What is AI Observability?

Revefi AI Observability gives Data and AI teams visibility into how their LLM applications behave in production. Every LLM call your application makes is captured as an OpenTelemetry span and streamed to Revefi, where it is parsed into structured metrics you can slice and chart:

• Cost — per request, model, user, agent, or any custom tag (uses provider-billed cost when available, otherwise derived from token counts).
• Token usage — prompt, completion, total, and cache read/creation tokens.
• Latency — request duration distributions.
• Reliability — error rate and failure breakdowns from span status.
• Tool / agent activity — tool calls, workflow and entity names for agentic apps.
• Prompts & completions — the actual request and response content for inspection and prompt-similarity analysis.

How it works

Revefi follows the OpenTelemetry GenAI semantic conventions (the same conventions emitted by OpenLLMetry). Spans are sent over OTLP/HTTP (protobuf) to Revefi's ingestion gateway, authenticated with a Revefi API token.

Your LLM app
   │  (instrumentation: Revefi LLM SDK or any OpenTelemetry GenAI instrumentation)
   ▼
OTLP/HTTP  POST https://gateway.revefi.com/api/v1/traces/ingest
   │  Authorization: Bearer <REVEFI_API_TOKEN>
   ▼
Revefi ingestion → parses gen_ai.* / traceloop.* attributes → structured LLM spans
   ▼
Revefi UI: cost, usage, latency, errors, tool calls, custom-tag breakdowns

Two ways to send data

Both paths hit the same endpoint and produce the same structured spans. The SDK is simply a thin wrapper that configures an OTLP exporter pointed at Revefi.

Prerequisites

• A Revefi account with AI Observability enabled for your tenant.

Span structure

To understand exactly which attributes Revefi reads (and what to set if you instrument manually), see the Span Structure reference.