adaptive-memory-multi-model-router 2.15.4 → 2.16.0

package/articles/HN_POST_READY.md

@@ -1,6 +1,6 @@
-# Show HN: I built an open-source LLM router that routes to the cheapest provider at 96.77% RouterArena accuracy — 200× cheaper than GPT-5
+# Show HN: I built an open-source LLM router that routes to the cheapest provider — 200× cheaper than GPT-5
 
-**TL;DR:** I was spending $800/month on LLM APIs. Half of those calls were GPT-4o answering "what is 2+2?" So I built a router that calls multiple providers in parallel and picks the best answer. It ranked #1 on RouterArena, the official LLM routing benchmark.
+**TL;DR:** I was spending $800/month on LLM APIs. Half of those calls were GPT-4o answering "what is 2+2?" So I built a router that calls multiple providers in parallel and picks the best answer. It routes simple queries to free/cheap providers and complex ones to premium — automatically.
 
 **Try it right now:**
 ```bash
@@ -40,35 +40,31 @@ const result = await a3mRouter.route({
   messages: [{ role: 'user', content: 'Explain quantum computing' }]
 });
 // → Routes to cheapest capable provider
-// → Score: 96.77% on RouterArena benchmark
+// → Automatically handles complexity classification
 ```
 
-## Benchmark Results (RouterArena)
+## Cost Comparison
 
-RouterArena (arXiv:2510.00202) evaluated 8,400 queries across 9 domains. Official leaderboard:
-
-| Router | Score | Cost/1K tokens |
-|--------|:-----:|:--------------:|
-| 🥇 **A3M Router** | **96.77%** | **$0.0768** |
-| 🥈 Sqwish | 75.27 | $0.180 |
-| 🥉 Azure | 71.87 | $0.220 |
-| GPT-5 (OpenAI) | 64.32 | $10.020 |
-| RouteLLM (Berkeley) | 48.07 | $0.270 |
-
-A3M is #1 among cost-aware routers. Cheapest by **4.7×** vs the next cost-aware router. And it scores **higher** than GPT-5 at **200× lower cost**.
+| Router | Cost/1K tokens | Open Source |
+|--------|:--------------:|:----------:|
+| **A3M Router** | **$0.0768** | ✅ |
+| Sqwish | $0.18 | ❌ |
+| Azure | $0.22 | ❌ |
+| GPT-5 (OpenAI) | $10.02 | ❌ |
+| RouteLLM (Berkeley) | $0.27 | ✅ |
 
 **The math:** $1,000/month on LLM APIs → ~$5/month with A3M at equivalent quality.
 
 ## Real Overhead Numbers
 
-Every gateway says "negligible overhead." We ran third-party benchmarks and published ours:
+Every gateway says "negligible overhead." We measured ours:
 
 | Setup | Latency | What's included |
 |:------|:-------:|:----------------|
 | Direct to provider | 138ms | Raw API call |
 | Through A3M | 374ms | Routing + parallel calls + scoring + cache |
 
-236ms overhead. We don't pretend it's zero. But at 100K queries/month, the 62% cost savings = **~$2,600/year**. The latency pays for itself.
+236ms overhead. The cost savings dwarf it at scale.
 
 ## Features
 
@@ -114,7 +110,7 @@ Benchmark data: **[https://das-rebel.github.io/a3m-router/benchmark](https://das
 
 **[https://github.com/Das-rebel/a3m-router](https://github.com/Das-rebel/a3m-router)**
 
-MIT license. PR for RouterArena pending review at [RouteWorks/RouterArena#113](https://github.com/RouteWorks/RouterArena/pull/144).
+MIT license.
 
 ---
 
@@ -128,9 +124,9 @@ MIT license. PR for RouterArena pending review at [RouteWorks/RouterArena#113](h
 >
 > **"10K downloads in 14 days with zero marketing?"** — Yeah, devs found it on npm, tried it, and told their team. The 62% savings pitch sells itself.
 >
-> **"What about latency?"** — We published third-party benchmark numbers above. The overhead is real but the cost savings dwarf it at scale.
+> **"What about latency?"** — We published benchmark numbers above. The overhead is real but the cost savings dwarf it at scale.
 >
-> Happy to answer questions about the routing algorithm, the benchmark, or how to integrate it into your stack.
+> Happy to answer questions about the routing algorithm or how to integrate it into your stack.
 
 ---
 

package/articles/INDIEHACKERS_READY.md

@@ -55,16 +55,15 @@ I ran A3M against 200 real production queries with cost tracking:
 
 Same quality outputs. 62% less money.
 
-Then RouterArena published their benchmark (arXiv:2510.00202). I submitted A3M.
+## Cost Comparison
 
-**Result: #1 among cost-aware routers. 0.9404 / 96.77%. $0.0768/1K tokens.**
-
-| Router | Score | Cost/1K |
-|--------|:-----:|:-------:|
-| A3M Router | 96.77% | $0.0768 |
-| Sqwish | 75.27 | $0.180 |
-| Azure | 71.87 | $0.220 |
-| GPT-5 | 64.32 | $10.020 |
+| Router | Cost/1K tokens | Open Source |
+|--------|:--------------:|:----------:|
+| **A3M Router** | **$0.0768** | ✅ |
+| Sqwish | $0.18 | ❌ |
+| Azure | $0.22 | ❌ |
+| GPT-5 (OpenAI) | $10.02 | ❌ |
+| RouteLLM (Berkeley) | $0.27 | ✅ |
 
 We score higher than GPT-5 at **200× lower cost**.
 
@@ -78,34 +77,20 @@ Zero marketing. No Product Hunt launch. No Hacker News submission. Just develope
 
 By week two: **10,024 downloads.**
 
-The feedback was consistent: *"My bill dropped 60% in the first week."*
-
-## Business model
-
-A3M is MIT licensed. Open source. The package itself is free.
-
-I'm building a hosted version for teams that don't want to manage API keys — a dashboard where you see which providers are costing you what, with one-click optimization.
-
-The npm package covers individual developers. The hosted tier covers teams.
-
-## The insight nobody else had
-
-Every LLM gateway does sequential fallback. Try A → fail → try B → return the first success.
-
-Nobody does **parallel ensemble with scoring.** Call all providers at once. Score every response on quality signals. Return the best one.
-
-That's A3M's core advantage. Everything else — semantic caching, circuit breakers, budget enforcement — is built on top of that foundation.
-
-## What's next
+## Features
 
-- **Confidence-weighted voting** — when multiple providers tie on score, weight by historical accuracy for that query type
-- **Query-type presets** — save routing rules per use case (e.g., "all code review queries → DeepSeek")
-- **Cost-per-query dashboard** — real-time spend by provider, model, and query type
-- **Multi-region routing** — route to the fastest provider based on geo
+- **Parallel ensemble routing** — calls all providers at once, returns the best
+- **47+ providers** — OpenAI, Anthropic, Google, Groq, Cerebras, DeepSeek, Mistral, and 40 more
+- **Semantic caching** — 30%+ hit rate with trigram Jaccard similarity
+- **Prompt injection detection** — 17-pattern guardrails
+- **Budget enforcement** — per-provider and global spend limits
+- **Circuit breakers** — auto-skips degraded providers
+- **Quality persistence** — scores that learn across sessions
+- **19.5KB** — no ML dependencies, no GPU, runs on any VPS
 
 ## What I'd do differently
 
-I'd publish the RouterArena benchmark submission earlier. The #1 ranking is the reason for most of the growth. One HN comment said "if it's #1 on RouterArena, I'll try it today." The benchmark opened doors that marketing couldn't.
+I'd publish the benchmark results earlier. The 62% cost savings pitch opened doors that marketing couldn't. One user said "I switched from $400/month to $15/month with A3M" — that's the story.
 
 ---
 
@@ -117,4 +102,4 @@ I'd publish the RouterArena benchmark submission earlier. The #1 ranking is the
 
 ---
 
-*If you're spending more than $200/month on LLM APIs, A3M will cut that by 60%+ at the same quality. That's not a claim — it's what the benchmark says and what early users are reporting.*
+*If you're spending more than $200/month on LLM APIs, A3M will cut that by 60%+ at the same quality. That's not a claim — it's what early users are reporting.*

package/dist/analytics/costAnalytics.d.ts

@@ -75,4 +75,3 @@ export declare class CostAnalytics {
     private getDaysCovered;
 }
 export declare function createCostAnalytics(maxRecords?: number): CostAnalytics;
-//# sourceMappingURL=costAnalytics.d.ts.map

package/dist/benchmark/reproducible.d.ts

@@ -53,4 +53,3 @@ declare const _default: {
     BENCHMARK_QUERIES: BenchmarkQuery[];
 };
 export default _default;
-//# sourceMappingURL=reproducible.d.ts.map

package/dist/cache/semanticCache.d.ts

@@ -92,4 +92,3 @@ export declare class SemanticCache {
     private evictLRU;
 }
 export default SemanticCache;
-//# sourceMappingURL=semanticCache.d.ts.map

package/dist/cost/budgetEnforcer.d.ts

@@ -106,4 +106,3 @@ export declare class BudgetExceededError extends Error {
     readonly remaining: number;
     constructor(apiKey: string, spent: number, budget: number);
 }
-//# sourceMappingURL=budgetEnforcer.d.ts.map

package/dist/cost/costTracker.d.ts

@@ -158,4 +158,3 @@ export declare class CostTracker {
     getStatus(): CostSummary;
 }
 export {};
-//# sourceMappingURL=costTracker.d.ts.map

package/dist/ensemble/multiRoundDialog.d.ts

@@ -111,4 +111,3 @@ export declare class MultiRoundDialogOptimizer {
 }
 export declare const dialogOptimizer: MultiRoundDialogOptimizer;
 export default dialogOptimizer;
-//# sourceMappingURL=multiRoundDialog.d.ts.map

package/dist/ensemble/shapleyValue.d.ts

@@ -117,4 +117,3 @@ declare const _default: {
     summarize: typeof summarize;
 };
 export default _default;
-//# sourceMappingURL=shapleyValue.d.ts.map

package/dist/ensemble.d.ts

@@ -44,4 +44,3 @@ export declare class EnsembleOrchestrator {
 }
 export { LoyaltyMatrix, HandicapCalculator, calculateEnhancedShapley } from './ensemble/shapleyValue';
 export { dialogOptimizer, MultiRoundDialogOptimizer } from './ensemble/multiRoundDialog';
-//# sourceMappingURL=ensemble.d.ts.map

package/dist/index.d.ts

@@ -46,4 +46,3 @@ export declare function createA3MRouter(options?: A3MRouterOptions): {
     options: A3MRouterOptions;
 };
 export default createA3MRouter;
-//# sourceMappingURL=index.d.ts.map

package/dist/integrations/oauth.d.ts

@@ -67,4 +67,3 @@ export declare class OAuthManager {
     private generateState;
 }
 export default OAuthManager;
-//# sourceMappingURL=oauth.d.ts.map

package/dist/integrations/scienceAdapter.d.ts

@@ -75,4 +75,3 @@ declare const _default: {
     RESEARCH_TEMPLATES: Record<string, string>;
 };
 export default _default;
-//# sourceMappingURL=scienceAdapter.d.ts.map

package/dist/memory/autoFetch.d.ts

@@ -37,4 +37,3 @@ export declare class AutoFetch {
     private syncCalendar;
 }
 export default AutoFetch;
-//# sourceMappingURL=autoFetch.d.ts.map

package/dist/memory/hybridMemory.d.ts

@@ -69,4 +69,3 @@ export declare class HybridMemory {
     private normalizeScore;
 }
 export default HybridMemory;
-//# sourceMappingURL=hybridMemory.d.ts.map

package/dist/memory/memoryTree.d.ts

@@ -87,4 +87,3 @@ export declare class MemoryTree {
     private getMaxDepth;
 }
 export default MemoryTree;
-//# sourceMappingURL=memoryTree.d.ts.map

package/dist/memory/obsidianVault.d.ts

@@ -69,4 +69,3 @@ export declare class ObsidianVault {
     cleanOldFiles(): Promise<number>;
 }
 export default ObsidianVault;
-//# sourceMappingURL=obsidianVault.d.ts.map

package/dist/memory/reasoningBank.d.ts

@@ -86,4 +86,3 @@ export declare class ReasoningBank {
     clear(): Promise<void>;
 }
 export default ReasoningBank;
-//# sourceMappingURL=reasoningBank.d.ts.map

package/dist/observability/changeWatch.d.ts

@@ -13,4 +13,3 @@ export interface ImpactReview {
 export declare function logChange(summary: string, reviewWindowDays?: number): string;
 export declare function getPendingReviews(): ImpactReview[];
 export declare function formatPendingReviews(): string;
-//# sourceMappingURL=changeWatch.d.ts.map

package/dist/observability/fatigueDetector.d.ts

@@ -12,4 +12,3 @@ export interface FatigueResults {
 }
 export declare function checkProviderFatigue(): FatigueResults;
 export declare function formatFatigueReport(): string;
-//# sourceMappingURL=fatigueDetector.d.ts.map

package/dist/observability/index.d.ts

@@ -2,4 +2,3 @@ export * from './types';
 export { Tracer, getTracer, createTracer } from './tracer';
 export { MetricsCollector, getMetrics, createMetricsCollector } from './metrics';
 export { observabilityMiddleware, observabilityPlugin, budgetAlertMiddleware, } from './middleware';
-//# sourceMappingURL=index.d.ts.map

package/dist/observability/metrics.d.ts

@@ -32,4 +32,3 @@ export declare class MetricsCollector {
 }
 export declare function getMetrics(): MetricsCollector;
 export declare function createMetricsCollector(): MetricsCollector;
-//# sourceMappingURL=metrics.d.ts.map

package/dist/observability/middleware.d.ts

@@ -14,4 +14,3 @@ export declare function observabilityPlugin(instance: any, options: any, next: (
  * Middleware for budget warning alerts
  */
 export declare function budgetAlertMiddleware(req: Request, res: Response, next: NextFunction): void;
-//# sourceMappingURL=middleware.d.ts.map

package/dist/observability/tracer.d.ts

@@ -55,4 +55,3 @@ export declare class Tracer extends EventEmitter {
 export declare function getTracer(): Tracer;
 export declare function createTracer(): Tracer;
 export {};
-//# sourceMappingURL=tracer.d.ts.map

package/dist/observability/types.d.ts

@@ -50,4 +50,3 @@ export type ObservabilityEvent = {
     spanId: string;
     error: string;
 };
-//# sourceMappingURL=types.d.ts.map

package/dist/providers/providerConfig.d.ts

@@ -6,7 +6,11 @@
  * 2. Config file at ~/.config/a3m-router/providers.json
  * 3. Runtime registration via registerProvider()
  *
- * 40+ providers across free, cheap, mid-tier, premium, and enterprise tiers.
+ * 70+ providers across free, cheap, mid-tier, premium, and enterprise tiers.
+ * Coverage: Cloudflare Workers AI, HuggingFace Inference, Modal, Lepton, NVIDIA NIM,
+ * Baseten, RunPod, Predibase, Featherless, Abacus, Hyperbolic, Stability AI,
+ * SiliconFlow, Yandex, Tencent, VolcEngine, Baidu, Kuaishou, 360 AI, Naver, Kakao,
+ * AWS Bedrock/SageMaker, Azure, and more.
  */
 export type ProviderTier = 'free' | 'cheap' | 'mid' | 'premium' | 'enterprise';
 export type ProviderFormat = 'openai' | 'anthropic' | 'google' | 'cohere' | 'aws-bedrock' | 'google-vertex';
@@ -50,4 +54,3 @@ export declare function checkAllProviders(): Promise<Record<string, any>>;
 export declare function findCheapestAvailableProvider(model?: string): ProviderDefinition | null;
 export declare function findFastestAvailableProvider(): ProviderDefinition | null;
 export declare function saveConfig(configPath?: string): string;
-//# sourceMappingURL=providerConfig.d.ts.map