Skip to main content

Instant regex, synthesized as you type

Try it here, nothing to install

Tools that generate small formal artifacts (regexes, SQL fragments, cron expressions, JSON schemas) feel laggy when each generation takes several seconds. The user types, waits, loses the thread, and falls back to writing the artifact by hand. When a synthesis lands in well under a second, the artifact can simply track the description: pause typing, see the result.

This example is a small web app that turns a plain-English description into a working regex on every typing pause, with a test panel that checks your sample strings against the regex instantly in the browser.

Run it​

cd examples/instant-suggest
npm install

export CELERIS_BASE_URL="https://inference.celeris.ai/celeris-1/v1"
export CELERIS_API_KEY="<your-api-key>"

npm start
# open http://localhost:8788

Type a description ("a US phone number with optional area code in parentheses"). About a fifth of a second after you stop typing, synthesis starts, and the test strings below re-check themselves against it.

Measured results​

Real runs against a live workspace (2026-07-21), end to end from the browser:

phone number pattern: synthesized in 839 ms (820 ms model time)
email pattern:        synthesized in 779 ms (767 ms model time)
date pattern:         synthesized in 844 ms (830 ms model time)

A regex plus a one-sentence explanation is a bigger reply than a routing label, so synthesis costs more than the roughly 300 ms single-label calls elsewhere in this cookbook. It is still fast enough to run on every typing pause. At several seconds per generation this interaction pattern does not work at all; nobody waits that long mid-keystroke.

How it works​

  • Debounce and cancel. The browser waits 200 ms after the last keystroke, then calls the server. A new keystroke aborts any in-flight request, so only the latest description ever renders.
  • Deterministic checking in the browser. The model proposes; the browser disposes. Test strings are evaluated locally with new RegExp(...), so match results are exact, instant, and free.
  • Strict output contract. The server asks for one JSON object ({"regex": ..., "explanation": ...}) with max_tokens at 256 (see the latency guide for the budgeting rules). Regex patterns are full of backslashes and the model sometimes emits them singly, which is invalid JSON, so the parser retries with lone backslashes doubled.
  • Same credential seam as the chatbot example. The browser only talks to the local server; config.ts resolves the endpoint and key from the environment, with an optional per-request user key header. Another deployment can replace that one module without changing the browser application.
  • An endpoint seam beside it. public/app.js reads an optional window.__CELERIS_APP_CONFIG__?.apiBase, defaulting to this server's own api/ route. Any host can inject another API-compatible endpoint without changing the application.
  • Prewarm before intent. When a host sets prewarm: true, the page sends a best-effort OPTIONS during startup. That establishes the browser connection and CORS cache while the user is still reading; a server can use the signal to warm its Celeris connection with /echo too.

Where this pattern fits​

Any "description to small artifact" surface: regexes, SQL WHERE clauses, cron schedules, glob patterns, color palettes, config snippets. The recipe is the same: debounce, cancel stale requests, validate the artifact deterministically on the client, and keep the reply small.

Runnable source for this example: examples/instant-suggest in the celeris-cookbook repository.