Cookbook
Practical, runnable examples for building with the Celeris API. Every example
is a standalone directory in the
celeris-cookbook repository. Clone it and run them with
just CELERIS_BASE_URL and CELERIS_API_KEY set.
celeris-1 delivers its whole reply as one burst. See it against token-by-token delivery (canned data, no live calls):
Recommended starting points
Begin with the API, then explore the patterns that fit your workload.
All examples
Search runnable examples, production patterns, and implementation guides.
19 of 19
Filter by topic
Jul 22, 2026
Rendering a one-burst reply
interactive · latency · typescript
Three ways to render a whole-reply-at-once model in a chat UI (instant, fast fade, fixed-duration typewriter), with reduced-motion handling and guidance on when each feels best.
Celeris
Jul 22, 2026
Shell tasks with the celeris CLI
guide · cli · shell
Put semantic operations in your pipes. Commit messages from diffs, per-line log triage, semantic filtering, extraction into jq, and the scripting patterns that make the official CLI dependable in loops.
Celeris
Jul 22, 2026
Framework compatibility
guide · integrations
Verified snippets for the OpenAI SDKs, Vercel AI SDK, LangChain, LiteLLM, and Instructor against the Celeris endpoint, with the max_tokens caveat per framework and the known limits.
Celeris
Jul 22, 2026
Extraction at scale
extraction · throughput · production
Structured JSON extraction over a document corpus with asyncio concurrency and a cheap-repair pass. Measured at 182 to 289 docs per minute with 99 percent field accuracy and zero failed documents.
Celeris
Jul 22, 2026
Instant query understanding for search
search · latency · production
Query rewriting in front of a search index, measured at about 300 ms per rewrite, with graceful fallback to the raw query. Every successful rewrite improved the results.
Celeris
Jul 22, 2026
Task fit, bring your own prompts
evals · benchmarking · production
Answer "is this model enough for my task?" with your own prompts. A JSONL evaluation harness with objective checkers, per-task verdicts, and latency percentiles. Committed runs show it telling the truth in both directions.
Celeris
Jul 22, 2026
The voice turn budget
guide · voice · latency
Where the milliseconds go in a voice agent turn. Measured Celeris-1 reply step at 288 ms median beside cited ranges for speech to text and text to speech, and the total both ways.
Celeris
Jul 21, 2026
Agent latency compounds
agents · latency · tool-use
Why per-step speed decides whether an agent is interactive. A measured multi-step tool loop on Celeris-1, and a table showing how per-step latency compounds over N steps.
Celeris
Jul 21, 2026
Alert triage agent
agents · tool-use · token-budgeting
A multi-step tool-use agent loop that triages a synthetic production alert with tools over fixture data and emits a structured verdict, all inside the 8192-token window.
Celeris
Jul 21, 2026
Benchmark harness
benchmarking · latency · evals
Measure your own workspace in under a minute. Latency percentiles (TTFT and total), throttle, retry, empty, and truncation counts, and objective correctness scoring in one Markdown report.
Celeris
Jul 21, 2026
The cascade, fast model first
guide · architecture · agents
A reference architecture. Celeris-1 answers in the hot path in about 300 ms; a slower, more intelligent model refines or handles escalations asynchronously. When to use which, with a runnable sketch.
Celeris
Jul 21, 2026
Chatbot with model timing
server-timing · chat-completions · typescript
See the speed in a browser. A web chatbot that shows model processing time from Server-Timing and end-to-end latency on every reply, built as a small TypeScript server plus a vanilla JS UI.
Celeris
Jul 21, 2026
FastAPI on Fly.io
deployment · security · python
Put Celeris behind your own FastAPI backend on Fly.io. One endpoint, with the API key in the encrypted secret store (fly secrets set), never in the browser, code, or git.
Celeris
Jul 21, 2026
Getting started with the Celeris API
getting-started · chat-completions · token-budgeting
Make your first Celeris chat completion with the OpenAI SDK, with correct 256-token output alignment and input budgeting for the 8192-token total window. Includes a streaming cell that times time-to-first-token.
Celeris
Jul 21, 2026
Guardrails on every request
guardrails · moderation · latency
A moderation and validation gate fast enough for the hot path. Structured allow, block, or flag verdicts with reasons, measured at about 290 ms median added latency per request.
Celeris
Jul 21, 2026
Instant regex, synthesized as you type
interactive · latency · typescript
A web app that turns a plain-English description into a working regex on every typing pause, with in-browser test validation. Measured at about 0.8 s per synthesis end to end.
Celeris
Jul 21, 2026
Understanding latency with Celeris-1
guide · latency · token-budgeting
How to measure, budget, and design for latency. Metric vocabulary, the one-burst delivery model, the 8192-token total window and 256-token output alignment, and what becomes possible around 300 ms.
Celeris
Jul 21, 2026
Next.js on Vercel
deployment · security · typescript
Put Celeris behind your own Next.js backend on Vercel. One route handler, with the API key in the managed environment store, never in the browser, code, or git.
Celeris
Jul 21, 2026
Real-time ticket routing
classification · latency · production
Route support tickets in the request path. Measured at 260 ms median per ticket with 100 percent routing accuracy on the sample set, fast enough to classify before the page loads.
Celeris