These packages give you instrumented wrappers for OpenAI, Anthropic, Vertex AI, and AWS Bedrock that emit OpenTelemetry GenAI semantic convention spans, plus deployable dashboards for Phoenix, Langfuse, and Cloud Trace. You would adopt them to get spec-compliant observability across multiple LLM providers without writing instrumentation code yourself. The packages are designed as independent, installable modules—core types, instrumentation framework, provider wrappers, utilities, and exporters—so you can compose exactly what you need rather than pulling in a monolithic library.
An OpenTelemetry instrumentation for the Anthropic Node.js SDK that wraps `client.messages.create()` to emit GenAI semantic convention spans with request metadata, token usage, cost tracking, and streaming metrics.
A zero-config OpenTelemetry instrumentation class for the AWS Bedrock Runtime SDK that wraps `client.send()` to emit GenAI semantic convention spans with model-family-aware attribute mapping for Anthropic Claude, Amazon Titan, Cohere Command, and AI21 Jurassic models.
Canonical TypeScript types, constants, Zod schemas, and span-builder utilities for the OpenTelemetry GenAI semantic conventions, providing the single source of truth for all `gen_ai.*` attribute names, event types, error classifications, and domain models used across the `@reaatech/otel-genai-semconv-*` ecosystem. It exports typed constants, interfaces like `LLMRequest` and `LLMResponse`, Zod validation schemas, and a `SpanBuilder` class that constructs OTel-compliant spans with automatic attribute mapping, requiring only `zod` and `@opentelemetry/api` at runtime.
Custom OpenTelemetry span exporters that convert GenAI spans into the native formats of Arize Phoenix, Langfuse, and Google Cloud Trace, each implementing the `SpanExporter` interface with a `get*Format()` method for direct format conversion.
Provides tracer management, lifecycle hooks, streaming response handling, error classification, retry with exponential backoff, per-provider circuit breaking, chunk aggregation, and span-level PII redaction as a set of classes (`TracerManager`, `HookManager`, `ErrorHandler`, `RetryHandler`, `CircuitBreaker`) for instrumenting LLM calls with OpenTelemetry.
A zero-config observability kit for GenAI workloads, providing structured Pino logging, OpenTelemetry SDK initialization with OTLP trace export, pre-built metrics counters/histograms for LLM requests, tokens, cost, errors, and TTFT, plus runtime health checks with a pluggable HTTP endpoint. Exports functions (`initOTelSDK`, `createLogger`, `initMetrics`, `logGenAIEvent`, etc.) that bundle Pino v10 and the OpenTelemetry JS SDK.
A function `instrument(client)` that wraps an OpenAI SDK client instance to automatically emit OpenTelemetry GenAI semantic convention spans for every `chat.completions.create()` call, capturing request metadata, token usage, cost, and streaming metrics.
Token counting, cost calculation, and PII redaction utilities for LLM observability. Provides a `TokenCounter` class with tiktoken-based estimation for OpenAI models and character-based fallbacks for Anthropic, Vertex AI, and Bedrock; a `CostCalculator` class with built-in pricing tables for GPT-4, GPT-4o, Claude 3, Gemini, and Bedrock model families; and a `PIIRedactor` class that recursively redacts emails, SSNs, credit cards, IP addresses, and phone numbers from nested objects.
A function that wraps a Google Generative Language (Vertex AI) SDK model instance to automatically emit OpenTelemetry GenAI semantic convention spans on every `generateContent()` call, including GCP metadata, generation config attributes, candidate events, and optional cost tracking.
