obsidian-hybrid-search 0.13.21 → 0.13.22

package/dist/src/chunker.js

@@ -184,6 +184,18 @@ export function slidingWindow(text, contextLength, overlap, headingChain = [], s
         ? chunks
         : [createSourceChunk(text, 0, text.length, headingChain, sectionOffset)];
 }
+/** Strip markdown syntax from a single line so it matches the textContent of a rendered DOM block. */
+function normalizeMatchLine(line) {
+    const stripped = stripMarkdownImages(stripHtmlTags(stripLeadingBlockquoteCallout(line)) // blockquote/callout markers and HTML tags
+        .replace(/\[\^[^\]]+\]/g, '') // footnote references ([^1])
+        .replace(/!\[\[[^\]]+\]\]/g, ''));
+    return replaceMarkdownInlineLinks(replaceWikilinkLabels(stripped))
+        .replace(/[*_]{1,2}([^*_]+)[*_]{1,2}/g, '$1') // bold / italic
+        .replace(/`([^`]+)`/g, '$1') // inline code
+        .replace(/^(?:[-*+]|\d+[.)]) \s*/m, '') // list markers
+        .replace(/^\[[xX ]\]\s*/m, '') // task checkboxes
+        .trim();
+}
 /**
  * Build a DOM-matchable string from chunk text.
  * Strips heading lines and markdown syntax so the result matches
@@ -210,20 +222,39 @@ export function buildMatchText(chunkText) {
     for (const line of [...bodyLines, fallback]) {
         if (!line.trim())
             continue;
-        const stripped = stripMarkdownImages(stripHtmlTags(stripLeadingBlockquoteCallout(line)) // blockquote/callout markers and HTML tags
-            .replace(/\[\^[^\]]+\]/g, '') // footnote references ([^1])
-            .replace(/!\[\[[^\]]+\]\]/g, ''));
-        const result = replaceMarkdownInlineLinks(replaceWikilinkLabels(stripped))
-            .replace(/[*_]{1,2}([^*_]+)[*_]{1,2}/g, '$1') // bold / italic
-            .replace(/`([^`]+)`/g, '$1') // inline code
-            .replace(/^(?:[-*+]|\d+[.)]) \s*/m, '') // list markers
-            .replace(/^\[[xX ]\]\s*/m, '') // task checkboxes
-            .trim();
+        const result = normalizeMatchLine(line);
         if (result)
             return result.slice(0, 80);
     }
     return '';
 }
+/** Remove sentinel markers from a snippet returned by SQLite's snippet() function. */
+export function stripSnippetMarkers(text, markStart, markEnd) {
+    return text.split(markStart).join('').split(markEnd).join('');
+}
+/**
+ * Build matchText from a BM25 snippet whose real FTS match is wrapped in sentinel
+ * markers (see searchBm25's snippet() call). A multi-line snippet window doesn't
+ * necessarily have the match on its first line, so — unlike buildMatchText, which
+ * assumes the chunk text itself starts at the relevant content — this picks the
+ * line that actually contains the marked match.
+ */
+export function buildMatchTextFromMarkedSnippet(markedSnippet, markStart, markEnd) {
+    const markIndex = markedSnippet.indexOf(markStart);
+    if (markIndex === -1) {
+        return buildMatchText(stripSnippetMarkers(markedSnippet, markStart, markEnd));
+    }
+    const lineStart = markedSnippet.lastIndexOf('\n', markIndex) + 1;
+    const lineEndRaw = markedSnippet.indexOf('\n', markIndex);
+    const lineEnd = lineEndRaw === -1 ? markedSnippet.length : lineEndRaw;
+    const line = stripSnippetMarkers(markedSnippet.slice(lineStart, lineEnd), markStart, markEnd);
+    const result = normalizeMatchLine(line);
+    if (result)
+        return result.slice(0, 80);
+    // Marked line stripped to nothing (e.g. a callout-type-only line) — fall back to
+    // the ordinary first-usable-line heuristic over the whole (unmarked) snippet.
+    return buildMatchText(stripSnippetMarkers(markedSnippet, markStart, markEnd));
+}
 function stripLeadingBlockquoteCallout(value) {
     let index = stripLeadingQuoteMarkers(value, 0);
     if (value.startsWith('[!', index)) {

package/dist/src/searcher.js

@@ -1,5 +1,5 @@
 import path from 'node:path';
-import { buildMatchText } from './chunker.js';
+import { buildMatchText, buildMatchTextFromMarkedSnippet, stripSnippetMarkers } from './chunker.js';
 import { config } from './config.js';
 import { filterNotePathsByFrontmatter, filterNotePathsByTag, getBacklinksForPaths, getChunkEmbeddingsByPath, getDb, getDbVersion, getLinksForPaths, getMarkdownLinksForPaths, getMatchingNotesByFrontmatter, getNoteByPath, getOutgoingLinks, getOutgoingLinksForPaths, getUrlsForPaths, hasVecTable, resolveNotePath, } from './db.js';
 import { embed } from './embedder.js';
@@ -114,12 +114,18 @@ function toFtsQuery(query, operator = 'AND') {
         return '""';
     return words.map((w) => `"${w}"*`).join(` ${operator} `);
 }
+// Sentinel markers bracketing the actual FTS match inside snippet() output. A snippet
+// window spans several sentences, so the true match isn't necessarily on its first
+// line — these let buildMatchTextFromMarkedSnippet find the right one. Stripped from
+// every snippet before it's exposed on RawResult; control chars won't appear in notes.
+const SNIPPET_MARK_START = '
';
+const SNIPPET_MARK_END = '';
 export function searchBm25(query, limit, snippetLength = 300, buildAnchors = false) {
     const db = getDb();
     const numTokens = Math.max(10, Math.ceil(snippetLength / 4));
     const stmt = db.prepare(`
       SELECT n.path, n.title, n.tags, n.aliases,
-             snippet(notes_fts_bm25, 2, '', '', '...', ?) AS snippet,
+             snippet(notes_fts_bm25, 2, ?, ?, '...', ?) AS snippet,
              bm25(notes_fts_bm25, 10.0, 5.0, 1.0) AS rank
       FROM notes_fts_bm25
       JOIN notes n ON n.id = notes_fts_bm25.rowid
@@ -133,13 +139,13 @@ export function searchBm25(query, limit, snippetLength = 300, buildAnchors = fal
         // relevant documents whenever a single query word (e.g. "between", "organize")
         // was absent from the document—even if that document was the best conceptual
         // match for every other term in the query.
-        const rows = stmt.all(numTokens, toFtsQuery(query, 'OR'), limit);
+        const rows = stmt.all(SNIPPET_MARK_START, SNIPPET_MARK_END, numTokens, toFtsQuery(query, 'OR'), limit);
         const results = rows.map((row) => ({
             path: row.path,
             title: row.title ?? '',
             tags: row.tags ?? '[]',
             aliases: row.aliases,
-            snippet: row.snippet ?? '',
+            snippet: stripSnippetMarkers(row.snippet ?? '', SNIPPET_MARK_START, SNIPPET_MARK_END),
             score: Math.max(0, Math.abs(row.rank) / (1 + Math.abs(row.rank))),
             scores: {
                 bm25: Math.max(0, Math.abs(row.rank) / (1 + Math.abs(row.rank))),
@@ -149,21 +155,21 @@ export function searchBm25(query, limit, snippetLength = 300, buildAnchors = fal
         // Skip when neither is needed (snippetLength=0 and buildAnchors=false).
         if (snippetLength > 0 || buildAnchors) {
             const chunkDataMap = getChunkDataForBm25Results(results);
-            for (const result of results) {
+            for (let i = 0; i < results.length; i++) {
+                const result = results[i];
                 const data = chunkDataMap.get(result.path);
                 if (data) {
                     if (snippetLength > 0 && data.headingPath) {
                         result.snippet = `${data.headingPath}\n${result.snippet}`;
                     }
                     if (buildAnchors) {
-                        // Use the BM25 snippet (positioned near the match) for matchText, not the full chunk.
-                        // Strip SQLite snippet ellipsis markers and HTML tags before extracting match text.
-                        // Remove all SQLite snippet ellipsis markers (may appear anywhere in the text)
-                        const snippetForMatch = result.snippet.replace(/\.\.\./g, '');
+                        // Use the still-marked raw snippet (positioned near the match) for matchText,
+                        // not the full chunk — the sentinel markers pin down which line is the real hit.
+                        const rawSnippet = (rows[i].snippet ?? '').replace(/\.\.\./g, '');
                         result.bm25Anchor = {
                             kind: 'bm25',
                             headingPath: data.headingPath,
-                            matchText: buildMatchText(snippetForMatch),
+                            matchText: buildMatchTextFromMarkedSnippet(rawSnippet, SNIPPET_MARK_START, SNIPPET_MARK_END),
                             charStart: data.charStart,
                             charEnd: data.charEnd,
                         };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "obsidian-hybrid-search",
-  "version": "0.13.21",
+  "version": "0.13.22",
   "mcpName": "io.github.flowing-abyss/obsidian-hybrid-search",
   "type": "module",
   "bin": {