phoenix-observability

Trace and debug LLM applications by capturing OpenTelemetry-based traces to inspect spans across frameworks like OpenAI, LangChain, and LlamaIndex. This resolves scattered trace data and provides visibility into failing pipeline components.

Evaluate LLM response quality by running automated checks with built-in or custom evaluators against spans, datasets, and experiment outputs. This allows you to score hallucinations and validate model behavior before deployment.

Yes, you can instrument OpenTelemetry traces for LangChain, LlamaIndex, Anthropic, and OpenAI pipelines. The system collects and inspects these traces to monitor chatbots and retrieval pipelines.

Monitor and compare prompt variants by tracking experiments and projects over time. You can evaluate different prompt and model variants against datasets to review real-time system health before shipping updates.

You can optionally use SQLite or PostgreSQL for LLM observability persistence. The system requires arize-phoenix, evaluator workflows, and compatible trace export endpoints to store and analyze monitoring data.

The best way to monitor retrieval pipelines for hallucinations is to instrument the application with OpenTelemetry, capture spans, and run automated quality checks against test sets to score and evaluate outputs.