📄 中文摘要
OpenTelemetry 在 LLM 追踪标准化方面取得了显著进展，提供了对延迟、令牌消耗、代理链的跨度树以及每一步的模型输入输出的可视化。这些功能确实有助于提高可观察性，但仅有可观察性而没有纠正措施，最终只会导致一个充满问题的仪表盘，仍需手动解决。虽然 OpenTelemetry 提供了有价值的信息，但它并不能检测输出是否存在幻觉，用户仍需依赖其他手段来解决这些问题。
OpenTelemetry 追踪你的 LLM，但无法修复它
出处: OpenTelemetry Traces Your LLM. It Does Not Fix It.
发布: 2026年3月25日
📄 English Summary
OpenTelemetry Traces Your LLM. It Does Not Fix It.
OpenTelemetry has made significant strides in standardizing LLM tracing, offering visibility into latency per call, token consumption, span trees across agent chains, and model inputs and outputs at each step. While these features are genuinely useful for enhancing observability, having observability without corrective measures results in a dashboard filled with problems that still require manual resolution. Although OpenTelemetry provides valuable insights, it does not offer detection of hallucinated outputs, leaving users to rely on additional methods to address these issues.