index.ts

import {createEnv} from "@t3-oss/env-core";
import {z} from "zod";

export const env = createEnv({
    server: {
        SUPABASE_URL: z.string().url(),
        SUPABASE_SERVICE_ROLE_KEY: z.string().min(1),
        INTELLIGENCE_PROVIDER: z.enum(["ollama", "gateway", "anthropic", "jev"]).default("ollama"),
        // Provider for features that generate text (reply suggestions, knowledge drafts). Defaults to INTELLIGENCE_PROVIDER, or ollama when that is jev.
        TEXT_PROVIDER: z.enum(["ollama", "gateway", "anthropic"]).optional(),
        AI_GATEWAY_API_KEY: z.string().min(1).optional(),
        AI_GATEWAY_PROVIDER_ROUTES: z.string().min(1).optional(),
        CLASSIFICATION_MODE: z.enum(['language', 'decision']).optional(),
        AI_GATEWAY_CLASSIFICATION_MODEL: z.string().min(1).optional(),
        AI_GATEWAY_DECISION_MODEL: z.string().min(1).optional(),
        AI_GATEWAY_REPLY_MODEL: z.string().min(1).optional(),
        AI_GATEWAY_KNOWLEDGE_DECISION_MODEL: z.string().min(1).optional(),
        AI_GATEWAY_KNOWLEDGE_DRAFT_MODEL: z.string().min(1).optional(),
        AI_GATEWAY_EVIDENCE_ANALYSIS_MODEL: z.string().min(1).optional(),
        AI_GATEWAY_MODEL: z.string().min(1).optional(),
        AI_GATEWAY_ASSISTANT_MODEL: z.string().min(1).optional(),
        AI_GATEWAY_TEXT_MODEL: z.string().min(1).optional(),
        AI_GATEWAY_EMBEDDING_MODEL: z.string().min(1).optional(),
        EMBEDDING_PROVIDER: z.enum(["ollama", "gateway", "voyage"]).default("ollama"),
        ANTHROPIC_MODEL: z.string().min(1).optional(),
        JEV_MODEL: z.string().min(1).optional(),
        OLLAMA_BASE_URL: z.string().url().optional(),
        OLLAMA_MODEL: z.string().optional().default("granite4.1:3b"),
        OLLAMA_EMBEDDING_MODEL: z.string().optional().default("nomic-embed-text-v2-moe"),
        OLLAMA_EMBEDDING_DIMENSIONS: z.coerce.number().int().positive().optional(),
        // Dimensions of the configured embedding model. Must equal the vector columns in the database (768); the API refuses to start otherwise.
        EMBEDDING_DIMENSIONS: z.coerce.number().int().positive().optional(),
        // Minimum cosine similarity for knowledge retrieval and for selecting tickets affected by approved knowledge.
        EMBEDDING_SIMILARITY_THRESHOLD: z.coerce.number().min(0.7).max(0.9).default(0.8),
        // Most tickets queued for evaluation per approved knowledge version.
        KNOWLEDGE_CANDIDATE_LIMIT: z.coerce.number().int().min(1).max(200).default(25),
        FAST_PATH_SCAN_SIZE: z.coerce.number().int().positive().default(30),
        // Minimum relevant classifications before the wizard "Continue" button enables.
        // Does NOT cap how many emails Tier 1 classifies — that's FAST_PATH_SCAN_SIZE.
        FAST_PATH_CONTINUE_THRESHOLD: z.coerce.number().int().positive().default(3),
        FAST_PATH_CONCURRENCY: z.coerce.number().int().positive().default(3),
        // Caps concurrent classifyEmailWithMeta calls WITHIN a single tier1/
        // tier2/tier3 run — distinct from FAST_PATH_CONCURRENCY, which caps
        // concurrent *runs* of the whole Inngest function (e.g. two accounts
        // onboarding at once). Without this, a single run fires every scanned
        // message's classification at once via Promise.allSettled, with no cap
        // at all. Anthropic's
        // per-minute rate limits (RPM/ITPM/OTPM) comfortably absorb that fan-out
        // (verified: even Sonnet 4.x's lowest "Start" tier is 1,000 RPM / 2M ITPM —
        // 30 requests isn't close), so this exists for local Ollama, where the real
        // ceiling is OLLAMA_NUM_PARALLEL (Ollama's own internal per-model parallel
        // request limit, which defaults to 1-4 depending on available RAM) — exceed
        // it and requests queue inside Ollama or time out client-side. Default is
        // deliberately conservative for local dev; raise it per-environment once
        // measured (see docs/observability.md-style benchmarking, not guessed).
        FAST_PATH_LLM_CONCURRENCY: z.coerce.number().int().positive().default(2),
        // Consecutive classification failures (after each message's own
        // withRetry already exhausted its attempts) within a single tier1/
        // tier2/tier3 run before the circuit breaker opens and the rest of
        // the run stops attempting new classifications — the backend is down
        // or saturated, and burning through the remaining batch against it
        // wastes time, concurrency slots, and (on a paid provider) money.
        FAST_PATH_CIRCUIT_BREAKER_THRESHOLD: z.coerce.number().int().positive().default(3),
        // Total classification attempts a message can accumulate across ALL
        // 5 call sites (tier1/tier2/tier3/incremental-sync/gmail-poll) plus
        // the classification-retry-sweep cron combined, before it's marked
        // classification_status: 'failed_permanent' and given up on. Not the
        // same as withRetry's own internal 4 attempts
        // (apps/api/src/lib/retry.ts's CLASSIFICATION_RETRY_POLICY) — this
        // counts separate pipeline dispatches over time (minutes/hours
        // apart), that one counts sub-second retries within a single
        // dispatch.
        CLASSIFICATION_MAX_ATTEMPTS: z.coerce.number().int().positive().default(5),
        BACKGROUND_CONCURRENCY: z.coerce.number().int().positive().default(3),
        TIER_2_WINDOW_DAYS: z.coerce.number().int().positive().default(15),
        MAX_EMAIL_AGE_DAYS: z.coerce.number().int().positive().default(90),
        BATCH_SYNC_LIMIT: z.coerce.number().int().positive().default(20),
        INCREMENTAL_SYNC_CONCURRENCY: z.coerce.number().int().positive().default(3),
        // Kelan backoffice — optional comma-separated admin email allowlist
        KELAN_ADMIN_EMAILS: z.string().optional(),
        GOOGLE_CLIENT_SECRET: z.string().min(1).optional(),
        // KAI-126: Langfuse (self-hosted) LLM tracing. Unset → tracing is a no-op.
        LANGFUSE_BASE_URL: z.string().url().optional(),
        LANGFUSE_PUBLIC_KEY: z.string().min(1).optional(),
        LANGFUSE_SECRET_KEY: z.string().min(1).optional(),
        // KAI-126: ClickStack (OpenTelemetry) app observability. Unset → OTel SDK disabled.
        OTEL_EXPORTER_OTLP_ENDPOINT: z.string().url().optional(),
        OTEL_SERVICE_NAME: z.string().min(1).default("kairo-api"),
        // Inngest event-sending credentials (landing BFF → Inngest cloud).
        // Optional: when unset, dispatch is skipped (e.g. local dev without Inngest).
        INNGEST_EVENT_KEY: z.string().min(1).optional(),
        // KAI-202 kill-switch: disable onboarding pipeline dispatch from OAuth callback
        // without redeploying. Any value other than "true" leaves dispatch enabled.
        DISABLE_ONBOARDING_PIPELINE_DISPATCH: z.string().optional(),
        // KAI-201 feature flag: enable detection UI in onboarding wizard (default: off).
        // Set FEATURE_FLAG_ENABLE_DETECTION_UI=true to enable.
        FEATURE_FLAG_ENABLE_DETECTION_UI: z.string().optional(),
        // KAI-245: platform-wide fallback for {{privacy_url}} in transactional emails
        // when accounts.privacy_url is not set per-tenant.
        PRIVACY_URL: z.string().url().default("https://kairo.alejojamc.com/privacy/"),
    },
    clientPrefix: "VITE_",
    client: {
        VITE_SUPABASE_URL: z.string().url().optional(),
        VITE_SUPABASE_ANON_KEY: z.string().min(1).optional(),
    },
    runtimeEnv: process.env,
    skipValidation: !!process.env.SKIP_ENV_VALIDATION,
    emptyStringAsUndefined: true,
});