AI configuration

Source: src/lib/ai/providers.ts, src/lib/ai/with-fallback.ts, src/lib/ai/generate.ts, src/lib/ai/embed.ts.

Provider IDs

const AI_PROVIDER_IDS = ["openai", "gemini"] as const;
type AiProvider = "openai" | "gemini";

Provider config

export const EMBEDDING_DIMENSIONS = 768;
export const DEFAULT_AI_PROVIDER: AiProvider = "openai";
export const AI_PROVIDER_STORAGE_KEY = "command-inbox-ai-provider-v2";

Client storage

User provider preference persists in localStorage under command-inbox-ai-provider-v2.

Client-safe helpers live in src/lib/ai/alternate-provider.ts (no server-only imports).

Provider chain

getProviderChain(preferred: AiProvider): AiProvider[]

Returns [preferred, alternate] — typically ["openai", "gemini"] or reversed if user selected Gemini.

withProviderFallback() wraps async AI calls:

1. Execute with primary provider
2. On rate limit / quota / retryable error → retry with fallback
3. Propagate error if both fail

Chat generation

Uses Vercel AI SDK generateText / streamText with model instances from @ai-sdk/openai and @ai-sdk/google.

Embeddings

embedText(text, provider) returns 768-dim vector.

Stored on classifications.embedding with embedding_provider column tracking source.

Re-embed: src/lib/embeddings/reembed-inbox.ts — batch updates all user classifications when provider switches.

Rate limiting

src/lib/ai/rate-limit.ts — in-memory sliding window for burst protection during development.

Draft fallback

src/lib/ai/drafts.ts — when AI generation fails, returns template HTML with source: "template".

Server-only boundary

These modules import server-only:

• src/lib/env.ts
• src/lib/ai/models.ts
• src/lib/ai/generate.ts
• src/lib/ai/embed.ts

Client components must use API routes or alternate-provider.ts helpers.

Env dependency

assertPhase2Env() // requires GOOGLE_GENERATIVE_AI_API_KEY || OPENAI_API_KEY

parseAiProvider

parseAiProvider(value: unknown, fallback = DEFAULT_AI_PROVIDER): AiProvider

Safely parses user/API input; invalid values fall back to openai.