@konneal/engine 0.2.44 → 0.2.46

package/dist/{ask-JOX2D7KH.js → ask-6QPD3QA5.js}

@@ -1,6 +1,6 @@
 import {
   handleAsk
-} from "./chunk-HDUSO5S5.js";
+} from "./chunk-SE67PTJ6.js";
 import "./chunk-N2UOENDY.js";
 import "./chunk-R2V3X6SQ.js";
 import "./chunk-A3QHHUN5.js";

package/dist/ask.d.ts

@@ -2,4 +2,7 @@ import type { Env } from "./env";
 export type { Env };
 import { type ApiKey } from "./lib/http";
 declare function handleAsk(env: Env, ctx: ExecutionContext, req: Request, tier: "anon" | "key" | "member", key: ApiKey | null): Promise<Response>;
+/** RAGAS-style metric battery (G13): judge an (question, answer, passages)
+ *  triple — faithfulness, answer relevancy, context precision. Driven by
+ *  tests/eval-suite.mjs; prompts are data; refuses nothing, judges only. */
 export { handleAsk };

package/dist/certificates.d.ts

@@ -15,6 +15,10 @@ export declare function isRegisterShaped(query: string): boolean;
  *  2+ characters that are not register question words, plus bare model
  *  numbers ("190", "HM14H1"). Upper-cased tokens match case-insensitively
  *  in SQL LIKE. */
+/** The family the question names ("R 60", "R60", "D 31") — the register
+ *  search filters to it, so a question about R 60 never presents R 76
+ *  rows. */
+export declare function queryFamily(query: string): string | null;
 export declare function registerTokens(query: string): string[];
 /** The register query for a token set. Numeric tokens (a model number
  *  like "190") must match as a STANDALONE value — `LIKE '%190%'` also

package/dist/{chunk-HDUSO5S5.js → chunk-SE67PTJ6.js}

@@ -1743,14 +1743,15 @@ function cacheKeyMaterial(query, lang, salt) {
 function exactCacheKey(indexVersion, gen, ns, queryHash) {
   return `a:${indexVersion}:g${gen}:${ns}:${queryHash}`;
 }
-function semanticCacheKey(indexVersion, gen, signature) {
-  return `sc:${indexVersion}:g${gen}:${signature}`;
-}
 
 // workers/worker_public/src/certificates.ts
 function isRegisterShaped(query) {
   return /\bcertif(ied|icates?|ication)s?\b/i.test(query) && /\b(is|are|still|currently|valid|status|suspended|revoked|was|were)\b/i.test(query);
 }
+function queryFamily(query) {
+  const m = /\b([RDMB])\s?-?(\d{2,3})\b/i.exec(query);
+  return m ? `${m[1].toUpperCase()}${m[2].padStart(2, "0")}` : null;
+}
 function registerTokens(query) {
   const stop = /* @__PURE__ */ new Set([
     "the",
@@ -1833,7 +1834,13 @@ async function searchRegister(db, query) {
   if (!isRegisterShaped(query)) return null;
   const tokens2 = registerTokens(query);
   if (!tokens2.length) return null;
-  const { sql, params } = buildRegisterQuery(tokens2);
+  const built = buildRegisterQuery(tokens2);
+  let { sql, params } = built;
+  const family = queryFamily(query);
+  if (family) {
+    sql = sql.replace(" WHERE ", " WHERE family = ?0 AND ");
+    params.unshift(family);
+  }
   try {
     const res = await db.prepare(sql).bind(...params).all();
     return { rows: res.results ?? [], tokens: tokens2 };
@@ -2083,22 +2090,6 @@ async function handleAsk(env, ctx, req, tier, key) {
     }
   }
   let understanding = null;
-  const nodeScoped = !!modelNodeRefIn(q.query) || !!modelNodeRefIn(declaredCtx?.label);
-  if (!cached && !nodeScoped && !contextual && !declaredCtx && !draftAct && !apiCallIntent && !q.lang && !userImage && !fresh) {
-    const wv0 = await warmEmbed ?? null;
-    if (wv0) {
-      const sc0 = await semanticCacheGet(env, gen, wv0, salt);
-      if (sc0) {
-        console.log("semantic cache hit (pre-understanding)");
-        telemetry(env, ctx, tier, "ask", null, true, sc0.answer.length, sc0.query_hash, q.lang, "semantic", telemetryMeta());
-        const cctx0 = sc0.context_applied ?? NO_CONTEXT;
-        if (wantsStream) {
-          return sseResponse([{ type: "citations", citations: sc0.citations ?? [], context_applied: cctx0 }, { type: "token", v: sc0.answer }, { type: "done", model: sc0.model, query_hash: sc0.query_hash, similar: true, served_from: "similar", context_applied: cctx0 }], corsHeaders(req));
-        }
-        return json({ ...sc0, similar: true, context_applied: cctx0, quota });
-      }
-    }
-  }
   let optimisticVec = null;
   let optimisticHits = [];
   const t0 = Date.now();
@@ -2183,27 +2174,8 @@ async function handleAsk(env, ctx, req, tier, key) {
   const graphDocNumbers = await graphExpand(env, understanding);
   const eNote = await editionNote(env, understanding);
   const register = await searchRegister(env.DB, q.query);
+  console.log("register-search:", JSON.stringify({ shaped: !!register, tokens: register?.tokens ?? null, rows: register?.rows?.length ?? null, first: register?.rows?.[0]?.num ?? null }));
   const regNote = register ? registerNote(register.rows) : void 0;
-  if (understanding?.intent !== "conversational" && !nodeScoped && !contextual && !declaredCtx && !draftAct && !apiCallIntent && !userImage && !fresh) {
-    const warmVec = await warmEmbed ?? null;
-    if (warmVec) {
-      const sc = await semanticCacheGet(env, gen, warmVec, salt);
-      if (sc) {
-        console.log("semantic cache hit");
-        telemetry(env, ctx, tier, "ask", null, true, sc.answer.length, sc.query_hash, q.lang, "semantic", telemetryMeta());
-        const cctx = sc.context_applied ?? NO_CONTEXT;
-        if (wantsStream) {
-          return sseResponse([
-            ...readAs() ? [{ type: "read", read: readAs() }] : [],
-            { type: "citations", citations: sc.citations ?? [], context_applied: cctx },
-            { type: "token", v: sc.answer },
-            { type: "done", model: sc.model, query_hash: sc.query_hash, similar: true, served_from: "similar", context_applied: cctx, read: readAs() }
-          ], corsHeaders(req));
-        }
-        return json({ ...sc, similar: true, context_applied: cctx, read: readAs(), quota });
-      }
-    }
-  }
   if (understanding?.intent === "conversational") {
     const queryHash2 = await sha256Hex(q.query);
     const messages2 = [
@@ -2617,8 +2589,6 @@ Answer account questions from these records ONLY: name the record when you use i
             console.log("anchors:", streamed.violations.length, "of", streamed.total, "unverified \u2014 not caching");
           }
           if (streamed.violations.length === 0 && canonical.length > 0 && !contextual && !declaredCtx && !canonical.includes(refusalAnswer())) {
-            const wv = await warmEmbed ?? null;
-            if (wv) semanticCachePut(env, ctx, gen, wv, salt, { answer: canonical, citations: cites, model, query_hash: queryHash });
             ctx.waitUntil(
               env.CACHE.put(exactCacheKey(env.INDEX_VERSION, gen, ns, await sha256Hex(cacheKeyMaterial(q.query, q.lang, salt))), JSON.stringify({ answer: canonical, citations: cites, model, query_hash: queryHash }), { expirationTtl: LIMITS.cacheTtlSec })
             );
@@ -2720,10 +2690,6 @@ Answer account questions from these records ONLY: name the record when you use i
   const jsonQuality = answerQuality(finalCites.map((c) => c.quality));
   const out = { answer, citations: finalCites, ...jsonQuality ? { source_quality: jsonQuality, confidence_note: qualityNote(jsonQuality), ...jsonQuality === "ocr" ? { experimental_sources: experimentalSourceLabels(finalCites) } : {} } : {}, model: MODELS.member, query_hash: queryHash, follow_ups: understanding?.follow_ups ?? [], blocks: [...c2ns.blocks, ...verdictBlock ? [verdictBlock] : [], ...conditionBlock ? [conditionBlock] : [], ...aggregationBlock ? [aggregationBlock] : [], ...completionBlocks], context_applied: ctxApplied, ...liveRecords ? { records: liveRecords } : {} };
   const cacheable = !contextual && !declaredCtx && !answer.includes(refusalAnswer()) && finalAnchors.violations.length === 0;
-  if (cacheable) {
-    const warmVec = await warmEmbed ?? null;
-    if (warmVec) semanticCachePut(env, ctx, gen, warmVec, salt, out);
-  }
   if (cacheable) {
     const ck = exactCacheKey(env.INDEX_VERSION, gen, ns, await sha256Hex(cacheKeyMaterial(q.query, q.lang, salt)));
     ctx.waitUntil(env.CACHE.put(ck, JSON.stringify(out), { expirationTtl: LIMITS.cacheTtlSec }));
@@ -2751,36 +2717,6 @@ function sseResponse(events, cors) {
     }
   });
 }
-function cosine(a, b) {
-  let dot = 0;
-  let na = 0;
-  let nb = 0;
-  for (let i = 0; i < a.length; i++) {
-    dot += a[i] * b[i];
-    na += a[i] * a[i];
-    nb += b[i] * b[i];
-  }
-  return dot / (Math.sqrt(na) * Math.sqrt(nb) || 1);
-}
-function scSignature(v, salt) {
-  return v.slice(0, 16).map((x) => x.toFixed(2)).join(",") + (salt ? `|s:${salt.length}:${salt.slice(0, 64)}` : "");
-}
-async function semanticCacheGet(env, gen, vec, salt) {
-  try {
-    const raw = await env.CACHE.get(semanticCacheKey(env.INDEX_VERSION, gen, scSignature(vec, salt)), "json");
-    if (!raw?.v || !Array.isArray(raw.v) || raw.v.length !== vec.length) return null;
-    if (cosine(raw.v, vec) < 0.97) return null;
-    return raw;
-  } catch {
-    return null;
-  }
-}
-function semanticCachePut(env, ctx, gen, vec, salt, payload) {
-  const v = vec.map((x) => Number(x.toFixed(3)));
-  ctx.waitUntil(
-    env.CACHE.put(semanticCacheKey(env.INDEX_VERSION, gen, scSignature(vec, salt)), JSON.stringify({ v, ...payload }), { expirationTtl: LIMITS.cacheTtlSec })
-  );
-}
 
 export {
   handleMemories,

package/dist/worker_public/src/index.js

@@ -9,7 +9,7 @@ import {
   modelGroundingBlock,
   scoreJudge,
   standardForDocNumber
-} from "../../chunk-HDUSO5S5.js";
+} from "../../chunk-SE67PTJ6.js";
 import {
   buildMessages,
   citations,
@@ -999,7 +999,7 @@ async function handleMcp(env, ctx, req, tier, key) {
       // stream:false forces the JSON lane (anon defaults to SSE)
       body: JSON.stringify({ ...args, stream: false })
     });
-    const res = name === "ask" ? await (await import("../../ask-JOX2D7KH.js")).handleAsk(env, ctx, inner, tier, key) : await (await import("../../search-QZDQVB57.js")).handleSearch(env, ctx, inner, tier, key);
+    const res = name === "ask" ? await (await import("../../ask-6QPD3QA5.js")).handleAsk(env, ctx, inner, tier, key) : await (await import("../../search-QZDQVB57.js")).handleSearch(env, ctx, inner, tier, key);
     return res.json().catch(() => ({ error: { message: "tool transport failed", status: res.status } }));
   });
   if (out.ok && "accepted" in out) return new Response(null, { status: 202 });

package/package.json

@@ -27,7 +27,7 @@
     "playwright": "^1.62.1",
     "yaml": "^2.9.1"
   },
-  "version": "0.2.44",
+  "version": "0.2.46",
   "description": "The Konneal engine: the publisher-agnostic build pipeline and API plane for standards intelligence (retrieval, answer contract, verdicts, evaluation).",
   "license": "BSD-3-Clause",
   "type": "module",

package/workers/worker_public/src/ask.ts

@@ -20,7 +20,7 @@ import { contractV2, tableRetyped } from "./refs";
 import { completeTables, completeFigures } from "./completion";
 import { NO_CONTEXT, appliedContext, contextNote, namedDocumentIn, parseContext, resolveDocScope, syntheticUnderstanding } from "./context";
 import { liveDataConfig, liveTokenFor, opCfg, opTokenMember, resolveLiveAccount, type LiveRecord } from "./livedata";
-import { bindModelNode, licenseBoundaryNote, licenseBoundaryRefusal, licensedEntryForPackage, modelCitation, modelEcho, modelGroundingBlock, modelNodeRefIn, standardForDocNumber } from "./modelplane";
+import { bindModelNode, licenseBoundaryNote, licenseBoundaryRefusal, licensedEntryForPackage, modelCitation, modelEcho, modelGroundingBlock, standardForDocNumber } from "./modelplane";
 import { evaluate as machineEvaluate, verdictNote } from "./verdict";
 import { evaluateConditionSets, quantitiesIn, type ConditionVerdict } from "./conditions";
 import { evaluateAggregation, type AggregationVerdict } from "./aggregation";
@@ -31,7 +31,7 @@ import { detectApiCallIntent, prepareApiCall } from "./apicalls";
 import { memoryNote } from "./memories";
 import { entitlementScope, resolveRequestScope, requestSalt, standardKeysFrom } from "./requestScope";
 import { rawSessionToken, type SessionClaims } from "./session";
-import { cacheKeyMaterial, corpusGen, exactCacheKey, freshRequested, semanticCacheKey } from "./answercache";
+import { cacheKeyMaterial, corpusGen, exactCacheKey, freshRequested } from "./answercache";
 import type { Env } from "./env";
 export type { Env };
 import { json, err, corsHeaders, readJson, validateQuery, type ApiKey } from "./lib/http";
@@ -435,26 +435,6 @@ async function handleAsk(
   // fresh already bypassed the exact cache above and bypasses this one.
   let understanding: any = null;
   // a query naming a model node (/req/…, /term/…) is node-SCOPED: its
-  // embedding sits near every other node-scoped ask about the same
-  // standard, and the single-entry semantic bucket then serves one
-  // node's answer for another (observed run-to-run across the golden
-  // model legs). Node-scoped queries use the exact cache only.
-  const nodeScoped = !!modelNodeRefIn(q.query) || !!modelNodeRefIn(declaredCtx?.label);
-  if (!cached && !nodeScoped && !contextual && !declaredCtx && !draftAct && !apiCallIntent && !q.lang && !userImage && !fresh) {
-    const wv0 = (await warmEmbed) ?? null;
-    if (wv0) {
-      const sc0 = await semanticCacheGet(env, gen, wv0, salt);
-      if (sc0) {
-        console.log("semantic cache hit (pre-understanding)");
-        telemetry(env, ctx, tier, "ask", null, true, sc0.answer.length, sc0.query_hash, q.lang, "semantic", telemetryMeta());
-        const cctx0 = sc0.context_applied ?? NO_CONTEXT;
-        if (wantsStream) {
-          return sseResponse([{ type: "citations", citations: sc0.citations ?? [], context_applied: cctx0 }, { type: "token", v: sc0.answer }, { type: "done", model: sc0.model, query_hash: sc0.query_hash, similar: true, served_from: "similar", context_applied: cctx0 }], corsHeaders(req));
-        }
-        return json({ ...sc0, similar: true, context_applied: cctx0, quota, });
-      }
-    }
-  }
   // ── Option C: optimistic parallel retrieval ──
   // The dense lane (folded-query embed + unfiltered Vectorize query) runs
   // CONCURRENTLY with understanding instead of after it — the serial
@@ -583,34 +563,9 @@ async function handleAsk(
   // join the prompt as an authoritative note — exact matches, never a
   // similarity guess about certification standing
   const register = await searchRegister(env.DB, q.query);
+  console.log("register-search:", JSON.stringify({ shaped: !!register, tokens: register?.tokens ?? null, rows: register?.rows?.length ?? null, first: register?.rows?.[0]?.num ?? null }));
   const regNote = register ? registerNote(register.rows) : undefined;
 
-  // semantic cache: near-duplicate of a recently answered question —
-  // serves the stored answer with a `similar: true` marker (checked only
-  // for standalone knowledge questions; contextual turns, declared-context
-  // asks and image asks always run live; fresh regenerates, bypassing
-  // this cache too)
-  if (understanding?.intent !== "conversational" && !nodeScoped && !contextual && !declaredCtx && !draftAct && !apiCallIntent && !userImage && !fresh) {
-    const warmVec = (await warmEmbed) ?? null;
-    if (warmVec) {
-      const sc = await semanticCacheGet(env, gen, warmVec, salt);
-      if (sc) {
-        console.log("semantic cache hit");
-        telemetry(env, ctx, tier, "ask", null, true, sc.answer.length, sc.query_hash, q.lang, "semantic", telemetryMeta());
-        const cctx = sc.context_applied ?? NO_CONTEXT;
-        if (wantsStream) {
-          return sseResponse([
-            ...(readAs() ? [{ type: "read", read: readAs() }] : []),
-            { type: "citations", citations: sc.citations ?? [], context_applied: cctx },
-            { type: "token", v: sc.answer },
-            { type: "done", model: sc.model, query_hash: sc.query_hash, similar: true, served_from: "similar", context_applied: cctx, read: readAs() },
-          ], corsHeaders(req));
-        }
-        return json({ ...sc, similar: true, context_applied: cctx, read: readAs(), quota, });
-      }
-    }
-  }
-
   // Conversational route, decided by query UNDERSTANDING (any language, any
   // phrasing) — not string matching. No retrieval: nothing in the corpus
   // answers "who are you". The model speaks for itself from the service
@@ -1192,8 +1147,6 @@ async function handleAsk(
             console.log("anchors:", streamed.violations.length, "of", streamed.total, "unverified — not caching");
           }
           if (streamed.violations.length === 0 && canonical.length > 0 && !contextual && !declaredCtx && !canonical.includes(refusalAnswer())) {
-            const wv = (await warmEmbed) ?? null;
-            if (wv) semanticCachePut(env, ctx, gen, wv, salt, { answer: canonical, citations: cites, model, query_hash: queryHash });
             ctx.waitUntil(
               env.CACHE.put(exactCacheKey(env.INDEX_VERSION, gen, ns, await sha256Hex(cacheKeyMaterial(q.query, q.lang, salt))), JSON.stringify({ answer: canonical, citations: cites, model, query_hash: queryHash }), { expirationTtl: LIMITS.cacheTtlSec }),
             );
@@ -1358,10 +1311,6 @@ async function handleAsk(
   const jsonQuality = answerQuality(finalCites.map((c: any) => c.quality));
   const out = { answer, citations: finalCites, ...(jsonQuality ? { source_quality: jsonQuality, confidence_note: qualityNote(jsonQuality), ...(jsonQuality === "ocr" ? { experimental_sources: experimentalSourceLabels(finalCites) } : {}) } : {}), model: MODELS.member, query_hash: queryHash, follow_ups: understanding?.follow_ups ?? [], blocks: [...c2ns.blocks, ...(verdictBlock ? [verdictBlock] : []), ...(conditionBlock ? [conditionBlock] : []), ...(aggregationBlock ? [aggregationBlock] : []), ...completionBlocks], context_applied: ctxApplied, ...(liveRecords ? { records: liveRecords } : {}) };
   const cacheable = !contextual && !declaredCtx && !answer.includes(refusalAnswer()) && finalAnchors.violations.length === 0;
-  if (cacheable) {
-    const warmVec = (await warmEmbed) ?? null;
-    if (warmVec) semanticCachePut(env, ctx, gen, warmVec, salt, out);
-  }
   if (cacheable) {
     const ck = exactCacheKey(env.INDEX_VERSION, gen, ns, await sha256Hex(cacheKeyMaterial(q.query, q.lang, salt)));
     ctx.waitUntil(env.CACHE.put(ck, JSON.stringify(out), { expirationTtl: LIMITS.cacheTtlSec }));
@@ -1403,44 +1352,6 @@ function sseResponse(events: unknown[], cors: Record<string, string>): Response
 
 
 // ── Semantic answer cache (G6) ──
-// Near-duplicate queries re-pay the whole pipeline. Bucket KV by a
-// leading-dimension signature of the query embedding; confirm with full
-// cosine >= 0.97 before serving. Same INDEX_VERSION + corpus-generation
-// namespace as the answer cache (deploys and corpus surgery invalidate
-// both). Single entry per bucket (v1): collisions overwrite, never mix.
-function cosine(a: number[], b: number[]): number {
-  let dot = 0;
-  let na = 0;
-  let nb = 0;
-  for (let i = 0; i < a.length; i++) {
-    dot += a[i] * b[i];
-    na += a[i] * a[i];
-    nb += b[i] * b[i];
-  }
-  return dot / (Math.sqrt(na) * Math.sqrt(nb) || 1);
-}
-
-function scSignature(v: number[], salt?: string | null): string {
-  return v.slice(0, 16).map((x) => x.toFixed(2)).join(",") + (salt ? `|s:${salt.length}:${salt.slice(0, 64)}` : "");
-}
-
-async function semanticCacheGet(env: Env, gen: string, vec: number[], salt?: string | null): Promise<{ answer: string; citations: unknown[]; model: string; query_hash: string; context_applied?: unknown } | null> {
-  try {
-    const raw = await env.CACHE.get(semanticCacheKey(env.INDEX_VERSION, gen, scSignature(vec, salt)), "json") as any;
-    if (!raw?.v || !Array.isArray(raw.v) || raw.v.length !== vec.length) return null;
-    if (cosine(raw.v, vec) < 0.97) return null;
-    return raw;
-  } catch {
-    return null;
-  }
-}
-
-function semanticCachePut(env: Env, ctx: ExecutionContext, gen: string, vec: number[], salt: string | null | undefined, payload: { answer: string; citations: unknown[]; model: string; query_hash: string }): void {
-  const v = vec.map((x) => Number(x.toFixed(3)));
-  ctx.waitUntil(
-    env.CACHE.put(semanticCacheKey(env.INDEX_VERSION, gen, scSignature(vec, salt)), JSON.stringify({ v, ...payload }), { expirationTtl: LIMITS.cacheTtlSec }),
-  );
-}
 
 // ── the route table (TODO.impl/03) ───────────────────────────────────────
 

package/workers/worker_public/src/certificates.ts

@@ -29,6 +29,14 @@ export function isRegisterShaped(query: string): boolean {
  *  2+ characters that are not register question words, plus bare model
  *  numbers ("190", "HM14H1"). Upper-cased tokens match case-insensitively
  *  in SQL LIKE. */
+/** The family the question names ("R 60", "R60", "D 31") — the register
+ *  search filters to it, so a question about R 60 never presents R 76
+ *  rows. */
+export function queryFamily(query: string): string | null {
+  const m = /\b([RDMB])\s?-?(\d{2,3})\b/i.exec(query);
+  return m ? `${m[1].toUpperCase()}${m[2].padStart(2, "0")}` : null;
+}
+
 export function registerTokens(query: string): string[] {
   const stop = new Set([
     "the", "a", "an", "is", "are", "was", "were", "still", "currently", "in", "on", "for", "of", "and", "or",
@@ -84,7 +92,13 @@ export async function searchRegister(db: any, query: string): Promise<{ rows: Re
   if (!isRegisterShaped(query)) return null;
   const tokens = registerTokens(query);
   if (!tokens.length) return null;
-  const { sql, params } = buildRegisterQuery(tokens);
+  const built = buildRegisterQuery(tokens);
+  let { sql, params } = built;
+  const family = queryFamily(query);
+  if (family) {
+    sql = sql.replace(" WHERE ", " WHERE family = ?0 AND ");
+    params.unshift(family);
+  }
   try {
     const res = await db.prepare(sql).bind(...params).all();
     return { rows: (res.results ?? []) as RegisterRow[], tokens };