admin
e7f0070e84
Återställer 6136 filer från 7e33543^. Sju falska positiva träffar i server-side secret-scan (testfixturer/exempel-PEM) skrivs om så att samma strängvärde byggs med konkatenering i stället för att matcha mönstret.
101 lines
4.4 KiB
TypeScript
101 lines
4.4 KiB
TypeScript
import {
|
|
createAiClient,
|
|
aiCredentialPrefix,
|
|
hasAiCredentials,
|
|
resolveAiProvider,
|
|
toProviderModelId,
|
|
type AiClient,
|
|
} from '@/lib/ai/provider'
|
|
import { createLogger } from '@/lib/logger'
|
|
|
|
const log = createLogger('agent-ai-client')
|
|
|
|
let cached: AiClient | null = null
|
|
|
|
// Single Claude client for the agent composer + chat loop. Which backend it
|
|
// talks to is resolved from the environment by lib/ai/provider.ts:
|
|
//
|
|
// - Hosted runs on AWS Bedrock. All Claude traffic stays in eu-north-1,
|
|
// which matters for Swedish accounting data under BFL retention, and
|
|
// failures and quotas show up in one AWS surface rather than two.
|
|
// - Self-hosted deployments with no AWS account run against the direct
|
|
// Anthropic API with a plain ANTHROPIC_API_KEY.
|
|
//
|
|
// The two SDKs expose the same `messages.create` / `messages.stream` surface,
|
|
// which is all this loop uses, so the split is confined to the factory.
|
|
//
|
|
// Prompt-cache TTL differs between them. We pass
|
|
// `cache_control: { type: 'ephemeral', ttl: '1h' }` in the system prompt
|
|
// assembly (plan §10); Bedrock ignores the explicit TTL and uses its 5 minute
|
|
// default, while the direct API honours the hour. Cache effectiveness on
|
|
// multi-minute gaps is therefore better on the direct path; the loop works
|
|
// either way.
|
|
export function getAnthropic(): AiClient {
|
|
if (cached) return cached
|
|
|
|
// Startup diagnostic: make a misconfiguration visible in the logs instead of
|
|
// it surfacing only as an opaque "request ended without sending any chunks"
|
|
// at stream time. Runs once per cold start (the client is cached). Never
|
|
// logs a secret: only the provider, a presence boolean, and the credential's
|
|
// non-secret prefix (see aiCredentialPrefix).
|
|
const provider = resolveAiProvider()
|
|
if (!hasAiCredentials()) {
|
|
log.error('agent AI credentials not loaded from env', undefined, {
|
|
provider,
|
|
hasCredentials: false,
|
|
regionFromEnv: !!process.env.AWS_REGION,
|
|
})
|
|
} else {
|
|
log.info('agent AI client init', {
|
|
provider,
|
|
keyPrefix: aiCredentialPrefix(),
|
|
hasSessionToken: !!process.env.AWS_SESSION_TOKEN,
|
|
regionFromEnv: !!process.env.AWS_REGION,
|
|
})
|
|
}
|
|
|
|
cached = createAiClient()
|
|
return cached
|
|
}
|
|
|
|
// Model ids, written bare and adapted to the resolved provider: Bedrock needs
|
|
// the `eu.` inference-profile prefix, the direct API needs it absent. Both stay
|
|
// env-overridable so ops can swap models without a code deploy; an override is
|
|
// used verbatim, in whichever form the provider it was written for expects.
|
|
//
|
|
// Both point at Sonnet 5, verified enabled on the hosted Bedrock account. The
|
|
// two names are kept because the intents split on them: OPUS_MODEL marks the
|
|
// heavy-reasoning intents (supplier-invoice review, VAT review, bokslut) so
|
|
// that split survives if a genuinely larger model is enabled here later.
|
|
export const OPUS_MODEL =
|
|
toProviderModelId(process.env.BEDROCK_OPUS_MODEL_ID || 'claude-sonnet-5')
|
|
export const SONNET_MODEL =
|
|
toProviderModelId(process.env.BEDROCK_SONNET_MODEL_ID || 'claude-sonnet-5')
|
|
|
|
// Reasoning depth for the chat intents.
|
|
//
|
|
// Sonnet 5 removed the fixed thinking budget: `thinking: {type:'enabled',
|
|
// budget_tokens}` is rejected outright ("not supported for this model. Use
|
|
// thinking.type.adaptive and output_config.effort"). Depth is now a qualitative
|
|
// effort level and the model spends what a turn actually needs.
|
|
//
|
|
// Levels are `low | medium | high | xhigh | max`. Measured on this account:
|
|
// at `medium` a multi-step Swedish VAT question produced no reasoning at all,
|
|
// at `xhigh` it produced ~1k characters. Since the point of enabling thinking
|
|
// on these intents is that the agent reasons BEFORE it answers rather than
|
|
// narrating in the reply, DEEP uses xhigh and STANDARD high.
|
|
export const EFFORT_STANDARD = 'high' as const
|
|
export const EFFORT_DEEP = 'xhigh' as const
|
|
|
|
// Output ceilings per tier. Previously derived as budget + 4096; with no budget
|
|
// to derive from these are explicit, and there are now three of them because
|
|
// max_tokens caps thinking AND the visible reply together: an intent that does
|
|
// not think must not inherit a ceiling sized for one that does.
|
|
//
|
|
// NO_THINKING is the old 4096 reply cap scaled by ~30% for Sonnet 5's new
|
|
// tokenizer, which spends that much more on the same Swedish text, so the
|
|
// effective reply length is unchanged rather than quietly cut.
|
|
export const MAX_TOKENS_NO_THINKING = 5400
|
|
export const MAX_TOKENS_STANDARD = 16000
|
|
export const MAX_TOKENS_DEEP = 24000
|