@arizeai/openinference-genai 0.3.1 → 0.3.2

package/src/attributes.ts

@@ -121,8 +121,21 @@ const ATTR_GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS = "gen_ai.usage.cache_read.input
 const ATTR_GEN_AI_USAGE_CACHE_CREATION_INPUT_TOKENS =
   "gen_ai.usage.cache_creation.input_tokens" as const;
 
+/**
+ * A reasoning (thinking) message part.
+ *
+ * Producers such as `@ai-sdk/otel` already emit `{ type: "reasoning", content: "..." }` parts in
+ * `gen_ai.input.messages` / `gen_ai.output.messages`, but the draft OTel GenAI message schemas the
+ * types in `__generated__` are derived from do not yet declare the part — hence the local type.
+ */
+interface ReasoningPart {
+  type: "reasoning";
+  content?: unknown;
+  [k: string]: unknown;
+}
+
 // Shared part parsing
-type AnyPart = GenAIInputMessagePart | GenAIOutputMessagePart;
+type AnyPart = GenAIInputMessagePart | GenAIOutputMessagePart | ReasoningPart;
 
 /**
  * Type guard for a GenAI chat message
@@ -241,6 +254,20 @@ const processMessageParts = ({
         }
         continue;
       }
+      case "reasoning": {
+        // MESSAGE_CONTENTS entry carrying the reasoning text itself, not the serialized part
+        const contentPrefix = `${msgPrefix}${SemanticConventions.MESSAGE_CONTENTS}.${contentIndex}.`;
+        set(attrs, `${contentPrefix}${SemanticConventions.MESSAGE_CONTENT_TYPE}`, "reasoning");
+        if (part.content != null) {
+          set(
+            attrs,
+            `${contentPrefix}${SemanticConventions.MESSAGE_CONTENT_TEXT}`,
+            toStringContent(part.content),
+          );
+        }
+        contentIndex += 1;
+        continue;
+      }
       case "tool_call": {
         const id = part.id ?? undefined;
         const name = part.name;
@@ -265,13 +292,14 @@ const processMessageParts = ({
         continue;
       }
       default: {
-        // Generic / unknown part type: capture as JSON text content
+        // Generic / unknown part type: capture as JSON text content. Only MESSAGE_CONTENTS is
+        // written — the flat MESSAGE_CONTENT is an alternative representation of the same message,
+        // so setting both duplicates the content in consumers that render each of them.
         const genericPart = part as GenericPart;
         const genericText = toStringContent(genericPart);
         const contentPrefix = `${msgPrefix}${SemanticConventions.MESSAGE_CONTENTS}.${contentIndex}.`;
         set(attrs, `${contentPrefix}${SemanticConventions.MESSAGE_CONTENT_TYPE}`, genericPart.type);
         set(attrs, `${contentPrefix}${SemanticConventions.MESSAGE_CONTENT_TEXT}`, genericText);
-        set(attrs, `${msgPrefix}${SemanticConventions.MESSAGE_CONTENT}`, genericText);
         contentIndex += 1;
       }
     }
@@ -519,13 +547,16 @@ export const mapSystemInstructions = (spanAttributes: Attributes): Attributes =>
   if (systemInstructions) {
     set(attrs, `${SemanticConventions.METADATA}.gen_ai.system_instructions`, systemInstructions);
 
-    const msgPrefix = `${SemanticConventions.LLM_INPUT_MESSAGES}.0.`;
-    set(attrs, `${msgPrefix}${SemanticConventions.MESSAGE_ROLE}`, "system");
-    processMessageParts({
-      attrs,
-      msgPrefix,
-      parts: getSystemInstructionParts(spanAttributes) ?? [],
-    });
+    // Only emit the synthetic system message when there are parts to put in it.
+    // mapInputMessages shifts its indexes by one on exactly this condition, so
+    // emitting here without parts would claim index 0 while the input messages
+    // still start at 0, overwriting the first message's role with "system".
+    const parts = getSystemInstructionParts(spanAttributes);
+    if (parts != null) {
+      const msgPrefix = `${SemanticConventions.LLM_INPUT_MESSAGES}.0.`;
+      set(attrs, `${msgPrefix}${SemanticConventions.MESSAGE_ROLE}`, "system");
+      processMessageParts({ attrs, msgPrefix, parts });
+    }
   }
   return attrs;
 };