rag-memory-epf-mcp 6.4.1 → 6.4.2

package/README.md

@@ -154,6 +154,19 @@ storeDocument(id, content, metadata)
 
 ## Changelog
 
+### v6.4.2 (2026-10-31)
+
+- **Fixed — the engine no longer keeps the peak memory of its largest embedding.** The ONNX session is opened with
+  the CPU memory arena off. With it on, one long input (12,000 characters, 5,774 tokens) left the process at about
+  8 GB until it exited (1.4 GB after model load), and every MCP client runs its own engine. With it off the same
+  input left 1.5-2.4 GB in a probe and 2.7-3.2 GB in the full engine (one Mac, one run each). Vectors are
+  byte-identical with the arena on and off. The transient peak while a long input is being embedded is not lowered.
+- **Changed — `@huggingface/transformers` 4.3.2** (was `^3.5.1`; onnxruntime-node 1.30.0). The model cache layout is
+  unchanged and an existing cache loads without a re-download. Vectors differ from 3.x only in the last bits (cosine
+  0.9999995 or higher on 5 inputs); the embedding profile is unchanged, so nothing is re-embedded. Measured on one
+  Mac: a 30-character query 128 ms -> 33 ms, a 484-token chunk 739 ms -> 639 ms (3.8.1 with the arena on vs 4.3.2
+  with it off). Not measured: fixed-query ranking of existing document vectors against new query vectors, Windows.
+
 ### v6.4.1 (2026-10-16)
 
 - **Changed — new opt-in manager method `generateRelationEvidenceChunks(items, { formatVersion: 'evidence-v1' })`.**

package/dist/index.d.ts

@@ -10,8 +10,12 @@ export declare function makeEmbedFn(model: (input: string, opts: any) => Promise
     data: any;
 }>, defaultModel: boolean): (text: string, dims: number, isQuery: boolean) => Promise<Float32Array<ArrayBuffer>>;
 export declare function compileFtsLiteralQuery(q: string): string | null;
+export declare function setPipelineImpl(fn: ((...args: any[]) => Promise<any>) | null): void;
 export declare function resolveDbFilePath(envValue: string | undefined, cwd: string, serverDir: string): string;
 export declare function parseMmapSize(raw: string | undefined): number;
+export declare const MODEL_SESSION_OPTIONS: Readonly<{
+    enableCpuMemArena: false;
+}>;
 interface Entity {
     name: string;
     entityType: string;

package/dist/index.js

@@ -95,6 +95,12 @@ export function compileFtsLiteralQuery(q) {
         return null;
     return terms.map(t => `"${t}"`).join(' OR ');
 }
+// Test-only seam for the real model loader: replaces `pipeline` so a test can see the exact
+// arguments buildRealLoader passes (same rule as setSyncFaultPoint: an argument, never an env var).
+let __pipelineImpl = pipeline;
+export function setPipelineImpl(fn) {
+    __pipelineImpl = (fn ?? pipeline);
+}
 // Configure Hugging Face transformers for better compatibility
 if (env.backends?.onnx?.wasm) {
     env.backends.onnx.wasm.wasmPaths = './node_modules/@huggingface/transformers/dist/';
@@ -142,6 +148,12 @@ const IS_DEFAULT_MODEL_CONFIG = !process.env.EMBEDDING_MODEL || process.env.EMBE
 // on 'main' (their vectors are never auto-grandfathered anyway).
 const MODEL_REVISION = IS_DEFAULT_MODEL_CONFIG ? '4de13258303883538bd53b696b452bf8099f0858' : 'main';
 const MODEL_DTYPE = 'fp16';
+// ONNX session options for the embedding model. The CPU memory arena is off: with it on, one long
+// input (12,000 chars = 5,774 tokens) left the process holding its peak (footprint 1.4 → 8.0 GB,
+// measured 2026-10-31 on transformers 3.8.1 and 4.3.2) and every MCP client runs its own engine.
+// Arena off returned to 1.5-2.4 GB with byte-identical vectors; the cost is per-call allocation
+// (query 29 → 33 ms, 484-token chunk 574 → 639 ms on 4.3.2). Not part of the cache artifact key.
+export const MODEL_SESSION_OPTIONS = Object.freeze({ enableCpuMemArena: false });
 // Entity embedding text builder version — mixed into entity input hashes so a
 // builder change re-backfills entities without touching chunk vectors (spec §6c).
 const TEXT_BUILDER_VERSION = 'tb1';
@@ -387,7 +399,8 @@ export class RAGKnowledgeGraphManager {
             env.allowRemoteModels = true;
             env.allowLocalModels = true;
             console.error(`🤖 Loading embedding model: ${EMBEDDING_MODEL} (1024-dim, cache=${cacheDir})...`);
-            const model = await pipeline('feature-extraction', EMBEDDING_MODEL, { revision: MODEL_REVISION, dtype: MODEL_DTYPE, cache_dir: cacheDir });
+            const model = await __pipelineImpl('feature-extraction', EMBEDDING_MODEL, { revision: MODEL_REVISION, dtype: MODEL_DTYPE, cache_dir: cacheDir,
+                session_options: { ...MODEL_SESSION_OPTIONS } });
             // dims fail-fast (spec §2 / beta B8): probe the ACTUAL output length — a
             // 384/768-dim custom model must fail here with a clear message, not at
             // every subsequent vector write.

package/docs/UPDATING.md

@@ -108,6 +108,13 @@ path and holder pid (e.g. `.download-<key>.lock`). Verify the holder process
 is genuinely gone or hung (`ps -p <pid>`), then remove the lock file manually;
 the next start becomes a clean download owner.
 
+## v6.4.2 (schema v14, unchanged): CPU memory arena off, transformers 4.3.2
+
+**What changes for an existing database.** Nothing is migrated and nothing is re-embedded (the embedding profile —
+model, revision, dtype, dimensions, pooling, normalize — is unchanged). The model cache is reused as is. What changes
+is the engine process: after a long embedding it returns to about 2-3 GB instead of holding about 8 GB. An engine that
+is already running keeps the old behaviour until it is restarted on 6.4.2. Details: README changelog.
+
 ## v6.4.1 (schema v14, unchanged): evidence-backed relationship chunks (opt-in), KG lifecycle fixes, faster oversize split
 
 **What changes for an existing database.** Nothing is migrated and search defaults are unchanged (the 113-query

package/package.json

@@ -1,6 +1,6 @@
 {
  "name": "rag-memory-epf-mcp",
- "version": "6.4.1",
+ "version": "6.4.2",
  "engines": {
   "node": ">=24"
  },
@@ -45,12 +45,12 @@
   "prepare": "npm run build",
   "watch": "tsc --watch",
   "verify:invariants": "node test/chunk-invariants.test.mjs",
-  "verify:engine": "node test/engine-smoke.test.mjs && node test/launch-smoke.test.mjs && node test/sync-atomicity.test.mjs && node test/dedup.test.mjs && node test/search-degradation.test.mjs && node test/entity-embed-cap.test.mjs && node test/migration12.test.mjs && node test/model-cache.test.mjs && node test/embedding-gate.test.mjs && node test/lazy-boot.test.mjs && node test/reconciliation.test.mjs && node test/backfill.test.mjs && node test/fts-query.test.mjs && node test/search-contracts.test.mjs && node test/tool-contracts.test.mjs && node test/bounded-exit.test.mjs && node test/observation-schema.test.mjs && node test/search-graph-default.test.mjs && node test/observation-migration.test.mjs && node test/observation-lifecycle.test.mjs && node test/observation-contracts.test.mjs && node test/observation-search.test.mjs && node test/observation-cascade.test.mjs && node test/observation-realdata.test.mjs && node --no-concurrent-sparkplug test/chunker-c.test.mjs && node test/migration14.test.mjs && node test/chunk-params-validation.test.mjs && node test/vector-reuse.test.mjs && node test/entity-range-linking.test.mjs && node test/alias-link-gate.test.mjs && node test/entity-name-boundary.test.mjs && node test/stats-chunking.test.mjs && node test/migration14-realdata.test.mjs && node test/migration14-realdata-sync.test.mjs && node test/search-summaries-off.test.mjs && node test/search-fusion-rrf.test.mjs && node test/query-input.test.mjs && node test/document-return-contracts.test.mjs && node test/observation-date-prefix.test.mjs && node test/delete-entities-cascade.test.mjs && node test/delete-entities-kg-hygiene.test.mjs && node test/backup-publish-portable.test.mjs && node test/graph-context-explain.test.mjs && node test/eval-graph-role-libs.test.mjs && node test/eval-graph-role-t5b.test.mjs && node --test test/eval-graph-role-t8-fix.test.mjs && node --test test/eval-graph-role-t7-upstream.test.mjs && node --test test/eval-graph-role-t11-decision.test.mjs && node --test test/eval-graph-role-prereq-fix.test.mjs && node test/db-path.test.mjs && node test/mmap-option.test.mjs && node test/wal-exit.test.mjs && node test/wal-sidecar-guard.test.mjs && node test/kg-evidence-core.test.mjs && node test/kg-delete-relation.test.mjs && node test/kg-confidence-copy.test.mjs && node test/kg-embed-cas.test.mjs && node test/kg-relation-evidence.test.mjs && node test/kg-detail-context.test.mjs && node test/kg-evidence-invalidation.test.mjs && node test/kg-evidence-cas.test.mjs && node test/kg-defaults.test.mjs && node test/chunker-oversize-cost.test.mjs && node test/chunker-structure-parity.test.mjs && node test/search-limit-depth.test.mjs && node test/kg-import-id-contract.test.mjs && node test/kg-evidence-guards.test.mjs",
+  "verify:engine": "node test/engine-smoke.test.mjs && node test/launch-smoke.test.mjs && node test/sync-atomicity.test.mjs && node test/dedup.test.mjs && node test/search-degradation.test.mjs && node test/entity-embed-cap.test.mjs && node test/migration12.test.mjs && node test/model-cache.test.mjs && node test/embedding-gate.test.mjs && node test/lazy-boot.test.mjs && node test/reconciliation.test.mjs && node test/backfill.test.mjs && node test/fts-query.test.mjs && node test/search-contracts.test.mjs && node test/tool-contracts.test.mjs && node test/bounded-exit.test.mjs && node test/observation-schema.test.mjs && node test/search-graph-default.test.mjs && node test/observation-migration.test.mjs && node test/observation-lifecycle.test.mjs && node test/observation-contracts.test.mjs && node test/observation-search.test.mjs && node test/observation-cascade.test.mjs && node test/observation-realdata.test.mjs && node --no-concurrent-sparkplug test/chunker-c.test.mjs && node test/migration14.test.mjs && node test/chunk-params-validation.test.mjs && node test/vector-reuse.test.mjs && node test/entity-range-linking.test.mjs && node test/alias-link-gate.test.mjs && node test/entity-name-boundary.test.mjs && node test/stats-chunking.test.mjs && node test/migration14-realdata.test.mjs && node test/migration14-realdata-sync.test.mjs && node test/search-summaries-off.test.mjs && node test/search-fusion-rrf.test.mjs && node test/query-input.test.mjs && node test/document-return-contracts.test.mjs && node test/observation-date-prefix.test.mjs && node test/delete-entities-cascade.test.mjs && node test/delete-entities-kg-hygiene.test.mjs && node test/backup-publish-portable.test.mjs && node test/graph-context-explain.test.mjs && node test/eval-graph-role-libs.test.mjs && node test/eval-graph-role-t5b.test.mjs && node --test test/eval-graph-role-t8-fix.test.mjs && node --test test/eval-graph-role-t7-upstream.test.mjs && node --test test/eval-graph-role-t11-decision.test.mjs && node --test test/eval-graph-role-prereq-fix.test.mjs && node test/db-path.test.mjs && node test/mmap-option.test.mjs && node test/wal-exit.test.mjs && node test/wal-sidecar-guard.test.mjs && node test/kg-evidence-core.test.mjs && node test/kg-delete-relation.test.mjs && node test/kg-confidence-copy.test.mjs && node test/kg-embed-cas.test.mjs && node test/kg-relation-evidence.test.mjs && node test/kg-detail-context.test.mjs && node test/kg-evidence-invalidation.test.mjs && node test/kg-evidence-cas.test.mjs && node test/kg-defaults.test.mjs && node test/chunker-oversize-cost.test.mjs && node test/chunker-structure-parity.test.mjs && node test/search-limit-depth.test.mjs && node test/kg-import-id-contract.test.mjs && node test/kg-evidence-guards.test.mjs && node test/model-session-options.test.mjs",
   "test": "npm run build && npm run verify:invariants && npm run verify:engine",
   "prepublishOnly": "npm run build && npm run verify:invariants && npm run verify:engine"
  },
  "dependencies": {
-  "@huggingface/transformers": "^3.5.1",
+  "@huggingface/transformers": "^4.3.2",
   "@modelcontextprotocol/sdk": "^1.27.1",
   "better-sqlite3": "^12.8.0",
   "graphology": "^0.26.0",