feat(coach): Lyra-LLM auf direkten Anthropic Haiku 4.5 + SOS-Fallback-Kette; ElevenLabs reaktiviert
- Chat (message.post.ts): neuer nativer anthropic-Branch in tryModel (api.anthropic.com/v1/messages, x-api-key, system top-level), führt die Fallback-Kette claude-haiku-4-5 → gemini-flash-lite → gemini-flash → gpt-4o-mini. - SOS (sos-stream.get.ts): Dispatch-Refactor mit buildUpstream() + Kandidaten- Fallback-Kette (anthropic → gemini → openai). Behebt strukturell den Bug "SOS liefert nur Krisen-Text" (vorher single fetch ohne Fallback). Nativer Anthropic-Stream: Delta-Parser liest content_block_delta.delta.text. - nitro.config.ts: anthropicApiKey deklariert (ANTHROPIC_API_KEY). - plan-features.ts: Legend-Voice zurück auf ElevenLabs eleven_turbo_v2_5 (Cartesia-Übergang nach Zahlungsproblem 2026-06-08 aufgehoben, neuer Key). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
chahinebrini
parent
21c1e31877
commit
1a270739bc
@ -81,6 +81,9 @@ export default defineNitroConfig({
|
||||
"",
|
||||
openaiApiKey:
|
||||
process.env.OPENAI_API_KEY ?? process.env.NUXT_OPENAI_API_KEY ?? "",
|
||||
// Direkter Anthropic-Pfad (native Messages-API) — primärer Lyra-Provider.
|
||||
anthropicApiKey:
|
||||
process.env.ANTHROPIC_API_KEY ?? process.env.NUXT_ANTHROPIC_API_KEY ?? "",
|
||||
groqApiKey: process.env.GROQ_API_KEY ?? process.env.NUXT_GROQ_API_KEY ?? "",
|
||||
googleAiApiKey: process.env.GOOGLE_AI_API_KEY ?? "",
|
||||
geminiApiKey: process.env.GEMINI_API_KEY ?? "",
|
||||
|
||||
@ -654,28 +654,69 @@ export default defineEventHandler(async (event) => {
|
||||
}
|
||||
|
||||
// ─── Tier-basiertes LLM-Routing (analog sos-stream.get.ts) ─────────────────
|
||||
// Default-Chain: Gemini Flash Lite (schnell+billig+warm) → Gemini 2.5 Flash
|
||||
// (smarter Fallback) → OpenAI gpt-4o-mini (Last-Resort, anderer Provider).
|
||||
// OpenRouter + Groq sind aktuell ohne Quota/Credits — entfernt aus Chain.
|
||||
// Default-Chain: Anthropic Haiku 4.5 (direkt, native Messages-API — warmer
|
||||
// Lyra-Ton, günstig+zuverlässig) → Gemini Flash Lite → Gemini 2.5 Flash →
|
||||
// OpenAI gpt-4o-mini (Last-Resort, anderer Provider). OpenRouter + Groq sind
|
||||
// aktuell ohne Quota/Credits — entfernt aus Chain.
|
||||
const planRaw = (profile?.plan ?? "free").toLowerCase();
|
||||
const plan =
|
||||
planRaw === "premium" ? "legend" : planRaw === "standard" ? "pro" : planRaw;
|
||||
const llmProvider = "gemini-flash-lite";
|
||||
const llmProvider = "anthropic-haiku";
|
||||
|
||||
type Candidate = {
|
||||
provider: "groq" | "openrouter" | "gemini" | "openai";
|
||||
provider: "anthropic" | "groq" | "openrouter" | "gemini" | "openai";
|
||||
model: string;
|
||||
};
|
||||
const candidates: Candidate[] = [
|
||||
{ provider: "anthropic", model: "claude-haiku-4-5" },
|
||||
{ provider: "gemini", model: "gemini-2.5-flash-lite" },
|
||||
{ provider: "gemini", model: "gemini-2.5-flash" },
|
||||
{ provider: "openai", model: "gpt-4o-mini" },
|
||||
];
|
||||
|
||||
async function tryModel(
|
||||
providerName: "groq" | "openrouter" | "gemini" | "openai",
|
||||
providerName: "anthropic" | "groq" | "openrouter" | "gemini" | "openai",
|
||||
model: string,
|
||||
) {
|
||||
// ── Direkter Anthropic-Pfad: native Messages-API (kein OpenAI-Shape) ──
|
||||
if (providerName === "anthropic") {
|
||||
const key = config.anthropicApiKey as string | undefined;
|
||||
if (!key) return null;
|
||||
// Anthropic verlangt user-first; .slice(-8) kann mit assistant beginnen.
|
||||
const aMsgs =
|
||||
trimmed[0]?.role === "assistant" ? trimmed.slice(1) : trimmed;
|
||||
try {
|
||||
const res = await $fetch<{ content: { type: string; text: string }[] }>(
|
||||
"https://api.anthropic.com/v1/messages",
|
||||
{
|
||||
method: "POST",
|
||||
headers: {
|
||||
"x-api-key": key,
|
||||
"anthropic-version": "2023-06-01",
|
||||
"Content-Type": "application/json",
|
||||
},
|
||||
body: {
|
||||
model,
|
||||
max_tokens: 500,
|
||||
system: systemPrompt,
|
||||
messages: aMsgs,
|
||||
},
|
||||
timeout: 15000,
|
||||
},
|
||||
);
|
||||
return (
|
||||
res.content?.find((b) => b.type === "text")?.text ?? null
|
||||
);
|
||||
} catch (err: any) {
|
||||
console.warn(
|
||||
`[coach/tryModel] anthropic:${model} FAIL:`,
|
||||
err?.statusCode ?? err?.status ?? "?",
|
||||
err?.data?.error?.message ?? err?.message ?? String(err).slice(0, 200),
|
||||
);
|
||||
return null;
|
||||
}
|
||||
}
|
||||
|
||||
const p = PROVIDER_CONFIG[providerName];
|
||||
const key = config[p.keyName];
|
||||
if (!key) return null;
|
||||
|
||||
@ -231,77 +231,158 @@ export default defineEventHandler(async (event) => {
|
||||
// Tier-based LLM: Pro=Groq (sachlich+schnell), Legend=Haiku 4.5 (warm+fast),
|
||||
// Free=Groq (kostenkontrolle). User kann via Toggle override (debug);
|
||||
// 'auto' (oder undefined) → plan-based default.
|
||||
// Default-Kette: Anthropic Haiku 4.5 (direkt, native Messages-API — warmer
|
||||
// Lyra-Ton, günstig+zuverlässig) → Gemini Flash Lite → OpenAI gpt-4o-mini.
|
||||
// Die ERSTE Kandidatin, die einen OK-Stream liefert, gewinnt — SOS darf im
|
||||
// Krisen-Kontext nie ohne Antwort bleiben (vorher: single fetch ohne Fallback
|
||||
// → bei Provider-Fail sofort Krisen-Text). Debug-Toggle override ohne Fallback.
|
||||
const userToggle = sessionData.llmProvider;
|
||||
let llmProvider: string;
|
||||
if (userToggle && userToggle !== "auto") {
|
||||
llmProvider = userToggle;
|
||||
} else {
|
||||
// Default chain: Gemini Flash Lite (schnell+billig+warm) für alle Pläne.
|
||||
// OpenRouter (Haiku/Sonnet) + Groq haben aktuell keine Quota/Credits.
|
||||
llmProvider = "gemini-flash-lite";
|
||||
}
|
||||
let upstreamUrl: string;
|
||||
let upstreamKey: string | undefined;
|
||||
let upstreamModel: string;
|
||||
const upstreamHeaders: Record<string, string> = { "Content-Type": "application/json" };
|
||||
let upstreamProviderField: { sort: string } | undefined;
|
||||
if (llmProvider === "gemini-flash-lite" || llmProvider === "gemini-flash") {
|
||||
upstreamUrl = "https://generativelanguage.googleapis.com/v1beta/openai/chat/completions";
|
||||
upstreamKey = config.geminiApiKey as string | undefined;
|
||||
upstreamModel = llmProvider === "gemini-flash" ? "gemini-2.5-flash" : "gemini-2.5-flash-lite";
|
||||
} else if (llmProvider === "groq-llama") {
|
||||
upstreamUrl = "https://api.groq.com/openai/v1/chat/completions";
|
||||
upstreamKey = config.groqApiKey as string | undefined;
|
||||
upstreamModel = "llama-3.3-70b-versatile";
|
||||
} else {
|
||||
upstreamUrl = "https://openrouter.ai/api/v1/chat/completions";
|
||||
upstreamKey = config.openrouterApiKey as string | undefined;
|
||||
upstreamModel = llmProvider === "openrouter-haiku"
|
||||
? "anthropic/claude-haiku-4.5"
|
||||
: "anthropic/claude-sonnet-4.5";
|
||||
upstreamHeaders["HTTP-Referer"] = "https://rebreak.org";
|
||||
upstreamHeaders["X-Title"] = "ReBreak SOS";
|
||||
upstreamProviderField = { sort: "latency" };
|
||||
}
|
||||
// Fallback: wenn gewählter Provider keinen Key hat → OpenAI gpt-4o-mini als Anker
|
||||
if (!upstreamKey && config.openaiApiKey) {
|
||||
console.warn(`[coach/sos-stream] ${llmProvider} key missing → fallback openai/gpt-4o-mini`);
|
||||
llmProvider = "openai-mini";
|
||||
upstreamUrl = "https://api.openai.com/v1/chat/completions";
|
||||
upstreamKey = config.openaiApiKey as string;
|
||||
upstreamModel = "gpt-4o-mini";
|
||||
upstreamProviderField = undefined;
|
||||
}
|
||||
if (!upstreamKey) {
|
||||
throw createError({ statusCode: 503, message: `API key for ${llmProvider} fehlt` });
|
||||
}
|
||||
upstreamHeaders.Authorization = `Bearer ${upstreamKey}`;
|
||||
console.log(`[coach/sos-stream] using provider=${llmProvider} model=${upstreamModel}`);
|
||||
type SosCandidate = { provider: string; model: string };
|
||||
const candidates: SosCandidate[] =
|
||||
userToggle && userToggle !== "auto"
|
||||
? [{ provider: userToggle, model: "" }]
|
||||
: [
|
||||
{ provider: "anthropic-haiku", model: "claude-haiku-4-5" },
|
||||
{ provider: "gemini-flash-lite", model: "gemini-2.5-flash-lite" },
|
||||
{ provider: "openai-mini", model: "gpt-4o-mini" },
|
||||
];
|
||||
|
||||
const upstream = await fetch(upstreamUrl, {
|
||||
method: "POST",
|
||||
headers: upstreamHeaders,
|
||||
body: JSON.stringify({
|
||||
model: upstreamModel,
|
||||
max_tokens: 400,
|
||||
stream: true,
|
||||
messages: [{ role: "system", content: systemPrompt }, ...trimmed],
|
||||
...(upstreamProviderField ? { provider: upstreamProviderField } : {}),
|
||||
}),
|
||||
});
|
||||
// Baut die Upstream-Request-Config für einen Provider; null wenn Key fehlt.
|
||||
// Anthropic nutzt die native Messages-API (system top-level, x-api-key);
|
||||
// alle anderen das OpenAI-kompatible /chat/completions-Shape.
|
||||
const buildUpstream = (
|
||||
provider: string,
|
||||
): { url: string; headers: Record<string, string>; body: string; label: string } | null => {
|
||||
const headers: Record<string, string> = { "Content-Type": "application/json" };
|
||||
if (provider === "anthropic-haiku") {
|
||||
const key = config.anthropicApiKey as string | undefined;
|
||||
if (!key) return null;
|
||||
headers["x-api-key"] = key;
|
||||
headers["anthropic-version"] = "2023-06-01";
|
||||
// Anthropic verlangt user-first; trimmed kann mit assistant beginnen.
|
||||
const aMsgs = trimmed[0]?.role === "assistant" ? trimmed.slice(1) : trimmed;
|
||||
return {
|
||||
url: "https://api.anthropic.com/v1/messages",
|
||||
headers,
|
||||
body: JSON.stringify({
|
||||
model: "claude-haiku-4-5",
|
||||
max_tokens: 400,
|
||||
stream: true,
|
||||
system: systemPrompt,
|
||||
messages: aMsgs,
|
||||
}),
|
||||
label: "anthropic:claude-haiku-4-5",
|
||||
};
|
||||
}
|
||||
if (provider === "gemini-flash-lite" || provider === "gemini-flash") {
|
||||
const key = config.geminiApiKey as string | undefined;
|
||||
if (!key) return null;
|
||||
headers.Authorization = `Bearer ${key}`;
|
||||
const model = provider === "gemini-flash" ? "gemini-2.5-flash" : "gemini-2.5-flash-lite";
|
||||
return {
|
||||
url: "https://generativelanguage.googleapis.com/v1beta/openai/chat/completions",
|
||||
headers,
|
||||
body: JSON.stringify({
|
||||
model,
|
||||
max_tokens: 400,
|
||||
stream: true,
|
||||
messages: [{ role: "system", content: systemPrompt }, ...trimmed],
|
||||
}),
|
||||
label: `gemini:${model}`,
|
||||
};
|
||||
}
|
||||
if (provider === "groq-llama") {
|
||||
const key = config.groqApiKey as string | undefined;
|
||||
if (!key) return null;
|
||||
headers.Authorization = `Bearer ${key}`;
|
||||
return {
|
||||
url: "https://api.groq.com/openai/v1/chat/completions",
|
||||
headers,
|
||||
body: JSON.stringify({
|
||||
model: "llama-3.3-70b-versatile",
|
||||
max_tokens: 400,
|
||||
stream: true,
|
||||
messages: [{ role: "system", content: systemPrompt }, ...trimmed],
|
||||
}),
|
||||
label: "groq:llama-3.3-70b",
|
||||
};
|
||||
}
|
||||
if (provider === "openrouter-haiku" || provider === "openrouter-sonnet") {
|
||||
const key = config.openrouterApiKey as string | undefined;
|
||||
if (!key) return null;
|
||||
headers.Authorization = `Bearer ${key}`;
|
||||
headers["HTTP-Referer"] = "https://rebreak.org";
|
||||
headers["X-Title"] = "ReBreak SOS";
|
||||
const model =
|
||||
provider === "openrouter-haiku"
|
||||
? "anthropic/claude-haiku-4.5"
|
||||
: "anthropic/claude-sonnet-4.5";
|
||||
return {
|
||||
url: "https://openrouter.ai/api/v1/chat/completions",
|
||||
headers,
|
||||
body: JSON.stringify({
|
||||
model,
|
||||
max_tokens: 400,
|
||||
stream: true,
|
||||
messages: [{ role: "system", content: systemPrompt }, ...trimmed],
|
||||
provider: { sort: "latency" },
|
||||
}),
|
||||
label: `openrouter:${model}`,
|
||||
};
|
||||
}
|
||||
// openai-mini — letzter Anker (anderer Provider).
|
||||
if (provider === "openai-mini") {
|
||||
const key = config.openaiApiKey as string | undefined;
|
||||
if (!key) return null;
|
||||
headers.Authorization = `Bearer ${key}`;
|
||||
return {
|
||||
url: "https://api.openai.com/v1/chat/completions",
|
||||
headers,
|
||||
body: JSON.stringify({
|
||||
model: "gpt-4o-mini",
|
||||
max_tokens: 400,
|
||||
stream: true,
|
||||
messages: [{ role: "system", content: systemPrompt }, ...trimmed],
|
||||
}),
|
||||
label: "openai:gpt-4o-mini",
|
||||
};
|
||||
}
|
||||
return null;
|
||||
};
|
||||
|
||||
// Erste Kandidatin, die einen OK-Stream liefert.
|
||||
let upstream: Response | undefined;
|
||||
let usedLabel = "none";
|
||||
for (const c of candidates) {
|
||||
const cfg = buildUpstream(c.provider);
|
||||
if (!cfg) continue;
|
||||
try {
|
||||
const resp = await fetch(cfg.url, {
|
||||
method: "POST",
|
||||
headers: cfg.headers,
|
||||
body: cfg.body,
|
||||
});
|
||||
if (resp.ok && resp.body) {
|
||||
upstream = resp;
|
||||
usedLabel = cfg.label;
|
||||
break;
|
||||
}
|
||||
const errText = await resp.text().catch(() => "");
|
||||
console.error(
|
||||
`[coach/sos-stream] ${cfg.label} upstream error:`,
|
||||
resp.status,
|
||||
errText.slice(0, 300),
|
||||
);
|
||||
} catch (e) {
|
||||
console.error(`[coach/sos-stream] ${cfg.label} fetch threw:`, e);
|
||||
}
|
||||
}
|
||||
|
||||
// ── Upstream-Fehler: SSE-Header trotzdem setzen, dann Fallback senden ──────
|
||||
// (b) Timeout/Leer-Fallback: Kein 502-Throw im Krisen-Kontext — User muss
|
||||
// immer eine Antwort sehen. Bei LLM-Fehler sofort Krisen-Fallback liefern.
|
||||
const upstreamFailed = !upstream.ok || !upstream.body;
|
||||
if (upstreamFailed) {
|
||||
const errText = await upstream.text().catch(() => "");
|
||||
console.error(
|
||||
"[coach/sos-stream] upstream error:",
|
||||
upstream.status,
|
||||
errText.slice(0, 300),
|
||||
);
|
||||
}
|
||||
const upstreamFailed = !upstream || !upstream.body;
|
||||
console.log(`[coach/sos-stream] using ${usedLabel}`);
|
||||
|
||||
// Direkt zu Node res schreiben — sendStream(ReadableStream) pumpt pull() in Nitro nicht zuverlässig
|
||||
const res = event.node.res;
|
||||
@ -352,7 +433,7 @@ export default defineEventHandler(async (event) => {
|
||||
return;
|
||||
}
|
||||
|
||||
const reader = upstream.body.getReader();
|
||||
const reader = upstream!.body!.getReader();
|
||||
const decoder = new TextDecoder();
|
||||
let buffer = "";
|
||||
let fullText = "";
|
||||
@ -382,8 +463,14 @@ export default defineEventHandler(async (event) => {
|
||||
try {
|
||||
const json = JSON.parse(payload) as {
|
||||
choices?: { delta?: { content?: string } }[];
|
||||
type?: string;
|
||||
delta?: { type?: string; text?: string };
|
||||
};
|
||||
const delta = json.choices?.[0]?.delta?.content;
|
||||
// OpenAI-kompatibel: choices[].delta.content
|
||||
// Anthropic nativ: event content_block_delta → delta.text
|
||||
const delta =
|
||||
json.choices?.[0]?.delta?.content ??
|
||||
(json.type === "content_block_delta" ? json.delta?.text : undefined);
|
||||
if (delta) {
|
||||
fullText += delta;
|
||||
chunkCount++;
|
||||
|
||||
@ -120,16 +120,12 @@ export const PLAN_LIMITS: Record<Exclude<Plan, "free">, PlanLimits> = {
|
||||
{ provider: "groq", model: "llama-3.3-70b-versatile" },
|
||||
],
|
||||
aiProvider: "openrouter",
|
||||
// ⚠️ TEMPORÄR (2026-06-08): ElevenLabs-Account hat ein Zahlungsproblem
|
||||
// (HTTP 401 payment_issue) → Legend würde sonst 502 bekommen. Übergangsweise
|
||||
// auf Cartesia (sonic-3) umgeleitet, damit Legend-User trotzdem Sprachausgabe
|
||||
// haben. ZURÜCKSETZEN sobald die ElevenLabs-Rechnung bezahlt ist:
|
||||
// provider: "elevenlabs", model: "eleven_turbo_v2_5", dailyQuotaSeconds: 0
|
||||
// (Hinweis: mit Cartesia greift der Legend-Voice-Picker / lyraVoiceId nicht —
|
||||
// Legend hört solange Cartesias Default-Stimme wie Pro.)
|
||||
// Legend → ElevenLabs Turbo v2.5, unlimited. Voice-Picker / lyraVoiceId greift.
|
||||
// (2026-06-08 war wg. ElevenLabs-Zahlungsproblem temporär auf Cartesia umgeleitet;
|
||||
// 2026-06-11 neuer Key + bezahlt → zurückgesetzt.)
|
||||
voice: {
|
||||
provider: "cartesia",
|
||||
model: "sonic-3",
|
||||
provider: "elevenlabs",
|
||||
model: "eleven_turbo_v2_5",
|
||||
dailyQuotaSeconds: 0, // Legend bleibt unlimited
|
||||
},
|
||||
},
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user