Syntopia BOX Guide
If you want AI that answers live stream comments the moment they land — by name, with the right price, the right size, the right shipping date — the feature that does it is called Smart Live Chat. It is the difference between a stream that looks pre-recorded and one that genuinely talks back.

Most sellers who want auto reply on TikTok LIVE picture a bot pasting canned lines. Smart Live Chat is the opposite: an LLM reads the live chat as it scrolls, finds the answer in a knowledge base you wrote, and speaks or types it back in real time. No two answers are identical, and the host never goes silent while a question sits unanswered.
This guide covers the two chat systems inside Syntopia BOX, the three ways an answer can reach the viewer (Host, Chat message, Assistant), and — the part that actually decides answer quality — how to build the ~20,000-character knowledge base so the host answers accurately instead of guessing. It is a deeper companion to the Syntopia BOX complete guide.
- Two systems, one tab. Proactive Live Chat starts conversations (welcomes, follow/like nudges, buy prompts); Smart Live Chat answers the questions viewers actually ask.
- Live Chat Method decides who delivers the answer: Host (spoken out loud), Chat message (typed into the chat), or Assistant (a second AI voice). Tick one, two, or all three.
- The knowledge base is the engine. Up to 20,000 characters of your own facts — prices, variants, shipping, returns, best answers. Better knowledge base, better answers.
- Answering in real time is the compliant way to run an AI live. Platforms restrict streams that loop and never interact; a host that replies instantly is the opposite of a loop.
Why real-time answers matter (not just for engagement)
Buyers in a live room ask the same high-intent questions over and over: “Does it come in black?”, “How much with shipping?”, “Is this the 500ml or the 250ml?” A host that ignores them loses the sale in the few seconds the viewer is deciding. A host that answers — by name, correctly — closes it.
There is a second reason, and it is about keeping the account alive. Selling platforms want real, interactive live broadcasts. A stream that reads as a recording on a loop — a voice that never responds to anyone — is exactly what gets restricted. Real-time interaction is the single clearest signal that a live is live. Smart Live Chat is how an AI-hosted stream produces that signal continuously: it reads every comment and answers it, so the room never looks unattended. (Pair it with the platform’s AI-content label, and you are running an AI live the way the policies describe.)
The rules are not anti-AI. They are anti-fake-live. A host that answers every comment in real time is the most live thing a stream can do.
The two chat systems: Proactive vs Smart Live Chat
Open your live stream in Syntopia BOX and go to the AI Chat tab. There are three settings plus the knowledge base. The first two are two different jobs, and you will usually want both on.
Proactive Live Chat
This is the host taking initiative with no comment required. It welcomes new viewers as they join, asks for follows and likes, and nudges people toward the product — the steady patter that keeps a room warm during quiet stretches. It is on by default, and you customise the wording under Engagement Prompts. Think of it as the host’s opening lines and filler; it does not need the knowledge base to work.

Smart Live Chat
Turn this on and an LLM reads the live chat in real time, matches each question to your knowledge base, and answers it. This is the feature that produces “AI that answers live stream comments.” It depends entirely on what you put in the knowledge base — an empty base means vague or made-up answers, a well-built one means accurate answers by name. Most of your setup time should go here.

| Proactive Live Chat | Smart Live Chat | |
|---|---|---|
| Job | Starts conversations | Answers questions |
| Trigger | New viewer, timed gap | A comment / question in chat |
| Typical output | Welcome, follow/like, buy nudge | Price, size, shipping, returns, product facts |
| Needs knowledge base? | No | Yes — it is the source of every answer |
| Where you edit it | Engagement Prompts | Knowledge base + Live Chat Method |
| Default state | On | Off until you turn it on |
Live Chat Method: who actually answers
Once Smart Live Chat is on, Live Chat Method decides how the answer reaches the viewer. You can tick one, two, or all three — they are not exclusive.
How to choose: for a solo digital-human stream, Host alone feels most natural. If you want the host to keep selling without interruption, add the Assistant so a second voice fields questions. Tick Chat message as well when answers carry precise data a viewer will want to re-read — sizes, prices, delivery dates. Running an AI Hosting Assistant stream (your own camera, no avatar, AI voice only)? There you tick Assistant so the AI answers when asked. To let the assistant answer inside a normal digital-human stream, go to AI Chat → Live Chat Method and tick Assistant.
- Host answers in the main host’s voice and face.
- Assistant is a separate voice you enable in Settings → Assistant Voice → Enable; pick a timbre different from the host. If you trained with two people, the assistant appears on the right of the screen; with one person, it is voice only.
- The assistant that answers chat is the same voice that reads your (round-bracket) script lines — one voice, two jobs.
Build the knowledge base so it answers accurately
This is where answer quality is won or lost. The host does not invent facts; it pulls them from the knowledge base you paste in — up to 20,000 characters. Treat those characters as a budget and spend them on the things buyers ask about. Below is a structure that uses the space well.
A knowledge-base template that works
Write it in plain text, grouped under clear headings so the model can find the right fact fast. A workable split of the 20,000 characters:
| Section | What goes in it | Rough budget |
|---|---|---|
| Product facts | Exact product names, every variant/colour/size, prices, what’s in the box, materials, dimensions, how to use it | ~8,000 chars |
| Shipping & delivery | Where you ship, cost, dispatch time, delivery windows per region, tracking | ~2,500 chars |
| Returns, refunds & warranty | Window, condition, who pays return postage, how refunds are issued, warranty length | ~2,500 chars |
| Offers & bundles | Today’s price, bundle deals, discount codes, quantity breaks, what’s live right now | ~2,000 chars |
| Best answers to FAQs | The 20–30 questions you hear every stream, each with your ideal one-line answer | ~3,500 chars |
| Off-limits / brand voice | Claims never to make, topics to deflect, tone and persona notes | ~1,500 chars |
Budget is a suggested split, not a hard limit — move characters to whatever your buyers ask about most.
Rules that make the answers better
- Write facts as the model should say them. If a question is “how long is delivery to the US?”, have a line that reads “US delivery: 5–8 business days, tracked.” The closer your text is to a good spoken answer, the better it sounds live.
- Name things exactly as buyers type them. Include nicknames and misspellings viewers use, so a question still matches the right fact.
- Front-load prices and variants. These are the questions that decide a sale; make them the easiest facts to find.
- Add the 20–30 real questions from your last streams. Pull them from chat logs or a recording — the Script Transcription tool turns a past live into text you can mine for them.
- State what not to say. Medical claims, guarantees you can’t keep, competitor comparisons — list them so the host deflects instead of improvising.
- Keep it current. Update prices and “what’s live today” before each stream; a stale base gives confidently wrong answers.
Step-by-step: turn it on
- Connect an AI model. Under AI model, paste an API key from a provider such as OpenAI or DeepSeek — this is the brain that reads chat and writes answers.
- Open your live stream → AI Chat tab. Leave Proactive Live Chat on and tune its lines under Engagement Prompts.
- Turn on Smart Live Chat.
- Paste your knowledge base (up to 20,000 characters), structured as above.
- Set Live Chat Method. Tick Host, Chat message, and/or Assistant. If you use Assistant, enable the Assistant Voice in Settings first and choose a different timbre.
- Save and Publish, then go live and route the picture and voice into your platform’s app as usual. Once connected, the host reads comments and answers viewers on its own.
- In a digital-human stream, the on-screen host is an AI avatar — an AI, which you disclose with your platform’s AI-content label.
- In Human Takeover the face is a real person on camera (the AI only drives lip-sync to the script and answers) — not an AI avatar in that mode.
- Either way, Smart Live Chat is what makes the mouth you see respond to real questions in real time.
Frequently asked questions
What is Smart Live Chat in Syntopia BOX?
Smart Live Chat is the setting that makes an AI host read your live chat and answer viewers’ questions in real time. When it is on, an LLM matches each incoming comment to your custom knowledge base and delivers the answer out loud, as a typed chat reply, or in a second assistant voice. It is separate from Proactive Live Chat, which starts conversations with welcomes and follow/like nudges rather than answering questions.
How does the AI know the right answer to a comment?
It pulls from a custom knowledge base you paste in — up to 20,000 characters of your own prices, variants, sizes, shipping, returns and best answers to common questions. The host does not invent facts; the quality of its answers tracks the quality of that knowledge base, so structure it clearly and keep prices and offers current before each stream.
Can the AI answer out loud and in the chat at the same time?
Yes. Live Chat Method lets you tick Host (spoken out loud), Chat message (typed into the live chat), and Assistant (a second AI voice) in any combination. A common setup is Host for natural spoken replies plus Chat message for precise details like sizes, prices and delivery windows that a viewer will want to re-read.
What is the difference between the Host and the Assistant answering?
Host means the main AI host speaks the answer in its own voice and face. Assistant means a second AI voice answers while the host keeps presenting; you enable it under Settings → Assistant Voice and give it a different timbre. If you trained your avatar with two people, the assistant appears on the right of the screen; with one person it is voice only. The same assistant voice also reads any script lines you write in round brackets.
Does answering comments in real time help with platform rules?
It is central to them. Selling platforms restrict streams that look pre-recorded — a loop that never responds to anyone. A host that reads every comment and answers it instantly is the clearest sign a stream is genuinely live. Run Smart Live Chat and switch on the platform’s AI-generated content label, and you are operating an AI live the way the policies describe: interactive and disclosed.
Do I need to write every answer myself?
No. You write the facts (the knowledge base) and the model phrases the answers from them on the fly, so no two replies are worded identically. You only pre-write the 20–30 recurring questions where you want an exact answer; everything else the host composes from your facts. See the Syntopia BOX complete guide for the full live-room setup, and pricing for plans.