Sasamigo

Friends who remember. Sasamigo, by Leinalok Labs, is an AI chat app built with Next.js + Vercel AI SDK + NVIDIA NIM and/or Gemini + Supabase, hosted on Vercel. It includes a public REST API (/api/v1) so a later app (mobile, desktop, another site) can use the same accounts and chats. See API.md.

Features: streaming replies · Markdown and code blocks with copy buttons · chats saved per user · auto-titled, renameable and deletable chats · model picker · starter prompts · stop and retry · friendly error messages · per-user rate limit · mobile layout · dark mode (follows the system setting).

1. Get your keys (free)

1. AI provider (one or both):
   • NVIDIA NIM (fast): https://build.nvidia.com → open any model → Get API Key (starts with nvapi-).
   • Gemini (free tier): https://aistudio.google.com → Get API key.
2. Supabase: create a project at https://supabase.com.
   • SQL Editor: paste the contents of supabase/schema.sql and click Run. Then do the same with supabase/personas-seed.sql to load the built-in personas. Both are safe to re-run, which is also how you upgrade after pulling new changes.
   • Project Settings → API: copy the Project URL and the publishable (or anon) key.
   • Authentication → URL Configuration: set Site URL to your app URL (e.g. http://localhost:3000, later your Vercel URL).

2. Run locally

cp .env.example .env.local   # then fill in the values
npm install
npm run dev

Open http://localhost:3000, sign up, and start chatting. By default Supabase emails a confirmation link on sign-up. For quicker local testing you can turn off Authentication → Sign In / Providers → Email → Confirm email.

3. Deploy to Vercel

1. Push this folder to a GitHub repository.
2. On https://vercel.com, click Add New → Project and import the repository.
3. Add the same environment variables from .env.local (Settings → Environment Variables).
4. Deploy, then put the Vercel URL into Supabase's Site URL (and Redirect URLs).

Project layout

src/
  app/
    page.tsx                         # chat page (requires sign-in)
    login/                           # sign in / sign up
    admin/stats/                     # persona stats (admins only)
    auth/callback/route.ts           # email confirmation links
    api/chat/route.ts                # streaming endpoint for the website UI
    api/v1/...                       # public REST API for other apps (see API.md)
  components/                        # chat UI (sidebar, messages, model picker, markdown)
  lib/
    chat-service.ts                  # shared chat logic: history, Gemini call, saving, rate limit
    models.ts                        # "Terra Fast" etc. → Gemini model IDs
    api.ts                           # API auth (Bearer token or cookie), errors, CORS
    reply-stats.ts, feedback.ts      # reply style metadata, thumbs up/down, admin stats
    supabase/                        # Supabase clients
  proxy.ts                           # keeps the login session fresh; redirects signed-out users
supabase/schema.sql                  # tables + Row Level Security
supabase/personas-seed.sql           # built-in personas (generated, don't edit)
personas/*.json                      # built-in persona character sheets (source of truth)
public/avatars/                      # built-in persona avatars

Personas

Each persona is a structured character sheet (src/lib/persona-schema.ts): identity, personality, behaviour rules, communication style with situational examples, knowledge, and relationship. Built-in personas live in personas/*.json. After editing one:

npm run personas:seed   # validates the JSON and regenerates supabase/personas-seed.sql

Then run supabase/personas-seed.sql in the Supabase SQL Editor.

Memory and check-ins: after each reply, a fast model extracts lasting facts (about the user, and details the persona said about itself) into memories; a grounding check drops anything that wasn't actually said. Users can see and delete them via the 🧠 Memory button. relationships tracks how much they've talked; the stage (new → warm → close) needs both messages and days. When a user reopens a persona chat after the persona's checkIn.afterHours, the persona texts first (never stacked; gentle wellbeing check only if they were struggling).

Safety layers: crisis language and "I'm under 18" statements are detected in code and add an urgent instruction to that reply; crisis messages also always show a helpline card in the UI (and crisis_support in the API).

Learning from user interaction

Every reply is saved with style metadata (words, emoji, bubbles, whether it asks a question, time to the first word, a prompt_version that changes when the persona or prompt builder changes) and a reply_id shared by its bubbles. Users can rate replies with 👍 / 👎 (optional reason). Admins get a stats page at /admin/stats with, per persona and model: thumbs-up rate, how often and how fast users answer, reply style, latency, day-1/day-7 return and messages per session; check-ins are shown separately.

• Only metadata is used: the stats functions never read message text and return no user ids.
• Messages around a crisis (and the 24 hours after) are left out. Groups with fewer than 5 users show counts only.
• Main measures are thumbs up and day-7 return, not message volume, so we don't optimise for clinginess.
• Make yourself an admin once in the SQL Editor (see the bottom of supabase/schema.sql): insert into public.admins (user_id) select id from auth.users where email = '[email protected]';
• Bump PROMPT_VERSION in src/lib/prompt-builder.ts when you change the prompt layers.

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗