fix: harden AI analysis and seed defaults

2026-07-08 04:38:22 -07:00
parent 9a437f4cbf
commit 6a1add3422
15 changed files with 653 additions and 167 deletions
--- a/modules/ai/components/KnowledgeManagementClient.tsx
+++ b/modules/ai/components/KnowledgeManagementClient.tsx
@@ -66,12 +66,6 @@ function matchesEntry(entry: KnowledgeEntry, query: string): boolean {
  return [entry.title, entry.category, entry.tags, entry.body].join(" ").toLowerCase().includes(normalized);
 }

-function formatKnowledgeFailureReason(reason: string): string {
-  if (reason.includes("AI_API_KEY") || reason.includes("向量生成失败")) {
-    return `保存知识需要可用的 AI 向量配置：${reason}`;
-  }
-  return reason;
-}

 export function KnowledgeManagementClient({ initialEntries, permissions }: KnowledgeManagementClientProps): React.ReactElement {
  const [entries, setEntries] = useState(initialEntries);
@@ -142,7 +136,7 @@ export function KnowledgeManagementClient({ initialEntries, permissions }: Knowl
    const result = (await response.json()) as ApiResult<{ entry: KnowledgeEntry }>;
    setIsPending(false);
    if (!result.success) {
-      setMessage(formatKnowledgeFailureReason(result.reason));
+      setMessage(result.reason);
      return;
    }

@@ -180,7 +174,7 @@ export function KnowledgeManagementClient({ initialEntries, permissions }: Knowl
    });
    const result = (await response.json()) as ApiResult<{ entry: KnowledgeEntry }>;
    if (!result.success) {
-      setMessage(formatKnowledgeFailureReason(result.reason));
+      setMessage(result.reason);
      return;
    }
    await refreshEntries();
@@ -192,7 +186,7 @@ export function KnowledgeManagementClient({ initialEntries, permissions }: Knowl
      <section className="flex flex-col gap-4 border-b pb-5 lg:flex-row lg:items-end lg:justify-between">
        <div>
          <h1 className="text-3xl font-semibold tracking-normal">知识库</h1>
-          <p className="mt-2 text-sm text-muted-foreground">维护平台共享和机构私有知识，保存后会重建检索向量。</p>
+          <p className="mt-2 text-sm text-muted-foreground">维护平台共享和机构私有知识，保存后会重建本地关键词检索分块。</p>
        </div>
        <Button disabled={!canManage} onClick={beginCreate} type="button">
          <Plus className="size-4" aria-hidden="true" />
@@ -277,7 +271,7 @@ export function KnowledgeManagementClient({ initialEntries, permissions }: Knowl
      </Card>

      <Dialog
-        description="保存时会写入数据库、重建分块和 embedding。"
+        description="保存时会写入数据库并重建分块；检索使用本地关键词匹配。"
        onClose={closeDialog}
        open={isDialogOpen}
        title={isEditing ? "编辑知识" : "新增知识"}
--- a/modules/ai/server/analysis.test.ts
+++ b/modules/ai/server/analysis.test.ts
@@ -96,8 +96,6 @@ const runtimeConfig: AiRuntimeConfigResult = {
    baseUrl: "https://ai.example.test/v1",
    apiKey: "test-api-key",
    chatModel: "gpt-test",
-    embeddingModel: "embedding-test",
-    embeddingDimensions: 1536,
  },
 };

@@ -234,7 +232,7 @@ function createModelOutput(overrides: Partial<ElderAiAnalysisOutput> = {}): Elde
    modelSummary: {
      provider: "openai-compatible",
      chatModel: "gpt-test",
-      embeddingModel: "embedding-test",
+      knowledgeRetrieval: "keyword",
    },
    ...overrides,
  };
@@ -387,7 +385,7 @@ describe("elder AI analysis service", () => {
  it("degrades analysis when knowledge retrieval fails and records the retrieval audit failure", async () => {
    const database = createDatabaseFake();
    useDatabase(database);
-    vi.mocked(retrieveKnowledge).mockResolvedValue({ success: false, reason: "vector store unavailable", status: 500 });
+    vi.mocked(retrieveKnowledge).mockResolvedValue({ success: false, reason: "knowledge retrieval unavailable", status: 500 });
    chatMocks.invoke.mockResolvedValue(modelMessage(createModelOutput({ dataGaps: ["No gait score available."] })));

    const result = await generateElderAiAnalysis(createAuthContext(), "elder-1");
--- a/modules/ai/server/analysis.ts
+++ b/modules/ai/server/analysis.ts
@@ -269,7 +269,7 @@ export async function generateElderAiAnalysis(
    "recommendations[] 字段：title, priority(low|normal|high|urgent), rationale, suggestedNextStep, citationIds。",
    "citations[] 只返回你实际引用过的 citation 对象，citationIds 必须来自可用 citations。",
    knowledgeUnavailableReason ? `知识库检索不可用：${knowledgeUnavailableReason}。请在 dataGaps 中包含“知识库检索不可用，分析未使用内部知识库”。` : "",
-    `modelSummary 固定为 {"provider":"openai-compatible","chatModel":"${runtimeConfig.config.chatModel}","embeddingModel":"${runtimeConfig.config.embeddingModel}"}。`,
+    `modelSummary 固定为 {"provider":"openai-compatible","chatModel":"${runtimeConfig.config.chatModel}","knowledgeRetrieval":"keyword"}。`,
    `可用 citations：${JSON.stringify(citations)}`,
    `住民上下文：\n${residentContext.context.promptContext}`,
    `知识库片段：\n${knowledgeItems.map((item, index) => `[kb-${index + 1}] ${item.title}\n${item.content}`).join("\n\n") || "无"}`,
--- a/modules/ai/server/config.ts
+++ b/modules/ai/server/config.ts
@@ -1,11 +1,8 @@
-import { AI_KNOWLEDGE_EMBEDDING_DIMENSIONS } from "@/modules/core/server/schema";

 export type AiRuntimeConfig = {
  baseUrl: string;
  apiKey: string;
  chatModel: string;
-  embeddingModel: string;
-  embeddingDimensions: number;
 };

 export type AiRuntimeConfigResult =
@@ -13,7 +10,6 @@ export type AiRuntimeConfigResult =
  | { success: false; reason: string };

 const DEFAULT_CHAT_MODEL = "gpt-4.1-mini";
-const DEFAULT_EMBEDDING_MODEL = "text-embedding-3-small";

 function readOptionalEnv(name: string): string {
  return process.env[name]?.trim() ?? "";
@@ -31,8 +27,6 @@ export function getAiRuntimeConfig(): AiRuntimeConfigResult {
      baseUrl: readOptionalEnv("AI_BASE_URL"),
      apiKey,
      chatModel: readOptionalEnv("AI_CHAT_MODEL") || DEFAULT_CHAT_MODEL,
-      embeddingModel: readOptionalEnv("AI_EMBEDDING_MODEL") || DEFAULT_EMBEDDING_MODEL,
-      embeddingDimensions: AI_KNOWLEDGE_EMBEDDING_DIMENSIONS,
    },
  };
 }
--- a/modules/ai/server/knowledge.test.ts
+++ b/modules/ai/server/knowledge.test.ts
@@ -1,7 +1,6 @@
 import { beforeEach, describe, expect, it, vi } from "vitest";

-import type { AiRuntimeConfigResult } from "@/modules/ai/server/config";
-import { getAiRuntimeConfig } from "@/modules/ai/server/config";
+import { OpenAIEmbeddings } from "@langchain/openai";
 import {
  createKnowledgeEntry,
  listKnowledgeEntries,
@@ -13,9 +12,6 @@ import type { AppDatabase } from "@/modules/core/server/db";
 import { getDatabase } from "@/modules/core/server/db";
 import type { AuthContext, Permission } from "@/modules/core/types";

-const embeddingMocks = vi.hoisted(() => ({
-  embedDocuments: vi.fn(),
-}));

 const drizzleDsl = vi.hoisted(() => {
  type TableName = "entries" | "chunks";
@@ -32,7 +28,6 @@ const drizzleDsl = vi.hoisted(() => {
    entryId: { table: "chunks", key: "entryId" } as ColumnRef,
    scope: { table: "chunks", key: "scope" } as ColumnRef,
    organizationId: { table: "chunks", key: "organizationId" } as ColumnRef,
-    embedding: { table: "chunks", key: "embedding" } as ColumnRef,
    sourceTitle: { table: "chunks", key: "sourceTitle" } as ColumnRef,
    sourceCategory: { table: "chunks", key: "sourceCategory" } as ColumnRef,
    content: { table: "chunks", key: "content" } as ColumnRef,
@@ -44,15 +39,10 @@ vi.mock("server-only", () => ({}));

 vi.mock("@langchain/openai", () => ({
  OpenAIEmbeddings: vi.fn(function MockOpenAIEmbeddings() {
-    return {
-      embedDocuments: embeddingMocks.embedDocuments,
-    };
+    return {};
  }),
 }));

-vi.mock("@/modules/ai/server/config", () => ({
-  getAiRuntimeConfig: vi.fn(),
-}));

 vi.mock("@/modules/core/server/db", () => ({
  getDatabase: vi.fn(),
@@ -110,7 +100,6 @@ type KnowledgeChunkRow = {
  scope: KnowledgeScope;
  chunkIndex: number;
  content: string;
-  embedding: number[];
  sourceTitle: string;
  sourceCategory: string;
  createdAt: Date;
@@ -130,28 +119,20 @@ type RetrievedChunkRow = {
  title: string;
  category: string;
  content: string;
-  embedding: number[];
 };

+type TransactionWrite = { table: "entries" | "chunks"; values: unknown };
+
 type KnowledgeDatabaseFake = {
  entries: KnowledgeRow[];
  chunks: KnowledgeChunkRow[];
+  transactionWrites: TransactionWrite[];
  transaction: ReturnlessFunction;
  select: ReturnlessFunction;
 };

 type ReturnlessFunction = (...args: unknown[]) => unknown;

-const runtimeConfig: AiRuntimeConfigResult = {
-  success: true,
-  config: {
-    baseUrl: "",
-    apiKey: "test-api-key",
-    chatModel: "gpt-test",
-    embeddingModel: "embedding-test",
-    embeddingDimensions: 1536,
-  },
-};

 const organization: NonNullable<AuthContext["organization"]> = {
  id: "org-1",
@@ -262,7 +243,6 @@ function createChunk(entry: KnowledgeRow, overrides: Partial<KnowledgeChunkRow>
    scope: entry.scope,
    chunkIndex: 0,
    content: entry.body,
-    embedding: [0.1, 0.2, 0.3],
    sourceTitle: entry.title,
    sourceCategory: entry.category,
    createdAt: new Date("2026-07-02T00:00:00.000Z"),
@@ -325,7 +305,6 @@ class SelectBuilder implements PromiseLike<SelectRow[]> {
          title: chunk.sourceTitle,
          category: chunk.sourceCategory,
          content: chunk.content,
-          embedding: chunk.embedding,
        });
      }
    }
@@ -336,10 +315,93 @@ class SelectBuilder implements PromiseLike<SelectRow[]> {
 function createDatabaseFake(input: { entries?: KnowledgeRow[]; chunks?: KnowledgeChunkRow[] } = {}): KnowledgeDatabaseFake {
  const entries = input.entries ?? [];
  const chunks = input.chunks ?? [];
+  const transactionWrites: TransactionWrite[] = [];
+  const tableName = (table: unknown): "entries" | "chunks" => (table === drizzleDsl.chunks ? "chunks" : "entries");
+  const readRecord = (value: unknown): Record<string, unknown> => {
+    if (value && typeof value === "object" && !Array.isArray(value)) {
+      return value as Record<string, unknown>;
+    }
+    return {};
+  };
+  const readString = (value: unknown, fallback: string): string => (typeof value === "string" ? value : fallback);
+  const readOrganizationId = (value: unknown, fallback: string | null): string | null => {
+    if (typeof value === "string") {
+      return value;
+    }
+    if (value === null) {
+      return null;
+    }
+    return fallback;
+  };
+  const buildEntryRow = (values: unknown, fallback?: KnowledgeRow): KnowledgeRow => {
+    const record = readRecord(values);
+    return createEntry({
+      id: fallback?.id ?? `entry-created-${entries.length + 1}`,
+      scope: record.scope === "platform" ? "platform" : "organization",
+      organizationId: readOrganizationId(record.organizationId, fallback?.organizationId ?? "org-1"),
+      title: readString(record.title, fallback?.title ?? "Created guidance"),
+      category: readString(record.category, fallback?.category ?? "Safety"),
+      tags: readString(record.tags, fallback?.tags ?? "fall"),
+      body: readString(record.body, fallback?.body ?? "Created body"),
+      status: record.status === "disabled" ? "disabled" : "enabled",
+      createdByAccountId: readString(record.createdByAccountId, fallback?.createdByAccountId ?? "account-1"),
+      updatedByAccountId: readString(record.updatedByAccountId, fallback?.updatedByAccountId ?? "account-1"),
+      createdAt: fallback?.createdAt ?? new Date("2026-07-02T00:00:00.000Z"),
+      updatedAt: new Date("2026-07-02T00:00:00.000Z"),
+    });
+  };
+  type TransactionFake = {
+    insert: ReturnlessFunction;
+    update: ReturnlessFunction;
+    delete: ReturnlessFunction;
+  };
+  const transactionFake: TransactionFake = {
+    insert: vi.fn((table: unknown) => ({
+      values: vi.fn((values: unknown) => {
+        transactionWrites.push({ table: tableName(table), values });
+        if (table === drizzleDsl.entries) {
+          const row = buildEntryRow(values);
+          entries.push(row);
+          return { returning: vi.fn(async () => [row]) };
+        }
+        return Promise.resolve();
+      }),
+    })),
+    update: vi.fn((table: unknown) => ({
+      set: vi.fn((values: unknown) => ({
+        where: vi.fn(() => ({
+          returning: vi.fn(async () => {
+            const fallback = entries[0] ?? createEntry();
+            const row = buildEntryRow(values, fallback);
+            transactionWrites.push({ table: tableName(table), values });
+            const existingIndex = entries.findIndex((entry) => entry.id === row.id);
+            if (existingIndex >= 0) {
+              entries[existingIndex] = row;
+            } else {
+              entries.push(row);
+            }
+            return [row];
+          }),
+        })),
+      })),
+    })),
+    delete: vi.fn((table: unknown) => ({
+      where: vi.fn(() => {
+        transactionWrites.push({ table: tableName(table), values: { delete: true } });
+      }),
+    })),
+  };
+  const runTransaction: ReturnlessFunction = async (callback: unknown) => {
+    if (typeof callback !== "function") {
+      return null;
+    }
+    return (callback as (transaction: TransactionFake) => Promise<KnowledgeRow | null>)(transactionFake);
+  };
  return {
    entries,
    chunks,
-    transaction: vi.fn(),
+    transactionWrites,
+    transaction: vi.fn(runTransaction),
    select: vi.fn(() => new SelectBuilder(entries, chunks)),
  };
 }
@@ -350,8 +412,6 @@ function useDatabase(fake: KnowledgeDatabaseFake): void {

 beforeEach(() => {
  vi.clearAllMocks();
-  vi.mocked(getAiRuntimeConfig).mockReturnValue(runtimeConfig);
-  embeddingMocks.embedDocuments.mockResolvedValue([[0.4, 0.5, 0.6]]);
  useDatabase(createDatabaseFake());
 });

@@ -404,32 +464,38 @@ describe("AI knowledge service", () => {
        ],
      },
    });
-    expect(embeddingMocks.embedDocuments).toHaveBeenCalledWith(["night fall risk"]);
+    expect(OpenAIEmbeddings).not.toHaveBeenCalled();
  });

-  it("orders retrieved chunks by cosine similarity after scope filtering", async () => {
-    const weakEntry = createEntry({ id: "entry-weak", title: "Weak guidance" });
-    const strongEntry = createEntry({ id: "entry-strong", title: "Strong guidance" });
+  it("ranks matching chunks by lexical overlap and drops nonmatching chunks after scope filtering", async () => {
+    const weakEntry = createEntry({ id: "entry-weak", title: "Weak guidance", body: "General fall prevention guidance." });
+    const strongEntry = createEntry({ id: "entry-strong", title: "Strong guidance", body: "Night fall risk guidance for wandering residents." });
+    const unrelatedEntry = createEntry({ id: "entry-unrelated", title: "Nutrition guidance", body: "Hydration and meals." });
+    const disabledEntry = createEntry({ id: "entry-disabled", status: "disabled", title: "Disabled night fall risk" });
+    const otherOrgEntry = createEntry({ id: "entry-other-org", organizationId: "org-2", title: "Other org night fall risk" });
    useDatabase(createDatabaseFake({
-      entries: [weakEntry, strongEntry],
+      entries: [weakEntry, strongEntry, unrelatedEntry, disabledEntry, otherOrgEntry],
      chunks: [
-        createChunk(weakEntry, { embedding: [0, 1, 0] }),
-        createChunk(strongEntry, { embedding: [1, 0, 0] }),
+        createChunk(weakEntry),
+        createChunk(strongEntry),
+        createChunk(unrelatedEntry),
+        createChunk(disabledEntry),
+        createChunk(otherOrgEntry, { organizationId: "org-2", scope: "organization" }),
      ],
    }));
-    embeddingMocks.embedDocuments.mockResolvedValueOnce([[1, 0, 0]]);

-    const result = await retrieveKnowledge("org-1", "night fall risk", { limit: 2 });
+    const result = await retrieveKnowledge("org-1", "night fall risk", { limit: 5 });

    expect(result).toEqual({
      success: true,
      data: {
        results: [
          expect.objectContaining({ entryId: "entry-strong", score: 1 }),
-          expect.objectContaining({ entryId: "entry-weak", score: 0 }),
+          expect.objectContaining({ entryId: "entry-weak", score: 1 / 3 }),
        ],
      },
    });
+    expect(OpenAIEmbeddings).not.toHaveBeenCalled();
  });

  it("requires platform management permission before mutating platform knowledge", async () => {
@@ -443,7 +509,7 @@ describe("AI knowledge service", () => {
    );

    expect(result).toEqual({ success: false, reason: "只有平台管理员可以维护平台知识", status: 403 });
-    expect(embeddingMocks.embedDocuments).not.toHaveBeenCalled();
+    expect(OpenAIEmbeddings).not.toHaveBeenCalled();
  });

  it("requires an active organization before mutating organization knowledge", async () => {
@@ -453,7 +519,7 @@ describe("AI knowledge service", () => {
    );

    expect(result).toEqual({ success: false, reason: "请选择机构后维护知识库", status: 400 });
-    expect(embeddingMocks.embedDocuments).not.toHaveBeenCalled();
+    expect(OpenAIEmbeddings).not.toHaveBeenCalled();
  });

  it("rejects attempts to mutate another organization's knowledge", async () => {
@@ -463,17 +529,36 @@ describe("AI knowledge service", () => {
    });

    expect(result).toEqual({ success: false, reason: "不能维护其他机构的知识库", status: 403 });
-    expect(embeddingMocks.embedDocuments).not.toHaveBeenCalled();
+    expect(OpenAIEmbeddings).not.toHaveBeenCalled();
  });

-  it("surfaces embedding failures without opening a write transaction", async () => {
-    const database = createDatabaseFake();
+  it("writes lexical chunks with empty embedding arrays when creating and updating knowledge", async () => {
+    const existingEntry = createEntry({ id: "entry-existing", title: "Existing guidance" });
+    const database = createDatabaseFake({ entries: [existingEntry] });
    useDatabase(database);
-    embeddingMocks.embedDocuments.mockRejectedValue(new Error("embedding provider unavailable"));

-    const result = await createKnowledgeEntry(createAuthContext(), baseInput);
+    const created = await createKnowledgeEntry(createAuthContext(), baseInput);

-    expect(result).toEqual({ success: false, reason: "知识库向量生成失败", status: 500 });
-    expect(database.transaction).not.toHaveBeenCalled();
+    expect(created.success).toBe(true);
+    const createdChunkWrite = database.transactionWrites.find((write) => write.table === "chunks" && Array.isArray(write.values));
+    expect(createdChunkWrite?.values).toEqual([
+      expect.objectContaining({
+        content: expect.stringContaining("Fall prevention"),
+        embedding: [],
+      }),
+    ]);
+
+    database.transactionWrites.length = 0;
+    const updated = await updateKnowledgeEntry(createAuthContext(), "entry-existing", baseInput);
+
+    expect(updated.success).toBe(true);
+    const updatedChunkWrite = database.transactionWrites.find((write) => write.table === "chunks" && Array.isArray(write.values));
+    expect(updatedChunkWrite?.values).toEqual([
+      expect.objectContaining({
+        content: expect.stringContaining("Fall prevention"),
+        embedding: [],
+      }),
+    ]);
+    expect(OpenAIEmbeddings).not.toHaveBeenCalled();
  });
 });
--- a/modules/ai/server/knowledge.ts
+++ b/modules/ai/server/knowledge.ts
@@ -1,9 +1,7 @@
 import "server-only";

-import { OpenAIEmbeddings } from "@langchain/openai";
 import { and, desc, eq, isNull, or } from "drizzle-orm";

-import { getAiRuntimeConfig } from "@/modules/ai/server/config";
 import type {
  AiKnowledgeScope,
  KnowledgeEntry,
@@ -41,22 +39,21 @@ function toKnowledgeEntry(row: KnowledgeRow): KnowledgeEntry {
  };
 }

-function createEmbeddings() {
-  const configResult = getAiRuntimeConfig();
-  if (!configResult.success) {
-    return configResult;
+function normalizeSearchText(value: string): string[] {
+  const tokens = new Set<string>();
+  const normalizedTokens = value.toLowerCase().match(/[\p{L}\p{N}]+/gu) ?? [];
+  for (const token of normalizedTokens) {
+    if (token.length >= 2) {
+      tokens.add(token);
+    }
+    const cjkSequences = token.match(/[\u3400-\u9fff]+/gu) ?? [];
+    for (const sequence of cjkSequences) {
+      for (let index = 0; index < sequence.length - 1; index += 1) {
+        tokens.add(sequence.slice(index, index + 2));
+      }
+    }
  }
-  const configuration = configResult.config.baseUrl ? { baseURL: configResult.config.baseUrl } : undefined;
-  return {
-    success: true as const,
-    config: configResult.config,
-    embeddings: new OpenAIEmbeddings({
-      apiKey: configResult.config.apiKey,
-      model: configResult.config.embeddingModel,
-      dimensions: configResult.config.embeddingDimensions,
-      configuration,
-    }),
-  };
+  return Array.from(tokens);
 }

 function chunkKnowledgeText(input: KnowledgeEntryInput): string[] {
@@ -118,19 +115,8 @@ function getEnabledChunkWhere(organizationId: string) {
  );
 }

-async function buildKnowledgeChunks(input: KnowledgeEntryInput): Promise<ServiceResult<{ chunks: string[]; vectors: number[][] }>> {
-  const embeddingsResult = createEmbeddings();
-  if (!embeddingsResult.success) {
-    return { success: false, reason: embeddingsResult.reason, status: 500 };
-  }
-
-  const chunks = chunkKnowledgeText(input);
-  try {
-    const vectors = await embeddingsResult.embeddings.embedDocuments(chunks);
-    return { success: true, data: { chunks, vectors } };
-  } catch {
-    return { success: false, reason: "知识库向量生成失败", status: 500 };
-  }
+function buildKnowledgeChunks(input: KnowledgeEntryInput): string[] {
+  return chunkKnowledgeText(input);
 }

 export async function listKnowledgeEntries(context: AuthContext): Promise<KnowledgeEntry[]> {
@@ -152,11 +138,8 @@ export async function createKnowledgeEntry(
    return scope;
  }

+  const chunks = buildKnowledgeChunks(input);
  const database = getDatabase();
-  const build = await buildKnowledgeChunks(input);
-  if (!build.success) {
-    return build;
-  }

  const row = await database.transaction(async (transaction) => {
    const rows = await transaction
@@ -178,15 +161,15 @@ export async function createKnowledgeEntry(
    if (!entry) {
      return null;
    }
-    if (build.data.chunks.length > 0) {
+    if (chunks.length > 0) {
      await transaction.insert(aiKnowledgeChunks).values(
-        build.data.chunks.map((content, index) => ({
+        chunks.map((content, index) => ({
          entryId: entry.id,
          organizationId: entry.organizationId,
          scope: entry.scope,
          chunkIndex: index,
          content,
-          embedding: build.data.vectors[index] ?? [],
+          embedding: [],
          sourceTitle: entry.title,
          sourceCategory: entry.category,
        })),
@@ -208,10 +191,7 @@ async function updateEntryWithChunks(
  input: KnowledgeEntryInput,
  organizationId: string | null,
 ): Promise<KnowledgeRow | null> {
-  const build = await buildKnowledgeChunks(input);
-  if (!build.success) {
-    throw new Error(build.reason);
-  }
+  const chunks = buildKnowledgeChunks(input);

  const database = getDatabase();
  return database.transaction(async (transaction) => {
@@ -236,15 +216,15 @@ async function updateEntryWithChunks(
    }

    await transaction.delete(aiKnowledgeChunks).where(eq(aiKnowledgeChunks.entryId, row.id));
-    if (build.data.chunks.length > 0) {
+    if (chunks.length > 0) {
      await transaction.insert(aiKnowledgeChunks).values(
-        build.data.chunks.map((content, index) => ({
+        chunks.map((content, index) => ({
          entryId: row.id,
          organizationId: row.organizationId,
          scope: row.scope,
          chunkIndex: index,
          content,
-          embedding: build.data.vectors[index] ?? [],
+          embedding: [],
          sourceTitle: row.title,
          sourceCategory: row.category,
        })),
@@ -284,7 +264,7 @@ export async function updateKnowledgeEntry(
  try {
    row = await updateEntryWithChunks(id, context, input, scope.data.organizationId);
  } catch {
-    return { success: false, reason: "AI_API_KEY 未配置或知识库向量生成失败", status: 500 };
+    return { success: false, reason: "知识库更新失败", status: 500 };
  }

  if (!row) {
@@ -316,27 +296,22 @@ export async function deleteKnowledgeEntry(context: AuthContext, id: string): Pr
  return { success: true, data: { entry: toKnowledgeEntry(row) } };
 }

-function calculateCosineSimilarity(left: readonly number[], right: readonly number[]): number {
-  if (left.length === 0 || right.length === 0 || left.length !== right.length) {
+function scoreKnowledgeChunk(
+  chunk: Pick<KnowledgeRetrievalResult, "title" | "category" | "content">,
+  queryTokens: readonly string[],
+): number {
+  if (queryTokens.length === 0) {
    return 0;
  }

-  let dotProduct = 0;
-  let leftMagnitude = 0;
-  let rightMagnitude = 0;
-  for (let index = 0; index < left.length; index += 1) {
-    const leftValue = left[index] ?? 0;
-    const rightValue = right[index] ?? 0;
-    dotProduct += leftValue * rightValue;
-    leftMagnitude += leftValue * leftValue;
-    rightMagnitude += rightValue * rightValue;
+  const chunkTokens = new Set(normalizeSearchText(`${chunk.title}\n${chunk.category}\n${chunk.content}`));
+  let matchedTokens = 0;
+  for (const token of queryTokens) {
+    if (chunkTokens.has(token)) {
+      matchedTokens += 1;
+    }
  }
-
-  if (leftMagnitude === 0 || rightMagnitude === 0) {
-    return 0;
-  }
-
-  return dotProduct / (Math.sqrt(leftMagnitude) * Math.sqrt(rightMagnitude));
+  return matchedTokens / queryTokens.length;
 }

 export async function retrieveKnowledge(
@@ -344,16 +319,7 @@ export async function retrieveKnowledge(
  query: string,
  options?: { limit?: number },
 ): Promise<ServiceResult<{ results: KnowledgeRetrievalResult[] }>> {
-  const embeddingsResult = createEmbeddings();
-  if (!embeddingsResult.success) {
-    return { success: false, reason: embeddingsResult.reason, status: 500 };
-  }
-
-  const vectors = await embeddingsResult.embeddings.embedDocuments([query]).catch(() => []);
-  const embedding = vectors[0];
-  if (!embedding) {
-    return { success: false, reason: "知识库向量生成失败", status: 500 };
-  }
+  const queryTokens = normalizeSearchText(query);

  const database = getDatabase();
  const rows = await database
@@ -363,7 +329,6 @@ export async function retrieveKnowledge(
      title: aiKnowledgeChunks.sourceTitle,
      category: aiKnowledgeChunks.sourceCategory,
      content: aiKnowledgeChunks.content,
-      embedding: aiKnowledgeChunks.embedding,
    })
    .from(aiKnowledgeChunks)
    .innerJoin(aiKnowledgeEntries, eq(aiKnowledgeEntries.id, aiKnowledgeChunks.entryId))
@@ -379,8 +344,9 @@ export async function retrieveKnowledge(
          title: row.title,
          category: row.category,
          content: row.content,
-          score: calculateCosineSimilarity(row.embedding, embedding),
+          score: scoreKnowledgeChunk(row, queryTokens),
        }))
+        .filter((row) => row.score > 0)
        .sort((left, right) => right.score - left.score)
        .slice(0, options?.limit ?? DEFAULT_RETRIEVAL_LIMIT),
    },
--- a/modules/ai/types.test.ts
+++ b/modules/ai/types.test.ts
@@ -51,7 +51,7 @@ const validAnalysisOutput: ElderAiAnalysisOutput = {
  modelSummary: {
    provider: "openai-compatible",
    chatModel: "gpt-test",
-    embeddingModel: "embedding-test",
+    knowledgeRetrieval: "keyword",
  },
 };

@@ -142,6 +142,13 @@ describe("AI validators", () => {
          modelSummary: { ...validAnalysisOutput.modelSummary, provider: "anthropic" },
        },
      },
+      {
+        name: "unexpected knowledge retrieval mode",
+        value: {
+          ...validAnalysisOutput,
+          modelSummary: { ...validAnalysisOutput.modelSummary, knowledgeRetrieval: "embedding" },
+        },
+      },
      {
        name: "non-finite confidence",
        value: { ...validAnalysisOutput, confidence: "not-a-number" },
--- a/modules/ai/types.ts
+++ b/modules/ai/types.ts
@@ -67,7 +67,7 @@ export type ElderAiRecommendation = {
 export type ElderAiModelSummary = {
  provider: "openai-compatible";
  chatModel: string;
-  embeddingModel: string;
+  knowledgeRetrieval: "keyword";
 };

 export type ElderAiAnalysisOutput = {
@@ -241,14 +241,14 @@ function validateModelSummary(value: unknown): ElderAiModelSummary | null {
  if (
    value.provider !== "openai-compatible" ||
    typeof value.chatModel !== "string" ||
-    typeof value.embeddingModel !== "string"
+    value.knowledgeRetrieval !== "keyword"
  ) {
    return null;
  }
  return {
    provider: "openai-compatible",
    chatModel: value.chatModel,
-    embeddingModel: value.embeddingModel,
+    knowledgeRetrieval: "keyword",
  };
 }