openrouter.ts
8,308 bytes
| 1 | /** |
|---|---|
| 2 | * OpenRouter-backed implementation of the core `LlmClient`, plus the REST |
| 3 | * helpers the UI needs (model catalog, key validation, credit balance). |
| 4 | * |
| 5 | * OpenRouter is OpenAI-compatible, so we drive it through the Vercel AI SDK's |
| 6 | * OpenAI provider pointed at the OpenRouter base URL. This is the ONLY place in |
| 7 | * the server that talks to a model provider; the orchestrator stays oblivious. |
| 8 | * |
| 9 | * Cost is computed locally from the model catalog's per-token pricing so we can |
| 10 | * attribute spend per model per stage without a second API round-trip. |
| 11 | */ |
| 12 | import { createOpenAI } from '@ai-sdk/openai'; |
| 13 | import { generateText, streamText } from 'ai'; |
| 14 | import { |
| 15 | LlmError, |
| 16 | type LlmClient, |
| 17 | type LlmMessage, |
| 18 | type LlmRequest, |
| 19 | type LlmResult, |
| 20 | type LlmStreamHandle, |
| 21 | } from '@/core/llm-client'; |
| 22 | import type { Usage } from '@/core/types'; |
| 23 | import { env } from './env'; |
| 24 | |
| 25 | // --------------------------------------------------------------------------- |
| 26 | // Catalog + pricing |
| 27 | // --------------------------------------------------------------------------- |
| 28 | |
| 29 | export interface OpenRouterModel { |
| 30 | id: string; |
| 31 | name: string; |
| 32 | description?: string; |
| 33 | context_length?: number; |
| 34 | pricing: { prompt: string; completion: string; request?: string; image?: string }; |
| 35 | top_provider?: { max_completion_tokens?: number | null }; |
| 36 | architecture?: { modality?: string; input_modalities?: string[] }; |
| 37 | } |
| 38 | |
| 39 | export interface ModelPricing { |
| 40 | /** USD per prompt token. */ |
| 41 | prompt: number; |
| 42 | /** USD per completion token. */ |
| 43 | completion: number; |
| 44 | } |
| 45 | export type PricingMap = Map<string, ModelPricing>; |
| 46 | |
| 47 | function authHeaders(apiKey?: string): Record<string, string> { |
| 48 | return { |
| 49 | ...(apiKey ? { Authorization: `Bearer ${apiKey}` } : {}), |
| 50 | 'HTTP-Referer': env.OPENROUTER_APP_URL, |
| 51 | 'X-Title': env.OPENROUTER_APP_TITLE, |
| 52 | 'Content-Type': 'application/json', |
| 53 | }; |
| 54 | } |
| 55 | |
| 56 | async function fetchWithRetry(url: string, init: RequestInit, attempts = 3): Promise<Response> { |
| 57 | let lastErr: unknown; |
| 58 | for (let i = 0; i < attempts; i++) { |
| 59 | try { |
| 60 | const res = await fetch(url, init); |
| 61 | if (res.status === 429 || res.status >= 500) { |
| 62 | if (i < attempts - 1) { |
| 63 | await sleep(250 * 2 ** i); |
| 64 | continue; |
| 65 | } |
| 66 | } |
| 67 | return res; |
| 68 | } catch (err) { |
| 69 | lastErr = err; |
| 70 | if (i < attempts - 1) await sleep(250 * 2 ** i); |
| 71 | } |
| 72 | } |
| 73 | throw new LlmError(`Network error contacting OpenRouter: ${String(lastErr)}`, { retryable: true, cause: lastErr }); |
| 74 | } |
| 75 | |
| 76 | export async function fetchOpenRouterModels(apiKey?: string): Promise<OpenRouterModel[]> { |
| 77 | const res = await fetchWithRetry(`${env.OPENROUTER_BASE_URL}/models`, { |
| 78 | method: 'GET', |
| 79 | headers: authHeaders(apiKey), |
| 80 | }); |
| 81 | if (!res.ok) { |
| 82 | throw new LlmError(`Failed to fetch model catalog (${res.status})`, { |
| 83 | status: res.status, |
| 84 | retryable: res.status >= 500, |
| 85 | }); |
| 86 | } |
| 87 | const json = (await res.json()) as { data: OpenRouterModel[] }; |
| 88 | return json.data ?? []; |
| 89 | } |
| 90 | |
| 91 | export function pricingFromModels(models: OpenRouterModel[]): PricingMap { |
| 92 | const map: PricingMap = new Map(); |
| 93 | for (const m of models) { |
| 94 | const prompt = Number(m.pricing?.prompt ?? '0'); |
| 95 | const completion = Number(m.pricing?.completion ?? '0'); |
| 96 | map.set(m.id, { |
| 97 | prompt: Number.isFinite(prompt) ? prompt : 0, |
| 98 | completion: Number.isFinite(completion) ? completion : 0, |
| 99 | }); |
| 100 | } |
| 101 | return map; |
| 102 | } |
| 103 | |
| 104 | // --------------------------------------------------------------------------- |
| 105 | // Key validation / credits |
| 106 | // --------------------------------------------------------------------------- |
| 107 | |
| 108 | export interface KeyValidation { |
| 109 | valid: boolean; |
| 110 | label?: string; |
| 111 | usage?: number; |
| 112 | limit?: number | null; |
| 113 | limitRemaining?: number | null; |
| 114 | isFreeTier?: boolean; |
| 115 | error?: string; |
| 116 | } |
| 117 | |
| 118 | export async function validateOpenRouterKey(apiKey: string): Promise<KeyValidation> { |
| 119 | try { |
| 120 | const res = await fetchWithRetry(`${env.OPENROUTER_BASE_URL}/key`, { |
| 121 | method: 'GET', |
| 122 | headers: authHeaders(apiKey), |
| 123 | }); |
| 124 | if (res.status === 401 || res.status === 403) return { valid: false, error: 'Invalid API key' }; |
| 125 | if (!res.ok) return { valid: false, error: `OpenRouter returned ${res.status}` }; |
| 126 | const json = (await res.json()) as { |
| 127 | data?: { label?: string; usage?: number; limit?: number | null; is_free_tier?: boolean }; |
| 128 | }; |
| 129 | const d = json.data ?? {}; |
| 130 | const limit = d.limit ?? null; |
| 131 | const usage = d.usage ?? 0; |
| 132 | return { |
| 133 | valid: true, |
| 134 | label: d.label, |
| 135 | usage, |
| 136 | limit, |
| 137 | limitRemaining: limit === null ? null : Math.max(0, limit - usage), |
| 138 | isFreeTier: d.is_free_tier, |
| 139 | }; |
| 140 | } catch (err) { |
| 141 | return { valid: false, error: err instanceof Error ? err.message : 'Network error' }; |
| 142 | } |
| 143 | } |
| 144 | |
| 145 | // --------------------------------------------------------------------------- |
| 146 | // LlmClient implementation |
| 147 | // --------------------------------------------------------------------------- |
| 148 | |
| 149 | function toCoreMessages(messages: LlmMessage[]) { |
| 150 | return messages.map((m) => ({ role: m.role, content: m.content }) as const); |
| 151 | } |
| 152 | |
| 153 | function buildUsage(model: string, raw: { promptTokens?: number; completionTokens?: number; totalTokens?: number } | undefined, pricing: PricingMap): Usage { |
| 154 | const promptTokens = raw?.promptTokens ?? 0; |
| 155 | const completionTokens = raw?.completionTokens ?? 0; |
| 156 | const totalTokens = raw?.totalTokens ?? promptTokens + completionTokens; |
| 157 | const price = pricing.get(model); |
| 158 | const costUsd = price ? promptTokens * price.prompt + completionTokens * price.completion : 0; |
| 159 | return { promptTokens, completionTokens, totalTokens, costUsd }; |
| 160 | } |
| 161 | |
| 162 | function toLlmError(err: unknown): LlmError { |
| 163 | if (err instanceof LlmError) return err; |
| 164 | const anyErr = err as { statusCode?: number; status?: number; name?: string; message?: string } | undefined; |
| 165 | const status = anyErr?.statusCode ?? anyErr?.status; |
| 166 | const retryable = status === 429 || (status !== undefined && status >= 500); |
| 167 | return new LlmError(anyErr?.message ?? String(err), { status, retryable, cause: err }); |
| 168 | } |
| 169 | |
| 170 | export interface OpenRouterClientOptions { |
| 171 | apiKey: string; |
| 172 | pricing?: PricingMap; |
| 173 | maxRetries?: number; |
| 174 | } |
| 175 | |
| 176 | export function createOpenRouterClient(opts: OpenRouterClientOptions): LlmClient { |
| 177 | const provider = createOpenAI({ |
| 178 | baseURL: env.OPENROUTER_BASE_URL, |
| 179 | apiKey: opts.apiKey, |
| 180 | name: 'openrouter', |
| 181 | headers: { |
| 182 | 'HTTP-Referer': env.OPENROUTER_APP_URL, |
| 183 | 'X-Title': env.OPENROUTER_APP_TITLE, |
| 184 | }, |
| 185 | }); |
| 186 | const pricing = opts.pricing ?? new Map(); |
| 187 | const maxRetries = opts.maxRetries ?? 4; |
| 188 | |
| 189 | return { |
| 190 | async complete(req: LlmRequest): Promise<LlmResult> { |
| 191 | const t0 = Date.now(); |
| 192 | try { |
| 193 | const res = await generateText({ |
| 194 | model: provider.chat(req.model), |
| 195 | messages: toCoreMessages(req.messages), |
| 196 | temperature: req.temperature, |
| 197 | maxTokens: req.maxTokens, |
| 198 | abortSignal: req.signal, |
| 199 | maxRetries, |
| 200 | }); |
| 201 | return { |
| 202 | text: res.text, |
| 203 | usage: buildUsage(req.model, res.usage, pricing), |
| 204 | model: req.model, |
| 205 | latencyMs: Date.now() - t0, |
| 206 | }; |
| 207 | } catch (err) { |
| 208 | throw toLlmError(err); |
| 209 | } |
| 210 | }, |
| 211 | |
| 212 | async streamComplete(req: LlmRequest): Promise<LlmStreamHandle> { |
| 213 | const t0 = Date.now(); |
| 214 | const result = streamText({ |
| 215 | model: provider.chat(req.model), |
| 216 | messages: toCoreMessages(req.messages), |
| 217 | temperature: req.temperature, |
| 218 | maxTokens: req.maxTokens, |
| 219 | abortSignal: req.signal, |
| 220 | maxRetries, |
| 221 | }); |
| 222 | |
| 223 | async function* stream(): AsyncGenerator<string> { |
| 224 | try { |
| 225 | for await (const delta of result.textStream) yield delta; |
| 226 | } catch (err) { |
| 227 | throw toLlmError(err); |
| 228 | } |
| 229 | } |
| 230 | |
| 231 | const final: Promise<LlmResult> = (async () => { |
| 232 | try { |
| 233 | const [text, usage] = await Promise.all([result.text, result.usage]); |
| 234 | return { |
| 235 | text, |
| 236 | usage: buildUsage(req.model, usage, pricing), |
| 237 | model: req.model, |
| 238 | latencyMs: Date.now() - t0, |
| 239 | }; |
| 240 | } catch (err) { |
| 241 | throw toLlmError(err); |
| 242 | } |
| 243 | })(); |
| 244 | |
| 245 | return { stream: stream(), result: final }; |
| 246 | }, |
| 247 | }; |
| 248 | } |
| 249 | |
| 250 | function sleep(ms: number): Promise<void> { |
| 251 | return new Promise((resolve) => setTimeout(resolve, ms)); |
| 252 | } |
| 253 | |