name="providerName">The name of the embedding generation provider, if applicable.</param>
+
+    /// <param name="providerName">
+    /// The name of the embedding generation provider, if applicable. Where possible, this should map to the
+    /// appropriate name defined in the OpenTelemetry Semantic Conventions for Generative AI systems.
+    /// </param>
     /// <param name="providerUri">The URL for accessing the embedding generation provider, if applicable.</param>
     /// <param name="modelId">The ID of the embedding generation model used, if applicable.</param>
     /// <param name="dimensions">The number of dimensions in vectors produced by this generator, if applicable.</param>
@@ -22,15 +26,26 @@ public EmbeddingGeneratorMetadata(string? providerName = null, Uri? providerUri
     }
 
     /// <summary>Gets the name of the embedding generation provider.</summary>
+    /// <remarks>
+    /// Where possible, this maps to the appropriate name defined in the
+    /// OpenTelemetry Semantic Conventions for Generative AI systems.
+    /// </remarks>
     public string? ProviderName { get; }
 
     /// <summary>Gets the URL for accessing the embedding generation provider.</summary>
     public Uri? ProviderUri { get; }
 
     /// <summary>Gets the ID of the model used by this embedding generation provider.</summary>
-    /// <remarks>This value can be null if either the name is unknown or there are multiple possible models associated with this instance.</remarks>
+    /// <remarks>
+    /// This value can be null if either the name is unknown or there are multiple possible models associated with this instance.
+    /// An individual request may override this value via <see cref="EmbeddingGenerationOptions.ModelId"/>.
+    /// </remarks>
     public string? ModelId { get; }
 
     /// <summary>Gets the number of dimensions in the embeddings produced by this instance.</summary>
+    /// <remarks>
+    /// This value can be null if either the number of dimensions is unknown or there are multiple possible lengths associated with this instance.
+    /// An individual request may override this value via <see cref="EmbeddingGenerationOptions.Dimensions"/>.
+    /// </remarks>
     public int? Dimensions { get; }
 }

src/Libraries/Microsoft.Extensions.AI.Abstractions/Microsoft.Extensions.AI.Abstractions.csproj

@@ -9,7 +9,7 @@
   <PropertyGroup>
     <Stage>preview</Stage>
     <SuppressFinalPackageVersion>true</SuppressFinalPackageVersion>
-    <MinCodeCoverage>83</MinCodeCoverage>
+    <MinCodeCoverage>84</MinCodeCoverage>
     <MinMutationScore>0</MinMutationScore>
   </PropertyGroup>
 

src/Libraries/Microsoft.Extensions.AI.AzureAIInference/Microsoft.Extensions.AI.AzureAIInference.csproj

@@ -9,7 +9,7 @@
   <PropertyGroup>
     <Stage>preview</Stage>
     <SuppressFinalPackageVersion>true</SuppressFinalPackageVersion>
-    <MinCodeCoverage>83</MinCodeCoverage>
+    <MinCodeCoverage>91</MinCodeCoverage>
     <MinMutationScore>0</MinMutationScore>
   </PropertyGroup>
 

src/Libraries/Microsoft.Extensions.AI/ChatCompletion/OpenTelemetryChatClient.cs

@@ -23,7 +23,7 @@ namespace Microsoft.Extensions.AI;
 
 /// <summary>Represents a delegating chat client that implements the OpenTelemetry Semantic Conventions for Generative AI systems.</summary>
 /// <remarks>
-/// The draft specification this follows is available at <see href="https://opentelemetry.io/docs/specs/semconv/gen-ai/" />.
+/// This class provides an implementation of the Semantic Conventions for Generative AI systems v1.29, defined at <see href="https://opentelemetry.io/docs/specs/semconv/gen-ai/" />.
 /// The specification is still experimental and subject to change; as such, the telemetry output by this client is also subject to change.
 /// </remarks>
 public sealed partial class OpenTelemetryChatClient : DelegatingChatClient
@@ -288,6 +288,19 @@ public override async IAsyncEnumerable<StreamingChatCompletionUpdate> CompleteSt
                         {
                             _ = activity.AddTag(OpenTelemetryConsts.GenAI.Request.PerProvider(_system, "seed"), seed);
                         }
+
+                        if (options.AdditionalProperties is { } props)
+                        {
+                            // Log all additional request options as per-provider tags. This is non-normative, but it covers cases where
+                            // there's a per-provider specification in a best-effort manner (e.g. gen_ai.openai.request.service_tier),
+                            // and more generally cases where there's additional useful information to be logged.
+                            foreach (KeyValuePair<string, object?> prop in props)
+                            {
+                                _ = activity.AddTag(
+                                    OpenTelemetryConsts.GenAI.Request.PerProvider(_system, JsonNamingPolicy.SnakeCaseLower.ConvertName(prop.Key)),
+                                    prop.Value);
+                            }
+                        }
                     }
                 }
             }
@@ -375,6 +388,22 @@ private void TraceCompletion(
                 {
                     _ = activity.AddTag(OpenTelemetryConsts.GenAI.Response.OutputTokens, outputTokens);
                 }
+
+                if (_system is not null)
+                {
+                    // Log all additional response properties as per-provider tags. This is non-normative, but it covers cases where
+                    // there's a per-provider specification in a best-effort manner (e.g. gen_ai.openai.response.system_fingerprint),
+                    // and more generally cases where there's additional useful information to be logged.
+                    if (completion.AdditionalProperties is { } props)
+                    {
+                        foreach (KeyValuePair<string, object?> prop in props)
+                        {
+                            _ = activity.AddTag(
+                                OpenTelemetryConsts.GenAI.Response.PerProvider(_system, JsonNamingPolicy.SnakeCaseLower.ConvertName(prop.Key)),
+                                prop.Value);
+                        }
+                    }
+                }
             }
         }
 

src/Libraries/Microsoft.Extensions.AI/Embeddings/OpenTelemetryEmbeddingGenerator.cs

@@ -6,6 +6,7 @@
 using System.Diagnostics;
 using System.Diagnostics.Metrics;
 using System.Linq;
+using System.Text.Json;
 using System.Threading;
 using System.Threading.Tasks;
 using Microsoft.Extensions.Logging;
@@ -15,8 +16,8 @@ namespace Microsoft.Extensions.AI;
 
 /// <summary>Represents a delegating embedding generator that implements the OpenTelemetry Semantic Conventions for Generative AI systems.</summary>
 /// <remarks>
-/// The draft specification this follows is available at <see href="https://opentelemetry.io/docs/specs/semconv/gen-ai/" />.
-/// The specification is still experimental and subject to change; as such, the telemetry output by this generator is also subject to change.
+/// This class provides an implementation of the Semantic Conventions for Generative AI systems v1.29, defined at <see href="https://opentelemetry.io/docs/specs/semconv/gen-ai/" />.
+/// The specification is still experimental and subject to change; as such, the telemetry output by this client is also subject to change.
 /// </remarks>
 /// <typeparam name="TInput">The type of input used to produce embeddings.</typeparam>
 /// <typeparam name="TEmbedding">The type of embedding generated.</typeparam>
@@ -29,6 +30,7 @@ public sealed class OpenTelemetryEmbeddingGenerator<TInput, TEmbedding> : Delega
     private readonly Histogram<int> _tokenUsageHistogram;
     private readonly Histogram<double> _operationDurationHistogram;
 
+    private readonly string? _system;
     private readonly string? _modelId;
     private readonly string? _modelProvider;
     private readonly string? _endpointAddress;
@@ -49,6 +51,7 @@ public OpenTelemetryEmbeddingGenerator(IEmbeddingGenerator<TInput, TEmbedding> i
         Debug.Assert(innerGenerator is not null, "Should have been validated by the base ctor.");
 
         EmbeddingGeneratorMetadata metadata = innerGenerator!.Metadata;
+        _system = metadata.ProviderName;
         _modelId = metadata.ModelId;
         _modelProvider = metadata.ProviderName;
         _endpointAddress = metadata.ProviderUri?.GetLeftPart(UriPartial.Path);
@@ -126,11 +129,11 @@ protected override void Dispose(bool disposing)
             string? modelId = options?.ModelId ?? _modelId;
 
             activity = _activitySource.StartActivity(
-                string.IsNullOrWhiteSpace(modelId) ? OpenTelemetryConsts.GenAI.Embed : $"{OpenTelemetryConsts.GenAI.Embed} {modelId}",
+                string.IsNullOrWhiteSpace(modelId) ? OpenTelemetryConsts.GenAI.Embeddings : $"{OpenTelemetryConsts.GenAI.Embeddings} {modelId}",
                 ActivityKind.Client,
                 default(ActivityContext),
                 [
-                    new(OpenTelemetryConsts.GenAI.Operation.Name, OpenTelemetryConsts.GenAI.Embed),
+                    new(OpenTelemetryConsts.GenAI.Operation.Name, OpenTelemetryConsts.GenAI.Embeddings),
                     new(OpenTelemetryConsts.GenAI.Request.Model, modelId),
                     new(OpenTelemetryConsts.GenAI.SystemName, _modelProvider),
                 ]);
@@ -148,6 +151,23 @@ protected override void Dispose(bool disposing)
                 {
                     _ = activity.AddTag(OpenTelemetryConsts.GenAI.Request.EmbeddingDimensions, dimensions);
                 }
+
+                if (options is not null &&
+                    _system is not null)
+                {
+                    // Log all additional request options as per-provider tags. This is non-normative, but it covers cases where
+                    // there's a per-provider specification in a best-effort manner (e.g. gen_ai.openai.request.service_tier),
+                    // and more generally cases where there's additional useful information to be logged.
+                    if (options.AdditionalProperties is { } props)
+                    {
+                        foreach (KeyValuePair<string, object?> prop in props)
+                        {
+                            _ = activity.AddTag(
+                                OpenTelemetryConsts.GenAI.Request.PerProvider(_system, JsonNamingPolicy.SnakeCaseLower.ConvertName(prop.Key)),
+                                prop.Value);
+                        }
+                    }
+                }
             }
         }
 
@@ -212,12 +232,26 @@ private void TraceCompletion(
             {
                 _ = activity.AddTag(OpenTelemetryConsts.GenAI.Response.Model, responseModelId);
             }
+
+            // Log all additional response properties as per-provider tags. This is non-normative, but it covers cases where
+            // there's a per-provider specification in a best-effort manner (e.g. gen_ai.openai.response.system_fingerprint),
+            // and more generally cases where there's additional useful information to be logged.
+            if (_system is not null &&
+                embeddings?.AdditionalProperties is { } props)
+            {
+                foreach (KeyValuePair<string, object?> prop in props)
+                {
+                    _ = activity.AddTag(
+                        OpenTelemetryConsts.GenAI.Response.PerProvider(_system, JsonNamingPolicy.SnakeCaseLower.ConvertName(prop.Key)),
+                        prop.Value);
+                }
+            }
         }
     }
 
     private void AddMetricTags(ref TagList tags, string? requestModelId, string? responseModelId)
     {
-        tags.Add(OpenTelemetryConsts.GenAI.Operation.Name, OpenTelemetryConsts.GenAI.Embed);
+        tags.Add(OpenTelemetryConsts.GenAI.Operation.Name, OpenTelemetryConsts.GenAI.Embeddings);
 
         if (requestModelId is not null)
         {

src/Libraries/Microsoft.Extensions.AI/Microsoft.Extensions.AI.csproj

@@ -11,7 +11,7 @@
   <PropertyGroup>
     <Stage>preview</Stage>
     <SuppressFinalPackageVersion>true</SuppressFinalPackageVersion>
-    <MinCodeCoverage>83</MinCodeCoverage>
+    <MinCodeCoverage>88</MinCodeCoverage>
     <MinMutationScore>0</MinMutationScore>
   </PropertyGroup>
 

src/Libraries/Microsoft.Extensions.AI/OpenTelemetryConsts.cs

@@ -30,7 +30,7 @@ public static class GenAI
         public const string SystemName = "gen_ai.system";
 
         public const string Chat = "chat";
-        public const string Embed = "embed";
+        public const string Embeddings = "embeddings";
 
         public static class Assistant
         {
@@ -81,6 +81,8 @@ public static class Response
             public const string InputTokens = "gen_ai.response.input_tokens";
             public const string Model = "gen_ai.response.model";
             public const string OutputTokens = "gen_ai.response.output_tokens";
+
+            public static string PerProvider(string providerName, string parameterName) => $"gen_ai.{providerName}.response.{parameterName}";
         }
 
         public static class System

test/Libraries/Microsoft.Extensions.AI.Abstractions.Tests/TestEmbeddingGenerator.cs

@@ -15,7 +15,7 @@ public TestEmbeddingGenerator()
         GetServiceCallback = DefaultGetServiceCallback;
     }
 
-    public EmbeddingGeneratorMetadata Metadata { get; } = new();
+    public EmbeddingGeneratorMetadata Metadata { get; set; } = new();
 
     public Func<IEnumerable<string>, EmbeddingGenerationOptions?, CancellationToken, Task<GeneratedEmbeddings<Embedding<float>>>>? GenerateAsyncCallback { get; set; }
 

test/Libraries/Microsoft.Extensions.AI.Tests/ChatCompletion/ChatClientBuilderTest.cs

@@ -78,6 +78,24 @@ public void DoesNotAllowFactoriesToReturnNull()
         Assert.Contains("entry at index 0", ex.Message);
     }
 
+    [Fact]
+    public void UsesEmptyServiceProviderWhenNoServicesProvided()
+    {
+        using var innerClient = new TestChatClient();
+        ChatClientBuilder builder = new(innerClient);
+        builder.Use((innerClient, serviceProvider) =>
+        {
+            Assert.Null(serviceProvider.GetService(typeof(object)));
+
+            var keyedServiceProvider = Assert.IsAssignableFrom<IKeyedServiceProvider>(serviceProvider);
+            Assert.Null(keyedServiceProvider.GetKeyedService(typeof(object), "key"));
+            Assert.Throws<InvalidOperationException>(() => keyedServiceProvider.GetRequiredKeyedService(typeof(object), "key"));
+
+            return innerClient;
+        });
+        builder.Build();
+    }
+
     private sealed class InnerClientCapturingChatClient(string name, IChatClient innerClient) : DelegatingChatClient(innerClient)
     {
 #pragma warning disable S3604 // False positive: Member initializer values should not be redundant