@modusensus/dsh-mneme 0.8.14 → 0.8.16

package/test/local-embedder.test.js

@@ -1,292 +1,292 @@
-import test from "node:test";
-import assert from "node:assert/strict";
-import {
-  LocalEmbedder,
-  OllamaEmbedder,
-  OpenAIEmbedder,
-  createEmbedderByProvider
-} from "../src/local-embedder.js";
-
-/** Build a transformers.js-like Tensor: [rows x dim] flat data + dims. */
-function fakeTensor(rows) {
-  const flat = Float32Array.from(rows.flat());
-  return { data: flat, dims: rows.length ? [rows.length, rows[0].length] : [0] };
-}
-
-/** Fake extractor: deterministic [batch, dim], records option passthrough. */
-function makeFakeExtractor(dim, calls = []) {
-  const fn = async (texts, opts) => {
-    calls.push({ count: texts.length, opts });
-    return fakeTensor(
-      texts.map((_, i) => Array.from({ length: dim }, (__, j) => (i + 1) * 0.1 + j * 0.01))
-    );
-  };
-  fn.dispose = () => {
-    fn.disposed = true;
-  };
-  return fn;
-}
-
-/** Stub globalThis.fetch, returning the original on restore. */
-function stubFetch(handler) {
-  const real = globalThis.fetch;
-  globalThis.fetch = async (url, init) => handler(url, init);
-  return () => {
-    globalThis.fetch = real;
-  };
-}
-
-function okJson(status, body) {
-  return {
-    ok: status >= 200 && status < 300,
-    status,
-    json: async () => body
-  };
-}
-
-test("LocalEmbedder init loads via injected engineFactory with cache_dir", async () => {
-  let seen = null;
-  const loader = async (task, model, options) => {
-    seen = { task, model, options };
-    return makeFakeExtractor(512);
-  };
-  const e = new LocalEmbedder({ cacheDir: "/tmp/model-cache", engineFactory: loader });
-  await e.init();
-  assert.equal(seen.task, "feature-extraction");
-  assert.equal(seen.model, "Xenova/bge-small-zh-v1.5");
-  assert.equal(seen.options.cache_dir, "/tmp/model-cache");
-  assert.equal(seen.options.dtype, "q8");
-  assert.equal(seen.options.device, "cpu");
-});
-
-test("LocalEmbedder init threads resilientModelDownload and logger to engineFactory", async () => {
-  // issue #194：开关与 logger 要能走到 defaultPipelineLoader（那里据此挂 env.fetch），
-  // 但不能混进 pipelineOptions 污染 transformers 的选项。
-  let seen = null;
-  const loader = async (_task, _model, options) => {
-    seen = options;
-    return makeFakeExtractor(512);
-  };
-  const logger = { info() {}, warn() {} };
-  const on = new LocalEmbedder({ cacheDir: "/tmp/model-cache", engineFactory: loader, resilientModelDownload: true, logger });
-  await on.init();
-  assert.equal(seen.resilientModelDownload, true);
-  assert.equal(seen.logger, logger);
-
-  seen = null;
-  const off = new LocalEmbedder({ engineFactory: loader });
-  await off.init();
-  assert.equal(seen.resilientModelDownload, false, "直连构造不显式给 true = 保持现状");
-});
-
-test("LocalEmbedder embed returns [n, dim] with the model's pooling and chunking", async () => {
-  const calls = [];
-  const e = new LocalEmbedder({
-    model: "Xenova/bge-small-zh-v1.5",
-    batchSize: 2,
-    engineFactory: async () => makeFakeExtractor(512, calls)
-  });
-  await e.init();
-  const vecs = await e.embed(["你好", "hello world", "测试文本"]);
-  assert.equal(vecs.length, 3);
-  for (const v of vecs) {
-    assert.equal(v.length, 512);
-    assert.ok(v.every(Number.isFinite));
-  }
-  // 3 texts at batchSize 2 -> two extractor calls.
-  assert.deepEqual(calls.map((c) => c.count), [2, 1]);
-  // BGE 系按 CLS 训练（模型自带 1_Pooling/config.json: pooling_mode_cls_token=true）
-  assert.deepEqual(calls[0].opts, { pooling: "cls", normalize: true });
-});
-
-test("LocalEmbedder keeps mean pooling for models outside the CLS families", async () => {
-  const calls = [];
-  const e = new LocalEmbedder({
-    model: "sentence-transformers/all-MiniLM-L6-v2",
-    engineFactory: async () => makeFakeExtractor(384, calls)
-  });
-  await e.init();
-  await e.embed(["x"]);
-  assert.deepEqual(calls[0].opts, { pooling: "mean", normalize: true });
-});
-
-test("LocalEmbedder embedSingle returns a single vector", async () => {
-  const e = new LocalEmbedder({ engineFactory: async () => makeFakeExtractor(384) });
-  await e.init();
-  const v = await e.embedSingle("single");
-  assert.equal(v.length, 384);
-});
-
-test("LocalEmbedder resolves pooling by model family, explicit value wins", async () => {
-  const cases = [
-    [{ model: "Xenova/bge-small-zh-v1.5" }, "cls", "BGE 默认 cls"],
-    [{ model: "BAAI/bge-m3" }, "cls", "BGE 大小写无关"],
-    [{ model: "sentence-transformers/all-MiniLM-L6-v2" }, "mean", "非 BGE 族保持 mean"],
-    [{ model: "Xenova/bge-small-zh-v1.5", pooling: "mean" }, "mean", "显式覆盖 auto"],
-    [{ model: "sentence-transformers/all-MiniLM-L6-v2", pooling: "cls" }, "cls", "显式覆盖非 BGE"],
-    [{ model: "Xenova/bge-small-zh-v1.5", pooling: "AUTO" }, "cls", "大小写/空白容错"],
-    [{ model: "Xenova/bge-small-zh-v1.5", pooling: " bogus " }, "cls", "未知值退回 auto"]
-  ];
-  for (const [opts, expected, why] of cases) {
-    const calls = [];
-    const e = new LocalEmbedder({ ...opts, engineFactory: async () => makeFakeExtractor(512, calls) });
-    assert.equal(e.pooling, expected, why);
-    await e.init();
-    await e.embed(["x"]);
-    assert.equal(calls[0].opts.pooling, expected, `${why}: 必须真的传进 extractor`);
-  }
-});
-
-test("LocalEmbedder dimension and modelHash are stable", () => {
-  const a = new LocalEmbedder({ model: "Xenova/bge-small-zh-v1.5" });
-  const b = new LocalEmbedder({ model: "Xenova/bge-small-zh-v1.5" });
-  assert.equal(a.dimension, 512);
-  assert.equal(a.modelHash, b.modelHash);
-  // BGE 解析为 cls，且池化进指纹（否则旧索引不会被判失配、不会重建）
-  assert.match(a.modelHash, /^Xenova\/bge-small-zh-v1\.5@cls#[0-9a-f]+$/);
-  // Different model -> different hash.
-  assert.notEqual(a.modelHash, new LocalEmbedder({ model: "other/model" }).modelHash);
-  // 显式 mean 保持历史指纹形状，不给未受影响的索引制造无谓重建
-  assert.match(
-    new LocalEmbedder({ model: "Xenova/bge-small-zh-v1.5", pooling: "mean" }).modelHash,
-    /^Xenova\/bge-small-zh-v1\.5#[0-9a-f]+$/
-  );
-  // 两种池化是不同向量空间，绝不能共用一个指纹
-  assert.notEqual(
-    a.modelHash,
-    new LocalEmbedder({ model: "Xenova/bge-small-zh-v1.5", pooling: "mean" }).modelHash
-  );
-});
-
-test("LocalEmbedder throws before init and after dispose", async () => {
-  const e = new LocalEmbedder({ engineFactory: async () => makeFakeExtractor(512) });
-  await assert.rejects(() => e.embed(["x"]), /not initialized/);
-  await e.init();
-  const disposed = makeFakeExtractor(512);
-  e.dispose();
-  assert.ok(disposed.disposed || e.extractor === null);
-  await assert.rejects(() => e.embed(["x"]), /not initialized/);
-});
-
-test("OllamaEmbedder init probes server and infers dimension", async () => {
-  const restore = stubFetch(async (url, init) => {
-    assert.equal(url, "http://localhost:11434/api/embeddings");
-    assert.equal(init.method, "POST");
-    const body = JSON.parse(init.body);
-    assert.equal(body.model, "nomic-embed-text");
-    assert.equal(body.prompt, "ping");
-    return okJson(200, { embedding: Array.from({ length: 768 }, (_, i) => i / 768) });
-  });
-  try {
-    const e = new OllamaEmbedder({ baseUrl: "http://localhost:11434", model: "nomic-embed-text" });
-    assert.equal(e.ready, false, "not ready before init");
-    await e.init();
-    assert.equal(e.ready, true, "ready after init");
-    assert.equal(e.dimension, 768);
-  } finally {
-    restore();
-  }
-});
-
-test("OllamaEmbedder embed loops single requests", async () => {
-  const prompts = [];
-  const restore = stubFetch(async (url, init) => {
-    const body = JSON.parse(init.body);
-    prompts.push(body.prompt);
-    return okJson(200, { embedding: [0.5, 0.25, 0.125] });
-  });
-  try {
-    const e = new OllamaEmbedder({});
-    await e.init();
-    const vecs = await e.embed(["one", "two"]);
-    assert.deepEqual(prompts, ["ping", "one", "two"]);
-    assert.equal(vecs.length, 2);
-    assert.deepEqual(vecs[0], [0.5, 0.25, 0.125]);
-    assert.equal(e.dimension, 3);
-  } finally {
-    restore();
-  }
-});
-
-test("OllamaEmbedder init throws when unreachable", async () => {
-  const restore = stubFetch(async () => {
-    throw new Error("ECONNREFUSED");
-  });
-  try {
-    const e = new OllamaEmbedder({});
-    await assert.rejects(() => e.init(), /ECONNREFUSED/);
-  } finally {
-    restore();
-  }
-});
-
-test("OllamaEmbedder embed throws on HTTP error", async () => {
-  const restore = stubFetch(async () => okJson(404, { error: "model not found" }));
-  try {
-    const e = new OllamaEmbedder({});
-    await assert.rejects(() => e.embedSingle("x"), /HTTP 404/);
-  } finally {
-    restore();
-  }
-});
-
-test("OllamaEmbedder modelHash is stable", () => {
-  const a = new OllamaEmbedder({ model: "nomic-embed-text" });
-  const b = new OllamaEmbedder({ model: "nomic-embed-text" });
-  assert.equal(a.modelHash, b.modelHash);
-  assert.notEqual(a.modelHash, new OllamaEmbedder({ model: "bge-m3" }).modelHash);
-});
-
-test("OpenAIEmbedder init requires config and calls /embeddings", async () => {
-  const e = new OpenAIEmbedder({ baseUrl: "http://localhost:8000/v1", apiKey: "k", model: "m" });
-  assert.equal(e._url, "http://localhost:8000/v1/embeddings");
-  await e.init();
-
-  const bad = new OpenAIEmbedder({ baseUrl: "", apiKey: "", model: "" });
-  await assert.rejects(() => bad.init(), /requires baseUrl, apiKey and model/);
-});
-
-test("OpenAIEmbedder embed batch posts array and parses data", async () => {
-  let captured = null;
-  const restore = stubFetch(async (url, init) => {
-    captured = { url, init };
-    return okJson(200, {
-      data: [
-        { embedding: [0.1, 0.2, 0.3] },
-        { embedding: [0.4, 0.5, 0.6] }
-      ]
-    });
-  });
-  try {
-    const e = new OpenAIEmbedder({ baseUrl: "https://x/v1", apiKey: "sk-abc", model: "text-embed" });
-    await e.init();
-    const vecs = await e.embed(["a", "b"]);
-    assert.equal(captured.url, "https://x/v1/embeddings");
-    assert.equal(captured.init.headers.Authorization, "Bearer sk-abc");
-    const body = JSON.parse(captured.init.body);
-    assert.equal(body.model, "text-embed");
-    assert.deepEqual(body.input, ["a", "b"]);
-    assert.equal(vecs.length, 2);
-    assert.equal(e.dimension, 3);
-  } finally {
-    restore();
-  }
-});
-
-test("OpenAIEmbedder throws on non-ok response", async () => {
-  const restore = stubFetch(async () => okJson(500, {}));
-  try {
-    const e = new OpenAIEmbedder({ baseUrl: "https://x/v1", apiKey: "k", model: "m" });
-    await assert.rejects(() => e.embed(["x"]), /HTTP 500/);
-  } finally {
-    restore();
-  }
-});
-
-test("createEmbedderByProvider returns the right class per provider", () => {
-  assert.ok(createEmbedderByProvider("local") instanceof LocalEmbedder);
-  assert.ok(createEmbedderByProvider("ollama") instanceof OllamaEmbedder);
-  assert.ok(createEmbedderByProvider("openai") instanceof OpenAIEmbedder);
-  assert.ok(createEmbedderByProvider("LOCAL") instanceof LocalEmbedder);
-  assert.throws(() => createEmbedderByProvider("watson"), /Unknown embedding provider/);
-});
+import test from "node:test";
+import assert from "node:assert/strict";
+import {
+  LocalEmbedder,
+  OllamaEmbedder,
+  OpenAIEmbedder,
+  createEmbedderByProvider
+} from "../src/local-embedder.js";
+
+/** Build a transformers.js-like Tensor: [rows x dim] flat data + dims. */
+function fakeTensor(rows) {
+  const flat = Float32Array.from(rows.flat());
+  return { data: flat, dims: rows.length ? [rows.length, rows[0].length] : [0] };
+}
+
+/** Fake extractor: deterministic [batch, dim], records option passthrough. */
+function makeFakeExtractor(dim, calls = []) {
+  const fn = async (texts, opts) => {
+    calls.push({ count: texts.length, opts });
+    return fakeTensor(
+      texts.map((_, i) => Array.from({ length: dim }, (__, j) => (i + 1) * 0.1 + j * 0.01))
+    );
+  };
+  fn.dispose = () => {
+    fn.disposed = true;
+  };
+  return fn;
+}
+
+/** Stub globalThis.fetch, returning the original on restore. */
+function stubFetch(handler) {
+  const real = globalThis.fetch;
+  globalThis.fetch = async (url, init) => handler(url, init);
+  return () => {
+    globalThis.fetch = real;
+  };
+}
+
+function okJson(status, body) {
+  return {
+    ok: status >= 200 && status < 300,
+    status,
+    json: async () => body
+  };
+}
+
+test("LocalEmbedder init loads via injected engineFactory with cache_dir", async () => {
+  let seen = null;
+  const loader = async (task, model, options) => {
+    seen = { task, model, options };
+    return makeFakeExtractor(512);
+  };
+  const e = new LocalEmbedder({ cacheDir: "/tmp/model-cache", engineFactory: loader });
+  await e.init();
+  assert.equal(seen.task, "feature-extraction");
+  assert.equal(seen.model, "Xenova/bge-small-zh-v1.5");
+  assert.equal(seen.options.cache_dir, "/tmp/model-cache");
+  assert.equal(seen.options.dtype, "q8");
+  assert.equal(seen.options.device, "cpu");
+});
+
+test("LocalEmbedder init threads resilientModelDownload and logger to engineFactory", async () => {
+  // issue #194：开关与 logger 要能走到 defaultPipelineLoader（那里据此挂 env.fetch），
+  // 但不能混进 pipelineOptions 污染 transformers 的选项。
+  let seen = null;
+  const loader = async (_task, _model, options) => {
+    seen = options;
+    return makeFakeExtractor(512);
+  };
+  const logger = { info() {}, warn() {} };
+  const on = new LocalEmbedder({ cacheDir: "/tmp/model-cache", engineFactory: loader, resilientModelDownload: true, logger });
+  await on.init();
+  assert.equal(seen.resilientModelDownload, true);
+  assert.equal(seen.logger, logger);
+
+  seen = null;
+  const off = new LocalEmbedder({ engineFactory: loader });
+  await off.init();
+  assert.equal(seen.resilientModelDownload, false, "直连构造不显式给 true = 保持现状");
+});
+
+test("LocalEmbedder embed returns [n, dim] with the model's pooling and chunking", async () => {
+  const calls = [];
+  const e = new LocalEmbedder({
+    model: "Xenova/bge-small-zh-v1.5",
+    batchSize: 2,
+    engineFactory: async () => makeFakeExtractor(512, calls)
+  });
+  await e.init();
+  const vecs = await e.embed(["你好", "hello world", "测试文本"]);
+  assert.equal(vecs.length, 3);
+  for (const v of vecs) {
+    assert.equal(v.length, 512);
+    assert.ok(v.every(Number.isFinite));
+  }
+  // 3 texts at batchSize 2 -> two extractor calls.
+  assert.deepEqual(calls.map((c) => c.count), [2, 1]);
+  // BGE 系按 CLS 训练（模型自带 1_Pooling/config.json: pooling_mode_cls_token=true）
+  assert.deepEqual(calls[0].opts, { pooling: "cls", normalize: true });
+});
+
+test("LocalEmbedder keeps mean pooling for models outside the CLS families", async () => {
+  const calls = [];
+  const e = new LocalEmbedder({
+    model: "sentence-transformers/all-MiniLM-L6-v2",
+    engineFactory: async () => makeFakeExtractor(384, calls)
+  });
+  await e.init();
+  await e.embed(["x"]);
+  assert.deepEqual(calls[0].opts, { pooling: "mean", normalize: true });
+});
+
+test("LocalEmbedder embedSingle returns a single vector", async () => {
+  const e = new LocalEmbedder({ engineFactory: async () => makeFakeExtractor(384) });
+  await e.init();
+  const v = await e.embedSingle("single");
+  assert.equal(v.length, 384);
+});
+
+test("LocalEmbedder resolves pooling by model family, explicit value wins", async () => {
+  const cases = [
+    [{ model: "Xenova/bge-small-zh-v1.5" }, "cls", "BGE 默认 cls"],
+    [{ model: "BAAI/bge-m3" }, "cls", "BGE 大小写无关"],
+    [{ model: "sentence-transformers/all-MiniLM-L6-v2" }, "mean", "非 BGE 族保持 mean"],
+    [{ model: "Xenova/bge-small-zh-v1.5", pooling: "mean" }, "mean", "显式覆盖 auto"],
+    [{ model: "sentence-transformers/all-MiniLM-L6-v2", pooling: "cls" }, "cls", "显式覆盖非 BGE"],
+    [{ model: "Xenova/bge-small-zh-v1.5", pooling: "AUTO" }, "cls", "大小写/空白容错"],
+    [{ model: "Xenova/bge-small-zh-v1.5", pooling: " bogus " }, "cls", "未知值退回 auto"]
+  ];
+  for (const [opts, expected, why] of cases) {
+    const calls = [];
+    const e = new LocalEmbedder({ ...opts, engineFactory: async () => makeFakeExtractor(512, calls) });
+    assert.equal(e.pooling, expected, why);
+    await e.init();
+    await e.embed(["x"]);
+    assert.equal(calls[0].opts.pooling, expected, `${why}: 必须真的传进 extractor`);
+  }
+});
+
+test("LocalEmbedder dimension and modelHash are stable", () => {
+  const a = new LocalEmbedder({ model: "Xenova/bge-small-zh-v1.5" });
+  const b = new LocalEmbedder({ model: "Xenova/bge-small-zh-v1.5" });
+  assert.equal(a.dimension, 512);
+  assert.equal(a.modelHash, b.modelHash);
+  // BGE 解析为 cls，且池化进指纹（否则旧索引不会被判失配、不会重建）
+  assert.match(a.modelHash, /^Xenova\/bge-small-zh-v1\.5@cls#[0-9a-f]+$/);
+  // Different model -> different hash.
+  assert.notEqual(a.modelHash, new LocalEmbedder({ model: "other/model" }).modelHash);
+  // 显式 mean 保持历史指纹形状，不给未受影响的索引制造无谓重建
+  assert.match(
+    new LocalEmbedder({ model: "Xenova/bge-small-zh-v1.5", pooling: "mean" }).modelHash,
+    /^Xenova\/bge-small-zh-v1\.5#[0-9a-f]+$/
+  );
+  // 两种池化是不同向量空间，绝不能共用一个指纹
+  assert.notEqual(
+    a.modelHash,
+    new LocalEmbedder({ model: "Xenova/bge-small-zh-v1.5", pooling: "mean" }).modelHash
+  );
+});
+
+test("LocalEmbedder throws before init and after dispose", async () => {
+  const e = new LocalEmbedder({ engineFactory: async () => makeFakeExtractor(512) });
+  await assert.rejects(() => e.embed(["x"]), /not initialized/);
+  await e.init();
+  const disposed = makeFakeExtractor(512);
+  e.dispose();
+  assert.ok(disposed.disposed || e.extractor === null);
+  await assert.rejects(() => e.embed(["x"]), /not initialized/);
+});
+
+test("OllamaEmbedder init probes server and infers dimension", async () => {
+  const restore = stubFetch(async (url, init) => {
+    assert.equal(url, "http://localhost:11434/api/embeddings");
+    assert.equal(init.method, "POST");
+    const body = JSON.parse(init.body);
+    assert.equal(body.model, "nomic-embed-text");
+    assert.equal(body.prompt, "ping");
+    return okJson(200, { embedding: Array.from({ length: 768 }, (_, i) => i / 768) });
+  });
+  try {
+    const e = new OllamaEmbedder({ baseUrl: "http://localhost:11434", model: "nomic-embed-text" });
+    assert.equal(e.ready, false, "not ready before init");
+    await e.init();
+    assert.equal(e.ready, true, "ready after init");
+    assert.equal(e.dimension, 768);
+  } finally {
+    restore();
+  }
+});
+
+test("OllamaEmbedder embed loops single requests", async () => {
+  const prompts = [];
+  const restore = stubFetch(async (url, init) => {
+    const body = JSON.parse(init.body);
+    prompts.push(body.prompt);
+    return okJson(200, { embedding: [0.5, 0.25, 0.125] });
+  });
+  try {
+    const e = new OllamaEmbedder({});
+    await e.init();
+    const vecs = await e.embed(["one", "two"]);
+    assert.deepEqual(prompts, ["ping", "one", "two"]);
+    assert.equal(vecs.length, 2);
+    assert.deepEqual(vecs[0], [0.5, 0.25, 0.125]);
+    assert.equal(e.dimension, 3);
+  } finally {
+    restore();
+  }
+});
+
+test("OllamaEmbedder init throws when unreachable", async () => {
+  const restore = stubFetch(async () => {
+    throw new Error("ECONNREFUSED");
+  });
+  try {
+    const e = new OllamaEmbedder({});
+    await assert.rejects(() => e.init(), /ECONNREFUSED/);
+  } finally {
+    restore();
+  }
+});
+
+test("OllamaEmbedder embed throws on HTTP error", async () => {
+  const restore = stubFetch(async () => okJson(404, { error: "model not found" }));
+  try {
+    const e = new OllamaEmbedder({});
+    await assert.rejects(() => e.embedSingle("x"), /HTTP 404/);
+  } finally {
+    restore();
+  }
+});
+
+test("OllamaEmbedder modelHash is stable", () => {
+  const a = new OllamaEmbedder({ model: "nomic-embed-text" });
+  const b = new OllamaEmbedder({ model: "nomic-embed-text" });
+  assert.equal(a.modelHash, b.modelHash);
+  assert.notEqual(a.modelHash, new OllamaEmbedder({ model: "bge-m3" }).modelHash);
+});
+
+test("OpenAIEmbedder init requires config and calls /embeddings", async () => {
+  const e = new OpenAIEmbedder({ baseUrl: "http://localhost:8000/v1", apiKey: "k", model: "m" });
+  assert.equal(e._url, "http://localhost:8000/v1/embeddings");
+  await e.init();
+
+  const bad = new OpenAIEmbedder({ baseUrl: "", apiKey: "", model: "" });
+  await assert.rejects(() => bad.init(), /requires baseUrl, apiKey and model/);
+});
+
+test("OpenAIEmbedder embed batch posts array and parses data", async () => {
+  let captured = null;
+  const restore = stubFetch(async (url, init) => {
+    captured = { url, init };
+    return okJson(200, {
+      data: [
+        { embedding: [0.1, 0.2, 0.3] },
+        { embedding: [0.4, 0.5, 0.6] }
+      ]
+    });
+  });
+  try {
+    const e = new OpenAIEmbedder({ baseUrl: "https://x/v1", apiKey: "sk-abc", model: "text-embed" });
+    await e.init();
+    const vecs = await e.embed(["a", "b"]);
+    assert.equal(captured.url, "https://x/v1/embeddings");
+    assert.equal(captured.init.headers.Authorization, "Bearer sk-abc");
+    const body = JSON.parse(captured.init.body);
+    assert.equal(body.model, "text-embed");
+    assert.deepEqual(body.input, ["a", "b"]);
+    assert.equal(vecs.length, 2);
+    assert.equal(e.dimension, 3);
+  } finally {
+    restore();
+  }
+});
+
+test("OpenAIEmbedder throws on non-ok response", async () => {
+  const restore = stubFetch(async () => okJson(500, {}));
+  try {
+    const e = new OpenAIEmbedder({ baseUrl: "https://x/v1", apiKey: "k", model: "m" });
+    await assert.rejects(() => e.embed(["x"]), /HTTP 500/);
+  } finally {
+    restore();
+  }
+});
+
+test("createEmbedderByProvider returns the right class per provider", () => {
+  assert.ok(createEmbedderByProvider("local") instanceof LocalEmbedder);
+  assert.ok(createEmbedderByProvider("ollama") instanceof OllamaEmbedder);
+  assert.ok(createEmbedderByProvider("openai") instanceof OpenAIEmbedder);
+  assert.ok(createEmbedderByProvider("LOCAL") instanceof LocalEmbedder);
+  assert.throws(() => createEmbedderByProvider("watson"), /Unknown embedding provider/);
+});