> ## Documentation Index
> Fetch the complete documentation index at: https://www.revve.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.
# Chat Agent Advanced Settings
> Every panel on a chat agent's Advanced Settings tab — defaults, trade-offs, and when to change each one — for the builder tuning agent behavior beyond the prompt.
Open your chat agent and select the **Advanced Settings** tab. Panels are listed in a left sidebar; the page opens on **Inactivity Follow-up**, and the active panel syncs to the URL so you can share a link straight to it.
Two things about how editing works here:
* **Changes auto-save into your draft.** Almost every panel saves as you type — there's no panel-level Save button (dialogs like Add Testimonial or Add Field have their own Save; the **Customize** modal in the Feedback panel is the one that saves manually). Note that testimonial content lives in a team-level library and saves immediately — only the agent's selection of testimonials is draft-gated. You need the draft edit lock (**Start Editing**) to change anything; viewing a historical version is read-only. Auto-saved changes still only affect the live agent once you publish — see [Publishing and Versioning](/docs/getting-started/publishing-and-versioning).
* **Some panels only appear for some agents.** Visibility depends on the agent's channel and engine:
| Panel | Appears when |
| ----------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Feedback**, **Daily Report** (in Others), Domain Allowed List (in Security) | Website channel only |
| **Social Feedback** | Messenger, Zalo OA, TikTok, or LinkedIn channel |
| **Schedule Meeting**, **Contact Collection**, **Message Approval** | Single Prompt engine only — Conversational Flow agents configure these per node instead ([Flow Node Features](/docs/conversational-flow/flow-node-features)) |
The guiding principle, as with [voice agents](/docs/voice-agents/voice-agent-settings): defaults are sensible — change a setting when you observe the specific problem it solves. Each section below tells you what that problem looks like.
## Inactivity Follow-up
When a customer goes quiet, the agent generates a follow-up message after a timeout instead of letting the thread die. Off by default; turn it on for sales and lead-capture agents where a stalled conversation is a lost lead.
| Setting | Default | When to change |
| -------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Inactivity Timeout (seconds)** | 300 (minimum 30) | Five minutes suits support chats where people step away. Shorten toward 120–180 for high-intent sales chats where momentum matters. Don't go below 60 — you'll nag people who are still typing or reading your last answer. |
| **Instructions** | "Generate a follow-up message when the customer is inactive. The message should be friendly and offer assistance. Consider the conversation context when generating the follow-up." | Required when enabled. This is an LLM instruction, not fixed text — the agent writes a context-aware message from it. Make it specific to your funnel: "If they were mid-checkout, offer to answer questions about pricing; otherwise ask if they'd like a summary emailed." |
## Human Escalation
Two cards: when to hand off, and what the bot says while the human is on the way. This is the configuration side of escalation — for how your team actually receives and works escalated threads, see [Human Takeover](/docs/inbox/human-takeover).
**Human Agent Escalation** — a master switch plus one textarea:
| Setting | Default | When to change |
| ----------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Escalation Criteria** | A three-bullet starter: "Escalate to human when customers: ask to speak with a human; express frustration or anger; have complex issues requiring human help" | Always rewrite this. The default catches the universal cases, but the escalations that hurt are product-specific: name *your* failure modes — "asks about a refund over \$200", "reports the widget not loading", "mentions a competitor's quote". Vague criteria produce both missed escalations and a bot that gives up too easily. |
The escalation criteria don't have to cover everything — when you need to escalate in other cases, your prompt or tools can call the `escalate_to_human` tool directly.
**Auto-Reply Messages** — what customers see between escalation and a human's first reply:
| Setting | Default | When to change |
| ---------------------------- | ----------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Waiting in Queue Message** | Empty — no auto-reply is sent until you set one | Sent when the customer messages after escalation, while the thread waits for agent assignment. Set it — silence after "let me connect you to a human" reads as being abandoned. The placeholder is a good starting point: "Please wait, our agent will be with you as soon as possible." Set expectations honestly — if your team replies within business hours only, say so here. |
| **Agent Joined Message** | Empty — no auto-reply is sent until you set one | Sent after an agent is assigned but before their first reply. The placeholder text works for most teams: "Please wait a moment, our agent has joined and will respond shortly." |
## Testimonials
A team-level library of customer testimonials the agent can surface in the chat widget. Click **Add** to open the **Add New Testimonial** dialog: **Client Name** and **Avatar Image** are required; **Client Title**, **Testimonial** text, **Business Description**, **Post URL**, and **Company Logo** are optional. Curate a handful of strong, relevant ones rather than dumping every review — the agent references the testimonials you select here.
## Tools
The agent's tool registry: custom API-call tools you define plus preset tools. This is where you give the agent abilities beyond text — checking order status, creating tickets, hitting your backend.
## Message Bundles
Manage reusable bundles of messages for this agent. Worth a look once your agent sends the same multi-message sequences repeatedly.
## Webhook
One field: a **URL** that receives real-time notifications about chat interactions. Must be `http`/`https`; leave it empty if you don't need it. Set it when your own systems need to react to conversations as they happen rather than polling.
## Analysis & Summary
Three cards controlling what happens *after* a conversation: the model that analyzes it, the summary it writes, and the structured fields it extracts. The output of all three lands in the [conversation analysis panel](/docs/inbox/conversation-analysis-panel).
**Analysis Model** — the model that runs post-conversation analysis. This is separate from the conversational model that talks to customers (see [Choosing an LLM Model](/docs/general-configuration/choosing-an-llm-model)).
| Setting | Default | When to change |
| -------------------- | ------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Provider** | openai | Rarely. |
| **Model** | o4-mini | The default is a good cost/quality balance for analysis. Consider a stronger model only if extracted fields keep coming back wrong on genuinely complex conversations. |
| **Reasoning effort** | medium | Shown for reasoning models. Raise for nuanced classification; lower to cut analysis cost on simple, high-volume chats. |
**Summary Generation Prompt** — customizes how conversation summaries are written. It ships with a long templated default (including a `{language}` placeholder) and a **Reset to Default** button, disabled while you're already on the default. Edit it when summaries consistently miss what your team scans for — e.g., "always state the customer's request in the first sentence and end with the resolution status."
**Analyze Fields Configuration** — structured data extracted from every conversation. Each field has:
| Field | Notes |
| -------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Name** (required) | Name them like variables — `customerName`, `orderNumber` — since they become data keys downstream, not display labels. |
| **Type** (required) | Text, Number, or Boolean (default Text). Use Boolean for yes/no facts ("mentioned a competitor") — they're far easier to filter on than free text. |
| **Description** (required) | Written for the analysis model. Be precise about what qualifies: "The order number the customer is asking about, digits only. Leave empty if none mentioned." |
| **Format** (optional) | A format hint like `YYYY-MM-DD` — set it for dates and anything you'll pipe into another system. |
## Conversation Categories
Define categories the agent uses to classify each conversation — one name plus a description per category, editable inline. A locked default category **Unknown** ("Not one of the above categories") always exists and can't be renamed or deleted; every other category needs both a name and a description. Write descriptions as decision rules, not labels — "Billing: questions about invoices, charges, or payment methods" — because the description is what the classifier actually reads. Keep the list short and mutually exclusive; overlapping categories produce noisy reports.
## Notification
Alerts your team when a condition you describe is met in a conversation. Off by default. When enabled, four things are required: the trigger description, a **Title**, a channel, and valid recipients.
| Setting | Default | When to change |
| ----------------------------------------------- | ---------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **When would you like to send a notification?** | Empty (required) | Write the trigger like an instruction to a colleague: "If a customer asks for sales information, notify the team." One condition per agent works best — a laundry list of triggers fires constantly and gets ignored. |
| **Channel** | Email | Email always works. **Slack** appears only if the Slack app is installed — otherwise you'll see an install prompt linking to the Apps page. |
| **Title** | Empty (required) | The subject line your team sees. Make it scannable: "Sales inquiry — website bot". |
| **Recipients** (Email) | Empty | Add at least one email, and/or use **Also send to contact** to pick a contact field holding an email — useful for routing the alert to an account owner stored on the contact. |
| **Slack Integration / Channel ID** (Slack) | — | Select the integration and paste the channel ID (e.g. `C01234ABCDE`). |
## Feedback
Website channel only.
Two independent switches:
* **Message Feedback** — when enabled, users see thumbs up/down buttons on chatbot messages. Cheap signal; turn it on unless your widget design forbids it.
* **Conversation Feedback** — lets users rate the overall conversation when the chat ends. When enabled, choose when the prompt appears (**Show feedback when user:** — **Click on ⭐**, the default, or **Close conversation**), and use the **Customize** modal to rewrite every string: thank-you message, the question text ("Please rate your experience"), input placeholder, skip/submit button labels, success message, the five rating labels (Very Dissatisfied through Very Satisfied), and the low-rating follow-up ("Thanks for your feedback! What could be improved?"). This modal is the one place on the tab that saves manually — click save before closing it.
## Social Feedback
Messenger, Zalo OA, TikTok, and LinkedIn channels only.
Automated feedback collection with a follow-up when the rating comes back low.
| Setting | Default | When to change |
| -------------------------------------- | ------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Rating Threshold** | 3 (range 1–5) | The follow-up is sent when the rating is at or below this. Lower it to 2 if you only want to chase genuinely unhappy customers; 3 also catches the lukewarm middle. |
| **Follow-up Message** | "We're sorry to hear that. Could you tell us more about what went wrong?" | Keep it open-ended — a yes/no question here wastes the one reply you'll get. |
| **Response Wait Time (minutes)** | 10 (minimum 1) | How long a reply still counts as feedback. After the window, a response starts a new conversation instead. Raise it for channels where people reply slowly (LinkedIn); keep it tight for live-feeling channels. |
| **Thank You Message (After Feedback)** | "Thank you for your feedback! This helps us improve our service." | Fine as-is. |
| **Thank You Message (High Ratings)** | "Thank you for your positive feedback! We're glad we could help you." | Sent for 4–5 star ratings if not empty. Clear it if you'd rather stay silent on high ratings. |
## Bridge Configuration
Technical settings for embedding the chat widget inside a native mobile app: a boot timeout and boot-message requirement, the iOS/Android handler function names (default `chatboxTriggerToApp`) and callback (`appTriggerToChatbox`), optional JWT authentication with an RS256 public key, and event mappings that turn external app events into chat messages. If you're not shipping the widget inside a native app, skip this panel entirely; if you are, your mobile developers set these values, not the agent builder.
## Security
**Domain Allowed List** (website channel only) — controls which domains can embed and use your chatbot widget. Empty means all domains are allowed. Two matching modes, and the distinction bites: `example.com` is an exact match — it allows `example.com` and nothing else; `*.example.com` allows `sub.example.com` but **not** `example.com` itself. To cover both the apex and subdomains, add both entries. Lock this down before launch on the website channel — an unrestricted widget can be embedded (and its message budget burned) by anyone.
**Message Limit** — when the limit is reached, the bot stops responding in that thread:
| Setting | Default | When to change |
| -------------------------------- | ----------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Max User Messages Per Thread** | 50 (range 1–500) | A cost and abuse cap. 50 is generous for support; drop it for public-facing marketing bots where long threads are almost always abuse or aimless chatter. Raise it only for genuinely long workflows. |
| **Limit Banner Text** | "You've reached the maximum number of messages for this conversation. Please start a new chat." | Match your brand voice; keep the instruction to start a new chat. |
| **Limit Input Placeholder** | "Message limit reached. Please start a new chat." | Shown in the disabled input box. |
## Schedule Meeting
Single Prompt engine only — Conversational Flow agents use a booking node instead ([Flow Node Features](/docs/conversational-flow/flow-node-features)).
A switch plus a **Calendar App** selector that lets the agent book meetings mid-conversation. The full booking setup — connecting a calendar, availability, event types — is covered in [Scheduling and Booking](/docs/integrations/scheduling-and-booking).
## Contact Collection
Single Prompt engine only — Conversational Flow agents use Contact Collection nodes instead ([Flow Node Features](/docs/conversational-flow/flow-node-features)).
Collects contact information and syncs it to your CRM. When enabled, at least one field is required.
| Setting | Default | When to change |
| -------------------------------------------------------------- | -------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| **Sync contact to** | — | Multi-select of installed apps with contact-sync capability (you'll see **Install App** if none are connected). Connect the CRM first — collected contacts with nowhere to go help nobody. |
| **How do you want to collect contact information?** | **Ask Questions** | **Ask Questions** collects details naturally through conversation — better completion rates, feels less like a gate. **Show Form** displays a form — pick it when you need structured, validated input up front (compliance, precise formats). |
| **Form Title** / **Submit Button Label** / **Success Message** | "Contact Information" / "Submit" / "Thank you! We've received your information." | Form mode only. |
| **Fields** | — | Each field has a name, type, label/question, required flag, description, and validation rules (min/max length, regex, error message). Field choices come from the standard contact fields plus your team's custom fields. Ask for the minimum you need — every required field costs completions. |
## Message Approval
Single Prompt engine only.
Holds outgoing messages for human approval before they're sent — for quality and compliance review of what the agent says at runtime.
**This is not the [publish approval workflow](/docs/governance-access/publish-approval-workflow).** That feature gates publishing a *draft agent version*; this one gates *individual messages* the live agent wants to send. You can use either or both.
| Setting | Default | When to change |
| ------------------------------- | ------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Approval Requirement** | Require approval for all outgoing messages | The strict mode — right for regulated industries or a brand-new agent you don't trust yet. Switch to **Require approval only for campaign messages** once organic replies have earned autonomy but bulk outreach still needs a human check. |
| **Connection Request Approval** | Off | Requires approval before sending connection requests to new leads. Turn on for LinkedIn-style outreach where a bad connection request burns the lead permanently. |
| **Slack Channel ID (Optional)** | Empty | Route approval requests to a Slack channel. If Slack isn't installed, an alert points you to Apps settings. Without a Slack channel, expect slower approvals — messages wait on someone checking the app. |
## Others
**AI Response Settings**
| Setting | Default | When to change |
| --------------------------------- | ---------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **AI Generation Delay (seconds)** | 5 (range 0–1000) | How long the agent waits before generating a response. The default gives customers a beat to finish a multi-message thought — reply too fast and you answer half a question. Lower toward 2–3 for snappy website support; raise on messaging channels where people habitually send four short messages in a row. |
**Automated Messages**
| Setting | Default | When to change |
| -------------------- | ------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Farewell Message** | Off | Sent when conversations end actively. Text is required when on — e.g. "Thank you for chatting with us!". Turn on to close threads politely instead of silently. |
| **Pending Timeout** | Off | Sent when pending conversations time out. Text is required when on — e.g. "We haven't heard from you. Closing soon...". Gives customers a warning before the thread closes on them. |
On Zalo OA and Messenger channels, each of these has a **Feedback Button** sub-switch that attaches a feedback prompt to the message.
**Daily Report** (website channel only) — an **Auto-generate Reports** switch that tracks daily performance metrics. See [Daily Reports](/docs/history-analytics/daily-reports) for what the report contains.
## A configuration pass that works
1. **Escalation first.** Rewrite the **Escalation Criteria** with your product's real failure modes, and set both **Auto-Reply Messages**. A bot that fails gracefully buys you time to tune everything else.
2. **Analysis fields next.** Define your **Analyze Fields** and **Conversation Categories** before launch — you can't retroactively extract structured data from conversations analyzed without them.
3. **Wire up Notification.** One specific trigger, a scannable title, the right channel. This is how you find out what's happening without reading every thread.
4. **Then limits and security.** Set the **Domain Allowed List** (website), sanity-check the **Message Limit**, and decide whether this agent needs **Message Approval** while it earns trust.
5. Leave the rest at defaults until a real conversation shows you the problem a setting solves.
## What's Next
* [Human Takeover](/docs/inbox/human-takeover) — how your team receives and works the conversations this tab escalates.
* [Conversation Analysis Panel](/docs/inbox/conversation-analysis-panel) — where your summary, analyze fields, and categories show up.