不會。智慧輔助只會選擇既有字卡組，再從字卡庫中抽取回覆。系統可能會在裝置端使用 Apple 提供的語言處理框架，例如 Natural Language framework，以及在相容裝置上的 Apple Intelligence / Foundation Models，協助選擇較合適的分類；它不會替角色生成新的對話內容，也不會把聊天訊息、字卡文字、觸發關鍵字或分類結果上傳至開發者伺服器。
不会。智能辅助只会选择既有字卡组，再从字卡库中抽取回复。系统可能会在设备端使用 Apple 提供的语言处理框架，例如 Natural Language framework，以及在兼容设备上的 Apple Intelligence / Foundation Models，协助选择较合适的分类；它不会替角色生成新的对话内容，也不会把聊天消息、字卡文字、触发关键词或分类结果上传至开发者服务器。
いいえ。スマートアシストは既存の定型文セットを選び、定型文集から返信を選ぶだけです。端末上で Apple の Natural Language framework や、対応端末では Apple Intelligence / Foundation Models を分類補助に使う場合がありますが、新しい会話文は生成せず、チャットメッセージ、定型文本文、トリガーキーワード、分類結果を開発者のサーバーへアップロードすることもありません。
This page helps you tune chat phrase libraries, phrase sets, and trigger keywords. Smart Assist uses the latest incoming message, phrase set name, usage context, trigger keywords, and chat phrase text to choose the most relevant phrase set. Trigger keywords are the clearest signal for guiding a topic.
1. Three Things to Remember
Write what you might send to the character, not what the character should say back.
Specific keywords that feel like a complete situation are easier to match.
Very short or very common words can make the decision ambiguous.
Smart Assist mainly evaluates the current user turn. If you send several short messages before the character replies, DreamChat treats those nearby user messages as the current context. Quoted replies or the character's previous message are used only as weaker context.
2. Choosing a Smart Assist Mode
Off: returns to regular phrase selection without Smart Assist phrase-set judgment.
Natural: the recommended default. Clear matches stay steady; close reasonable contexts can branch naturally.
Precise: favors the top-scoring phrase set, useful when you want the most stable classification.
Varied: allows a wider reasonable candidate pool so open-ended messages can vary more.
Smart Assist decides which primary phrase set to use. Randomness applies after that choice, controlling whether replies can mix in General or other phrase sets. Lower randomness keeps replies steadier; Natural or Varied can still choose between several reasonable primary sets when their scores are close.
3. Writing Better Keywords
Use words that would appear in a real message
Think like the person chatting. Instead of only writing "comfort," use phrases like I'm tired today, I'm stressed, I want to cry, or I got scolded. Instead of only writing "sleep," use good night, I can't sleep, stayed up late, or dreamed about you.
Prioritize unique nouns and short complete phrases
Character names, nicknames, titles, places, events, and words that only belong to one phrase set are strong clues. Phrases like are you jealous, I dreamed about you, my manager scolded me, or exam tomorrow usually work better than single common words.
Add several angles to one phrase set
Instead of one long sentence, add 5 to 12 natural variations that point to the same situation. For a bedtime set, you might include good night, I can't sleep, staying up late, I'm sleepy, dream about you, and come to my dream tonight.
Give similar sets clear differences
If two sets are similar, write the words that separate them. For Chinese interaction direction, write the subject and object clearly: you comfort me and I comfort you, or you help me and I help you, are different situations.
Use context for background and keywords for decisive clues
Usage Context is best for a readable description of the emotion, scene, or relationship where a phrase set should be used. Trigger Keywords are best for short phrases, nouns, nicknames, and situation clues that may appear directly in a message.
Do not rely only on chat phrase text. Chat phrases are what the character says back; they may not match what you would send to the character. For important phrase sets, fill in the set name, usage context, and trigger keywords.
4. Quick Checklist
Are these keywords things you or the other chatter might actually type?
Do you have several specific keywords longer than one character or word?
Are you avoiding relying only on common words like okay, you, me, or today?
Do similar phrase sets include the words that make them different?
Does the phrase set have enough chat phrases to draw from?
5. FAQ
I set it up, but this phrase set was not chosen.
Add the actual message that triggered the reply to Trigger Keywords, or split it into a few natural short phrases. Also check whether the set only uses broad words like "tired," "today," or "sad"; those usually need more specific context.
The right phrase set is chosen, but some replies still drift.
That is usually related to Randomness. Smart Assist chooses the best phrase set first; Randomness then decides whether General or other phrase sets can mix in. Lower Randomness and add more usable phrases inside the matched set for steadier replies.
Does Apple Intelligence generate replies?
No. Smart Assist only chooses from existing phrase sets, then DreamChat draws from your chat phrase library. The app may use Apple-provided language processing frameworks, such as the Natural Language framework, and on compatible devices Apple Intelligence / Foundation Models, on device to help choose a category. It does not generate new conversation text, and it does not upload chat messages, chat phrase text, trigger keywords, or classification results to the developer's servers.
