Docs/Get Started/Introduction
Introduction
Prove cheaper models on your real traffic before a single production request moves.
What is valite?
valite is a drop-in gateway for your LLM traffic. You point your existing OpenAI or Anthropic SDK at valite's base URL and keep shipping — no request or response is modified, and your provider API key passes through untouched. From that moment, every call is logged: model, tokens, latency, cost, full request and response bodies.
On top of that call log, valite runs an offline evidence pipeline. It fingerprints your calls into callsites (the distinct prompts your code makes), replays captured traffic against cheaper candidate models, grades every response against your own captured baseline with a calibrated pairwise judge, and produces a per-workflow routing plan with the evidence attached. You apply the plan when you choose — behind shadow mode, canary percentages, a kill switch, and an automatic rollback tripwire.
What valite is not
How it works
The entire integration is one environment variable:
export OPENAI_BASE_URL=https://gateway.valite.ai/openai/v1
From there, the lifecycle runs in four phases — observe, prove, plan, promote:
1 · Observe→
The gateway relays your traffic byte-for-byte and logs every call off the hot path — zero added time-to-first-token.
2 · Prove→
Replay jobs re-send your golden-set requests to candidate models and grade every response against your captured baseline with a blinded, calibrated pairwise judge.
3 · Plan→
The plan job picks the cheapest candidate whose judged worse-rate stays under the ceiling per callsite and emits an evidence-carrying routing config.
4 · Promote→
Activate routes in shadow or canary mode. A kill switch and an error-rate tripwire guard every promotion.
Explore the docs
Quickstart→
First logged call in under five minutes: swap the base URL, tag a workflow, watch calls arrive.
Core concepts→
Callsites, workflows, runs, and the artifact pipeline that turns a call log into routing evidence.
Integrations→
OpenAI, Anthropic, and any OpenAI-compatible host — Together, Groq, vLLM, Ollama, OpenRouter.
API reference→
The gateway's HTTP surface: proxy semantics, the jobs API, and the workflows read API.