اپلیکیشن بدون کدنویسی
Animam.ai
An agent that answers from your site's content, captures leads, quotes and books · server-side.
نتیجه ارزیابی
بخش چت در یک آخر هفته ساخته میشود؛ تمام چیزهایی که آن را برای ارائه مستقیم به مشتری امن میکند، خیر. ایندکس کردن سایت، جستجو و استریم کردن پاسخ امروزه یک کار ساده و در دسترس است و پرامپت زیر واقعاً آن را انجام میدهد. آنچه مقاومت میکند نیمه ملالآور و سخت ماجراست: مبالغی که فقط سمت سرور محاسبه میشوند و هرگز توسط مدل حدس زده نمیشوند، ایمیل کاربری که قبل از اجرای هر عملیاتی تایید میشود، وبهوکهای امضاشده همراه با تلاش مجدد، و دامنه ارسالی که اعتبار آن را نمیشود در یک بعدازظهر ایجاد کرد.
- زمان تقریبی ساخت
- یک آخر هفته
- هزینه تقریبی اجرا
- نامشخص
- اشتراک محصول اصلی
- ۳۳.۵۲ دلار در ماه
چه چیزی را از دست میدهی؟
- محاسبه قیمتها در سمت سرور بر اساس لیست قیمت؛ مدل فقط کالا را انتخاب میکند و هیچوقت عدد نمیسازد
- تایید هویت کاربر با کد یکبارمصرف قبل از اجرای هرگونه عملیات سرور به سرور
- وبهوکهای خروجی امضاشده با تأخیر تصاعدی، گزارش ارسال و ارسال مجدد دستی
- حفاظت SSRF با تحلیل DNS روی هر نشانی وب که عامل مجاز به فراخوانی آن است
- رمزنگاری اطلاعات محرمانه در حالت سکون برای هر مشتری و جداسازی کامل محیطها بدون دغدغه فکری
- نرخ تحویل ایمیل؛ اعتبارسنجی و گرم کردن دامنه چیزی نیست که با یک پرامپت آماده شود
- آن شبی که ارائهدهنده مدل تنظیمات پیشفرض را تغییر میدهد و ابزار شما شروع به پاسخهای ساختگی و اشتباه میکند
پرامپت شروع
برای شروع ساخت، مستقیم کپیاش کن.
Build a website chat widget that answers strictly from a site's own content. Node 20 + TypeScript, no framework. 1. Install hono, @anthropic-ai/sdk, cheerio, turndown, better-sqlite3. 2. scripts/ingest.ts takes a sitemap.xml URL as argv and fetches every page, concurrency 4, skipping non-HTML. 3. Strip nav, header, footer and script tags with cheerio, then convert what is left to markdown with turndown. 4. Upsert into SQLite pages(url PRIMARY KEY, title, body, fetched_at) so re-running updates rows instead of duplicating them. 5. src/search.ts takes the 3 to 6 most distinctive words of the question and runs SELECT url, title, substr(body,1,1200) FROM pages WHERE body LIKE ? for each. 6. Rank by number of hits and return the top 5. No embeddings: under ~500 pages LIKE is cheaper, instant, and you can read why a passage was picked. 7. src/server.ts is a Hono app with POST /chat {message, sessionId}. 8. Put the 5 passages and their URLs in the system prompt, call claude-haiku with stream:true, and pipe SSE to the client. 9. System prompt: "Answer only from the CONTEXT below, and cite the page you used." 10. Continue it: "If the answer is not in the context, say you do not know and offer to take a message." 11. And the line that matters: "Never state a price, a date or an availability that is not written in the context." 12. public/widget.js is an IIFE that injects a bubble into a shadow DOM and keeps sessionId in localStorage. 13. It consumes the SSE and writes every message with textContent, never innerHTML · the text comes from a model that read the internet. 14. Store sessions and messages in the same SQLite file, so you can read back what people actually asked. 15. ANTHROPIC_API_KEY lives in .env, server-side only, and never reaches the bundle. 16. No telemetry, no third-party script in the widget. 17. Serve /widget.js with CORS * and a 5 minute cache so a single <script> tag installs it on any page. Out of scope on purpose, and each of these is a week rather than a line: booking on a real calendar, payments, quotes with amounts computed server-side, human handover, transactional email that reaches the inbox, multi-tenancy, GDPR consent, rate limiting and abuse protection.
برای ساختش کمک میخوای؟
به درد یکی دیگه هم میخوره؟
