Aime is a chatbot that replicates a real person's texting style — not by fine-tuning, but by statistically analyzing thousands of real chat messages and encoding the patterns into a system prompt.
Aime 是一个复刻真人聊天风格的聊天机器人——不靠微调，而是通过统计分析数千条真实聊天记录，将模式编码进 system prompt。
Multi-bubble replies arrive with natural typing delays proportional to message length, simulating real human texting rhythm.
多条消息气泡以与消息长度成正比的打字延迟逐条到达，模拟真实的聊天节奏。
No model training required. Instead, we extract statistical patterns from real WeChat conversations and encode them directly into a structured system prompt.
无需模型训练。我们从真实微信对话中提取统计模式，直接编码进结构化的 system prompt。
Parse raw HTML exports with BeautifulSoup. Merge consecutive messages, convert emoji tags, filter media.
用 BeautifulSoup 解析原始 HTML 导出文件。合并连续消息，转换 emoji 标签，过滤媒体消息。
Extract (context, response) pairs as JSONL — up to 5 context messages per response.
提取 (上下文, 回复) 对，以 JSONL 格式存储——每条回复最多配 5 条上下文消息。
Quantify 7 dimensions: message length, punctuation, filler words, emoji usage, multiline tendency, language mix, and overall style summary.
量化 7 个维度：消息长度、标点习惯、口头禅、emoji 使用、换行倾向、中英混杂度和整体风格总结。
Algorithmically pick 20 diverse examples: 1/3 catchphrases, 1/3 length-balanced, 1/3 random.
算法选取 20 条多样化示例：1/3 口头禅对话、1/3 长度均衡、1/3 随机采样。
Merge QA profile + style rules + examples + guardrails into a ~4000-char structured prompt. No weights modified.
合并 QA 问卷 + 风格规则 + 对话示例 + 安全边界，生成约 4000 字的结构化 prompt。无需修改任何模型权重。
From the moment you hit send to the typing dots disappearing — every message travels through this serverless pipeline.
从你按下发送到打字指示器消失——每条消息都经过这个 serverless 管道。
React frontend captures message + session context
React 前端捕获消息和会话上下文
Builds prompt with sliding context window (last 20 turns)
用滑动上下文窗口（最近 20 轮）构建 prompt
Generates reply with persona system prompt + conversation history
基于人格 system prompt 和对话历史生成回复
Logs conversation for review & rating (90-day TTL)
记录对话用于审查和评分（90 天 TTL）
Multi-bubble display with staggered typing delays
多气泡展示，配合交错的打字延迟
Minimal dependencies, maximum capability. Each piece chosen for simplicity and performance.
极简依赖，最大能力。每个组件都为简洁和性能而选。
Latest React with full type safety. Vite-powered dev with instant HMR.
最新 React，完整类型安全。Vite 驱动开发，即时 HMR。
FrontendGoogle's fast multimodal model. Low latency for chat-like response times.
Google 快速多模态模型。低延迟，适合聊天级响应速度。
AI / LLMServerless Redis for conversation logging. Per-request pricing, zero cold starts.
Serverless Redis 用于对话日志。按请求计费，零冷启动。
DatabaseServerless API endpoints. Auto-scaling, edge-optimized, zero config deployment.
Serverless API 端点。自动扩缩容，边缘优化，零配置部署。
BackendLightning-fast build tool. ESM-native, sub-second cold starts in dev.
极速构建工具。原生 ESM，开发环境亚秒级冷启动。
BuildBuilt-in conversation review page for rating quality by session. Export to JSON for iterative prompt tuning.
内置对话审查页面，按会话评分质量。导出 JSON 用于迭代优化 prompt。
ToolingThe details that make an AI chatbot feel like a real person texting you back.
这些细节让 AI 聊天机器人的回复像真人在发消息。
Instead of handwriting style rules, every number in the prompt (14.2 avg chars, 97% no punctuation, 18.2% emoji) comes from statistical analysis of real chat data.
不靠手写风格规则，prompt 中的每个数字（14.2 平均字符、97% 无标点、18.2% emoji）都来自真实聊天数据的统计分析。
Responses split into 1-3 short messages with staggered typing delays proportional to message length. Mimics real texting rhythm instead of a single wall of text.
回复拆分为 1-3 条短消息，配合与消息长度成正比的交错打字延迟。模拟真实聊天节奏，而非一整段文字。
Sliding window of last 20 messages sent to the model. Older messages referenced via a system note to maintain conversational continuity without hitting token limits.
最近 20 条消息的滑动窗口发送给模型。更早的消息通过系统备注引用，在不超 token 限制的前提下保持对话连贯性。
Standalone review page for rating conversations by session. Filter by quality, export JSON. Creates a feedback loop for iteratively improving the system prompt.
独立的审查页面，按会话评分对话质量。按质量筛选，导出 JSON。形成反馈闭环，持续迭代优化 system prompt。
Strict rules against leaking personal info (address, phone, ID), company internals, or fabricating facts. Graceful deflection with humor when pushed.
严格禁止泄露个人信息（地址、电话、身份证）、公司内部信息或编造事实。被追问时以幽默方式优雅回避。
Casual mode (default): short, playful, 5-20 chars. Knowledge mode: triggered by tech questions, allows longer answers with the same personality voice.
闲聊模式（默认）：简短、俏皮、5-20 字。知识模式：由技术问题触发，允许更长回答但保持同样的个性语气。
The Python prototype includes a full RAG pipeline — but is it worth the complexity?
Python 原型包含完整的 RAG 管道——但值得增加这个复杂度吗？
Niche topics. Surfacing specific shared memories or rare catchphrases that the static prompt can't cover.
小众话题。挖掘静态 prompt 无法覆盖的特定共同记忆或罕见口头禅。
Casual chat. Statistical style rules already handle common patterns. RAG adds latency without meaningful improvement.
日常闲聊。统计风格规则已能处理常见模式。RAG 增加延迟却无明显提升。
Verdict: Prompt-only wins for now. Single API call (~800ms), zero infrastructure, edit one file to improve. RAG reserved for when the bot needs factual recall across conversation partners.
结论：目前 Prompt-only 方案胜出。单次 API 调用约 800ms，零基础设施，改一个文件就能迭代。RAG 留作后备——等到机器人需要跨对话伙伴的事实回忆能力时再启用。