@arizeai/openinference-instrumentation-openai 4.0.6 → 4.1.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@arizeai/openinference-instrumentation-openai",
-  "version": "4.0.6",
+  "version": "4.1.0",
   "private": false,
   "description": "OpenInference instrumentation for OpenAI",
   "keywords": [],
@@ -28,8 +28,8 @@
     "@opentelemetry/api": "^1.9.0",
     "@opentelemetry/core": "^1.25.1",
     "@opentelemetry/instrumentation": "^0.46.0",
-    "@arizeai/openinference-core": "2.0.6",
-    "@arizeai/openinference-semantic-conventions": "2.2.0"
+    "@arizeai/openinference-core": "2.0.8",
+    "@arizeai/openinference-semantic-conventions": "2.3.0"
   },
   "devDependencies": {
     "@opentelemetry/exporter-trace-otlp-proto": "^0.50.0",

package/src/instrumentation.ts

@@ -51,6 +51,45 @@ const MODULE_NAME = "openai";
 
 const INSTRUMENTATION_NAME = "@arizeai/openinference-instrumentation-openai";
 
+/**
+ * Maps hostname suffixes to their corresponding LLM provider value.
+ */
+export const HOST_SUFFIX_TO_PROVIDER: Record<string, LLMProvider> = {
+  "api.openai.com": LLMProvider.OPENAI,
+  "openai.azure.com": LLMProvider.AZURE,
+  "services.ai.azure.com": LLMProvider.AZURE,
+  "cognitiveservices.azure.com": LLMProvider.AZURE,
+  "api.anthropic.com": LLMProvider.ANTHROPIC,
+  "api.cohere.com": LLMProvider.COHERE,
+  "api.cohere.ai": LLMProvider.COHERE,
+  "api.mistral.ai": LLMProvider.MISTRALAI,
+  "generativelanguage.googleapis.com": LLMProvider.GOOGLE,
+  "aiplatform.googleapis.com": LLMProvider.GOOGLE,
+  "amazonaws.com": LLMProvider.AWS,
+  "api.x.ai": LLMProvider.XAI,
+  "api.deepseek.com": LLMProvider.DEEPSEEK,
+  "api.groq.com": LLMProvider.GROQ,
+  "api.fireworks.ai": LLMProvider.FIREWORKS,
+  "api.moonshot.cn": LLMProvider.MOONSHOT,
+  "api.cerebras.ai": LLMProvider.CEREBRAS,
+  "api.perplexity.ai": LLMProvider.PERPLEXITY,
+  "api.together.ai": LLMProvider.TOGETHER,
+  "api.together.xyz": LLMProvider.TOGETHER,
+};
+
+/**
+ * Return the LLM provider name for the given API hostname.
+ */
+export function getProviderFromHost(host: string): LLMProvider | undefined {
+  const normalised = host.toLowerCase().trim();
+  for (const [suffix, provider] of Object.entries(HOST_SUFFIX_TO_PROVIDER)) {
+    if (normalised.endsWith(suffix)) {
+      return provider;
+    }
+  }
+  return undefined;
+}
+
 /**
  * Flag to check if the openai module has been patched
  * Note: This is a fallback in case the module is made immutable (e.x. Deno, webpack, etc.)
@@ -84,9 +123,9 @@ function getExecContext(span: Span) {
  * Gets the appropriate LLM provider based on the OpenAI client instance
  * Follows the same logic as the Python implementation by checking the baseURL host
  * @param clientInstance The OpenAI client instance
- * @returns LLMProvider.AZURE for Azure OpenAI, LLMProvider.OPENAI for regular OpenAI
+ * @returns LLM provider based on the API hostname
  */
-function getLLMProvider(clientInstance: unknown): LLMProvider {
+function getLLMProvider(clientInstance: unknown): LLMProvider | undefined {
   try {
     // The clientInstance might be a sub-object (like Completions) that has a _client property
     // pointing to the actual OpenAI/AzureOpenAI client
@@ -124,24 +163,13 @@ function getLLMProvider(clientInstance: unknown): LLMProvider {
     }
 
     if (host && typeof host === "string") {
-      // Follow the same pattern as Python implementation
-      if (host.includes("api.openai.com")) {
-        return LLMProvider.OPENAI;
-      } else if (host.includes("openai.azure.com")) {
-        return LLMProvider.AZURE;
-      } else if (host.includes("api.microsoft.com")) {
-        // Additional Azure endpoint pattern
-        return LLMProvider.AZURE;
-      }
+      return getProviderFromHost(host);
     }
   } catch (error) {
-    // If we can't determine, default to regular OpenAI
     diag.debug("Failed to determine LLM provider from instance", error);
   }
-
-  // Default to OpenAI if we can't determine
-  return LLMProvider.OPENAI;
 }
+
 /**
  * An auto instrumentation class for OpenAI that creates {@link https://github.com/Arize-ai/openinference/blob/main/spec/semantic_conventions.md|OpenInference} Compliant spans for the OpenAI API
  * @param instrumentationConfig The config for the instrumentation @see {@link InstrumentationConfig}
@@ -294,6 +322,7 @@ export class OpenAIInstrumentation extends InstrumentationBase<typeof openai> {
                 // Override the model from the value sent by the server
                 [SemanticConventions.LLM_MODEL_NAME]: result.model,
                 ...getChatCompletionLLMOutputMessagesAttributes(result),
+                ...getChatCompletionFinishReasonAttributes(result),
                 ...getUsageAttributes(result),
               });
               span.setStatus({ code: SpanStatusCode.OK });
@@ -400,6 +429,7 @@ export class OpenAIInstrumentation extends InstrumentationBase<typeof openai> {
               [SemanticConventions.EMBEDDING_MODEL_NAME]: body.model,
               [SemanticConventions.INPUT_VALUE]: isStringInput ? input : JSON.stringify(input),
               [SemanticConventions.INPUT_MIME_TYPE]: isStringInput ? MimeType.TEXT : MimeType.JSON,
+              [SemanticConventions.LLM_SYSTEM]: LLMSystem.OPENAI,
               [SemanticConventions.LLM_PROVIDER]: getLLMProvider(this),
               ...getEmbeddingTextAttributes(body),
             },
@@ -762,6 +792,19 @@ function getChatCompletionLLMOutputMessagesAttributes(chatCompletion: ChatComple
   }, {} as Attributes);
 }
 
+/**
+ * Extracts the llm.finish_reason attribute from a chat completion response.
+ */
+function getChatCompletionFinishReasonAttributes(chatCompletion: ChatCompletion): Attributes {
+  const choice = chatCompletion.choices[0];
+  if (!choice?.finish_reason) {
+    return {};
+  }
+  return {
+    [SemanticConventions.LLM_FINISH_REASON]: choice.finish_reason,
+  };
+}
+
 function getChatCompletionOutputMessageAttributes(message: ChatCompletionMessage): Attributes {
   const role = message.role;
   const attributes: Attributes = {
@@ -856,6 +899,7 @@ async function consumeChatCompletionStreamChunks(stream: Stream<ChatCompletionCh
   // So the final tool and function calls need to be aggregated
   // across chunks
   const toolAndFunctionCallAttributes: Attributes = {};
+  let finishReason: string | undefined;
   // The first message is for the assistant response so we start at 1
   for await (const chunk of stream) {
     if (chunk.choices.length <= 0) {
@@ -865,6 +909,9 @@ async function consumeChatCompletionStreamChunks(stream: Stream<ChatCompletionCh
     if (choice.delta.content) {
       streamResponse += choice.delta.content;
     }
+    if (choice.finish_reason) {
+      finishReason = choice.finish_reason;
+    }
     // Accumulate the tool and function call attributes
     const toolAndFunctionCallAttributesDiff =
       getToolAndFunctionCallAttributesFromStreamChunk(chunk);
@@ -885,6 +932,9 @@ async function consumeChatCompletionStreamChunks(stream: Stream<ChatCompletionCh
     [`${messageIndexPrefix}${SemanticConventions.MESSAGE_CONTENT}`]: streamResponse,
     [`${messageIndexPrefix}${SemanticConventions.MESSAGE_ROLE}`]: "assistant",
   };
+  if (finishReason) {
+    attributes[SemanticConventions.LLM_FINISH_REASON] = finishReason;
+  }
   // Add the tool and function call attributes
   for (const [key, value] of Object.entries(toolAndFunctionCallAttributes)) {
     attributes[`${messageIndexPrefix}${key}`] = value;

package/src/version.ts

@@ -1,2 +1,2 @@
 // this is an auto-generated file. see js/scripts/version-update.js
-export const VERSION = "4.0.6";
+export const VERSION = "4.1.0";