How ReplyRabbit learns from your past conversations
ReplyRabbit doesn't train a model on your ticket archive. For each draft it looks up the most relevant material at that moment and gives it to the AI model alongside the conversation. Past conversations come in through two routes: the replies your team already sent to that customer, and the style examples it has learned from your edits. The limits below apply to each draft, so they aren't a cap on how much history your helpdesk can hold.
What goes into one draft
| Source | What ReplyRabbit looks through | What reaches the draft | Plan |
|---|---|---|---|
| The current conversation | The whole thread, including your team's earlier replies | All of it, trimmed only when very long | All |
| Customer history | The last 120 replies your team sent to this customer, in any mailbox, with no age limit | The 4 most relevant passages | Pro, Team |
| Learned style | Up to 200 recent examples per mailbox, kept for 24 months | Up to 8 examples | Pro, Team |
| Knowledge Base | Every indexed article | The 5 most relevant articles | All |
| Documentation sources | Every indexed page | The 5 most relevant pages | Pro, Team (Outline wikis: Team only) |
| Product catalog | Every indexed product | Up to 8 products | Pro, Team |
Outline is a team wiki (getoutline.com). Using an Outline wiki as a documentation source needs the Team plan: see Connect your Outline knowledge base.
Long pages send their most relevant passage
When a matching Knowledge Base article, documentation page or product page is long, the draft gets the passage that best answers the customer's question, not just the start of the page. That works whether the page matched by meaning or only by keywords, so it applies with semantic retrieval switched off too.
ReplyRabbit also leaves out articles that only loosely match, so drafts stay on topic. When a customer is asking about one of your products, drafts stay on that product, and greetings or names in the message don't change what gets looked up.
Customer history is counted per customer
The 120-reply window is per customer, not for the mailbox as a whole. A helpdesk with 500 regular customers has up to 120 replies searchable for each of them, which is every reply most customers have ever received.
For each draft ReplyRabbit picks the 4 passages from that window that best match the new message, and labels them with their dates so an old answer isn't presented as current. It searches every mailbox on your helpdesk for the same customer, which keeps a reply consistent with what a colleague said from a different inbox.
Customer history is on by default on Pro and Team. You can turn it off per mailbox under Customer history in the mailbox's ReplyRabbit settings. With semantic retrieval switched on, passages are matched by meaning as well as by keywords.
Learned style is counted per mailbox
When an agent edits a draft and sends it, ReplyRabbit keeps the pair as an example of how your team writes. Up to 8 examples guide each draft. They're chosen in this order: same kind of request, same customer, same agent, then most recent.
You can review, edit, add or remove examples on the Learned style screen. Adding a few of your best real replies by hand is the fastest way to give a new install a head start.
ReplyRabbit also learns two smaller things from edits: phrases your agents keep deleting from drafts, and corrections to suggested next actions.
What ReplyRabbit doesn't use
By default, replies you sent to other customers aren't searched as reference answers for a draft. If customer A asks something you answered last month for customer B, that earlier reply isn't looked up. Learned style examples carry your team's tone and phrasing into drafts, not the facts from those replies.
From version 1.42.0, Pro and Team mailboxes can switch on Learn from past replies, a Beta option that checks each draft against answers sent to other customers and adds a fact the draft was missing. It's off until an administrator turns it on.
General answers reach every draft when they live in a shared source:
- A FreeScout Knowledge Base article. See Connect the FreeScout Knowledge Base.
- Your public documentation or help pages, added as a documentation source.
- Company Context in the mailbox settings, for policies, terms and opening hours.
If the same question keeps coming back, turning the best past reply into a Knowledge Base article gives it to every future draft.
Why a draft may miss something you expected
- Customer history, documentation sources and the product catalog need a Pro or Team plan.
- Knowledge Base, documentation and product sources are each off until you switch them on for the mailbox.
- A newly added source has to finish indexing before drafts can use it.
- Past replies are indexed the first time they're needed, so the first draft for a customer with a long history may use only part of it. Later drafts use the rest.
Troubleshooting ReplyRabbit covers the other usual causes. Keep a short set of approved examples in the mailbox settings when you need drafts to follow a particular support voice.
Next