promptimizer 0.1.36 → 0.1.37

package/dist/classify.d.ts

@@ -1,4 +1,5 @@
 import type { ChatMessage, Classification, Tier } from "./types";
 export declare function classifyMessages(messages: ChatMessage[]): Classification;
+/** Production classifier — prefer over-routing hard work over silent quality loss. */
 export declare function classifyText(text: string): Classification;
 export declare function difficultyTier(complexity: number): Tier;

package/dist/index.cjs

@@ -136,6 +136,7 @@ var Promptimizer = class _Promptimizer {
 
 // src/classify.ts
 var HIGH_RISK = /* @__PURE__ */ new Set(["system_design", "safety_sensitive", "code_debug", "reasoning"]);
+var TIER_RANK = { economy: 0, standard: 1, frontier: 2 };
 function textFrom(messages) {
   return messages.map((m) => {
     if (typeof m.content === "string") return m.content;
@@ -148,34 +149,46 @@ function classifyMessages(messages) {
 function classifyText(text) {
   const words = text.split(/\s+/).filter(Boolean).length;
   const lines = text.split("\n").length;
+  const numbered = (text.match(/\(\s*\d+\s*\)/g) ?? []).length + (text.match(/^\s*\d+[\).:]/gm) ?? []).length + (text.match(/\b\d+\.\s+[A-Za-z]/g) ?? []).length;
+  const constraintWords = (text.match(
+    /\b(must|include|constraint|requirement|requirements|need to|ensure|exactly-once|idempotent)\b/gi
+  ) ?? []).length;
   const features = {
     code_fence: /```/.test(text),
-    code_kw: /\b(def |class |function |import |fn |pub |async |SELECT |goroutine|mutex|traceback)\b/i.test(
+    code_kw: /\b(def |class |function |import |fn |pub |async |goroutine|mutex|traceback)\b/i.test(text) || /\b(SELECT\b.+\bFROM\b|\bJOIN\b.+\bON\b|postgresql|\bpsql\b)/i.test(text),
+    math: /(\$\$|\\frac|prove that|expected value|O\([nN]\)|\d+\s*[\*\^×x]\s*\d+|infinitely many|factorial|prime)/i.test(
       text
     ),
-    math: /(\$\$|\\frac|prove that|expected value|O\([nN]\)|\d+\s*[\*\^]\s*\d+)/i.test(text),
-    design: /\b(design|architect|rate limiter|distributed|consistency|shard|1 million QPS)\b/i.test(
+    design: /\b(design(?:ing)?|architect(?:ure|ing)?|rate[\s-]?limiter|distributed|consistency|shard(?:ing)?|multi-?region|multi-?tenant|qps|edge regions?|data[\s-]?plane|failure modes?|outbox|ledger|webhook|saas|billing pipeline|consensus|raft|paxos|quorum|leader election|log replication|network partition|cap theorem|eventual consistency|token bucket|sliding window)\b/i.test(
       text
     ),
-    reason: /\b(prove|why does|walk through|step by step|derive|contradiction|p-value)\b/i.test(
+    reason: /\b(prove|proof|theorem|lemma|why does|walk through|step by step|derive|contradiction|p-value|induction|euclid|sketch (?:a |the )?proof)\b/i.test(
+      text
+    ),
+    debug: /\b(bug|race|panic|fails on|diagnose|deadlock|concurrency hazard|code review|correctness)\b/i.test(
       text
     ),
-    debug: /\b(bug|race|panic|fails on|diagnose|deadlock)\b/i.test(text),
     summarize: /\b(summarize|tl;dr|in two sentences|eli5)\b/i.test(text),
     translate: /\b(translate|traduce)\b/i.test(text),
     creative: /\b(write a (poem|story|song)|haiku)\b/i.test(text),
-    safety: /\b(refund|legal|medical|hipaa|lawsuit|diagnosis)\b/i.test(text),
-    analysis: /\b(compare|trade-?off|versus|analyse|analyze|evaluate|should we)\b/i.test(text),
-    constraints: (text.match(/\b(must|include|constraints?|requirements?)\b/gi) ?? []).length,
+    safety: /\b(refund|legal|medical|hipaa|lawsuit|diagnosis|gdpr|erasure|pii|compliance|audit trail)\b/i.test(
+      text
+    ),
+    analysis: /\b(compare|trade-?off|versus|analyse|analyze|evaluate|should we|outline architecture)\b/i.test(
+      text
+    ),
+    multi_part: numbered >= 3 || numbered >= 2 && constraintWords >= 1,
+    constraints: constraintWords + Math.min(4, numbered),
     words,
     lines,
-    question_marks: (text.match(/\?/g) ?? []).length
+    question_marks: (text.match(/\?/g) ?? []).length,
+    long_form: words >= 55 || lines >= 8
   };
   const category = categoryOf(features);
   const complexity = complexityOf(features, category);
   const p_small_quality = pSmallQuality(features, category, complexity);
   const quality_risk = riskOf(category, complexity, p_small_quality);
-  const recommended_tier = tierFromP(p_small_quality);
+  const recommended_tier = recommendTier(p_small_quality, complexity, category, quality_risk);
   const signals = [
     features.code_fence,
     features.code_kw,
@@ -183,9 +196,10 @@ function classifyText(text) {
     features.design,
     features.reason,
     features.debug,
-    features.safety
+    features.safety,
+    features.multi_part
   ].filter(Boolean).length;
-  const confidence = category === "factual_recall" && words < 16 ? 0.9 : Math.min(0.95, 0.55 + 0.12 * signals);
+  const confidence = category === "factual_recall" && words < 16 ? 0.92 : Math.min(0.96, 0.55 + 0.1 * signals);
   return {
     complexity,
     category,
@@ -201,47 +215,64 @@ function classifyText(text) {
   };
 }
 function categoryOf(h) {
-  if (h.safety) return "safety_sensitive";
   if (h.design) return "system_design";
+  if (h.safety && !h.design) return "safety_sensitive";
   if (h.debug && (h.code_fence || h.code_kw)) return "code_debug";
-  if (h.code_fence || h.code_kw) return "code_generation";
+  if (h.debug && h.long_form) return "code_debug";
   if (h.math && h.reason) return "reasoning";
-  if (h.math) return "math";
   if (h.reason) return "reasoning";
+  if (h.code_fence || h.code_kw) return "code_generation";
+  if (h.math) return "math";
+  if (h.safety) return "safety_sensitive";
   if (h.translate) return "translation";
   if (h.summarize) return "summarization";
   if (h.creative) return "creative";
-  if (h.analysis) return "analysis";
-  if (h.words < 24 && h.question_marks >= 1) return "factual_recall";
+  if (h.analysis || h.multi_part) return "analysis";
+  if (h.words < 28 && h.question_marks >= 1) return "factual_recall";
   return h.words > 80 ? "analysis" : "factual_recall";
 }
 function complexityOf(h, category) {
   let score = 1;
-  if (h.words > 40) score += 1;
-  if (h.words > 120) score += 1;
-  if (h.lines > 12 || h.code_fence) score += 1;
+  if (h.words > 35) score += 1;
+  if (h.words > 90) score += 1;
+  if (h.lines > 10 || h.code_fence) score += 1;
   if (h.constraints >= 2) score += 1;
-  if (h.design || h.reason) score += 1;
+  if (h.constraints >= 4 || h.multi_part) score += 1;
+  if (h.design) score += 1;
+  if (h.reason) score += 1;
   if (h.debug) score += 1;
-  if (["system_design", "reasoning", "safety_sensitive"].includes(category)) score = Math.max(score + 1, 4);
-  if (category === "code_generation") score = Math.max(score, 3);
+  if (h.safety && (h.design || h.multi_part)) score += 1;
+  if (category === "system_design") score = Math.max(score, 4);
+  if (category === "reasoning") score = Math.max(score, 4);
+  if (category === "safety_sensitive") score = Math.max(score, 4);
   if (category === "code_debug") score = Math.max(score, 4);
-  if (category === "factual_recall" && h.words < 20) score = Math.min(score, 2);
+  if (category === "code_generation") score = Math.max(score, 3);
+  if (category === "analysis" && (h.multi_part || h.long_form)) score = Math.max(score, 3);
+  if (h.multi_part && (h.design || h.safety || h.analysis)) score = Math.max(score, 4);
+  if (category === "factual_recall" && h.words < 28 && !h.design && !h.reason && !h.debug) {
+    score = Math.min(score, 2);
+  }
+  if (category === "math" && h.words < 24 && !h.reason) {
+    score = Math.min(Math.max(score, 1), 2);
+  }
   return Math.max(1, Math.min(5, score));
 }
 function pSmallQuality(h, category, complexity) {
   let p = 0.96;
-  if (h.design || category === "system_design") p -= 0.28;
-  if (h.safety || category === "safety_sensitive") p -= 0.3;
-  if (h.debug || category === "code_debug") p -= 0.22;
-  if (h.reason || category === "reasoning") p -= 0.18;
-  if (complexity >= 5) p -= 0.22;
-  else if (complexity >= 4) p -= 0.14;
-  else if (complexity === 3) p -= 0.06;
-  if (h.words > 120) p -= 0.07;
-  if (h.constraints >= 2) p -= 0.07;
-  if (category === "code_generation") p -= 0.05;
-  if (category === "factual_recall" && h.words < 24) p = Math.max(p, 0.94);
+  if (h.design || category === "system_design") p -= 0.3;
+  if (h.safety || category === "safety_sensitive") p -= 0.32;
+  if (h.debug || category === "code_debug") p -= 0.24;
+  if (h.reason || category === "reasoning") p -= 0.22;
+  if (h.multi_part) p -= 0.1;
+  if (complexity >= 5) p -= 0.24;
+  else if (complexity >= 4) p -= 0.16;
+  else if (complexity === 3) p -= 0.07;
+  if (h.words > 100) p -= 0.07;
+  if (h.constraints >= 3) p -= 0.08;
+  if (category === "code_generation") p -= 0.06;
+  if (category === "analysis" && complexity >= 3) p -= 0.08;
+  if (category === "factual_recall" && h.words < 28) p = Math.max(p, 0.94);
+  if (category === "math" && h.words < 24 && !h.reason) p = Math.max(p, 0.9);
   return Number(Math.min(0.99, Math.max(0.05, p)).toFixed(3));
 }
 function riskOf(category, complexity, p) {
@@ -254,6 +285,21 @@ function tierFromP(p) {
   if (p >= 0.72) return "standard";
   return "frontier";
 }
+function recommendTier(p, complexity, category, risk) {
+  const fromP = tierFromP(p);
+  const fromComplexity = difficultyTier(complexity);
+  let tier = maxTier(fromP, fromComplexity);
+  if (risk === "high") tier = maxTier(tier, "standard");
+  if (["system_design", "reasoning", "safety_sensitive"].includes(category) && complexity >= 4) {
+    tier = maxTier(tier, complexity >= 5 ? "frontier" : "standard");
+  }
+  if (category === "system_design" && complexity >= 4) tier = maxTier(tier, "frontier");
+  if (category === "reasoning" && complexity >= 4) tier = maxTier(tier, "standard");
+  return tier;
+}
+function maxTier(a, b) {
+  return TIER_RANK[a] >= TIER_RANK[b] ? a : b;
+}
 function difficultyTier(complexity) {
   if (complexity <= 2) return "economy";
   if (complexity === 3) return "standard";

package/dist/index.js

@@ -101,6 +101,7 @@ var Promptimizer = class _Promptimizer {
 
 // src/classify.ts
 var HIGH_RISK = /* @__PURE__ */ new Set(["system_design", "safety_sensitive", "code_debug", "reasoning"]);
+var TIER_RANK = { economy: 0, standard: 1, frontier: 2 };
 function textFrom(messages) {
   return messages.map((m) => {
     if (typeof m.content === "string") return m.content;
@@ -113,34 +114,46 @@ function classifyMessages(messages) {
 function classifyText(text) {
   const words = text.split(/\s+/).filter(Boolean).length;
   const lines = text.split("\n").length;
+  const numbered = (text.match(/\(\s*\d+\s*\)/g) ?? []).length + (text.match(/^\s*\d+[\).:]/gm) ?? []).length + (text.match(/\b\d+\.\s+[A-Za-z]/g) ?? []).length;
+  const constraintWords = (text.match(
+    /\b(must|include|constraint|requirement|requirements|need to|ensure|exactly-once|idempotent)\b/gi
+  ) ?? []).length;
   const features = {
     code_fence: /```/.test(text),
-    code_kw: /\b(def |class |function |import |fn |pub |async |SELECT |goroutine|mutex|traceback)\b/i.test(
+    code_kw: /\b(def |class |function |import |fn |pub |async |goroutine|mutex|traceback)\b/i.test(text) || /\b(SELECT\b.+\bFROM\b|\bJOIN\b.+\bON\b|postgresql|\bpsql\b)/i.test(text),
+    math: /(\$\$|\\frac|prove that|expected value|O\([nN]\)|\d+\s*[\*\^×x]\s*\d+|infinitely many|factorial|prime)/i.test(
       text
     ),
-    math: /(\$\$|\\frac|prove that|expected value|O\([nN]\)|\d+\s*[\*\^]\s*\d+)/i.test(text),
-    design: /\b(design|architect|rate limiter|distributed|consistency|shard|1 million QPS)\b/i.test(
+    design: /\b(design(?:ing)?|architect(?:ure|ing)?|rate[\s-]?limiter|distributed|consistency|shard(?:ing)?|multi-?region|multi-?tenant|qps|edge regions?|data[\s-]?plane|failure modes?|outbox|ledger|webhook|saas|billing pipeline|consensus|raft|paxos|quorum|leader election|log replication|network partition|cap theorem|eventual consistency|token bucket|sliding window)\b/i.test(
       text
     ),
-    reason: /\b(prove|why does|walk through|step by step|derive|contradiction|p-value)\b/i.test(
+    reason: /\b(prove|proof|theorem|lemma|why does|walk through|step by step|derive|contradiction|p-value|induction|euclid|sketch (?:a |the )?proof)\b/i.test(
+      text
+    ),
+    debug: /\b(bug|race|panic|fails on|diagnose|deadlock|concurrency hazard|code review|correctness)\b/i.test(
       text
     ),
-    debug: /\b(bug|race|panic|fails on|diagnose|deadlock)\b/i.test(text),
     summarize: /\b(summarize|tl;dr|in two sentences|eli5)\b/i.test(text),
     translate: /\b(translate|traduce)\b/i.test(text),
     creative: /\b(write a (poem|story|song)|haiku)\b/i.test(text),
-    safety: /\b(refund|legal|medical|hipaa|lawsuit|diagnosis)\b/i.test(text),
-    analysis: /\b(compare|trade-?off|versus|analyse|analyze|evaluate|should we)\b/i.test(text),
-    constraints: (text.match(/\b(must|include|constraints?|requirements?)\b/gi) ?? []).length,
+    safety: /\b(refund|legal|medical|hipaa|lawsuit|diagnosis|gdpr|erasure|pii|compliance|audit trail)\b/i.test(
+      text
+    ),
+    analysis: /\b(compare|trade-?off|versus|analyse|analyze|evaluate|should we|outline architecture)\b/i.test(
+      text
+    ),
+    multi_part: numbered >= 3 || numbered >= 2 && constraintWords >= 1,
+    constraints: constraintWords + Math.min(4, numbered),
     words,
     lines,
-    question_marks: (text.match(/\?/g) ?? []).length
+    question_marks: (text.match(/\?/g) ?? []).length,
+    long_form: words >= 55 || lines >= 8
   };
   const category = categoryOf(features);
   const complexity = complexityOf(features, category);
   const p_small_quality = pSmallQuality(features, category, complexity);
   const quality_risk = riskOf(category, complexity, p_small_quality);
-  const recommended_tier = tierFromP(p_small_quality);
+  const recommended_tier = recommendTier(p_small_quality, complexity, category, quality_risk);
   const signals = [
     features.code_fence,
     features.code_kw,
@@ -148,9 +161,10 @@ function classifyText(text) {
     features.design,
     features.reason,
     features.debug,
-    features.safety
+    features.safety,
+    features.multi_part
   ].filter(Boolean).length;
-  const confidence = category === "factual_recall" && words < 16 ? 0.9 : Math.min(0.95, 0.55 + 0.12 * signals);
+  const confidence = category === "factual_recall" && words < 16 ? 0.92 : Math.min(0.96, 0.55 + 0.1 * signals);
   return {
     complexity,
     category,
@@ -166,47 +180,64 @@ function classifyText(text) {
   };
 }
 function categoryOf(h) {
-  if (h.safety) return "safety_sensitive";
   if (h.design) return "system_design";
+  if (h.safety && !h.design) return "safety_sensitive";
   if (h.debug && (h.code_fence || h.code_kw)) return "code_debug";
-  if (h.code_fence || h.code_kw) return "code_generation";
+  if (h.debug && h.long_form) return "code_debug";
   if (h.math && h.reason) return "reasoning";
-  if (h.math) return "math";
   if (h.reason) return "reasoning";
+  if (h.code_fence || h.code_kw) return "code_generation";
+  if (h.math) return "math";
+  if (h.safety) return "safety_sensitive";
   if (h.translate) return "translation";
   if (h.summarize) return "summarization";
   if (h.creative) return "creative";
-  if (h.analysis) return "analysis";
-  if (h.words < 24 && h.question_marks >= 1) return "factual_recall";
+  if (h.analysis || h.multi_part) return "analysis";
+  if (h.words < 28 && h.question_marks >= 1) return "factual_recall";
   return h.words > 80 ? "analysis" : "factual_recall";
 }
 function complexityOf(h, category) {
   let score = 1;
-  if (h.words > 40) score += 1;
-  if (h.words > 120) score += 1;
-  if (h.lines > 12 || h.code_fence) score += 1;
+  if (h.words > 35) score += 1;
+  if (h.words > 90) score += 1;
+  if (h.lines > 10 || h.code_fence) score += 1;
   if (h.constraints >= 2) score += 1;
-  if (h.design || h.reason) score += 1;
+  if (h.constraints >= 4 || h.multi_part) score += 1;
+  if (h.design) score += 1;
+  if (h.reason) score += 1;
   if (h.debug) score += 1;
-  if (["system_design", "reasoning", "safety_sensitive"].includes(category)) score = Math.max(score + 1, 4);
-  if (category === "code_generation") score = Math.max(score, 3);
+  if (h.safety && (h.design || h.multi_part)) score += 1;
+  if (category === "system_design") score = Math.max(score, 4);
+  if (category === "reasoning") score = Math.max(score, 4);
+  if (category === "safety_sensitive") score = Math.max(score, 4);
   if (category === "code_debug") score = Math.max(score, 4);
-  if (category === "factual_recall" && h.words < 20) score = Math.min(score, 2);
+  if (category === "code_generation") score = Math.max(score, 3);
+  if (category === "analysis" && (h.multi_part || h.long_form)) score = Math.max(score, 3);
+  if (h.multi_part && (h.design || h.safety || h.analysis)) score = Math.max(score, 4);
+  if (category === "factual_recall" && h.words < 28 && !h.design && !h.reason && !h.debug) {
+    score = Math.min(score, 2);
+  }
+  if (category === "math" && h.words < 24 && !h.reason) {
+    score = Math.min(Math.max(score, 1), 2);
+  }
   return Math.max(1, Math.min(5, score));
 }
 function pSmallQuality(h, category, complexity) {
   let p = 0.96;
-  if (h.design || category === "system_design") p -= 0.28;
-  if (h.safety || category === "safety_sensitive") p -= 0.3;
-  if (h.debug || category === "code_debug") p -= 0.22;
-  if (h.reason || category === "reasoning") p -= 0.18;
-  if (complexity >= 5) p -= 0.22;
-  else if (complexity >= 4) p -= 0.14;
-  else if (complexity === 3) p -= 0.06;
-  if (h.words > 120) p -= 0.07;
-  if (h.constraints >= 2) p -= 0.07;
-  if (category === "code_generation") p -= 0.05;
-  if (category === "factual_recall" && h.words < 24) p = Math.max(p, 0.94);
+  if (h.design || category === "system_design") p -= 0.3;
+  if (h.safety || category === "safety_sensitive") p -= 0.32;
+  if (h.debug || category === "code_debug") p -= 0.24;
+  if (h.reason || category === "reasoning") p -= 0.22;
+  if (h.multi_part) p -= 0.1;
+  if (complexity >= 5) p -= 0.24;
+  else if (complexity >= 4) p -= 0.16;
+  else if (complexity === 3) p -= 0.07;
+  if (h.words > 100) p -= 0.07;
+  if (h.constraints >= 3) p -= 0.08;
+  if (category === "code_generation") p -= 0.06;
+  if (category === "analysis" && complexity >= 3) p -= 0.08;
+  if (category === "factual_recall" && h.words < 28) p = Math.max(p, 0.94);
+  if (category === "math" && h.words < 24 && !h.reason) p = Math.max(p, 0.9);
   return Number(Math.min(0.99, Math.max(0.05, p)).toFixed(3));
 }
 function riskOf(category, complexity, p) {
@@ -219,6 +250,21 @@ function tierFromP(p) {
   if (p >= 0.72) return "standard";
   return "frontier";
 }
+function recommendTier(p, complexity, category, risk) {
+  const fromP = tierFromP(p);
+  const fromComplexity = difficultyTier(complexity);
+  let tier = maxTier(fromP, fromComplexity);
+  if (risk === "high") tier = maxTier(tier, "standard");
+  if (["system_design", "reasoning", "safety_sensitive"].includes(category) && complexity >= 4) {
+    tier = maxTier(tier, complexity >= 5 ? "frontier" : "standard");
+  }
+  if (category === "system_design" && complexity >= 4) tier = maxTier(tier, "frontier");
+  if (category === "reasoning" && complexity >= 4) tier = maxTier(tier, "standard");
+  return tier;
+}
+function maxTier(a, b) {
+  return TIER_RANK[a] >= TIER_RANK[b] ? a : b;
+}
 function difficultyTier(complexity) {
   if (complexity <= 2) return "economy";
   if (complexity === 3) return "standard";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "promptimizer",
-  "version": "0.1.36",
+  "version": "0.1.37",
   "description": "OpenAI-compatible SDK for Promptimizer — BYOK model routing, classification, and prompt caching.",
   "type": "module",
   "main": "./dist/index.cjs",