Momento/memento-note/lib/ai/router.ts

/**
 * Central synchronous AI gateway routing (Story 3.2 — FR17 / NFR-P3).
 *
 * Future (Story 3.5 BYOK): plug user-scoped API keys into resolveAiRoute output / factory instantiation.
 *
 * Non-goals here (by design):
 * - Multi-provider HTTP fallback on 429/500 → Story 3.3
 * - BYOK / UserAPIKey decryption → Story 3.5 (extension seam: same resolve output + key source later)
 */

export type AiFeatureLane = 'chat' | 'tags' | 'embedding'

export type AiGatewayProvider =
  | 'ollama'
  | 'openai'
  | 'google'
  | 'minimax'
  | 'glm'
  | 'custom'
  | 'deepseek'
  | 'openrouter'
  | 'mistral'
  | 'zai'
  | 'lmstudio'
  | 'anthropic'
  | 'anthropic_custom'
  | 'custom_openai'
  | 'custom_anthropic'

export interface ResolvedAiRoute {
  lane: AiFeatureLane
  providerType: AiGatewayProvider
  modelName: string
  embeddingModelName: string
  ollamaBaseUrl?: string
  meta: {
    resolveMs?: number
  }
}

export const VALID_PROVIDERS = new Set<string>([
  'ollama', 'openai', 'google', 'minimax', 'glm', 'custom',
  'deepseek', 'openrouter', 'mistral', 'zai', 'lmstudio',
  'anthropic', 'anthropic_custom',
  'custom_openai', 'custom_anthropic',
])

const PROVIDER_MODEL_DEFAULTS: Record<string, { model: string; embeddingModel: string }> = {
  ollama:            { model: 'granite4:latest',              embeddingModel: 'embeddinggemma:latest' },
  openai:            { model: 'gpt-4o-mini',                  embeddingModel: 'text-embedding-3-small' },
  anthropic:         { model: 'claude-sonnet-4-6-20250514',   embeddingModel: '' },
  anthropic_custom:  { model: 'claude-sonnet-4-6-20250514',   embeddingModel: '' },
  deepseek:          { model: 'deepseek-chat',                embeddingModel: '' },
  openrouter:        { model: 'openai/gpt-4o-mini',          embeddingModel: 'openai/text-embedding-3-small' },
  google:            { model: 'gemini-1.5-flash',             embeddingModel: 'text-embedding-004' },
  mistral:           { model: 'mistral-small-latest',         embeddingModel: 'mistral-embed' },
  zai:               { model: 'gpt-4o-mini',                  embeddingModel: 'text-embedding-3-small' },
  minimax:           { model: 'abab6.5-chat',                 embeddingModel: '' },
  glm:               { model: 'glm-4',                        embeddingModel: 'embedding-2' },
  lmstudio:          { model: '',                              embeddingModel: '' },
  custom:            { model: '',                              embeddingModel: '' },
  custom_openai:     { model: 'gpt-4o-mini',                    embeddingModel: 'text-embedding-3-small' },
  custom_anthropic:  { model: 'claude-sonnet-4-6-20250514',   embeddingModel: '' },
}

function pick(config: Record<string, string>, key: string): string | undefined {
  const v = config[key]
  if (v != null && v !== '') return v
  const e = process.env[key]
  return e != null && e !== '' ? e : undefined
}

function cfgOnly(config: Record<string, string>, key: string): string | undefined {
  const v = config[key]
  return v != null && v !== '' ? v : undefined
}

const VALID_PROVIDER_LIST = [...VALID_PROVIDERS].join(', ')

export function resolveAiRoute(lane: AiFeatureLane, config: Record<string, string>): ResolvedAiRoute {
  let providerRaw: string | undefined
  let modelKey: string
  let ollamaBaseUrl: string | undefined

  if (lane === 'tags') {
    providerRaw =
      pick(config, 'AI_PROVIDER_TAGS') ||
      pick(config, 'AI_PROVIDER_EMBEDDING') ||
      pick(config, 'AI_PROVIDER')
    modelKey = 'AI_MODEL_TAGS'
    ollamaBaseUrl = cfgOnly(config, 'OLLAMA_BASE_URL_TAGS') || cfgOnly(config, 'OLLAMA_BASE_URL')
    if (!providerRaw) {
      throw new Error(
        'AI_PROVIDER_TAGS is not configured. Please set it in the admin settings or environment variables. ' +
          'Options: ' + VALID_PROVIDER_LIST
      )
    }
  } else if (lane === 'embedding') {
    providerRaw =
      pick(config, 'AI_PROVIDER_EMBEDDING') ||
      pick(config, 'AI_PROVIDER_TAGS') ||
      pick(config, 'AI_PROVIDER')
    modelKey = 'AI_MODEL_EMBEDDING'
    ollamaBaseUrl = cfgOnly(config, 'OLLAMA_BASE_URL_EMBEDDING') || cfgOnly(config, 'OLLAMA_BASE_URL')
    if (!providerRaw) {
      throw new Error(
        'AI_PROVIDER_EMBEDDING is not configured. Please set it in the admin settings or environment variables. ' +
          'Options: ' + VALID_PROVIDER_LIST
      )
    }
  } else {
    providerRaw =
      pick(config, 'AI_PROVIDER_CHAT') ||
      pick(config, 'AI_PROVIDER_TAGS') ||
      pick(config, 'AI_PROVIDER_EMBEDDING') ||
      pick(config, 'AI_PROVIDER')
    modelKey = 'AI_MODEL_CHAT'
    ollamaBaseUrl =
      cfgOnly(config, 'OLLAMA_BASE_URL_CHAT') ||
      cfgOnly(config, 'OLLAMA_BASE_URL_TAGS') ||
      cfgOnly(config, 'OLLAMA_BASE_URL_EMBEDDING') ||
      cfgOnly(config, 'OLLAMA_BASE_URL')
    if (!providerRaw) {
      throw new Error(
        'AI_PROVIDER_CHAT is not configured. Please set it in the admin settings or environment variables. ' +
          'Options: ' + VALID_PROVIDER_LIST
      )
    }
  }

  const providerType = providerRaw.toLowerCase()

  if (!VALID_PROVIDERS.has(providerType)) {
    throw new Error(
      `Unknown AI provider '${providerRaw}'. Valid options: ${VALID_PROVIDER_LIST}`
    )
  }

  if (lane === 'embedding' && (providerType === 'anthropic' || providerType === 'anthropic_custom')) {
    throw new Error(
      'AI_PROVIDER_EMBEDDING cannot use "anthropic" or "anthropic_custom": these gateways use the Anthropic Messages API only (no embeddings in Memento). Use ollama, openai, or "custom" with MiniMax OpenAI URL https://api.minimax.io/v1 for embeddings.'
    )
  }

  const defaults = PROVIDER_MODEL_DEFAULTS[providerType] || { model: '', embeddingModel: '' }
  const modelName = pick(config, modelKey) || defaults.model
  const embeddingModelName = pick(config, 'AI_MODEL_EMBEDDING') || defaults.embeddingModel

  return {
    lane,
    providerType: providerType as AiGatewayProvider,
    modelName,
    embeddingModelName,
    ollamaBaseUrl,
    meta: {},
  }
}

export function resolveAiRouteWithTiming(lane: AiFeatureLane, config: Record<string, string>): ResolvedAiRoute {
  const t0 = performance.now()
  const route = resolveAiRoute(lane, config)
  const resolveMs = performance.now() - t0
  return {
    ...route,
    meta: { ...route.meta, resolveMs },
  }
}

export function formatAiRouteDebug(route: ResolvedAiRoute): string {
  return JSON.stringify({
    lane: route.lane,
    providerType: route.providerType,
    modelId: route.modelName,
    embeddingModelId: route.embeddingModelName,
    resolveMs: route.meta.resolveMs,
  })
}