inference_tracing

OpenTelemetry tracer for LangChain/LangGraph inference aligned with GenAI spec.

This implementation emits spans for agent execution, model invocations, tool calls, and retriever activity and records attributes in line with the OpenTelemetry GenAI semantic conventions. It supports simultaneous export to Azure Monitor (when a connection string is supplied) and to any OTLP collector configured via environment variables (for example::

export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317
export OTEL_EXPORTER_OTLP_PROTOCOL=grpc

The tracer focuses on three design goals:

1. Emit spec-compliant spans with the required/conditional GenAI attributes.
2. Capture as much context as LangChain exposes (messages, tool schemas, model parameters, token usage) while allowing content redaction.
3. Provide safe defaults that work across Azure OpenAI, public OpenAI, GitHub Models, Ollama, and other OpenAI-compatible deployments.

The module exports the AzureAIOpenTelemetryTracer callback handler. Attach an instance to LangChain run configs (for example in config["callbacks"]) to instrument your applications.