-
[NA] [SDK] feat: enrich opik-vercel exporter for AI SDK v7 / Vercel eve (#7177)
frostbyte_neo 发布于
2026-06-22 12:27:56 +00:00 - [NA] [SDK] feat: enrich opik-vercel exporter for AI SDK v7 / Vercel eve
The opik-vercel OpenTelemetry exporter only understood AI SDK v4/v5/v6
(ai.*scope) and produced nearly-empty traces for AI SDK v7 / Vercel eve,
which emit the OpenTelemetry GenAIgen_ai.*scope (plus eve'seve.*
context) and no parent-child links.Exporter changes:
- Ingest
gen_aiandevescopes in addition toai. - Capture prompts, answers, tool args/results, token usage (incl. cache
tokens), thread id (from eve session id) andeve.*metadata. - Classify span types: chat -> llm, execute_tool -> tool, agent/step/turn ->
general. - Reconstruct the hierarchy by time containment (eve emits no parent links)
and set trace timestamps to span the whole turn. - Normalize messages to OpenAI chat shape (messages/choices/tool_calls) so the
Opik UI pretty-renders the conversation. - Emit create-only (upsert) operations from stable ids; never update()/end().
- Extract all attribute parsing into attributes.ts, split by convention
(aiSdk vs genAi), leaving the exporter to do OTel orchestration only.
Tests & CI:
- gen-ai.test.ts: deterministic tests over captured v7/eve spans (input,
output, usage, error info, span types, containment nesting, create-only). - real-llm.test.ts: gated live Anthropic test (generateText, streamText,
error) asserting input present / some output / usage / error info. - Add TypeScript SDK Library Integration Tests workflow that runs these on
TS SDK PRs with the ANTHROPIC_API_KEY secret.
No regressions to the AI SDK v4/v5/v6 path (existing ai-sdk/otel tests pass).
Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
- fix(opik-vercel): attach token usage only to LLM spans
eve repeats each model call's gen_ai.usage on the parent invoke_agent/step
spans as well as the chat span. Logging usage on those non-LLM wrapper spans
double-counts tokens once the trace total is aggregated across spans (a real
turn showed prompt_tokens 20842 instead of 10421). Emit usage only on llm-typed
spans. Verified against a local Opik backend via the Opik MCP.Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
- fix(opik-vercel): record token usage only on LLM spans, not traces
Token usage is now emitted exclusively on llm-typed spans. It is no longer set
on the trace (the backend aggregates the trace total from the LLM spans) nor on
tool/general spans. Update the v5/v6, gen_ai and real-LLM tests to assert usage
lives on the LLM span and is absent from the trace.Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
- test(opik-vercel): mirror real invoke_agent usage in double-count fixture
The invoke_agent wrapper span carries the model call's full gen_ai.usage in real
eve telemetry (input/output + cache tokens), not just input/output. Match that
in the fixture so the test proves the exporter drops the complete usage payload
from non-LLM spans.Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
- perf(opik-vercel): bound exporter memory with TTL pruning of stale traces
The exporter accumulates a trace's spans and stable id mappings across export
batches and never released them, so a long-running process (e.g. an eve server)
retained every trace for its lifetime. Add TTL pruning: track each trace's
last-seen time and, on each export, drop the accumulated spans / trace id /
span ids for traces idle longer than traceTtlMs (default 10 min). The window is
generous so in-flight turns are never dropped; eager 'delete after flush' would
break cross-batch coalescing and the create-only upsert. Configurable via the
new traceTtlMs option.Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
- fix(opik-vercel): set model & provider on LLM spans so cost is computed
The backend estimates span cost from the dedicated top-level model + provider +
usage fields, but the exporter only put model/provider in metadata, so no cost
was computed (trace providers was empty). Set model (gen_ai.response/request.model)
and provider (gen_ai.system / gen_ai.provider.name) on LLM spans only. Verified
via the Opik MCP: a real eve turn now reports per-span and aggregated trace cost.Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
Co-authored-by: Claude Opus 4.8 (1M context) noreply@anthropic.com
下载附件