How AI replies work
ReviewBox uses a 3-tier pipeline to generate reply drafts. About 85% of replies are served from tiers 1–2 at zero cost. Only ~15% of reviews ever touch an AI model.
The 3-tier pipeline
Template match
$0 · instantReviewBox checks your review against 25 built-in reply templates. Templates match on rating + detected tags (crash, billing, login, feature request, etc.). If a template matches, the reply is returned instantly with zero AI tokens consumed.
~70% of all reviews are handled here.
Reply cache
$0 · ~5msIf no template matches, ReviewBox checks a Redis cache for a previously-generated reply to an identical or near-identical review. Cache keys are SHA-256 hashes of the review text + rating + tone. Cached replies are served instantly.
~15% of reviews are handled here. Cache entries live for 7 days.
AI generation
~$0.001 · ~600msIf neither tier 1 nor 2 matches, ReviewBox generates a new reply using Groq's Llama 3.3 70B model. Before sending, the review text is compressed (stripping filler phrases) to reduce token usage by ~73%. The generated reply is stored in the cache so identical future reviews are free.
~12–18% of reviews reach AI generation.
Tone options
Formal, concise, corporate-safe.
Warm, conversational, approachable.
Leads with understanding before solutions.
2–3 sentences max. Respect the reader's time.
Write your own persona description (up to 200 chars). Applied to every AI-generated reply.
Tone is applied to AI-generated replies only. Template replies use their own pre-written tone.
Auto-publish rules
On the Team plan, you can configure auto-publish rules. A rule defines conditions (e.g. “5-star review + positive sentiment + reply from template”) — when all conditions match, ReviewBox publishes the reply without human approval.
- Disabled by default — opt-in per workspace
- Only template-tier replies can be auto-published (not AI-generated ones)
- All auto-published replies are logged in the audit trail
- You can pause or disable auto-publish at any time from Settings → Automations
Does ReviewBox learn from edits?
Not yet. When you edit an AI draft before publishing, your edit is saved but not currently used to fine-tune the model. This is on the roadmap — the goal is for ReviewBox to learn your preferred phrasing over time and reduce AI-generated replies in favour of cached edits.