mcp-memory-bucket 0.14.0 → 0.14.1

package/dist/src/store/embeddings.js

@@ -40,7 +40,12 @@ export function blobToVector(blob) {
     // Copy: a Buffer's byteOffset into its pool is not guaranteed to be 4-byte aligned.
     return new Float32Array(blob.buffer.slice(blob.byteOffset, blob.byteOffset + blob.byteLength));
 }
-const BATCH_SIZE = 16;
+/**
+ * One chunk per model call. A batch is padded to its longest member, and onnxruntime keeps its
+ * peak allocation for the life of the process: on a 1,700-chunk corpus, batches of 16 left the
+ * server holding 1.6 GB and ran slower (45 ms/chunk) than single inputs do (26 ms/chunk, ~490 MB).
+ */
+const BATCH_SIZE = 1;
 const RETRY_AFTER_FAILURE_MS = 5 * 60_000;
 /**
  * Embeds every chunk that has no vector for the current model, in the background, a batch at a

package/dist/src/store/recall.js

@@ -20,14 +20,16 @@ export function questionToFtsQuery(question, corrections) {
     return terms.size > 0 ? [...terms].map((t) => `"${t}"`).join(' OR ') : null;
 }
 const WORD_RE = /[\p{L}\p{N}_]+/gu;
-/** `query` with each misspelled word swapped for its correction — what gets embedded, since a typo shifts an embedding toward nothing in particular. */
-function respell(query, corrections) {
-    return corrections.size === 0 ? query : query.replace(WORD_RE, (word) => corrections.get(word.toLowerCase()) ?? word);
-}
 const CANDIDATES_PER_LIST = 30;
 const DOC_CANDIDATES = 10;
 /** Reciprocal-rank-fusion constant — the conventional 60; damps how much a rank-1 hit outweighs a rank-5 one. */
 const RRF_K = 60;
+/**
+ * The semantic list counts double: two of the three fused lists are keyword-based (chunk and doc
+ * BM25), so at equal weight a chunk that is the clear best match by meaning loses to one that
+ * merely shares common words with the question in both keyword lists.
+ */
+const SEMANTIC_WEIGHT = 2;
 const MAX_CHUNKS_PER_DOC = 3;
 /** Fused scores below this fraction of the best hit's are dropped rather than used to fill the token budget. */
 const RELATIVE_SCORE_FLOOR = 0.3;
@@ -72,8 +74,7 @@ const SEMANTIC_MAX_DOCS = 5;
  * of recall()'s vector list, for the web UI's result list (which shows docs, not passages).
  */
 export async function semanticDocMatches(db, embedder, query) {
-    const corrections = correctWords(db, query.toLowerCase().match(WORD_RE) ?? []);
-    const [queryVector] = await embedder.embed([respell(query, corrections)], 'query');
+    const [queryVector] = await embedder.embed([query], 'query');
     const hits = vectorSearch(db, embedder.model, queryVector, 50);
     const floor = Math.max(SEMANTIC_MIN_SIMILARITY, (hits[0]?.similarity ?? 0) - SEMANTIC_WINDOW);
     const chunkDoc = db.prepare(`SELECT ref_table, ref_id, ref_folder FROM chunks WHERE id = ?`);
@@ -98,7 +99,7 @@ const docKey = (c) => `${c.ref_table}\0${c.ref_folder}\0${c.ref_id}`;
 export async function recall(db, query, opts = {}) {
     const { maxTokens = 2000, table, hidden, embedder } = opts;
     const fused = new Map();
-    const add = (chunkId, rank) => fused.set(chunkId, (fused.get(chunkId) ?? 0) + 1 / (RRF_K + rank));
+    const add = (chunkId, rank, weight = 1) => fused.set(chunkId, (fused.get(chunkId) ?? 0) + weight / (RRF_K + rank));
     const corrections = correctWords(db, query.toLowerCase().match(WORD_RE) ?? []);
     const ftsQuery = questionToFtsQuery(query, corrections);
     const tableFilter = table ? `AND c.ref_table = '${table}'` : '';
@@ -118,16 +119,16 @@ export async function recall(db, query, opts = {}) {
     }
     let mode = 'lexical';
     if (embedder) {
-        const [queryVector] = await embedder.embed([respell(query, corrections)], 'query');
+        const [queryVector] = await embedder.embed([query], 'query');
         const semantic = vectorSearch(db, embedder.model, queryVector, CANDIDATES_PER_LIST * (table ? 4 : 1));
         if (semantic.length > 0)
             mode = 'hybrid';
-        semantic.forEach((hit, i) => add(hit.id, i + 1));
+        semantic.forEach((hit, i) => add(hit.id, i + 1, SEMANTIC_WEIGHT));
     }
     // Doc-level boost: every candidate chunk of a doc-level hit gains that doc's rank; a doc with no
     // candidate chunk yet (the query matched only its name/key/tags) contributes its opening chunk.
     const firstChunkOf = db.prepare(`SELECT id FROM chunks WHERE ref_table = ? AND ref_id = ? AND ref_folder = ? ORDER BY ord LIMIT 1`);
-    const chunkById = db.prepare(`SELECT id, ref_table, ref_id, ref_folder, ord, heading, text FROM chunks WHERE id = ?`);
+    const chunkById = db.prepare(`SELECT id, ref_table, ref_id, ref_folder, ord, heading, text, hash FROM chunks WHERE id = ?`);
     const rows = new Map();
     for (const id of fused.keys()) {
         const row = chunkById.get(id);
@@ -169,6 +170,8 @@ export async function recall(db, query, opts = {}) {
         .sort((a, b) => b.score - a.score);
     const floor = (ranked[0]?.score ?? 0) * RELATIVE_SCORE_FLOOR;
     const perDoc = new Map();
+    // The same skill synced into two folders yields byte-identical passages — return each once.
+    const seenText = new Set();
     const chunks = [];
     let budget = maxTokens;
     let truncated = false;
@@ -176,7 +179,7 @@ export async function recall(db, query, opts = {}) {
         if (score < floor)
             break;
         const key = docKey(row);
-        if ((perDoc.get(key) ?? 0) >= MAX_CHUNKS_PER_DOC)
+        if ((perDoc.get(key) ?? 0) >= MAX_CHUNKS_PER_DOC || seenText.has(row.hash))
             continue;
         let text = row.text;
         const cost = () => HEADER_TOKENS + Math.ceil((row.heading.length + text.length) / CHARS_PER_TOKEN);
@@ -189,6 +192,7 @@ export async function recall(db, query, opts = {}) {
         }
         budget -= cost();
         perDoc.set(key, (perDoc.get(key) ?? 0) + 1);
+        seenText.add(row.hash);
         chunks.push({ ref_table: row.ref_table, id: doc.id, folder: doc.folder, key: doc.key, heading: row.heading, ord: row.ord, text, score });
     }
     return { mode, chunks, truncated, corrections: Object.fromEntries(corrections) };

package/dist/src/store/spelling.js

@@ -14,7 +14,9 @@ function loadVocabulary(db) {
  * The corpus word a misspelled query word most likely meant, or undefined when the word is fine
  * (it, or a stemmed form of it, occurs somewhere) or nothing is close enough. Short words and
  * anything with a digit are left alone — an edit away from a 3-letter word or a ticket number is
- * usually a different word, not a typo. Ties go to the word found in more chunks.
+ * usually a different word, not a typo. A candidate must also start with the same letter: typos
+ * rarely fall on the first letter, while unrelated real words one edit apart often differ exactly
+ * there ("night" is not a misspelled "right"). Ties go to the word found in more chunks.
  */
 export function suggestCorrection(db, word) {
     const lower = word.toLowerCase();
@@ -32,7 +34,7 @@ export function suggestCorrection(db, word) {
     const maxDistance = lower.length <= 5 ? 1 : 2;
     let best;
     for (const { term, docs } of terms) {
-        if (Math.abs(term.length - lower.length) > maxDistance)
+        if (term[0] !== lower[0] || Math.abs(term.length - lower.length) > maxDistance)
             continue;
         const d = distance(lower, term);
         if (d > maxDistance)

package/dist/src/web/routes.js

@@ -7,7 +7,7 @@ import { saveFolder, saveRemoteFolder, removeFolder as removeFolderFromConfig, u
 import { stripKey } from '../memory/repository.js';
 import { pruneOrphanChunks } from '../store/chunks.js';
 import { initialScan, walkMarkdownFiles } from '../store/sync.js';
-import { SEARCH_BM25, withOrFallback, hasExplicitSyntax } from '../store/search.js';
+import { SEARCH_BM25, withOrFallback, hasExplicitSyntax, sanitizeFtsQuery } from '../store/search.js';
 import { semanticDocMatches } from '../store/recall.js';
 import { resolveWithinBase } from '../store/safe-path.js';
 import { attachmentsDirFor, guessMimeType, ATTACHMENT_MAX_BYTES, listSkillSourceFiles } from '../attachments/storage.js';
@@ -314,6 +314,7 @@ function matchDateRange(db, from, to) {
  * reciprocal rank fusion, so a doc matching in meaning but sharing no keyword still appears; the
  * stored rank is then the negated fused score, keeping "lower is better".
  */
+const LOOSENED_RESULT_LIMIT = 15;
 function matchSearch(db, q, semantic = []) {
     const skills = new Set();
     const memory_docs = new Set();
@@ -321,7 +322,16 @@ function matchSearch(db, q, semantic = []) {
     let rows;
     try {
         const statement = db.prepare(`SELECT ref_table, ref_id, ref_folder, ${SEARCH_BM25} AS score FROM search_index WHERE search_index MATCH ? ORDER BY score`);
-        rows = withOrFallback(db, q, 0, (ftsQuery) => statement.all(ftsQuery));
+        const strict = sanitizeFtsQuery(q);
+        let loosened = false;
+        rows = withOrFallback(db, q, 0, (ftsQuery) => {
+            loosened = ftsQuery !== strict;
+            return statement.all(ftsQuery);
+        });
+        // Any-word and typo-corrected matches are a best guess, not what was typed — a query of common
+        // words would otherwise list half the bucket. Keep only the strongest.
+        if (loosened)
+            rows = rows.slice(0, LOOSENED_RESULT_LIMIT);
     }
     catch {
         // Bad FTS5 query syntax (e.g. a bare quote) — treat as no matches rather than 500ing.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mcp-memory-bucket",
-  "version": "0.14.0",
+  "version": "0.14.1",
   "type": "module",
   "description": "MCP server exposing skill_* (reusable coding patterns) and memory_* (point-in-time working context) tools over a markdown+frontmatter source, cached into SQLite at runtime, with hybrid keyword + local-embedding passage retrieval.",
   "repository": {