@hevmind/ask 0.3.2 → 0.3.4

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hevmind/ask",
-  "version": "0.3.2",
+  "version": "0.3.4",
   "type": "module",
   "description": "hev ask: a heading-anchored, agentic search overlay for Astro docs sites.",
   "keywords": [
@@ -28,11 +28,11 @@
     "ask": "./bin/ask.mjs"
   },
   "optionalDependencies": {
-    "@hevmind/ask-darwin-arm64": "0.3.2",
-    "@hevmind/ask-linux-x64": "0.3.2",
-    "@hevmind/ask-linux-arm64": "0.3.2",
-    "@hevmind/ask-darwin-x64": "0.3.2",
-    "@hevmind/ask-win32-x64": "0.3.2"
+    "@hevmind/ask-darwin-arm64": "0.3.4",
+    "@hevmind/ask-linux-arm64": "0.3.4",
+    "@hevmind/ask-darwin-x64": "0.3.4",
+    "@hevmind/ask-win32-x64": "0.3.4",
+    "@hevmind/ask-linux-x64": "0.3.4"
   },
   "exports": {
     ".": "./src/index.ts",

package/src/digest/expand.ts

@@ -2,23 +2,28 @@ import { tokenize } from '../search/chunk.ts';
 import type { GlossaryEntry } from './schema';
 
 export function expandQueryTerms(query: string, glossary: GlossaryEntry[], cap = 24): string[] {
-  const terms = new Set(tokenize(query));
-  if (!terms.size) return [];
+  const queryTokens = new Set(tokenize(query));
+  if (!queryTokens.size) return [];
+  const terms = new Set(queryTokens);
 
   for (const entry of glossary) {
     if (terms.size >= cap) break;
-    const entryTerms = new Set([...tokenize(entry.term), ...entry.aliases.flatMap((alias) => tokenize(alias))]);
-    if (!intersects(terms, entryTerms)) continue;
-    for (const term of entryTerms) {
-      terms.add(term);
+    // Expand only when the query contains a full glossary phrase — the term or one
+    // of its aliases, every token present. Matching on any shared token (e.g. the
+    // ubiquitous "authentication") drags in every entry that merely mentions it,
+    // which floods results once the glossary is large.
+    const phrases = [entry.term, ...entry.aliases].map((phrase) => tokenize(phrase)).filter((tokens) => tokens.length);
+    const matched = phrases.some((phrase) => phrase.every((token) => queryTokens.has(token)));
+    if (!matched) continue;
+    for (const phrase of phrases) {
+      for (const token of phrase) {
+        if (token.length < 3) continue; // skip noisy short tokens like "ad", "id"
+        terms.add(token);
+        if (terms.size >= cap) break;
+      }
       if (terms.size >= cap) break;
     }
   }
 
   return [...terms];
 }
-
-function intersects(a: Set<string>, b: Set<string>): boolean {
-  for (const item of b) if (a.has(item)) return true;
-  return false;
-}

package/src/search/prefilter.ts

@@ -81,11 +81,16 @@ export function prefilter(
   const scored = chunks
     .map((chunk) => {
       const boost = signal.get(chunk.id);
+      // A query term in the section heading or page title is a strong topical
+      // signal — the page titled "OIDC Authentication" is what someone asking
+      // about "oidc" wants, far more than a page that merely mentions it.
+      const headingTokens = new Set([...tokenize(chunk.heading ?? ''), ...tokenize(chunk.docTitle ?? '')]);
       let raw = 0;
       for (const term of terms) {
         const weight = weights.get(term) ?? 0;
         if (chunk.tokens.has(term)) raw += weight;
         if (boost?.has(term)) raw += weight;
+        if (headingTokens.has(term)) raw += weight * 3;
       }
       // Length-penalize (not -reward): a huge page (e.g. an autogenerated CLI flag
       // dump that mentions nearly every term) is divided down, but short sections