chattercatcher 0.1.16 → 0.1.18

package/dist/cli.js

@@ -3,12 +3,12 @@
 // src/cli.ts
 import { input, password, select, confirm, number } from "@inquirer/prompts";
 import { Command } from "commander";
-import fs13 from "fs/promises";
+import fs14 from "fs/promises";
 
 // package.json
 var package_default = {
   name: "chattercatcher",
-  version: "0.1.16",
+  version: "0.1.18",
   description: "\u672C\u5730\u4F18\u5148\u7684\u98DE\u4E66/Lark \u5BB6\u5EAD\u7FA4\u77E5\u8BC6\u5E93\u673A\u5668\u4EBA",
   type: "module",
   main: "dist/index.js",
@@ -102,6 +102,13 @@ var appConfigSchema = z.object({
     model: z.string().default(""),
     dimension: z.number().int().positive().nullable().default(null)
   }),
+  multimodal: z.preprocess(
+    (value) => value ?? {},
+    z.object({
+      baseUrl: z.string().url().or(z.literal("")).default(""),
+      model: z.string().default("")
+    })
+  ),
   storage: z.object({
     dataDir: z.string().default(defaultDataDir)
   }),
@@ -126,13 +133,20 @@ var appSecretsSchema = z.object({
   }),
   embedding: z.object({
     apiKey: z.string().default("")
-  })
+  }),
+  multimodal: z.preprocess(
+    (value) => value ?? {},
+    z.object({
+      apiKey: z.string().default("")
+    })
+  )
 });
 function createDefaultConfig() {
   return appConfigSchema.parse({
     feishu: {},
     llm: {},
     embedding: {},
+    multimodal: {},
     storage: {},
     web: {},
     schedules: {},
@@ -143,7 +157,8 @@ function createDefaultSecrets() {
   return appSecretsSchema.parse({
     feishu: {},
     llm: {},
-    embedding: {}
+    embedding: {},
+    multimodal: {}
   });
 }
 
@@ -478,6 +493,24 @@ function migrateDatabase(database) {
       created_at TEXT NOT NULL,
       updated_at TEXT NOT NULL
     );
+
+    CREATE TABLE IF NOT EXISTS image_multimodal_tasks (
+      id TEXT PRIMARY KEY,
+      source_message_id TEXT NOT NULL REFERENCES messages(id) ON DELETE CASCADE,
+      platform_message_id TEXT NOT NULL,
+      image_key TEXT NOT NULL,
+      stored_path TEXT NOT NULL,
+      mime_type TEXT NOT NULL,
+      status TEXT NOT NULL CHECK(status IN ('pending','running','succeeded','skipped','failed')),
+      attempts INTEGER NOT NULL DEFAULT 0,
+      last_error TEXT,
+      derived_message_id TEXT REFERENCES messages(id) ON DELETE SET NULL,
+      created_at TEXT NOT NULL,
+      updated_at TEXT NOT NULL,
+      UNIQUE(source_message_id, image_key)
+    );
+
+    CREATE INDEX IF NOT EXISTS image_multimodal_tasks_status_idx ON image_multimodal_tasks(status, updated_at);
   `);
 }
 
@@ -894,6 +927,40 @@ function getGatewayStatus(config, secrets) {
 function normalizeBaseUrl(baseUrl) {
   return baseUrl.replace(/\/+$/, "");
 }
+function toOpenAIMessage(message) {
+  return {
+    role: message.role,
+    content: message.content,
+    ...message.toolCallId ? { tool_call_id: message.toolCallId } : {},
+    ...message.toolCalls ? {
+      tool_calls: message.toolCalls.map((toolCall) => ({
+        id: toolCall.id,
+        type: "function",
+        function: {
+          name: toolCall.name,
+          arguments: JSON.stringify(toolCall.input)
+        }
+      }))
+    } : {}
+  };
+}
+function toOpenAITool(tool) {
+  return {
+    type: "function",
+    function: {
+      name: tool.name,
+      description: tool.description,
+      parameters: tool.inputSchema
+    }
+  };
+}
+function parseToolCalls(message) {
+  return message?.tool_calls?.map((toolCall) => ({
+    id: toolCall.id,
+    name: toolCall.function.name,
+    input: JSON.parse(toolCall.function.arguments)
+  })) ?? [];
+}
 var OpenAICompatibleChatModel = class {
   constructor(options) {
     this.options = options;
@@ -911,7 +978,7 @@ var OpenAICompatibleChatModel = class {
       },
       body: JSON.stringify({
         model: this.options.model,
-        messages,
+        messages: messages.map(toOpenAIMessage),
         temperature: this.options.temperature ?? 0.2
       })
     });
@@ -926,6 +993,35 @@ var OpenAICompatibleChatModel = class {
     }
     return content;
   }
+  async completeWithTools(messages, tools) {
+    if (!this.options.baseUrl || !this.options.apiKey || !this.options.model) {
+      throw new Error("LLM \u914D\u7F6E\u4E0D\u5B8C\u6574\u3002\u8BF7\u8FD0\u884C chattercatcher setup \u6216 chattercatcher settings\u3002");
+    }
+    const response = await fetch(`${normalizeBaseUrl(this.options.baseUrl)}/chat/completions`, {
+      method: "POST",
+      headers: {
+        authorization: `Bearer ${this.options.apiKey}`,
+        "content-type": "application/json"
+      },
+      body: JSON.stringify({
+        model: this.options.model,
+        messages: messages.map(toOpenAIMessage),
+        tools: tools.map(toOpenAITool),
+        tool_choice: "auto",
+        temperature: this.options.temperature ?? 0.2
+      })
+    });
+    if (!response.ok) {
+      const body = await response.text();
+      throw new Error(`LLM \u8BF7\u6C42\u5931\u8D25\uFF1A${response.status} ${body}`);
+    }
+    const data2 = await response.json();
+    const message = data2.choices?.[0]?.message;
+    return {
+      content: message?.content ?? "",
+      toolCalls: parseToolCalls(message)
+    };
+  }
 };
 var OpenAICompatibleEmbeddingModel = class {
   constructor(options) {
@@ -1081,6 +1177,7 @@ var MessageRepository = class {
           )
           ON CONFLICT(platform, platform_message_id)
           DO UPDATE SET
+            message_type = excluded.message_type,
             text = excluded.text,
             raw_payload_json = excluded.raw_payload_json,
             received_at = excluded.received_at
@@ -1125,6 +1222,48 @@ var MessageRepository = class {
     transaction();
     return messageId;
   }
+  createImageSummaryMessage(input2) {
+    const source = this.database.prepare(
+      `
+        SELECT
+          m.platform AS platform,
+          m.platform_message_id AS platformMessageId,
+          m.chat_id AS chatId,
+          m.sender_id AS senderId,
+          m.sender_name AS senderName,
+          m.sent_at AS sentAt,
+          c.platform_chat_id AS platformChatId,
+          c.name AS chatName
+        FROM messages m
+        JOIN chats c ON c.id = m.chat_id
+        WHERE m.id = ?
+      `
+    ).get(input2.sourceMessageId);
+    if (!source) {
+      throw new Error("\u539F\u59CB\u56FE\u7247\u6D88\u606F\u4E0D\u5B58\u5728\u3002");
+    }
+    const derivedPlatformMessageId = `${source.platformMessageId}:image-summary:${input2.imageKey}`;
+    return this.ingest({
+      platform: source.platform,
+      platformChatId: source.platformChatId,
+      chatName: source.chatName,
+      platformMessageId: derivedPlatformMessageId,
+      senderId: source.senderId,
+      senderName: source.senderName,
+      messageType: "image_summary",
+      text: `[\u56FE\u7247\u8F6C\u8FF0] ${input2.summary.trim()}`,
+      sentAt: source.sentAt,
+      rawPayload: {
+        derivedFromMessageId: input2.sourceMessageId,
+        sourceAttachmentKind: "image",
+        sourceResourceKey: input2.imageKey,
+        multimodalModel: input2.multimodalModel,
+        isMeaningful: true,
+        ...input2.reason?.trim() ? { reason: input2.reason.trim() } : {},
+        generatedAt: input2.generatedAt
+      }
+    });
+  }
   listRecentMessages(limit = 20) {
     return this.database.prepare(
       `
@@ -1444,6 +1583,69 @@ var EpisodeRepository = class {
       messageIds: window.messages.map((message) => message.id)
     };
   }
+  async refreshWindowForMessage(input2) {
+    const target = this.database.prepare(
+      `
+          SELECT chat_id AS chatId, sent_at AS sentAt
+          FROM messages
+          WHERE id = ?
+        `
+    ).get(input2.messageId);
+    if (!target) {
+      return void 0;
+    }
+    const existingWindow = this.database.prepare(
+      `
+          SELECT e.started_at AS startedAt, e.ended_at AS endedAt
+          FROM messages target
+          JOIN messages source
+            ON source.id = json_extract(target.raw_payload_json, '$.derivedFromMessageId')
+          JOIN memory_episode_messages mem ON mem.message_id = source.id
+          JOIN memory_episodes e ON e.id = mem.episode_id
+          WHERE target.id = ?
+          LIMIT 1
+        `
+    ).get(input2.messageId);
+    if (!existingWindow) {
+      return void 0;
+    }
+    const messageTime = toMillis(target.sentAt);
+    const windowStart = toMillis(existingWindow.startedAt);
+    const windowEnd = Math.max(toMillis(existingWindow.endedAt), messageTime);
+    const rows = this.database.prepare(
+      `
+          SELECT
+            m.id,
+            m.chat_id AS chatId,
+            c.name AS chatName,
+            m.sender_name AS senderName,
+            m.text,
+            m.sent_at AS sentAt
+          FROM messages m
+          JOIN chats c ON c.id = m.chat_id
+          WHERE m.chat_id = ?
+          ORDER BY m.sent_at ASC
+        `
+    ).all(target.chatId);
+    const windowMessages = rows.filter((message) => {
+      const time = toMillis(message.sentAt);
+      return time >= windowStart && time <= windowEnd;
+    });
+    const first = windowMessages[0];
+    const last = windowMessages.at(-1);
+    if (!first || !last) {
+      return void 0;
+    }
+    const window = {
+      chatId: first.chatId,
+      chatName: first.chatName,
+      startedAt: first.sentAt,
+      endedAt: last.sentAt,
+      messages: windowMessages
+    };
+    const summary = await input2.summarize(window);
+    return this.insertEpisode(window, summary);
+  }
   getEpisodeCount() {
     const row = this.database.prepare("SELECT count(*) AS count FROM memory_episodes").get();
     return row.count;
@@ -1612,6 +1814,73 @@ var MessageFtsRetriever = class {
   }
 };
 
+// src/rag/search-tools.ts
+var searchInputSchema = {
+  type: "object",
+  properties: {
+    query: { type: "string", description: "Search query written by the model." },
+    limit: { type: "number", description: "Maximum number of evidence blocks to return." }
+  },
+  required: ["query"],
+  additionalProperties: false
+};
+function parseSearchInput(input2) {
+  const rawQuery = typeof input2 === "object" && input2 !== null && "query" in input2 ? input2.query : void 0;
+  if (typeof rawQuery !== "string") {
+    throw new Error("\u641C\u7D22 query \u5FC5\u987B\u662F\u975E\u7A7A\u5B57\u7B26\u4E32\u3002");
+  }
+  const query = rawQuery.trim();
+  if (!query) {
+    throw new Error("\u641C\u7D22 query \u5FC5\u987B\u662F\u975E\u7A7A\u5B57\u7B26\u4E32\u3002");
+  }
+  const rawLimit = typeof input2 === "object" && input2 !== null && "limit" in input2 ? input2.limit : void 0;
+  const numericLimit = typeof rawLimit === "number" && Number.isFinite(rawLimit) ? rawLimit : 5;
+  const limit = Math.min(12, Math.max(1, Math.floor(numericLimit)));
+  return { query, limit };
+}
+async function runRetriever(retriever, input2) {
+  const { query, limit } = parseSearchInput(input2);
+  const results = await retriever.retrieve(query);
+  return results.slice(0, limit);
+}
+function createSearchTool(name, description, retriever) {
+  return {
+    name,
+    description,
+    inputSchema: searchInputSchema,
+    execute: (input2) => runRetriever(retriever, input2)
+  };
+}
+function createRagSearchTools(input2) {
+  const tools = [
+    createSearchTool(
+      "hybrid_search",
+      "Search across all indexed RAG evidence using the default hybrid retrieval strategy.",
+      input2.hybrid
+    ),
+    createSearchTool(
+      "search_messages",
+      "Search chat messages only when the answer likely depends on message-level evidence.",
+      input2.messages
+    ),
+    createSearchTool(
+      "search_episodes",
+      "Search episode summaries only when the answer likely depends on longer-running context.",
+      input2.episodes
+    )
+  ];
+  if (input2.semantic) {
+    tools.push(
+      createSearchTool(
+        "semantic_search",
+        "Search semantic vector evidence only when broader conceptual recall is needed.",
+        input2.semantic
+      )
+    );
+  }
+  return tools;
+}
+
 // src/rag/embedding.ts
 function cosineSimilarity(left, right) {
   if (left.length === 0 || right.length === 0 || left.length !== right.length) {
@@ -1766,6 +2035,20 @@ async function createHybridRetriever(input2) {
     }
   };
 }
+async function createAgenticRagSearchTools(input2) {
+  const episodes = new EpisodeFtsRetriever(new EpisodeRepository(input2.database));
+  const messages = new MessageFtsRetriever(input2.messages, { excludeMessageIds: input2.excludeMessageIds });
+  const semantic = hasEmbeddingConfig(input2.config, input2.secrets) ? new VectorRetriever(
+    createEmbeddingModel(input2.config, input2.secrets),
+    new SqliteVectorStore(input2.database, { model: input2.config.embedding.model })
+  ) : void 0;
+  const hybrid = new HybridRetriever(semantic ? [episodes, messages, semantic] : [episodes, messages]);
+  return {
+    tools: createRagSearchTools({ hybrid, messages, episodes, semantic }),
+    close: () => {
+    }
+  };
+}
 
 // src/doctor/checks.ts
 function pass(name, message) {
@@ -2334,6 +2617,268 @@ async function ensureFeishuBotOpenId(config, secrets, options = {}) {
 // src/feishu/gateway.ts
 import * as lark2 from "@larksuiteoapi/node-sdk";
 
+// src/multimodal/tasks.ts
+import crypto4 from "crypto";
+function nowIso4() {
+  return (/* @__PURE__ */ new Date()).toISOString();
+}
+function stableId3(sourceMessageId, imageKey) {
+  return crypto4.createHash("sha256").update(`${sourceMessageId}${imageKey}`).digest("hex").slice(0, 32);
+}
+function mapRow(row) {
+  if (!row) {
+    return void 0;
+  }
+  return {
+    id: row.id,
+    sourceMessageId: row.source_message_id,
+    platformMessageId: row.platform_message_id,
+    imageKey: row.image_key,
+    storedPath: row.stored_path,
+    mimeType: row.mime_type,
+    status: row.status,
+    attempts: row.attempts,
+    ...row.last_error ? { lastError: row.last_error } : {},
+    ...row.derived_message_id ? { derivedMessageId: row.derived_message_id } : {},
+    createdAt: row.created_at,
+    updatedAt: row.updated_at
+  };
+}
+var ImageMultimodalTaskRepository = class {
+  constructor(database) {
+    this.database = database;
+  }
+  database;
+  enqueue(input2) {
+    const id = stableId3(input2.sourceMessageId, input2.imageKey);
+    const timestamp = nowIso4();
+    this.database.prepare(
+      `
+          INSERT INTO image_multimodal_tasks (
+            id,
+            source_message_id,
+            platform_message_id,
+            image_key,
+            stored_path,
+            mime_type,
+            status,
+            attempts,
+            created_at,
+            updated_at
+          )
+          VALUES (
+            @id,
+            @sourceMessageId,
+            @platformMessageId,
+            @imageKey,
+            @storedPath,
+            @mimeType,
+            'pending',
+            0,
+            @createdAt,
+            @updatedAt
+          )
+          ON CONFLICT(source_message_id, image_key)
+          DO UPDATE SET
+            platform_message_id = excluded.platform_message_id,
+            stored_path = excluded.stored_path,
+            mime_type = excluded.mime_type,
+            status = 'pending',
+            attempts = 0,
+            last_error = NULL,
+            derived_message_id = NULL,
+            updated_at = excluded.updated_at
+        `
+    ).run({
+      id,
+      sourceMessageId: input2.sourceMessageId,
+      platformMessageId: input2.platformMessageId,
+      imageKey: input2.imageKey,
+      storedPath: input2.storedPath,
+      mimeType: input2.mimeType,
+      createdAt: timestamp,
+      updatedAt: timestamp
+    });
+    const record = this.getById(id);
+    if (!record) {
+      throw new Error(`\u56FE\u7247\u591A\u6A21\u6001\u4EFB\u52A1\u5199\u5165\u5931\u8D25\uFF1A${id}`);
+    }
+    return record;
+  }
+  listPending(limit = 10) {
+    const rows = this.database.prepare(
+      `
+          SELECT
+            id,
+            source_message_id,
+            platform_message_id,
+            image_key,
+            stored_path,
+            mime_type,
+            status,
+            attempts,
+            last_error,
+            derived_message_id,
+            created_at,
+            updated_at
+          FROM image_multimodal_tasks
+          WHERE status = 'pending'
+          ORDER BY updated_at ASC
+          LIMIT ?
+        `
+    ).all(limit);
+    return rows.map((row) => mapRow(row)).filter((row) => Boolean(row));
+  }
+  markRunning(id) {
+    const result = this.database.prepare(
+      `
+          UPDATE image_multimodal_tasks
+          SET status = 'running',
+              attempts = attempts + 1,
+              last_error = NULL,
+              updated_at = @updatedAt
+          WHERE id = @id AND status = 'pending'
+        `
+    ).run({ id, updatedAt: nowIso4() });
+    if (result.changes === 0) {
+      throw new Error(`\u56FE\u7247\u591A\u6A21\u6001\u4EFB\u52A1\u72B6\u6001\u65E0\u6CD5\u66F4\u65B0\uFF1A${id}`);
+    }
+    return this.requireById(id);
+  }
+  markSucceeded(id, derivedMessageId) {
+    this.database.prepare(
+      `
+          UPDATE image_multimodal_tasks
+          SET status = 'succeeded',
+              last_error = NULL,
+              derived_message_id = @derivedMessageId,
+              updated_at = @updatedAt
+          WHERE id = @id
+        `
+    ).run({ id, derivedMessageId, updatedAt: nowIso4() });
+    return this.requireById(id);
+  }
+  markSkipped(id, reason) {
+    this.database.prepare(
+      `
+          UPDATE image_multimodal_tasks
+          SET status = 'skipped',
+              last_error = @reason,
+              derived_message_id = NULL,
+              updated_at = @updatedAt
+          WHERE id = @id
+        `
+    ).run({ id, reason, updatedAt: nowIso4() });
+    return this.requireById(id);
+  }
+  markFailed(id, error, finalFailure) {
+    this.database.prepare(
+      `
+          UPDATE image_multimodal_tasks
+          SET status = @status,
+              last_error = @error,
+              derived_message_id = NULL,
+              updated_at = @updatedAt
+          WHERE id = @id
+        `
+    ).run({ id, status: finalFailure ? "failed" : "pending", error, updatedAt: nowIso4() });
+    return this.requireById(id);
+  }
+  getById(id) {
+    const row = this.database.prepare(
+      `
+          SELECT
+            id,
+            source_message_id,
+            platform_message_id,
+            image_key,
+            stored_path,
+            mime_type,
+            status,
+            attempts,
+            last_error,
+            derived_message_id,
+            created_at,
+            updated_at
+          FROM image_multimodal_tasks
+          WHERE id = ?
+        `
+    ).get(id);
+    return mapRow(row);
+  }
+  requireById(id) {
+    const record = this.getById(id);
+    if (!record) {
+      throw new Error(`\u56FE\u7247\u591A\u6A21\u6001\u4EFB\u52A1\u4E0D\u5B58\u5728\uFF1A${id}`);
+    }
+    return record;
+  }
+};
+
+// src/multimodal/worker.ts
+var ImageMultimodalWorker = class {
+  constructor(options) {
+    this.options = options;
+  }
+  options;
+  async processPending(limit = 10) {
+    const result = { processed: 0, succeeded: 0, skipped: 0, failed: 0 };
+    const pending = this.options.tasks.listPending(limit);
+    for (const task of pending) {
+      result.processed += 1;
+      await this.processTask(task, result);
+    }
+    return result;
+  }
+  async processTask(task, result) {
+    let running;
+    try {
+      running = this.options.tasks.markRunning(task.id);
+    } catch (error) {
+      const message = error instanceof Error ? error.message : String(error);
+      if (message.startsWith("\u56FE\u7247\u591A\u6A21\u6001\u4EFB\u52A1\u72B6\u6001\u65E0\u6CD5\u66F4\u65B0\uFF1A")) {
+        return;
+      }
+      throw error;
+    }
+    try {
+      const described = await this.options.model.describeImage({
+        imagePath: running.storedPath,
+        mimeType: running.mimeType
+      });
+      if (!described.isMeaningful) {
+        this.options.tasks.markSkipped(running.id, described.reason || "\u591A\u6A21\u6001\u6A21\u578B\u5224\u5B9A\u56FE\u7247\u65E0\u610F\u4E49\u3002");
+        result.skipped += 1;
+        return;
+      }
+      const derivedMessageId = this.options.messages.createImageSummaryMessage({
+        sourceMessageId: running.sourceMessageId,
+        imageKey: running.imageKey,
+        summary: described.summary,
+        reason: described.reason,
+        multimodalModel: this.options.multimodalModelName,
+        generatedAt: (/* @__PURE__ */ new Date()).toISOString()
+      });
+      if (this.options.vectorIndexMessage) {
+        await this.options.vectorIndexMessage(derivedMessageId);
+      }
+      if (this.options.episodes && this.options.summarizeEpisode) {
+        await this.options.episodes.refreshWindowForMessage({
+          messageId: derivedMessageId,
+          windowMs: this.options.config.episodes.windowMinutes * 60 * 1e3,
+          summarize: this.options.summarizeEpisode
+        });
+      }
+      this.options.tasks.markSucceeded(running.id, derivedMessageId);
+      result.succeeded += 1;
+    } catch (error) {
+      const message = error instanceof Error ? error.message : String(error);
+      this.options.tasks.markFailed(running.id, message, running.attempts >= 3);
+      result.failed += 1;
+    }
+  }
+};
+
 // src/rag/citations.ts
 function isOpaqueId(value) {
   return Boolean(value && /^(ou|oc|om|cli|on|un|uid)_?[a-z0-9]+/i.test(value));
@@ -2454,12 +2999,99 @@ async function generateGroundedAnswer(input2) {
   };
 }
 
-// src/rag/qa-service.ts
-async function askWithRag(input2) {
-  const evidence = await input2.retriever.retrieve(input2.question);
+// src/rag/agentic-qa-service.ts
+var DEFAULT_MAX_MODEL_TURNS = 4;
+var DEFAULT_MAX_TOOL_CALLS = 8;
+var DEFAULT_MAX_EVIDENCE = 12;
+var NO_EVIDENCE_ANSWER = "\u4E0D\u77E5\u9053\u3002\u5F53\u524D\u672C\u5730\u77E5\u8BC6\u5E93\u6CA1\u6709\u68C0\u7D22\u5230\u8DB3\u591F\u8BC1\u636E\u3002";
+var AGENTIC_SYSTEM_PROMPT = "\u4F60\u662F\u672C\u5730\u77E5\u8BC6\u4FE1\u606F\u6536\u96C6\u4EE3\u7406\u3002\u4F60\u7684\u804C\u8D23\u662F\u56F4\u7ED5\u7528\u6237\u95EE\u9898\u51B3\u5B9A\u662F\u5426\u8C03\u7528\u641C\u7D22\u5DE5\u5177\u3001\u9009\u62E9\u5408\u9002\u7684\u5DE5\u5177\u548C\u67E5\u8BE2\u8BCD\uFF0C\u5E76\u6839\u636E\u5F53\u524D\u7ED3\u679C\u51B3\u5B9A\u662F\u5426\u7EE7\u7EED\u641C\u7D22\u3002\u4E0D\u8981\u7F16\u9020\u4EFB\u4F55\u8BC1\u636E\u6216\u58F0\u79F0\u770B\u8FC7\u672A\u68C0\u7D22\u5230\u7684\u5185\u5BB9\u3002\u4F60\u7684\u8F93\u51FA\u53EA\u7528\u4E8E\u6536\u96C6\u8BC1\u636E\uFF0C\u6700\u7EC8\u7B54\u6848\u4F1A\u7531\u53E6\u4E00\u4E2A\u57FA\u4E8E\u8BC1\u636E\u7684\u6B65\u9AA4\u751F\u6210\u3002";
+function toToolResultContent(results) {
+  return JSON.stringify(
+    results.map((item) => ({
+      id: item.id,
+      text: item.text,
+      score: item.score,
+      source: item.source
+    }))
+  );
+}
+function toToolErrorContent(message) {
+  return JSON.stringify({ error: message });
+}
+function dedupeEvidence(evidence, maxEvidence) {
+  const deduped = [];
+  const seen = /* @__PURE__ */ new Set();
+  for (const item of evidence) {
+    if (seen.has(item.id)) {
+      continue;
+    }
+    seen.add(item.id);
+    deduped.push(item);
+    if (deduped.length >= maxEvidence) {
+      break;
+    }
+  }
+  return deduped;
+}
+async function askWithAgenticRag(input2) {
+  if (!input2.model.completeWithTools) {
+    throw new Error("\u5F53\u524D LLM \u5BA2\u6237\u7AEF\u4E0D\u652F\u6301\u5DE5\u5177\u8C03\u7528\u3002");
+  }
+  const maxModelTurns = input2.maxModelTurns ?? DEFAULT_MAX_MODEL_TURNS;
+  const maxToolCalls = input2.maxToolCalls ?? DEFAULT_MAX_TOOL_CALLS;
+  const maxEvidence = input2.maxEvidence ?? DEFAULT_MAX_EVIDENCE;
+  const messages = [
+    { role: "system", content: AGENTIC_SYSTEM_PROMPT },
+    { role: "user", content: input2.question }
+  ];
+  const toolsByName = new Map(input2.tools.map((tool) => [tool.name, tool]));
+  let evidence = [];
+  let toolCallsUsed = 0;
+  for (let turn = 0; turn < maxModelTurns; turn += 1) {
+    const assistantResult = await input2.model.completeWithTools(messages, input2.tools);
+    messages.push({
+      role: "assistant",
+      content: assistantResult.content,
+      toolCalls: assistantResult.toolCalls
+    });
+    if (assistantResult.toolCalls.length === 0) {
+      break;
+    }
+    for (const toolCall of assistantResult.toolCalls) {
+      if (toolCallsUsed >= maxToolCalls) {
+        break;
+      }
+      toolCallsUsed += 1;
+      const tool = toolsByName.get(toolCall.name);
+      if (!tool) {
+        messages.push({
+          role: "tool",
+          toolCallId: toolCall.id,
+          content: toToolErrorContent(`\u672A\u77E5\u5DE5\u5177\uFF1A${toolCall.name}`)
+        });
+        continue;
+      }
+      try {
+        const results = await tool.execute(toolCall.input);
+        evidence = dedupeEvidence([...evidence, ...results], maxEvidence);
+        messages.push({
+          role: "tool",
+          toolCallId: toolCall.id,
+          content: toToolResultContent(results)
+        });
+      } catch (error) {
+        const message = error instanceof Error ? error.message : String(error);
+        messages.push({
+          role: "tool",
+          toolCallId: toolCall.id,
+          content: toToolErrorContent(message)
+        });
+      }
+    }
+  }
   if (evidence.length === 0) {
     return {
-      answer: "\u4E0D\u77E5\u9053\u3002\u5F53\u524D\u672C\u5730\u77E5\u8BC6\u5E93\u6CA1\u6709\u68C0\u7D22\u5230\u8DB3\u591F\u8BC1\u636E\u3002",
+      answer: NO_EVIDENCE_ANSWER,
       citations: []
     };
   }
@@ -2577,7 +3209,7 @@ var FeishuQuestionHandler = class {
     }
     const questionMessageId = payload.event?.message?.message_id;
     await this.acknowledgeQuestion(decision.chatId, questionMessageId);
-    const { retriever, close } = await createHybridRetriever({
+    const { tools, close } = await createAgenticRagSearchTools({
       config: this.options.config,
       secrets: this.options.secrets,
       database: this.options.database,
@@ -2586,9 +3218,9 @@ var FeishuQuestionHandler = class {
     });
     try {
       try {
-        const result = await askWithRag({
+        const result = await askWithAgenticRag({
           question: decision.question,
-          retriever,
+          tools,
           model: this.options.model
         });
         const citations = formatCitations(result.citations);
@@ -2724,6 +3356,7 @@ function createFeishuEventDispatcher(options) {
         payload,
         downloader: options.resourceDownloader,
         config: options.config,
+        secrets: options.secrets,
         vectorIndexMessage: options.attachmentVectorIndexer
       }) : options.ingestor.ingestFeishuEvent(payload);
       if (!result.accepted) {
@@ -2749,6 +3382,23 @@ function createFeishuEventDispatcher(options) {
       }
       if (result.attachment?.downloaded) {
         console.log(`\u98DE\u4E66\u9644\u4EF6\u5DF2\u4E0B\u8F7D\uFF1A${result.attachment.downloaded.storedPath}`);
+        if (options.imageMultimodalProcessor && result.attachment.imageTask) {
+          void new ImageMultimodalWorker({
+            config: options.config,
+            messages: new MessageRepository(options.imageMultimodalProcessor.database),
+            tasks: new ImageMultimodalTaskRepository(options.imageMultimodalProcessor.database),
+            model: options.imageMultimodalProcessor.model,
+            multimodalModelName: options.config.multimodal.model,
+            vectorIndexMessage: options.attachmentVectorIndexer
+          }).processPending().then((imageResult) => {
+            console.log(
+              `\u98DE\u4E66\u56FE\u7247\u591A\u6A21\u6001\u5904\u7406\u5B8C\u6210\uFF1Aprocessed=${imageResult.processed}, succeeded=${imageResult.succeeded}, skipped=${imageResult.skipped}, failed=${imageResult.failed}`
+            );
+          }).catch((error) => {
+            const message = error instanceof Error ? error.message : String(error);
+            console.error(`\u98DE\u4E66\u56FE\u7247\u591A\u6A21\u6001\u5904\u7406\u5931\u8D25\uFF1A${message}`);
+          });
+        }
         if (result.attachment.indexedMessageId) {
           console.log(`\u98DE\u4E66\u9644\u4EF6\u5DF2\u8FDB\u5165 RAG\uFF1A${result.attachment.indexedMessageId}`);
           if (result.attachment.vectorIndexed) {
@@ -2799,7 +3449,8 @@ function createFeishuGateway(options) {
     questionHandler: options.questionHandler,
     resourceDownloader: options.resourceDownloader,
     attachmentVectorIndexer: options.attachmentVectorIndexer,
-    episodeProcessor: options.episodeProcessor
+    episodeProcessor: options.episodeProcessor,
+    imageMultimodalProcessor: options.imageMultimodalProcessor
   });
   return {
     async start() {
@@ -2881,7 +3532,7 @@ var FeishuResourceDownloader = class _FeishuResourceDownloader {
 };
 
 // src/files/ingest.ts
-import crypto4 from "crypto";
+import crypto5 from "crypto";
 import fs11 from "fs/promises";
 import path13 from "path";
 
@@ -2945,7 +3596,7 @@ function ensureSupportedTextFile(filePath) {
   }
 }
 function stableStoredName(sourcePath, fileName) {
-  const digest = crypto4.createHash("sha256").update(sourcePath).digest("hex").slice(0, 16);
+  const digest = crypto5.createHash("sha256").update(sourcePath).digest("hex").slice(0, 16);
   return `${digest}-${fileName}`;
 }
 async function ingestLocalFile(input2) {
@@ -3183,12 +3834,17 @@ function extractAttachment(message) {
   }
   return candidate;
 }
+function isMultimodalReady(config, secrets) {
+  return Boolean(config.multimodal.baseUrl && config.multimodal.model && secrets.multimodal.apiKey);
+}
 var GatewayIngestor = class {
   messages;
   jobs;
+  imageTasks;
   constructor(database) {
     this.messages = new MessageRepository(database);
     this.jobs = new FileJobRepository(database);
+    this.imageTasks = new ImageMultimodalTaskRepository(database);
   }
   ingestFeishuEvent(payload) {
     const normalized = normalizeFeishuReceiveMessageEvent(payload);
@@ -3220,6 +3876,23 @@ var GatewayIngestor = class {
       messageId: result.message.platformMessageId,
       attachment
     });
+    if (attachment.kind === "image") {
+      const imageTask = isMultimodalReady(input2.config, input2.secrets) ? this.imageTasks.enqueue({
+        sourceMessageId: result.messageId,
+        platformMessageId: result.message.platformMessageId,
+        imageKey: attachment.fileKey,
+        storedPath: downloaded.storedPath,
+        mimeType: attachment.mimeType || "image/jpeg"
+      }) : void 0;
+      return {
+        ...result,
+        attachment: {
+          downloaded,
+          ...imageTask ? { imageTask } : {},
+          skippedReason: imageTask ? "\u56FE\u7247\u5DF2\u4E0B\u8F7D\uFF0C\u7B49\u5F85\u591A\u6A21\u6001\u540E\u53F0\u5904\u7406\u3002" : "\u56FE\u7247\u5DF2\u4E0B\u8F7D\uFF0C\u4F46\u591A\u6A21\u6001\u672A\u914D\u7F6E\u3002"
+        }
+      };
+    }
     if (!isSupportedTextFile(downloaded.storedPath)) {
       return {
         ...result,
@@ -3355,6 +4028,96 @@ async function startDetachedGateway(input2) {
   }
 }
 
+// src/multimodal/openai-compatible.ts
+import fs13 from "fs/promises";
+function normalizeBaseUrl2(baseUrl) {
+  return baseUrl.replace(/\/+$/, "");
+}
+function buildPrompt(context) {
+  const contextText = context?.trim();
+  return [
+    "\u8BF7\u7406\u89E3\u8FD9\u5F20\u56FE\u7247\uFF0C\u5224\u65AD\u5B83\u662F\u5426\u5305\u542B\u503C\u5F97\u8FDB\u5165\u77E5\u8BC6\u5E93\u548C\u4F1A\u8BDD\u8BB0\u5FC6\u7684\u6709\u610F\u4E49\u4FE1\u606F\u3002",
+    '\u8BF7\u53EA\u8F93\u51FA JSON\uFF0C\u683C\u5F0F\u4E3A {"summary": string, "isMeaningful": boolean, "reason": string}\u3002',
+    "summary \u4F7F\u7528\u7B80\u6D01\u4E2D\u6587\u8F6C\u8FF0\u56FE\u7247\u4E2D\u7684\u5173\u952E\u4FE1\u606F\uFF1B\u65E0\u610F\u4E49\u56FE\u7247\u4E5F\u8981\u7ED9\u51FA\u7B80\u77ED summary\u3002",
+    contextText ? `\u4E0A\u4E0B\u6587\uFF1A${contextText}` : void 0
+  ].filter(Boolean).join("\n");
+}
+function parseDescribeImageResult(content) {
+  let data2;
+  try {
+    data2 = JSON.parse(content);
+  } catch {
+    throw new Error("\u591A\u6A21\u6001\u6A21\u578B\u8FD4\u56DE\u7684 JSON \u65E0\u6CD5\u89E3\u6790\u3002");
+  }
+  if (!data2 || typeof data2 !== "object") {
+    throw new Error("\u591A\u6A21\u6001\u6A21\u578B\u8FD4\u56DE\u683C\u5F0F\u4E0D\u6B63\u786E\u3002");
+  }
+  const result = data2;
+  const summary = typeof result.summary === "string" ? result.summary.trim() : "";
+  if (!summary) {
+    throw new Error("\u591A\u6A21\u6001\u6A21\u578B\u8FD4\u56DE\u7684 summary \u4E3A\u7A7A\u3002");
+  }
+  if (typeof result.isMeaningful !== "boolean") {
+    throw new Error("\u591A\u6A21\u6001\u6A21\u578B\u8FD4\u56DE\u7684 isMeaningful \u4E0D\u662F\u5E03\u5C14\u503C\u3002");
+  }
+  const reason = typeof result.reason === "string" ? result.reason.trim() : "";
+  return {
+    summary,
+    isMeaningful: result.isMeaningful,
+    ...reason ? { reason } : {}
+  };
+}
+var OpenAICompatibleMultimodalModel = class {
+  constructor(options) {
+    this.options = options;
+  }
+  options;
+  async describeImage(input2) {
+    if (!this.options.baseUrl || !this.options.apiKey || !this.options.model) {
+      throw new Error("\u591A\u6A21\u6001\u914D\u7F6E\u4E0D\u5B8C\u6574\u3002\u8BF7\u8FD0\u884C chattercatcher setup \u6216 chattercatcher settings\u3002");
+    }
+    const image = await fs13.readFile(input2.imagePath);
+    const response = await fetch(`${normalizeBaseUrl2(this.options.baseUrl)}/chat/completions`, {
+      method: "POST",
+      headers: {
+        authorization: `Bearer ${this.options.apiKey}`,
+        "content-type": "application/json"
+      },
+      body: JSON.stringify({
+        model: this.options.model,
+        messages: [
+          {
+            role: "user",
+            content: [
+              { type: "text", text: buildPrompt(input2.context) },
+              { type: "image_url", image_url: { url: `data:${input2.mimeType};base64,${image.toString("base64")}` } }
+            ]
+          }
+        ],
+        response_format: { type: "json_object" },
+        temperature: this.options.temperature ?? 0.2
+      })
+    });
+    if (!response.ok) {
+      const body = await response.text();
+      throw new Error(`\u591A\u6A21\u6001\u8BF7\u6C42\u5931\u8D25\uFF1A${response.status} ${body}`);
+    }
+    const data2 = await response.json();
+    const content = data2.choices?.[0]?.message?.content?.trim();
+    if (!content) {
+      throw new Error("\u591A\u6A21\u6001\u6A21\u578B\u8FD4\u56DE\u4E3A\u7A7A\u3002");
+    }
+    return parseDescribeImageResult(content);
+  }
+};
+function createMultimodalModel(config, secrets) {
+  return new OpenAICompatibleMultimodalModel({
+    baseUrl: config.multimodal.baseUrl,
+    apiKey: secrets.multimodal.apiKey,
+    model: config.multimodal.model
+  });
+}
+
 // src/rag/indexer.ts
 async function indexMessageChunks(input2) {
   const chunks = input2.messageIds ? input2.messages.listMessageChunksByMessageIds(input2.messageIds, input2.limit ?? 1e4) : input2.messages.listAllMessageChunks(input2.limit ?? 1e4);
@@ -3433,6 +4196,22 @@ async function processMessagesNow(input2) {
   };
 }
 
+// src/rag/qa-service.ts
+async function askWithRag(input2) {
+  const evidence = await input2.retriever.retrieve(input2.question);
+  if (evidence.length === 0) {
+    return {
+      answer: "\u4E0D\u77E5\u9053\u3002\u5F53\u524D\u672C\u5730\u77E5\u8BC6\u5E93\u6CA1\u6709\u68C0\u7D22\u5230\u8DB3\u591F\u8BC1\u636E\u3002",
+      citations: []
+    };
+  }
+  return generateGroundedAnswer({
+    question: input2.question,
+    evidence,
+    model: input2.model
+  });
+}
+
 // src/update/npm-updater.ts
 import { execFile } from "child_process";
 import { promisify } from "util";
@@ -4065,12 +4844,31 @@ async function promptForConfiguration(config, secrets) {
     llmApiKey: secrets.llm.apiKey
   });
   config.embedding.model = await input({ message: "Embedding Model", default: config.embedding.model });
+  const multimodalBaseUrl = await input({
+    message: "Multimodal Base URL\uFF08OpenAI-compatible\uFF0C\u53EF\u7559\u7A7A\uFF09",
+    default: config.multimodal.baseUrl
+  });
+  const multimodalApiKey = await password({
+    message: "Multimodal API Key\uFF08\u53EF\u7559\u7A7A\uFF09",
+    mask: "*"
+  });
+  const multimodalModel = await input({
+    message: "Multimodal Model\uFF08\u53EF\u7559\u7A7A\uFF09",
+    default: config.multimodal.model
+  });
   const dimension = await number({
     message: "Embedding \u7EF4\u5EA6\uFF08\u4E0D\u77E5\u9053\u53EF\u5148\u7559\u7A7A\uFF09",
     default: config.embedding.dimension ?? void 0,
     required: false
   });
   config.embedding.dimension = dimension ?? null;
+  config.multimodal = {
+    baseUrl: multimodalBaseUrl,
+    model: multimodalModel
+  };
+  secrets.multimodal = {
+    apiKey: multimodalApiKey || secrets.multimodal.apiKey
+  };
   config.web.port = await number({ message: "Web UI \u7AEF\u53E3", default: config.web.port, required: true }) ?? config.web.port;
   config.feishu.requireMention = await confirm({
     message: "\u7FA4\u804A\u56DE\u7B54\u662F\u5426\u8981\u6C42 @ \u673A\u5668\u4EBA\uFF1F",
@@ -4098,7 +4896,8 @@ function printSettings(config, secrets) {
       secrets: {
         feishu: { appSecret: maskSecret(secrets.feishu.appSecret) },
         llm: { apiKey: maskSecret(secrets.llm.apiKey) },
-        embedding: { apiKey: maskSecret(secrets.embedding.apiKey) }
+        embedding: { apiKey: maskSecret(secrets.embedding.apiKey) },
+        multimodal: { apiKey: maskSecret(secrets.multimodal.apiKey) }
       }
     },
     null,
@@ -4210,6 +5009,10 @@ async function startGatewayForegroundCommand() {
       database,
       model: createChatModel(config, secrets)
     },
+    imageMultimodalProcessor: config.multimodal.baseUrl && config.multimodal.model && secrets.multimodal.apiKey ? {
+      database,
+      model: createMultimodalModel(config, secrets)
+    } : void 0,
     questionHandler: new FeishuQuestionHandler({
       config,
       secrets,
@@ -4577,7 +5380,7 @@ dev.command("ingest-feishu-event").description("\u4ECE JSON \u6587\u4EF6\u6A21\u
   const config = await loadConfig();
   const database = openDatabase(config);
   try {
-    const raw = await fs13.readFile(options.file, "utf8");
+    const raw = await fs14.readFile(options.file, "utf8");
     const payload = JSON.parse(raw);
     const result = new GatewayIngestor(database).ingestFeishuEvent(payload);
     if (!result.accepted) {