Skip to content

🪢 Open source agent evals & observability: Trace, evaluate, and improve LLM applications with one open platform.

fix(otel): normalize official cache-write usage (#17129)

Hassieb Pakzadddf44348a3fc
diff --git a/packages/shared/src/server/otel/OtelIngestionProcessor.ts b/packages/shared/src/server/otel/OtelIngestionProcessor.ts
index d93dedfbd9..c5b25da912 100644
--- a/packages/shared/src/server/otel/OtelIngestionProcessor.ts
+++ b/packages/shared/src/server/otel/OtelIngestionProcessor.ts
@@ -3065,6 +3065,7 @@ export class OtelIngestionProcessor {
rawUsageDetails["input_cached_tokens"];
const cacheCreationTokens =
rawUsageDetails["cache_creation.input_tokens"] ??
+ rawUsageDetails["cache_write.input_tokens"] ??
rawUsageDetails["cache_creation_input_tokens"] ??
rawUsageDetails["cache_write_tokens"] ??
rawUsageDetails["details.cache_write_tokens"] ??
@@ -3098,6 +3099,7 @@ export class OtelIngestionProcessor {
"prompt_details.cache_read",
"input_cached_tokens",
"cache_creation.input_tokens",
+ "cache_write.input_tokens",
"cache_creation_input_tokens",
"cache_write_tokens",
"details.cache_write_tokens",
diff --git a/web/src/__tests__/server/api/otel/otelMapping.servertest.ts b/web/src/__tests__/server/api/otel/otelMapping.servertest.ts
index d105411aaf..681986e74c 100644
--- a/web/src/__tests__/server/api/otel/otelMapping.servertest.ts
+++ b/web/src/__tests__/server/api/otel/otelMapping.servertest.ts
@@ -8109,6 +8109,77 @@ describe("OTel Resource Span Mapping", () => {
expect(usageDetails.output_reasoning_tokens).toBe(25);
expect(usageDetails["reasoning.output_tokens"]).toBeUndefined();
});
+
+ it("should normalize the official cache-write token attribute", async () => {
+ const traceId = "abcdef1234567890abcdef1234567893";
+ const officialCacheUsageSpan = {
+ resource: {
+ attributes: [
+ {
+ key: "service.name",
+ value: { stringValue: "test-service" },
+ },
+ ],
+ },
+ scopeSpans: [
+ {
+ scope: {
+ name: "opentelemetry",
+ version: "1.0.0",
+ },
+ spans: [
+ {
+ traceId: Buffer.from(traceId, "hex"),
+ spanId: Buffer.from("1234567890abcde2", "hex"),
+ name: "official-cache-usage",
+ kind: 1,
+ startTimeUnixNano: {
+ low: 1000000,
+ high: 406528574,
+ unsigned: true,
+ },
+ endTimeUnixNano: {
+ low: 2000000,
+ high: 406528574,
+ unsigned: true,
+ },
+ attributes: [
+ {
+ key: "gen_ai.usage.input_tokens",
+ value: { intValue: { low: 300, high: 0, unsigned: false } },
+ },
+ {
+ key: "gen_ai.usage.cache_read.input_tokens",
+ value: { intValue: { low: 40, high: 0, unsigned: false } },
+ },
+ {
+ key: "gen_ai.usage.cache_write.input_tokens",
+ value: { intValue: { low: 25, high: 0, unsigned: false } },
+ },
+ ],
+ status: {},
+ },
+ ],
+ },
+ ],
+ };
+
+ const events = await convertOtelSpanToIngestionEvent(
+ officialCacheUsageSpan,
+ new Set(),
+ );
+ const observationEvent = events.find(
+ (event) =>
+ event.type === "generation-create" || event.type === "span-create",
+ );
+
+ expect(observationEvent?.body.usageDetails).toEqual({
+ input: 235,
+ input_cached_tokens: 40,
+ input_cache_creation: 25,
+ });
+ });
+
it("should normalize raw Anthropic cache_read_input_tokens / cache_creation_input_tokens into Langfuse canonical keys", async () => {
// flat Anthropic cache spellings must map to cache aliases, not opaque passthrough buckets
const traceId = "abcdef1234567890abcdef1234567893";
diff --git a/worker/src/services/IngestionService/tests/calculateTokenCost.unit.test.ts b/worker/src/services/IngestionService/tests/calculateTokenCost.unit.test.ts
index 59b038e21a..ab65d01760 100644
--- a/worker/src/services/IngestionService/tests/calculateTokenCost.unit.test.ts
+++ b/worker/src/services/IngestionService/tests/calculateTokenCost.unit.test.ts
@@ -190,6 +190,22 @@ describe("Token Cost Calculation", () => {
expect(costs.total_cost).toBe(9.0);
});
+ it("should apply the configured cache-creation price", () => {
+ const costs = (IngestionService as any).calculateUsageCosts(
+ [
+ {
+ price: new Decimal(0.00000375),
+ usageType: "input_cache_creation",
+ },
+ ],
+ { provided_cost_details: {} },
+ { input_cache_creation: 25 },
+ );
+
+ expect(costs.cost_details.input_cache_creation).toBe(0.00009375);
+ expect(costs.total_cost).toBe(0.00009375);
+ });
+
it("should correctly calculate token costs with user provided costs", async () => {
const prices = await prisma.price.findMany({
where: {