@core / ai

1.0.0 ▾

1.0.0

vérifiéMIT

GitHub

Chat IA en streaming : prompts versionnés, appels d’outils, suivi d’usage et de coût ; modèle via env

Code11 fichiersContexte~668 tokensanalyse réussie

~668 tokens→ src/lib/ai/AGENTS.md→ .cursor/rules/genpm-core-ai.mdc

Voici exactement ce que lit votre IA quand elle travaille dans src/lib/ai. Rien d’autre n’est ajouté à son contexte.

@core/ai — rules for AI agents

Purpose

Streaming chat endpoint on the Vercel AI SDK, with versioned system prompts, a tool-calling scaffold and per-response usage logging (tokens and cost) in the database. Provider and model come from env: nothing is hardcoded. No RAG, embeddings or chat history storage (store UIMessage[] yourself if you need it).

Map

index.ts — public API: streamChat, getModel, setModel, recordUsage, prompts, tools, PRICES.
prompts.ts — your config: system prompts with version. Bump version whenever you change a prompt.
tools.ts — your config: tools the model can call (zod inputSchema + server-side execute).
prices.ts — your config: USD per million tokens per AI_MODEL, for cost logging.
schema.ts — ai_usage table. Depends on ../db (@core/db).
adapters/hono.ts, adapters/next.ts — POST /ai/chat returning a UI message stream for useChat.

Integration

Requires Node 22+ (AI SDK 7). Env: AI_MODEL as <provider>:<model-id> plus AI_PROVIDER_API_KEY, and install that provider package: anthropic → @ai-sdk/anthropic, openai → @ai-sdk/openai, also google, mistral, groq, xai. Or AI_MODEL=<provider>/<model-id> with AI_GATEWAY_API_KEY (Vercel AI Gateway, no extra package). Ask the user which provider and model to use; do not pick one for them.
Generate and apply migrations (see src/lib/db/AGENTS.md).
Hono: app.route('/ai', aiRoutes({ getUserId: (c) => c.get('user')?.id, requireUser: true })) after @core/auth's sessionMiddleware (omit both options without auth). Next.js: app/api/ai/chat/route.ts → export const POST = chatRoute({ getUserId, requireUser: true }). Delete the adapter of the framework you don't use.
Frontend: useChat({ transport: new DefaultChatTransport({ api: '/ai/chat' }) }) from @ai-sdk/react.
Verify: curl -N -X POST localhost:3000/ai/chat -H 'content-type: application/json' -d '{"messages":[{"id":"1","role":"user","parts":[{"type":"text","text":"hi"}]}]}'.

Conventions

One prompt per use case in prompts.ts; pass it with streamChat({ prompt: 'assistant', … }).
Tools authorize inside execute using the user id; never trust ids sent by the model.
Query ai_usage for quotas (e.g. tokens per user per day) before calling streamChat.
Tests: setModel(new MockLanguageModelV4(...)) from ai/test.

Don't

Don't hardcode a model or API key, and don't call the provider from the browser.
Don't log prompts or completions with user data in production.
Don't let tools run destructive actions without an explicit user confirmation step.

# @core/ai — rules for AI agents

## Purpose
Streaming chat endpoint on the Vercel AI SDK, with versioned system prompts, a tool-calling scaffold and per-response
usage logging (tokens and cost) in the database. Provider and model come from env: nothing is hardcoded. No RAG,
embeddings or chat history storage (store `UIMessage[]` yourself if you need it).

## Map
- `index.ts` — public API: `streamChat`, `getModel`, `setModel`, `recordUsage`, `prompts`, `tools`, `PRICES`.
- `prompts.ts` — **your config**: system prompts with `version`. Bump `version` whenever you change a prompt.
- `tools.ts` — **your config**: tools the model can call (zod `inputSchema` + server-side `execute`).
- `prices.ts` — **your config**: USD per million tokens per `AI_MODEL`, for cost logging.
- `schema.ts` — `ai_usage` table. Depends on `../db` (@core/db).
- `adapters/hono.ts`, `adapters/next.ts` — `POST /ai/chat` returning a UI message stream for `useChat`.

## Integration
1. Requires Node 22+ (AI SDK 7). Env: `AI_MODEL` as `<provider>:<model-id>` plus `AI_PROVIDER_API_KEY`, and install
   that provider package: `anthropic` → `@ai-sdk/anthropic`, `openai` → `@ai-sdk/openai`, also `google`, `mistral`,
   `groq`, `xai`. Or `AI_MODEL=<provider>/<model-id>` with `AI_GATEWAY_API_KEY` (Vercel AI Gateway, no extra package).
   Ask the user which provider and model to use; do not pick one for them.
2. Generate and apply migrations (see `src/lib/db/AGENTS.md`).
3. Hono: `app.route('/ai', aiRoutes({ getUserId: (c) => c.get('user')?.id, requireUser: true }))` after
   `@core/auth`'s `sessionMiddleware` (omit both options without auth).
   Next.js: `app/api/ai/chat/route.ts` → `export const POST = chatRoute({ getUserId, requireUser: true })`.
   Delete the adapter of the framework you don't use.
4. Frontend: `useChat({ transport: new DefaultChatTransport({ api: '/ai/chat' }) })` from `@ai-sdk/react`.
5. Verify: `curl -N -X POST localhost:3000/ai/chat -H 'content-type: application/json' -d '{"messages":[{"id":"1","role":"user","parts":[{"type":"text","text":"hi"}]}]}'`.

## Conventions
- One prompt per use case in `prompts.ts`; pass it with `streamChat({ prompt: 'assistant', … })`.
- Tools authorize inside `execute` using the user id; never trust ids sent by the model.
- Query `ai_usage` for quotas (e.g. tokens per user per day) before calling `streamChat`.
- Tests: `setModel(new MockLanguageModelV4(...))` from `ai/test`.

## Don't
- Don't hardcode a model or API key, and don't call the provider from the browser.
- Don't log prompts or completions with user data in production.
- Don't let tools run destructive actions without an explicit user confirmation step.

L’arborescence exacte qui sera injectée, après .genpmignore. Épinglée à

src/lib/ai/model.tslecture seule · 5b1e62b

// Modelo por entorno, nunca fijo en el código:
//   AI_MODEL="anthropic:<model-id>" (o "openai:…", "google:…", "mistral:…", "groq:…", "xai:…") + AI_PROVIDER_API_KEY
//   e instala el paquete del proveedor (`@ai-sdk/anthropic`, `@ai-sdk/openai`…).
//   AI_MODEL="<provider>/<model>" usa el AI Gateway de Vercel (AI_GATEWAY_API_KEY), sin paquete extra.
import type { LanguageModel } from 'ai';

const FACTORIES: Record<string, { pkg: string; fn: string }> = {
  anthropic: { pkg: '@ai-sdk/anthropic', fn: 'createAnthropic' },
  openai: { pkg: '@ai-sdk/openai', fn: 'createOpenAI' },
  google: { pkg: '@ai-sdk/google', fn: 'createGoogleGenerativeAI' },
  mistral: { pkg: '@ai-sdk/mistral', fn: 'createMistral' },
  groq: { pkg: '@ai-sdk/groq', fn: 'createGroq' },
  xai: { pkg: '@ai-sdk/xai', fn: 'createXai' },
};

let override: LanguageModel | null = null;

/** Fija el modelo a mano (tests, o un proveedor que no está en la lista). */
export function setModel(model: LanguageModel | null): void {
  override = model;
}

export function modelId(env: NodeJS.ProcessEnv = process.env): string {
  if (override) return typeof override === 'string' ? override : `${override.provider}:${override.modelId}`;
  const id = env.AI_MODEL;
  if (!id) throw new Error('AI_MODEL is not set, e.g. "anthropic:<model-id>" (see src/lib/ai/AGENTS.md)');
  return id;
}

export async function getModel(env: NodeJS.ProcessEnv = process.env): Promise<LanguageModel> {
  if (override) return override;
  const id = modelId(env);
  const sep = id.indexOf(':');
  if (sep < 0) return id; // "<provider>/<model>" → AI Gateway
  const provider = id.slice(0, sep);
  const name = id.slice(sep + 1);
  const f = FACTORIES[provider];
  if (!f) throw new Error(`Unknown AI provider "${provider}". Use one of ${Object.keys(FACTORIES).join(', ')} or setModel().`);
  let mod: Record<string, unknown>;
  try {
    mod = (await import(f.pkg)) as Record<string, unknown>;
  } catch {
    throw new Error(`Install ${f.pkg} to use AI_MODEL=${id}`);
  }
  const create = mod[f.fn] as (opts: { apiKey?: string }) => (model: string) => LanguageModel;
  return create({ apiKey: env.AI_PROVIDER_API_KEY })(name);
}