@cerefox/memory 1.0.0-rc.2 → 1.0.0-rc.3

package/dist/bin/cerefox.js

@@ -7184,7 +7184,7 @@ var exports_meta = {};
 __export(exports_meta, {
   PKG_VERSION: () => PKG_VERSION
 });
-var PKG_VERSION = "1.0.0-rc.2";
+var PKG_VERSION = "1.0.0-rc.3";
 var init_meta = () => {};
 
 // ../../node_modules/.bun/tslib@2.8.1/node_modules/tslib/tslib.js
@@ -25532,9 +25532,22 @@ async function ensurePipeline() {
 async function warmup() {
   await ensurePipeline();
 }
+function onnxBatchSize() {
+  const env4 = globalThis.process?.env ?? {};
+  const n = Number.parseInt(env4.CEREFOX_ONNX_BATCH ?? "", 10);
+  return Number.isFinite(n) && n > 0 ? n : DEFAULT_ONNX_BATCH;
+}
 async function onnxEmbed(texts, role) {
   if (texts.length === 0)
     return [];
+  const batch = onnxBatchSize();
+  if (texts.length > batch) {
+    const out2 = [];
+    for (let i = 0;i < texts.length; i += batch) {
+      out2.push(...await onnxEmbed(texts.slice(i, i + batch), role));
+    }
+    return out2;
+  }
   const pipeline = await ensurePipeline();
   const inputs = buildPrefixedInputs(texts, role);
   const out = await pipeline(inputs, { pooling: "mean", normalize: true });
@@ -25549,7 +25562,7 @@ async function onnxEmbed(texts, role) {
   }
   return vectors;
 }
-var ONNX_MODEL_ID = "nomic-ai/nomic-embed-text-v1.5", ONNX_MODEL_NAME = "nomic-embed-text-v1.5", ONNX_MODEL_DTYPE = "q8", ONNX_MODEL_DIM = 768, ONNX_MODEL_APPROX_MB = 130, transformersModule = null, pipelinePromise = null;
+var ONNX_MODEL_ID = "nomic-ai/nomic-embed-text-v1.5", ONNX_MODEL_NAME = "nomic-embed-text-v1.5", ONNX_MODEL_DTYPE = "q8", ONNX_MODEL_DIM = 768, ONNX_MODEL_APPROX_MB = 130, transformersModule = null, pipelinePromise = null, DEFAULT_ONNX_BATCH = 4;
 var init_onnx_embedder = () => {};
 
 // ../../_shared/embeddings/index.ts
@@ -74646,7 +74659,7 @@ import { homedir as homedir6 } from "node:os";
 import { join as join9 } from "node:path";
 
 // ../../_shared/ef-meta/index.ts
-var EF_VERSION = "1.0.0-beta.4";
+var EF_VERSION = "1.0.0-rc.3";
 
 // src/cli/util/checks.ts
 init_config();

package/dist/server-assets/_shared/ef-meta/index.ts

@@ -18,7 +18,7 @@
  * doesn't touch `supabase/functions/` leaves it alone).
  */
 
-export const EF_VERSION = "1.0.0-beta.4";
+export const EF_VERSION = "1.0.0-rc.3";
 
 /**
  * The 8 peer EFs the cerefox-mcp aggregator probes (excludes cerefox-mcp

package/dist/server-assets/_shared/embeddings/onnx-embedder.ts

@@ -208,8 +208,32 @@ export async function warmup(): Promise<void> {
  * Mean pooling + L2 normalisation (sentence-transformers convention; nomic
  * expects both). Returns plain `number[][]` to match the OpenAI path.
  */
+/**
+ * Per-inference sub-batch. Peak tensor memory scales with the batch, and the
+ * container shares a (often small) Docker VM with Postgres + PostgREST + the
+ * web server — a 12-text single call was OOM-killed on Colima's default 2 GB
+ * VM (rc.2 dogfood, exit 137). 4 keeps peak memory flat at personal scale;
+ * override with CEREFOX_ONNX_BATCH.
+ */
+const DEFAULT_ONNX_BATCH = 4;
+
+function onnxBatchSize(): number {
+  const env = (globalThis as { process?: { env?: Record<string, string | undefined> } })
+    .process?.env ?? {};
+  const n = Number.parseInt(env.CEREFOX_ONNX_BATCH ?? "", 10);
+  return Number.isFinite(n) && n > 0 ? n : DEFAULT_ONNX_BATCH;
+}
+
 export async function onnxEmbed(texts: string[], role: EmbedRole): Promise<number[][]> {
   if (texts.length === 0) return [];
+  const batch = onnxBatchSize();
+  if (texts.length > batch) {
+    const out: number[][] = [];
+    for (let i = 0; i < texts.length; i += batch) {
+      out.push(...(await onnxEmbed(texts.slice(i, i + batch), role)));
+    }
+    return out;
+  }
   const pipeline = await ensurePipeline();
   const inputs = buildPrefixedInputs(texts, role);
   const out = await pipeline(inputs, { pooling: "mean", normalize: true });

package/docs/guides/setup-local.md

@@ -49,6 +49,10 @@ install; switching later requires a re-index (see below).
 The local model (~130 MB) downloads once — at install/init when selected — into the
 data volume, so it survives `cerefox-local upgrade`.
 
+> **Memory**: give the Docker VM **≥ 4 GB** for comfortable local-embedder use
+> (Colima defaults to 2 GB: `colima start --memory 4`). Inference is
+> sub-batched to keep peak memory flat, so smaller VMs work — just slower.
+
 > **Switching embedders on existing data is breaking**: the two models produce
 > incompatible vector spaces, so documents embedded with one are invisible to
 > semantic search under the other. `cerefox-local init` warns and requires

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@cerefox/memory",
-  "version": "1.0.0-rc.2",
+  "version": "1.0.0-rc.3",
   "description": "Cerefox — user-owned shared memory for AI agents. The local TypeScript runtime: stdio MCP server in v0.4; CLI binary added in v0.5; in-process web server in v0.6; ingestion pipeline in v0.7.",
   "license": "Apache-2.0",
   "homepage": "https://github.com/fstamatelopoulos/cerefox",