@wirecat/cli-core 0.450.0 → 0.451.0

package/CHANGELOG.md

@@ -5,6 +5,29 @@ Notable changes to `@wirecat/cli-core` (`@wirecat/cli-messaging` up to 0.261.0,
 minor version may break callers, and says how under "Changed — may break callers". `pnpm docs:check` checks the
 shape of this file.
 
+## Unreleased
+
+## 0.451.0 — 27.10.2026
+
+### Added
+
+- **Picture models run on the native ONNX Runtime too**, where text models do: on one machine CLIP embedded 12
+  pictures in 0.5 s against 1.1 s through WebAssembly, and text queries in 0.15 s against 0.26 s, with vectors at
+  cosine 0.99 or more to WebAssembly's, so pictures embedded before still match. `openSession` in
+  `src/embeddings/onnx-runtime.ts` is now the one place that picks the runtime and falls back per model.
+- **Text models and rerankers run natively on Windows too** (x64 and arm64): `models download` fetches
+  `onnxruntime.dll` out of `onnxruntime-node`'s npm tarball (113.5 MB once, checked by sha256), because the Windows
+  binding needs its DirectML build, which Microsoft's release archive lacks. Needs `@wirecat/cli-core-onnx` 1.2.0.
+  A new native-runtime workflow downloads and runs the tiny test models natively on Linux x64 and arm64, macOS,
+  and Windows x64 and arm64, and checks that Alpine stays on WebAssembly.
+
+### Fixed
+
+- **Meaning search over documents and notes uses the vector index**, as message search already did: it scanned every
+  chunk of the store in pages (on one archive 600,000 chunks, mostly mail, for 23,000 document chunks), 7–18 s a
+  query; now about 0.05 s, with the same results, ties included. It still scans when the index is not ready or cannot
+  give an exact answer. `memo search all` gets most of the gain.
+
 ## 0.450.0 — 27.10.2026
 
 ### Added

package/README.md

@@ -131,8 +131,8 @@ code and its document must agree. A search plan that reads an index once per mes
 - **Presets for what people actually look for.** `preset:password`, `api-key`, `card`, `phone`, `email`, `url` and
   seven more find candidates, labelled as candidates.
 - **Search by meaning, opt in.** Conversations are split into chunks and embedded locally by ONNX Runtime — native
-  on Linux (glibc) and Apple-silicon Macs once `models download` has fetched it, WebAssembly elsewhere, on Node and
-  Bun alike — e5-small by default, or any OpenAI-compatible API with your own key, after
+  on Linux (glibc), Apple-silicon Macs and Windows once `models download` has fetched it, WebAssembly elsewhere, on
+  Node and Bun alike — e5-small by default, or any OpenAI-compatible API with your own key, after
   you see the cost. Results by meaning and by words are merged by reciprocal rank fusion.
 - **The archive and the server together.** `--backend both` also asks the messenger's own search; its hits are saved
   and filtered by the same query, so nothing is counted twice. `--thread` brings the reply thread around a hit.
@@ -226,9 +226,11 @@ they deliberately differ. A rule changes in its own pull request, before the cod
   over an in-memory fake or a real client.
 - **Checked across tools and runtimes.** CI runs on Linux: the test suite on Node 24, smoke checks on Bun and on
   Node 22.16, a parity workflow that builds the Telegram, MAX and WhatsApp CLIs against this repository, and
-  gitleaks over the full history. macOS and Windows run only three workflows: the SQLite library build (macOS) and
-  the ONNX runtime probe (macOS and Windows) run monthly and when started by hand; the scheduler test (macOS and
-  Windows) runs only when started by hand. The rest of the suite does not run on macOS or Windows.
+  gitleaks over the full history. macOS and Windows run only four workflows: the SQLite library build (macOS) and
+  the ONNX runtime probe (macOS and Windows) run monthly and when started by hand; the native-runtime check (Linux,
+  macOS and Windows, x64 and arm64, and Alpine) runs monthly and on every pull request that changes how models
+  load; the scheduler test (macOS and Windows) runs only when started by hand. The rest of the suite does not run on
+  macOS or Windows.
 - **One shared toolchain.** Biome, markdownlint, CSpell and Lefthook configs come from the
   [WireCat standards](https://github.com/WireCatLabs/community/blob/main/standards/README.md), and CI fails when a
   copy drifts.

package/dist/embeddings/embed.js

@@ -2,7 +2,7 @@ import { readFileSync } from "node:fs";
 import { availableParallelism } from "node:os";
 import { basename, join } from "node:path";
 import { filesPresent, sharedModelsDirectory } from "../speech/install.js";
-import { onnxRuntime } from "./onnx-runtime.js";
+import { openSession } from "./onnx-runtime.js";
 export const textModelsDirectory = (env = process.env) => sharedModelsDirectory("text", env);
 export const placedText = (model, directory) => model.files.map((file) => [file, join(directory, model.id, file.name)]);
 export const isTextModelInstalled = (model, directory) => filesPresent(placedText(model, directory));
@@ -123,29 +123,12 @@ export const meanOf = (data, tokens, dims) => {
 };
 /** A model's tokenizer and ONNX session, from its folder. */
 export const loadModel = async (model, directory, threads) => {
-    const [{ Tokenizer }, runtime] = await Promise.all([import("@huggingface/tokenizers"), onnxRuntime()]);
+    const { Tokenizer } = await import("@huggingface/tokenizers");
     const path = (name) => join(directory, model.id, name);
     const tokenizer = new Tokenizer(JSON.parse(readFileSync(path("tokenizer.json"), "utf8")), JSON.parse(readFileSync(path("tokenizer_config.json"), "utf8")));
-    const bytes = readFileSync(path(model.onnx));
-    const external = model.externalData
+    const { ort, session } = await openSession(readFileSync(path(model.onnx)), threads, model.externalData
         ? { externalData: [{ path: basename(model.externalData), data: readFileSync(path(model.externalData)) }] }
-        : {};
-    if (runtime.native) {
-        try {
-            const session = await runtime.ort.InferenceSession.create(bytes, {
-                ...external,
-                intraOpNumThreads: threads,
-                interOpNumThreads: 1,
-            });
-            return { ort: runtime.ort, tokenizer, session };
-        }
-        catch {
-            // A model the native runtime will not open still runs through WebAssembly.
-        }
-    }
-    const ort = await import("@wirecat/cli-core-onnx");
-    ort.env.wasm.numThreads = threads;
-    const session = await ort.InferenceSession.create(bytes, external);
+        : {});
     return { ort, tokenizer, session };
 };
 export const openEmbedder = async (model, directory, { threads = defaultThreads() } = {}) => {

package/dist/embeddings/onnx-runtime.d.ts

@@ -1,7 +1,7 @@
 import { type Fetch } from "../speech/install.js";
 import type { ModelFile } from "../speech/models.js";
 export declare const ONNX_RUNTIME_VERSION = "1.30.0";
-/** The library of Microsoft's release, in its archive, and the name the binding looks for beside itself. */
+/** The library in the archive it comes in, and the name the binding looks for beside itself. */
 export interface NativeRuntime {
     readonly target: string;
     readonly archive: ModelFile;
@@ -40,3 +40,16 @@ export interface OnnxRuntime {
  * process: a binding, once loaded, stays.
  */
 export declare const onnxRuntime: (env?: NodeJS.ProcessEnv) => Promise<OnnxRuntime>;
+/**
+ * A session on the native runtime when it is installed and opens this model, else on WebAssembly, with the `ort`
+ * whose tensors that session takes.
+ */
+export declare const openSession: (bytes: Uint8Array, threads: number, options?: {
+    externalData?: {
+        path: string;
+        data: Uint8Array;
+    }[];
+}) => Promise<{
+    ort: typeof import("@wirecat/cli-core-onnx");
+    session: import("@wirecat/cli-core-onnx").InferenceSession;
+}>;

package/dist/embeddings/onnx-runtime.js

@@ -12,9 +12,16 @@ const release = (name, sha256, bytes) => ({
     bytes,
 });
 /**
- * The systems the ONNX workflow runs natively; everywhere else, and on musl, the models run through WebAssembly.
- * Windows too: its binding needs the DirectML build of onnxruntime.dll, which Microsoft's release does not carry.
+ * onnxruntime-node's own npm tarball, for Windows: its binding imports the DirectML build of onnxruntime.dll by
+ * ordinal, which Microsoft's CPU release lacks; DirectML.dll is loaded only when used, so the DLL alone runs on CPU.
  */
+const npmTarball = {
+    name: `onnxruntime-node-${ONNX_RUNTIME_VERSION}.tgz`,
+    url: `https://registry.npmjs.org/onnxruntime-node/-/onnxruntime-node-${ONNX_RUNTIME_VERSION}.tgz`,
+    sha256: "6e3390d6b783e7be946fad629292799da28d0b42f84856e50d2c1b0383291e75",
+    bytes: 113_507_888,
+};
+/** The systems the native-runtime workflow runs natively; elsewhere, and on musl, the models run through WebAssembly. */
 const NATIVE = {
     "linux-x64": {
         archive: release("onnxruntime-linux-x64-1.30.0.tgz", "a5ed5a3cac51fbb2e90da632ae43d19212faaa20e76484e62bcb7c23ddb3b3fd", 11_306_877),
@@ -43,6 +50,24 @@ const NATIVE = {
             bytes: 43_879_424,
         },
     },
+    "win32-x64": {
+        archive: npmTarball,
+        member: "package/bin/napi-v6/win32/x64/onnxruntime.dll",
+        library: {
+            name: "onnxruntime.dll",
+            sha256: "508c362f5673483dd3a086379c392795b2e42d10d5e6f3f90ebd7ac21c97af67",
+            bytes: 28_754_232,
+        },
+    },
+    "win32-arm64": {
+        archive: npmTarball,
+        member: "package/bin/napi-v6/win32/arm64/onnxruntime.dll",
+        library: {
+            name: "onnxruntime.dll",
+            sha256: "6c3a2abf5c6aca11c48a0d338a91af309f2266f25ce2e4084e6a87bbcb0fbfd0",
+            bytes: 29_815_136,
+        },
+    },
 };
 const onMusl = () => {
     try {
@@ -159,3 +184,26 @@ export const onnxRuntime = (env = process.env) => {
     }
     return found;
 };
+/**
+ * A session on the native runtime when it is installed and opens this model, else on WebAssembly, with the `ort`
+ * whose tensors that session takes.
+ */
+export const openSession = async (bytes, threads, options = {}) => {
+    const runtime = await onnxRuntime();
+    if (runtime.native) {
+        try {
+            const session = await runtime.ort.InferenceSession.create(bytes, {
+                ...options,
+                intraOpNumThreads: threads,
+                interOpNumThreads: 1,
+            });
+            return { ort: runtime.ort, session };
+        }
+        catch {
+            // A model the native runtime will not open still runs through WebAssembly.
+        }
+    }
+    const ort = await import("@wirecat/cli-core-onnx");
+    ort.env.wasm.numThreads = threads;
+    return { ort, session: await ort.InferenceSession.create(bytes, options) };
+};

package/dist/images/picture/embedder.js

@@ -1,22 +1,22 @@
 import { readFileSync } from "node:fs";
 import { join } from "node:path";
 import { truncated, unit } from "../../embeddings/embed.js";
+import { openSession } from "../../embeddings/onnx-runtime.js";
 import { preprocess, SIZE } from "./preprocess.js";
 export const openPictureEmbedder = async (model, directory, { threads = 4, canvas } = {}) => {
-    const [{ Tokenizer }, ort] = await Promise.all([import("@huggingface/tokenizers"), import("@wirecat/cli-core-onnx")]);
+    const { Tokenizer } = await import("@huggingface/tokenizers");
     const path = (name) => join(directory, model.id, name);
     const tokenizer = new Tokenizer(JSON.parse(readFileSync(path("tokenizer.json"), "utf8")), JSON.parse(readFileSync(path("tokenizer_config.json"), "utf8")));
-    ort.env.wasm.numThreads = threads;
     const sessions = [];
     // A search needs the text side alone and `images embed` the vision side alone: each loads on first use.
     const lazy = (name) => {
-        let session;
+        let opened;
         return () => {
-            session ??= ort.InferenceSession.create(readFileSync(path(name))).then((created) => {
-                sessions.push(created);
+            opened ??= openSession(readFileSync(path(name)), threads).then((created) => {
+                sessions.push(created.session);
                 return created;
             });
-            return session;
+            return opened;
         };
     };
     const vision = lazy(model.vision);
@@ -25,13 +25,14 @@ export const openPictureEmbedder = async (model, directory, { threads = 4, canva
         model,
         embedImage: async (bytes) => {
             const pixels = await preprocess(bytes, canvas);
-            const output = await (await vision()).run({ pixel_values: new ort.Tensor("float32", pixels, [1, 3, SIZE, SIZE]) });
+            const { ort, session } = await vision();
+            const output = await session.run({ pixel_values: new ort.Tensor("float32", pixels, [1, 3, SIZE, SIZE]) });
             return unit(Float32Array.from(output.image_embeds?.data));
         },
         embedQuery: async (query) => {
             const ids = truncated(tokenizer.encode(query.toLowerCase()).ids, model.maxTokens);
             const shape = [1, ids.length];
-            const session = await text();
+            const { ort, session } = await text();
             const feeds = {
                 input_ids: new ort.Tensor("int64", BigInt64Array.from(ids, BigInt), shape),
             };

package/dist/store/sqlite/note-search.js

@@ -12,6 +12,7 @@ import { documentOf, noteOf } from "./notes.js";
 import { stemmerCache } from "./stems.js";
 import { canonicalTagNamed } from "./tags.js";
 import { read } from "./transaction.js";
+import { searchIndex } from "./vector-index.js";
 import { dot } from "./vectors.js";
 /** The fields a note has. A message-only field (`from:`, `chat:`, …) is refused rather than matching nothing. */
 export const NOTE_FIELDS = [
@@ -375,29 +376,72 @@ export const noteSearchOver = (context) => {
         nearest: async (model, query, { limit, folderIds, withInternal, source }) => {
             drain();
             const scope = live(folderIds, source, withInternal);
-            const page = database.prepare(`SELECT k.id, k.chunkable_type AS type, k.chunkable_id AS owner, k.start_offset AS start, k.end_offset AS end, v.vector
-           FROM chunks k JOIN embeddings v ON v.model = ? AND v.content_hash = k.content_hash
-          WHERE k.id > ?${scope.sql}
-          ORDER BY k.id LIMIT ?`);
-            const best = new Map();
-            let after = 0;
-            for (;;) {
-                const rows = page.all(model, after, ...scope.params, SCAN_PAGE);
-                for (const row of rows) {
-                    const score = dot(query, row.vector);
-                    const key = `${row.type}:${row.owner}`;
-                    if (score > (best.get(key)?.score ?? Number.NEGATIVE_INFINITY))
-                        best.set(key, { score, start: Number(row.start), end: Number(row.end) });
+            const columns = `k.id, k.chunkable_type AS type, k.chunkable_id AS owner, k.start_offset AS start,
+        k.end_offset AS end, k.content_hash AS hash, v.vector
+           FROM chunks k JOIN embeddings v ON v.model = ? AND v.content_hash = k.content_hash`;
+            const keep = (best, row) => {
+                const score = dot(query, row.vector);
+                const key = `${row.type}:${row.owner}`;
+                const held = best.get(key);
+                if (!held || score > held.score)
+                    best.set(key, {
+                        score,
+                        first: held?.first ?? Number(row.id),
+                        chunk: Number(row.id),
+                        start: Number(row.start),
+                        end: Number(row.end),
+                    });
+            };
+            // The scan meets each note first at its lowest chunk, and keeps that order among equal scores.
+            const ranked = (best) => [...best.entries()].sort(([, a], [, b]) => b.score - a.score || a.first - b.first).slice(0, limit);
+            const indexed = () => {
+                const types = source === undefined ? ["document", "note"] : source === "file" ? ["document"] : ["note"];
+                const accounts = database.prepare("SELECT DISTINCT coalesce(account_id, 0) AS account FROM chunks WHERE chunkable_type = ?");
+                const wanted = folderIds?.length ? new Set(folderIds.map(Number)) : undefined;
+                const candidates = database.prepare(`SELECT ${columns} WHERE k.id IN (SELECT value FROM json_each(?))${scope.sql} ORDER BY k.id`);
+                const first = database.prepare(`SELECT min(k.id) AS first FROM chunks k JOIN embeddings v ON v.model = ? AND v.content_hash = k.content_hash
+            WHERE k.chunkable_type = ? AND k.chunkable_id = ?${scope.sql}`);
+                const merged = new Map();
+                for (const type of types)
+                    for (const { account } of accounts.all(type)) {
+                        if (type === "document" && wanted && !wanted.has(Number(account)))
+                            continue;
+                        const found = searchIndex(database, model, { kind: type, account: Number(account) }, query, limit, (near) => {
+                            const hashes = new Map(near.map(({ id, hash }) => [id, hash]));
+                            const best = new Map();
+                            for (const row of candidates.all(model, JSON.stringify([...hashes.keys()]), ...scope.params))
+                                if (hashes.get(Number(row.id)) === row.hash)
+                                    keep(best, row);
+                            for (const [key, held] of best) {
+                                const [kind, owner] = key.split(":");
+                                held.first = Number(first.get(model, String(kind), Number(owner), ...scope.params)?.first);
+                            }
+                            const results = ranked(best);
+                            return { results, best: results.map(([, { chunk }]) => chunk) };
+                        });
+                        if (!found)
+                            return undefined;
+                        for (const [key, held] of found)
+                            merged.set(key, held);
+                    }
+                return ranked(merged);
+            };
+            const scanned = () => {
+                const page = database.prepare(`SELECT ${columns} WHERE k.id > ?${scope.sql} ORDER BY k.id LIMIT ?`);
+                const best = new Map();
+                let after = 0;
+                for (;;) {
+                    const rows = page.all(model, after, ...scope.params, SCAN_PAGE);
+                    for (const row of rows)
+                        keep(best, row);
+                    const last = rows.at(-1);
+                    if (!last || rows.length < SCAN_PAGE)
+                        break;
+                    after = Number(last.id);
                 }
-                const last = rows.at(-1);
-                if (!last || rows.length < SCAN_PAGE)
-                    break;
-                after = Number(last.id);
-            }
-            return [...best.entries()]
-                .sort(([, a], [, b]) => b.score - a.score)
-                .slice(0, limit)
-                .flatMap(([key, { score, start, end }]) => {
+                return ranked(best);
+            };
+            return (indexed() ?? scanned()).flatMap(([key, { score, start, end }]) => {
                 const [type, id] = key.split(":");
                 const note = read(String(type), Number(id));
                 return note ? [{ ref: note.ref, note, score, range: { start, end } }] : [];

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@wirecat/cli-core",
-  "version": "0.450.0",
+  "version": "0.451.0",
   "description": "The messenger-neutral half of a messaging CLI: domain model, message locators, rendering, name resolution, the SQLite seam, the command skeleton and shared read commands, the send guard, run records and the message store",
   "license": "Apache-2.0",
   "author": "Viacheslav Ptsarev",
@@ -176,7 +176,7 @@
     "@modelcontextprotocol/server": "^2.4.0",
     "@resvg/resvg-js": "2.6.2",
     "@valibot/to-json-schema": "1.8.0",
-    "@wirecat/cli-core-onnx": "1.1.0",
+    "@wirecat/cli-core-onnx": "1.2.0",
     "@wirecat/cli-core-sqlite": "1.1.0",
     "@xmldom/xmldom": "0.9.12",
     "chardet": "2.2.0",