RR LLM
The model that writes your review replies
RR LLM is a single-purpose generation model. It takes a guest review and your settings, and returns one editable management reply. It is built on a Qwen3 8B base model, fine-tuned on more than one million real reviews — thirteen million training data points — and served as an HTTPS API at around 255 output tokens per second.
Same model that powers the Review Reply extension. One job: review in, reply out.
// request
{
"review": "Room was spotless, but the wifi
kept dropping in the evening.",
"rating": 3,
"reviewerName": "Alex",
"tone": "professional",
"language": "en"
}
// response
{
"reply": "Thank you, Alex — we are glad the
room met your expectations, and
sorry the connection was unstable...",
"model": "qwen3-8b-review-v5",
"provider": "local-trained-qwen",
"createdAt": "2026-08-25T16:30:00Z"
}Answer first
What is the RR LLM?
RR LLM is a language model trained for one task: writing the management response to a guest review. You send the review text and, when you have them, the star rating, the reviewer name, a tone, and a language. It returns a single reply draft that names the guest when a name is available, matches the review's facts and sentiment, and stays inside the length and tone you asked for.
It is deliberately not a general chat assistant. It does not answer questions about the world, it does not hold a conversation, and it does not post anything. Its output is a draft that a person edits and publishes. That narrow focus is what makes it small, fast, and consistent on the review-reply task.
Model card
Specification
What the model is, what it takes in, and how it is served.
| Base model | Qwen3 8B (qwen3-8b-review-v5), fine-tuned for review-reply generation |
|---|---|
| Task | Guest review → one management reply draft |
| Training | 1,000,000+ real reviews; 13,000,000+ labelled data points |
| Throughput | ~255 output tokens per second, sustained |
| Concurrency | Horizontally scaled; a deployment serves 10,000+ concurrent generations |
| Languages | 29 languages, or match the review's language automatically |
| Tones | Professional · Friendly · Casual |
| Context inputs | Review text, star rating, reviewer name, tone, language, optional manager signature |
| Response | reply, model, provider, createdAt |
| Serving | HTTPS API at rr-api.reputationsystems.ai/api/v1/rr/replies |
| Auth | X-Api-Key or Bearer; up to five keys per account, or a private 30-day extension session |
| Rate limits | Per plan (requests per minute / per day); 429 with Retry-After, and X-RateLimit-* headers on every reply |
| Errors | 400 invalid review · 401 missing or revoked key · 503 model unavailable |
These examples match the published RR API reference; the setup guide covers access, keys, limits, and errors in full.
Why fine-tuned
Built for one job, not for everything
Grounded in review language
Trained on reviews and their real responses, so it recognises the patterns that matter in guest feedback — apologies that need to stay vague, promises that should not be made, and the greeting conventions of each platform.
Fast and predictable
A focused model returns a reply quickly and in a consistent shape, so you can draft across a whole backlog without post-processing. Same inputs, same tone — every time.
Controlled output
Tone and language are explicit request fields, not prompt suggestions. The client can constrain length, strip unsupported symbols, and run Truth Check on the draft before a person posts it.
Quickstart
From a review to a draft in three calls
The shape below is what a technical user does: get a key, send the review, use the draft. Every call is a single HTTPS POST.
Get access
Create an RR account and generate an API key, or connect the extension and use its 30-day session. Keys are sent in the X-Api-Key header. The setup guide walks through access, keys, and your first call.
Send the review
POST the review text with the rating, reviewer name, tone, and language you want. Missing optional fields are fine — the model works from what it has.
Use the draft
The response carries the reply text, the model name, the provider, and a timestamp. Show it to a person, run Truth Check, let them edit it, then post through the platform yourself.
curl -X POST \
https://rr-api.reputationsystems.ai/api/v1/rr/replies \
-H "X-Api-Key: $RR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"review": "Room was spotless, but the
wifi kept dropping.",
"rating": 3,
"reviewerName": "Alex",
"tone": "professional",
"language": "en"
}'import os, requests
res = requests.post(
"https://rr-api.reputationsystems.ai/api/v1/rr/replies",
headers={"X-Api-Key": os.environ["RR_API_KEY"]},
json={
"review": review_text,
"rating": 3,
"reviewerName": "Alex",
"tone": "professional",
"language": "en",
},
timeout=60,
)
res.raise_for_status()
reply = res.json()["reply"]const res = await fetch(
"https://rr-api.reputationsystems.ai/api/v1/rr/replies",
{
method: "POST",
headers: {
"X-Api-Key": process.env.RR_API_KEY,
"Content-Type": "application/json",
},
body: JSON.stringify({
review,
rating: 3,
reviewerName: "Alex",
tone: "professional",
language: "en",
}),
}
);
if (!res.ok) throw new Error(`RR API ${res.status}`);
const { reply } = await res.json();| review | required Guest review text, 1–10,000 characters. |
|---|---|
| rating | A rating such as 4, 4/5, or 8/10 (number or string). |
| reviewerName | Guest name used for a personalised greeting. |
| tone | Reply style — the extension uses professional, friendly, or casual. |
| managerName | Name included in the reply signature. |
| managerTitle | Title included in the reply signature. |
| includeSignature | Whether to include the manager signature. Defaults to true. |
| hotelName | Property name used for context. |
| platform | Source platform, such as Google, Booking.com, Expedia, or Tripadvisor. |
| language | Reply language code, for example en, es, or fr. |
| allowEmoji | Whether the reply may include emoji. Defaults to false. |
Scale
What 255 tokens per second means in practice
A reply in a blink
A typical 90-word reply is around 120 output tokens. At 255 tokens per second that draft is generated in well under a second, so the panel feels instant rather than waiting on a spinner.
A backlog clears fast
Throughput is per request and the service scales horizontally. A single deployment serves 10,000+ concurrent generations, enough for a portfolio answering its review queue at the same time.
Cost stays predictable
A focused model produces short, bounded replies instead of open-ended text, so output — and therefore the cost of each draft — stays close to what a reply actually needs.
Throughput and concurrency describe the serving platform as configured for RR. Your plan sets the request limits that apply to your account.
Guardrails
Draft, then verify
A generator can be wrong. Everything built around RR LLM assumes it will be, and keeps a person in the loop.
Truth Check on every draft
Before a reply is treated as ready, Truth Check flags concrete claims the review does not support, unverified promises, missed complaints, and confident language that does not match the review. It can propose a safer acknowledgement in place of a flagged sentence.
Platform rules enforced
The client applies platform constraints around the model — for example, Booking.com replies are cleaned of emoji and unsupported symbols, and common contractions are expanded — so a draft is ready for the specific form it will be posted into.
A human makes the final click
RR LLM never posts. The draft reaches a person, who edits it and submits it through the platform's own reply form. If Truth Check still finds something, publishing takes a second explicit confirmation.
Privacy
What an API call carries
A request carries only what is needed to write the reply: the review content you deliberately send, the displayed or typed reviewer name and rating, your reply settings, and a platform label — the review hostname, or the fixed label manual for a pasted review. Nothing is sent until you invoke an RR action or make the call yourself.
Requests go over HTTPS to the Reputation Systems API at rr-api.reputationsystems.ai. Review data is not sold and is not used for personalised advertising, retargeting, creditworthiness, or lending. Processing follows the Chrome Web Store Limited Use requirements. Read the privacy and security summary or the full policy at rr.reputationsystems.ai.
FAQ
Questions a developer asks first
Is this the same model the extension uses?
Yes. RR LLM is the generation model behind Review Reply. When you click RR in the extension, the draft you get comes from this model; this page is for teams who want to call it directly with their own key.
Can I use it for something other than review replies?
It is tuned narrowly for the review-to-reply task, so that is where it performs best. Using it as a general assistant is out of scope and not what it was trained to do.
How much does a call cost?
Every successfully generated reply uses one credit, and failed requests are not charged. Each plan includes a monthly number of reply credits that reset each period; top-up credits cost $1.00 per 10 replies after that. See the plan table and limits.
How do I get an API key?
Create an RR account and generate a key, or connect the extension and reuse its session. Keys are sent in the X-Api-Key header on every request. The setup guide covers the full walkthrough. For higher volumes or a service agreement, book a demo.
How fast and how large is it?
The served model runs at around 255 output tokens per second, and the platform scales horizontally to 10,000+ concurrent generations. Replies are short by design, so a single draft is normally well under a second of generation time.
What happens to the reviews I send?
A call carries only the review content needed to draft a reply, over HTTPS, and only when you ask for it. What is sent, stored, and used is defined by the privacy policy; review data is not sold and is not used for personalised advertising.
Related
Documentation and related pages
Everything that goes with the model: the API reference, the privacy policy, pricing, and the product itself.
Setup guide and API reference
How to get access, generate a key, and make your first call, plus the endpoint, request fields, and limits.
Privacy policy
What a request carries, how it is stored, and what it is used for — the full published policy.
Privacy and security summary
A short, plain-language summary of the same policy, covering the extension and the API.
How the extension works
The end-to-end workflow — draft, edit, Truth Check, and your final click — on each platform.
Reply guides
How to answer guest reviews, with sample replies and a worked Truth Check example.
Get started
Build review replies into your own product
Call the same model behind Review Reply, or start with the extension and connect a key when you are ready.