feat(langchain): emit GenAI request attributes

CHANGELOG.md

@@ -2,6 +2,9 @@
 
 ## [Unreleased]
 
+### Features Added
+- Emit `gen_ai.output.type` and available GenAI request parameters on LangChain chat spans. [#213](https://github.com/microsoft/opentelemetry-distro-javascript/pull/213)
+
 ### Bugs Fixed
 - Fix duplicate Bunyan logs, missing HTTP duration metrics, duplicate request filtering, and incorrect performance-counter values. [#212](https://github.com/microsoft/opentelemetry-distro-javascript/pull/212)
 

src/genai/instrumentations/langchain/tracer.ts

@@ -228,6 +228,7 @@ export class LangChainTracer extends BaseTracer {
       }
       Utils.setModelAttribute(run, span);
       Utils.setChoiceCountAttribute(run, span);
+      Utils.setRequestAttributes(run, span);
       Utils.setResponseIdAttribute(run, span);
       Utils.setFinishReasonsAttribute(run, span);
       Utils.setProviderNameAttribute(run, span);

src/genai/instrumentations/langchain/utils.ts

@@ -9,10 +9,20 @@ import {
   ATTR_GEN_AI_CONVERSATION_ID,
   ATTR_GEN_AI_INPUT_MESSAGES,
   ATTR_GEN_AI_OPERATION_NAME,
+  ATTR_GEN_AI_OUTPUT_TYPE,
   ATTR_GEN_AI_OUTPUT_MESSAGES,
   ATTR_GEN_AI_PROVIDER_NAME,
   ATTR_GEN_AI_REQUEST_CHOICE_COUNT,
+  ATTR_GEN_AI_REQUEST_FREQUENCY_PENALTY,
+  ATTR_GEN_AI_REQUEST_MAX_TOKENS,
   ATTR_GEN_AI_REQUEST_MODEL,
+  ATTR_GEN_AI_REQUEST_PRESENCE_PENALTY,
+  ATTR_GEN_AI_REQUEST_SEED,
+  ATTR_GEN_AI_REQUEST_STOP_SEQUENCES,
+  ATTR_GEN_AI_REQUEST_STREAM,
+  ATTR_GEN_AI_REQUEST_TEMPERATURE,
+  ATTR_GEN_AI_REQUEST_TOP_K,
+  ATTR_GEN_AI_REQUEST_TOP_P,
   ATTR_GEN_AI_RESPONSE_ID,
   ATTR_GEN_AI_RESPONSE_FINISH_REASONS,
   ATTR_GEN_AI_RESPONSE_MODEL,
@@ -566,6 +576,124 @@ export function setChoiceCountAttribute(run: Run, span: Span) {
   }
 }
 
+// LangChain exposes provider-specific invocation params as unknown values. Match the upstream
+// OpenTelemetry GenAI instrumentations by accepting typed SDK values without coercing strings.
+function firstDefined(params: Record<string, unknown>, keys: string[]): unknown {
+  for (const key of keys) {
+    if (params[key] !== undefined && params[key] !== null) {
+      return params[key];
+    }
+  }
+  return undefined;
+}
+
+function integer(value: unknown): number | undefined {
+  return typeof value === "number" && Number.isFinite(value) && Number.isInteger(value)
+    ? value
+    : undefined;
+}
+
+function stringArray(value: unknown): string[] | undefined {
+  if (isString(value) && value.length > 0) return [value];
+  if (!Array.isArray(value)) return undefined;
+  const strings = value.filter((item): item is string => isString(item) && item.length > 0);
+  return strings.length === value.length && strings.length > 0 ? strings : undefined;
+}
+
+const OUTPUT_TYPES: readonly string[] = ["text", "json", "image"];
+
+const OUTPUT_TYPE_ALIASES: Readonly<Record<string, string>> = {
+  json_object: "json",
+  json_schema: "json",
+  b64_json: "image",
+  url: "image",
+};
+
+function normalizeOutputType(value: unknown): string | undefined {
+  if (!isString(value)) return undefined;
+  const normalized = value.trim().toLowerCase();
+  return OUTPUT_TYPES.includes(normalized) ? normalized : OUTPUT_TYPE_ALIASES[normalized];
+}
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+  return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+
+function getOutputType(params: Record<string, unknown>): string | undefined {
+  const explicit = firstDefined(params, ["output_type", "outputType"]);
+  const explicitType = normalizeOutputType(explicit);
+  if (explicitType) return explicitType;
+
+  const responseFormat = firstDefined(params, ["response_format", "responseFormat"]);
+  if (isRecord(responseFormat)) {
+    const formatType = normalizeOutputType(responseFormat.type);
+    if (formatType) return formatType;
+  }
+  const responseFormatType = normalizeOutputType(responseFormat);
+  if (responseFormatType) return responseFormatType;
+
+  const text = params.text;
+  if (isRecord(text)) {
+    const format = text.format;
+    if (isRecord(format)) {
+      return normalizeOutputType(format.type);
+    }
+  }
+  return undefined;
+}
+
+// Request attributes surfaced by LangChain under extra.invocation_params.
+// Both OpenAI-style snake_case and LangChain/provider camelCase aliases are
+// accepted because callback payloads vary by model integration and API path.
+export function setRequestAttributes(run: Run, span: Span): void {
+  const params = run.extra?.invocation_params;
+  if (!isRecord(params)) return;
+
+  const outputType = getOutputType(params);
+  if (outputType) span.setAttribute(ATTR_GEN_AI_OUTPUT_TYPE, outputType);
+
+  const numberAttributes: Array<[string, string[]]> = [
+    [ATTR_GEN_AI_REQUEST_FREQUENCY_PENALTY, ["frequency_penalty", "frequencyPenalty"]],
+    [ATTR_GEN_AI_REQUEST_PRESENCE_PENALTY, ["presence_penalty", "presencePenalty"]],
+    [ATTR_GEN_AI_REQUEST_TEMPERATURE, ["temperature"]],
+    [ATTR_GEN_AI_REQUEST_TOP_P, ["top_p", "topP"]],
+  ];
+  for (const [attribute, keys] of numberAttributes) {
+    const value = firstDefined(params, keys);
+    if (typeof value === "number" && Number.isFinite(value)) {
+      span.setAttribute(attribute, value);
+    }
+  }
+
+  const maxTokens = integer(
+    firstDefined(params, [
+      "max_tokens",
+      "maxTokens",
+      "max_completion_tokens",
+      "maxCompletionTokens",
+      "max_output_tokens",
+      "maxOutputTokens",
+    ]),
+  );
+  if (maxTokens !== undefined) {
+    span.setAttribute(ATTR_GEN_AI_REQUEST_MAX_TOKENS, maxTokens);
+  }
+
+  const seed = integer(firstDefined(params, ["seed"]));
+  if (seed !== undefined) span.setAttribute(ATTR_GEN_AI_REQUEST_SEED, seed);
+
+  const topK = integer(firstDefined(params, ["top_k", "topK"]));
+  if (topK !== undefined) span.setAttribute(ATTR_GEN_AI_REQUEST_TOP_K, topK);
+
+  const stopSequences = stringArray(
+    firstDefined(params, ["stop", "stop_sequences", "stopSequences"]),
+  );
+  if (stopSequences) span.setAttribute(ATTR_GEN_AI_REQUEST_STOP_SEQUENCES, stopSequences);
+
+  const stream = firstDefined(params, ["stream", "streaming"]);
+  if (typeof stream === "boolean") span.setAttribute(ATTR_GEN_AI_REQUEST_STREAM, stream);
+}
+
 // Response identifier - Helper to extract the unique response id returned by
 // the underlying provider (e.g. OpenAI chat completion id). LangChain.js
 // typically surfaces this as the AIMessage id (top-level for v1, nested

src/genai/semconv.ts

@@ -26,8 +26,18 @@ export const ATTR_ERROR_MESSAGE = "error.message" as const;
 
 // GenAI core
 export const ATTR_GEN_AI_OPERATION_NAME = "gen_ai.operation.name" as const;
+export const ATTR_GEN_AI_OUTPUT_TYPE = "gen_ai.output.type" as const;
 export const ATTR_GEN_AI_REQUEST_MODEL = "gen_ai.request.model" as const;
 export const ATTR_GEN_AI_REQUEST_CHOICE_COUNT = "gen_ai.request.choice.count" as const;
+export const ATTR_GEN_AI_REQUEST_FREQUENCY_PENALTY = "gen_ai.request.frequency_penalty" as const;
+export const ATTR_GEN_AI_REQUEST_MAX_TOKENS = "gen_ai.request.max_tokens" as const;
+export const ATTR_GEN_AI_REQUEST_PRESENCE_PENALTY = "gen_ai.request.presence_penalty" as const;
+export const ATTR_GEN_AI_REQUEST_SEED = "gen_ai.request.seed" as const;
+export const ATTR_GEN_AI_REQUEST_STOP_SEQUENCES = "gen_ai.request.stop_sequences" as const;
+export const ATTR_GEN_AI_REQUEST_STREAM = "gen_ai.request.stream" as const;
+export const ATTR_GEN_AI_REQUEST_TEMPERATURE = "gen_ai.request.temperature" as const;
+export const ATTR_GEN_AI_REQUEST_TOP_K = "gen_ai.request.top_k" as const;
+export const ATTR_GEN_AI_REQUEST_TOP_P = "gen_ai.request.top_p" as const;
 export const ATTR_GEN_AI_RESPONSE_MODEL = "gen_ai.response.model" as const;
 export const ATTR_GEN_AI_RESPONSE_ID = "gen_ai.response.id" as const;
 export const ATTR_GEN_AI_RESPONSE_FINISH_REASONS = "gen_ai.response.finish_reasons" as const;

test/internal/unit/a365/agent365Exporter.test.ts

@@ -374,6 +374,35 @@ describe("Agent365Exporter", () => {
       assert.deepStrictEqual(exportedSpan.attributes["number_array_attr"], [1, 2, 3]);
     });
 
+    it("should preserve GenAI request attributes in the A365 payload", async () => {
+      const { exportedSpan } = await exportAndGetPayload(fetchSpy, {
+        "gen_ai.output.type": "json",
+        "gen_ai.request.frequency_penalty": 0.1,
+        "gen_ai.request.max_tokens": 512,
+        "gen_ai.request.presence_penalty": -0.2,
+        "gen_ai.request.seed": 42,
+        "gen_ai.request.stop_sequences": ["DONE", "STOP"],
+        "gen_ai.request.stream": false,
+        "gen_ai.request.temperature": 0.2,
+        "gen_ai.request.top_k": 40,
+        "gen_ai.request.top_p": 0.8,
+      });
+
+      assert.strictEqual(exportedSpan.attributes["gen_ai.output.type"], "json");
+      assert.strictEqual(exportedSpan.attributes["gen_ai.request.frequency_penalty"], 0.1);
+      assert.strictEqual(exportedSpan.attributes["gen_ai.request.max_tokens"], 512);
+      assert.strictEqual(exportedSpan.attributes["gen_ai.request.presence_penalty"], -0.2);
+      assert.strictEqual(exportedSpan.attributes["gen_ai.request.seed"], 42);
+      assert.deepStrictEqual(exportedSpan.attributes["gen_ai.request.stop_sequences"], [
+        "DONE",
+        "STOP",
+      ]);
+      assert.strictEqual(exportedSpan.attributes["gen_ai.request.stream"], false);
+      assert.strictEqual(exportedSpan.attributes["gen_ai.request.temperature"], 0.2);
+      assert.strictEqual(exportedSpan.attributes["gen_ai.request.top_k"], 40);
+      assert.strictEqual(exportedSpan.attributes["gen_ai.request.top_p"], 0.8);
+    });
+
     it("should partition spans by identity and export separately", async () => {
       const exporter = new Agent365Exporter({
         tokenResolver: () => "test-token",

test/internal/unit/genai/langchain/tracer.test.ts

@@ -16,10 +16,13 @@ import {
   ATTR_ERROR_MESSAGE,
   ATTR_GEN_AI_INPUT_MESSAGES,
   ATTR_GEN_AI_OPERATION_NAME,
+  ATTR_GEN_AI_OUTPUT_TYPE,
   ATTR_GEN_AI_OUTPUT_MESSAGES,
   ATTR_GEN_AI_PROVIDER_NAME,
   ATTR_GEN_AI_REQUEST_CHOICE_COUNT,
   ATTR_GEN_AI_REQUEST_MODEL,
+  ATTR_GEN_AI_REQUEST_TEMPERATURE,
+  ATTR_GEN_AI_REQUEST_TOP_P,
   ATTR_GEN_AI_RESPONSE_ID,
   ATTR_GEN_AI_RESPONSE_MODEL,
   ATTR_GEN_AI_TOOL_CALL_ARGUMENTS,
@@ -426,7 +429,13 @@ describe("LangChainTracer", () => {
         },
         extra: {
           metadata: { ls_model_name: "deployment-o4-mini", ls_provider: "openai" },
-          invocation_params: { model: "deployment-o4-mini", n: 3 },
+          invocation_params: {
+            model: "deployment-o4-mini",
+            n: 3,
+            temperature: 0.2,
+            top_p: 0.8,
+            response_format: { type: "json_object" },
+          },
         },
         inputs: {
           messages: [[{ role: "user", content: "hello" }]],
@@ -481,6 +490,9 @@ describe("LangChainTracer", () => {
         3,
         "request choice count should come from invocation_params.n when >1",
       );
+      assert.strictEqual(got(ATTR_GEN_AI_REQUEST_TEMPERATURE), 0.2, "request temperature");
+      assert.strictEqual(got(ATTR_GEN_AI_REQUEST_TOP_P), 0.8, "request top_p");
+      assert.strictEqual(got(ATTR_GEN_AI_OUTPUT_TYPE), "json", "requested output type");
       assert.strictEqual(
         got(ATTR_GEN_AI_RESPONSE_MODEL),
         "o4-mini-2025-04-16",