7.4 billion tokens per month. 34 free LLM providers. 635 free model endpoints. All behind one /v1 endpoint, plus any custom OpenAI-compatible endpoint. Smart routing, automatic failover, encrypted keys. Personal experimentation only.
TypeScript
fix(router): exploration never promotes a model that cannot serve the request (#847)
熬夜砍树4270280ac33f
diff --git a/server/src/__tests__/services/router-bandit.test.ts b/server/src/__tests__/services/router-bandit.test.tsindex 6ec300b3..87ed6d49 100644--- a/server/src/__tests__/services/router-bandit.test.ts+++ b/server/src/__tests__/services/router-bandit.test.ts@@ -131,19 +131,31 @@ describe('bandit router', () => {expect(counts['y'] ?? 0).toBeGreaterThan(0);});- it('vision requests never route to a text-only model, even with exploration on', () => {+ it('exploration draws are never wasted on a model that cannot serve the request', () => {+ // vision-a is measured and wins every ordinary bandit draw. text-b and+ // vision-c are both unmeasured, but only vision-c can serve a vision+ // request. The pre-fix router put BOTH in the explore pool, so ~half the+ // explore draws promoted text-b, which the main loop's vision gate then+ // skipped — a wasted draw. With the pool filtered, vision-c receives+ // every explore draw (~10% of requests); unfixed it got only ~5%, far+ // below the 150/2000 bound asserted here.setExploreEnabled(true);- addModel({ platform: 'google', modelId: 'vision-a', name: 'Vision A', intelligenceRank: 3, sizeLabel: 'Large', budget: '~50M', priority: 1, vision: true });- addModel({ platform: 'groq', modelId: 'text-b', name: 'Text B', intelligenceRank: 3, sizeLabel: 'Large', budget: '~50M', priority: 2 });+ addModel({ platform: 'google', modelId: 'vision-a', name: 'Vision A', intelligenceRank: 1, sizeLabel: 'Frontier', budget: '~50M', priority: 1, vision: true });+ addModel({ platform: 'groq', modelId: 'text-b', name: 'Text B', intelligenceRank: 2, sizeLabel: 'Frontier', budget: '~50M', priority: 2 });+ addModel({ platform: 'google', modelId: 'vision-c', name: 'Vision C', intelligenceRank: 30, sizeLabel: 'Small', budget: '~50M', priority: 3, vision: true });+ addHistory('google', 'vision-a', { successes: 500, failures: 0, outTokens: 1000, latencyMs: 300, ttfbMs: 100 });setRoutingStrategy('balanced');refreshStatsCache(getDb(), true);+const counts: Record<string, number> = {};- for (let i = 0; i < 300; i++) {+ for (let i = 0; i < 2000; i++) {const r = routeRequest(100, undefined, undefined, /*requireVision=*/ true);counts[r.modelId] = (counts[r.modelId] ?? 0) + 1;}+ // The loop gate already guarantees this half; the pool filter is what the+ // vision-c bound below actually pins down.expect(counts['text-b'] ?? 0).toBe(0);- expect(counts['vision-a'] ?? 0).toBe(300);+ expect(counts['vision-c'] ?? 0).toBeGreaterThan(150);});it('smartest vs fastest flips which model wins, at equal reliability', () => {diff --git a/server/src/services/router.ts b/server/src/services/router.tsindex d90e4d47..1063836d 100644--- a/server/src/services/router.ts+++ b/server/src/services/router.ts@@ -1575,6 +1575,7 @@ export function routeRequest(estimatedTokens = 1000, skipKeys?: Set<string>, pre// Mirror the main loop's gates below so exploration only samples// candidates that can actually serve this request.if (skipModels?.has(e.model_db_id)) return false;+ if (skipPlatforms?.has(e.platform)) return false;if (requireVision && !e.supports_vision) return false;if (requireTools && !e.supports_tools) return false;if (requireStructured && platformDropsResponseFormat(e.platform)) return false;