MCP Docs RAG Server

Instructions

Implementation Reference

• src/index.ts:454-511 (handler)

  Main handler for 'rag_query' tool call. Validates inputs, checks document existence, loads/indexes document, sets up Gemini LLM and query engine using LlamaIndex, executes query, handles errors.

  case "rag_query": {
    const documentId = String(request.params.arguments?.document_id);
    const query = String(request.params.arguments?.query);
    
    if (!documentId || !query) {
      throw new Error("Document ID and query are required");
    }
    
    try {
      // ドキュメントが存在するか確認し、存在しなければ自動的に作成を試みる
      let documents = await listDocuments();
      let document = documents.find(c => c.id === documentId);
      
      if (!document) {
        return {
          content: [{
            type: "text",
            text: `Document '${documentId}' not found. Please add it manually using add_git_repository or add_text_file tools.`
          }]
        };
      }
      
      // Load and index document if needed
      const index = await loadDocument(documentId);
    
    // 一時的にGemini LLMを設定
    const originalLLM = Settings.llm;
    const gemini = new Gemini({
      model: GEMINI_MODEL.GEMINI_2_0_FLASH
    });
    
    // グローバル設定に設定
    Settings.llm = gemini;
    
    // クエリエンジンの作成
    const queryEngine = index.asQueryEngine();
    
    // クエリの実行
    const response = await queryEngine.query({
      query
    });
    
    return {
      content: [{
        type: "text",
        text: response.toString()
      }]
    };
    } catch (error: any) {
      console.error(`Error in rag_query:`, error.message);
      return {
        content: [{
          type: "text",
          text: `Error processing query: ${error.message}`
        }]
      };
    }
  }

• src/index.ts:371-388 (registration)

  Registration of 'rag_query' tool in ListToolsRequestSchema handler, including name, description, and input schema definition.

  {
    name: "rag_query",
    description: "Query a document using RAG. Note: If the index does not exist, it will be created when you query, which may take some time.",
    inputSchema: {
      type: "object",
      properties: {
        document_id: {
          type: "string",
          description: "ID of the document to query"
        },
        query: {
          type: "string",
          description: "Query to run against the document"
        }
      },
      required: ["document_id", "query"]
    }
  },

• src/index.ts:374-387 (schema)

  Input schema for rag_query tool defining required document_id (string) and query (string) parameters.

  inputSchema: {
    type: "object",
    properties: {
      document_id: {
        type: "string",
        description: "ID of the document to query"
      },
      query: {
        type: "string",
        description: "Query to run against the document"
      }
    },
    required: ["document_id", "query"]
  }

• src/index.ts:160-222 (helper)

  Key helper function called by rag_query handler. Loads document files recursively, creates and persists VectorStoreIndex using LlamaIndex with Gemini embedding/LLM, caches in global indices.

  export async function loadDocument(documentId: string): Promise<VectorStoreIndex> {
    if (indices[documentId]?.index) {
      return indices[documentId].index;
    }
  
    let documents = await listDocuments();
    let document = documents.find(c => c.id === documentId);
    
    // ドキュメントが存在しない場合はエラーをスロー
    if (!document) {
      throw new Error(`Document not found: ${documentId}`);
    }
  
    let documentItems: Document[] = [];
    
    if (fs.statSync(document.path).isDirectory()) {
      // ディレクトリを再帰的に処理
      documentItems = await readDirectoryRecursively(document.path);
      
      // 空のドキュメントリストの場合にフォールバックメッセージを追加
      if (documentItems.length === 0) {
        console.warn(`No documents found in document: ${documentId}`);
        documentItems.push(new Document({ 
          text: `This document (${document.name}) appears to be empty. Please check if files exist at path: ${document.path}`, 
          metadata: { name: 'empty-notice', source: document.path } 
        }));
      }
    } else {
      // Process single file
      const text = fs.readFileSync(document.path, 'utf-8');
      documentItems = [new Document({ text, metadata: { name: document.id, source: document.path } })];
    }
    
    // Gemini埋め込みモデルを設定
    const geminiEmbed = new GeminiEmbedding();
    
    // Gemini LLMモデルを設定
    const gemini = new Gemini({
      model: GEMINI_MODEL.GEMINI_2_0_FLASH
    });
    
    // グローバル設定に埋め込みモデルとLLMを設定（これをグローバルに保持）
    Settings.embedModel = geminiEmbed;
    Settings.llm = gemini;
    
    // Create storage context
    const storageContext = await storageContextFromDefaults({
      persistDir: path.join(DOCS_PATH, '.indices', documentId),
    });
    
    // Create index
    const index = await VectorStoreIndex.fromDocuments(documentItems, {
      storageContext,
    });
    
    // Save index for future use
    indices[documentId] = { 
      index, 
      description: document.description 
    };
    
    return index;
  }