search target fix

lib/text-highlight.ts

@@ -206,28 +206,46 @@ export function extractMatchedTextFromSnippet(
   const firstStart = first.index ?? 0;
   const lastEnd = (last.index ?? 0) + last[0].length;
 
-  const span = snippetHtml.slice(firstStart, lastEnd);
-  const plain = normalizeQuotes(
-    decodeSnippetEntities(span.replace(/<[^>]*>/g, ""))
+  // Marked-span text — just the tokens Typesense highlighted.
+  const markedSpan = snippetHtml.slice(firstStart, lastEnd);
+  const markedPlain = normalizeQuotes(
+    decodeSnippetEntities(markedSpan.replace(/<[^>]*>/g, ""))
       .replace(/\s+/g, " ")
       .trim(),
   );
 
-  if (!plain) return null;
+  // Full-snippet text — includes Typesense's surrounding context. Typesense
+  // snippets are direct slices of the source field, so the surrounding text
+  // exists verbatim in the destination page. Using it as the prefix makes
+  // single-token matches (e.g. `<mark>Therefore</mark> material senses...`)
+  // land on the *correct* occurrence on the destination page rather than the
+  // first instance of that token alphabetically.
+  const fullPlain = normalizeQuotes(
+    decodeSnippetEntities(snippetHtml.replace(/<[^>]*>/g, ""))
+      .replace(/\s+/g, " ")
+      // strip leading/trailing ellipsis Typesense sometimes adds
+      .replace(/^(?:…|\.{3,})\s*/, "")
+      .replace(/\s*(?:…|\.{3,})$/, "")
+      .trim(),
+  );
 
-  // If the span between first and last mark is unreasonably long
-  // (e.g. snippet has marks at far ends of a long passage), fall back to
-  // just the first marked term — that's still a real match in the doc.
-  if (plain.length > 200) {
-    const firstPlain = normalizeQuotes(
-      decodeSnippetEntities(first[1].replace(/<[^>]*>/g, ""))
-        .replace(/\s+/g, " ")
-        .trim(),
-    );
-    return firstPlain || null;
+  // Pick the best candidate:
+  //   - if the marked span alone is already ≥6 words, it's a multi-token phrase
+  //     and is specific enough — use it.
+  //   - otherwise prefer the full snippet (more context = more uniquely placed),
+  //     capped at 200 chars to keep URLs sane.
+  //   - fall back to marked span if the full snippet is empty / unusable.
+  const markedWordCount = markedPlain
+    ? markedPlain.split(/\s+/).filter(Boolean).length
+    : 0;
+
+  if (markedWordCount >= 6) {
+    return markedPlain || null;
   }
-
-  return plain;
+  if (fullPlain && fullPlain.length <= 200) {
+    return fullPlain;
+  }
+  return markedPlain || null;
 }
 
 // Decode the HTML entities Typesense commonly emits in match snippets.