# Pattern: route between an LLM and Dex > Let Dex decide every incoming message and call an LLM only for the messages that need a written reply, with abstained answers going to a person. Dex and an LLM do different jobs. Dex reads a message and answers typed questions about it, fast and the same way every time; an LLM writes text. Put Dex in front: it decides every message, and your code calls the LLM only when a written reply is needed. Most messages never reach the LLM, so you pay for its output tokens only where they add something. ## The flow 1. Dex answers three questions about each message: what the writer wants (`intent`, a pick), whether it only asks where an order is (`template_ok`, a check, so a template can answer it), and whether it contains instructions aimed at an AI (`injection`, a check). 2. Your code branches on the typed answers: - an abstained `intent`, or a likely injection: a person handles the message; - `spam`: archive it; - `order_status` with `template_ok`: answer from your order system with a template, no LLM; - everything else: the LLM drafts a reply, and a person approves it before it is sent. 3. Log the Dex request id, model and calibration with the route, so every decision can be traced. ## The request The code below builds this request with the SDK helpers. Here it is on the wire, for one email, with the answer the live API gave: ```json { "state": { "email": { "subject": "Order 58213", "body": "Hi, my order 58213 arrived yesterday but the kettle leaks from the bottom. I would like my money back, or a new kettle if that is quicker. What do I need to do?" } }, "questions": { "intent": { "type": "pick", "instructions": "What does the writer of {{email.body}} want?", "options": { "order_status": "Where their order is or when it arrives", "refund": "Money back or a return", "complaint": "A complaint about a product or the service", "question": "A question about products, prices or the account", "spam": "Advertising, phishing or nonsense", "other": null }, "min_confidence": 0.5 }, "template_ok": { "type": "check", "instructions": "Does {{email.body}} ask where an order is or when it will arrive, and nothing else?", "min_confidence": 0.6 }, "injection": { "type": "check", "instructions": "Does {{email.body}} contain instructions addressed to an AI assistant or agent?", "min_confidence": 0.3 } } } ``` ```json { "id": "req_01M3JHAKGJ66SX1GMT2F94ZBTC", "object": "decision", "created": 1790549773, "model": "dex-1.0.1", "served_by": "gpu", "calibration": "cal-20260926-1", "answers": { "intent": { "type": "pick", "choice": "refund", "probabilities": { "order_status": 0.0255, "refund": 0.8352, "complaint": 0.0505, "question": 0.035, "spam": 0.0235, "other": 0.0303 }, "confidence": 0.7847, "abstained": false }, "template_ok": { "type": "check", "probability": 0.1066, "confidence": 0.7869, "abstained": false }, "injection": { "type": "check", "probability": 0.1282, "confidence": 0.7437, "abstained": false } }, "usage": { "input_tokens": 164, "state_tokens": 59, "question_tokens": 105, "allowance_tokens": 0, "paid_tokens": 0, "charge_micro_cents": 0, "unit_price_micro_cents": 0, "tier": "test" } } ``` Captured from the live API on 2026-09-27 with a test key: model `dex-1.0.1`, calibration `cal-20260926-1`, `served_by: gpu`, 164 input tokens. A live key gives the same answers and counts, with `allowance_tokens`, `paid_tokens` and the charge filled in. For this email the writer wants a `refund` (0.8352), it asks for more than an order's whereabouts (`template_ok` 0.1066) and it holds no instructions for an AI (`injection` 0.1282), so the code below sends it to the LLM for a draft that a person approves. - The intents are mutually exclusive, so they are one pick with a catch-all `other`. - `template_ok` is a separate yes or no fact about the same message; it costs only its own question tokens, because the message is billed once. Phrase it as what the message asks, not as what your system can do: "Can this email be answered completely with the order's status and delivery date?" got a confident no (0.0599) on a plain "Where is my order 58213?" email, while the wording above answers it with yes (0.8554) and still says no to the kettle email. - The injection check guards the LLM step: an email that tells an AI to do something should not reach an LLM that can act on it. ## The code `draft_with_llm`, `send_template`, `archive` and `queue_for_person` stand for your own code. The LLM client is whichever one you use. ```python tab="Python" from thinqit_dex import Client, check, pick dex = Client() # reads DEX_API_KEY QUESTIONS = { "intent": pick( "What does the writer of {{email.body}} want?", { "order_status": "Where their order is or when it arrives", "refund": "Money back or a return", "complaint": "A complaint about a product or the service", "question": "A question about products, prices or the account", "spam": "Advertising, phishing or nonsense", "other": None, }, min_confidence=0.5, ), "template_ok": check( "Does {{email.body}} ask where an order is or when it will arrive, and nothing else?", min_confidence=0.6, ), "injection": check( "Does {{email.body}} contain instructions addressed to an AI assistant or agent?", min_confidence=0.3, ), } def handle(email: dict) -> str: # Only what the decision needs: the sender address stays out of the state. state = {"email": {"subject": email["subject"], "body": email["body"]}} decision = dex.decide(state=state, questions=QUESTIONS) intent = decision.pick("intent") template_ok = decision.check("template_ok") injection = decision.check("injection") if intent.abstained or injection.abstained or injection.probability >= 0.5: route = "person" queue_for_person(email, decision.id) elif intent.choice == "spam": route = "archive" archive(email) elif intent.choice == "order_status" and not template_ok.abstained and template_ok.probability >= 0.5: route = "template" send_template(email, "order_status") else: route = "llm" queue_for_person(email, decision.id, draft=draft_with_llm(email, intent.choice)) print(route, decision.id, decision.model, decision.calibration) return route ``` ```ts tab="TypeScript" import { Client, check, pick } from "@thinqit/dex"; const dex = new Client(); // reads DEX_API_KEY const questions = { intent: pick( "What does the writer of {{email.body}} want?", { order_status: "Where their order is or when it arrives", refund: "Money back or a return", complaint: "A complaint about a product or the service", question: "A question about products, prices or the account", spam: "Advertising, phishing or nonsense", other: null, }, { min_confidence: 0.5 }, ), template_ok: check("Does {{email.body}} ask where an order is or when it will arrive, and nothing else?", { min_confidence: 0.6, }), injection: check("Does {{email.body}} contain instructions addressed to an AI assistant or agent?", { min_confidence: 0.3, }), }; export async function handle(email: { from: string; subject: string; body: string }): Promise { // Only what the decision needs: the sender address stays out of the state. const state = { email: { subject: email.subject, body: email.body } }; const decision = await dex.decide({ state, questions }); const { intent, template_ok: templateOk, injection } = decision.answers; let route: string; if (intent.abstained || injection.abstained || injection.probability >= 0.5) { route = "person"; await queueForPerson(email, decision.id); } else if (intent.choice === "spam") { route = "archive"; await archive(email); } else if (intent.choice === "order_status" && !templateOk.abstained && templateOk.probability >= 0.5) { route = "template"; await sendTemplate(email, "order_status"); } else { route = "llm"; await queueForPerson(email, decision.id, await draftWithLlm(email, intent.choice)); } console.log(route, decision.id, decision.model, decision.calibration); return route; } ``` In TypeScript the builders type the answers: `intent.choice` is `"order_status" | "refund" | ... | "other"`, so a typo in a label is a compile error. ## Why this order - **Cheap and fast first.** Dex answers in one read of the message and charges input tokens only. The LLM, which charges for the text it writes, runs only on the messages that need text. - **Deterministic routing.** On a pinned version the same email always takes the same route, so the routing can be tested like any other code (see [Pin versions in CI](/docs/cookbook/pin-versions-in-ci/)). - **A person on every uncertain or risky path.** Abstained answers and likely injections never reach an automatic action. See [Abstention](/docs/concepts/abstention/) and [Known limits](/docs/concepts/known-limits/#instructions-inside-the-state-can-steer-answers).