Docs-MCP

Instructions

Args:
    query: 検索クエリ
    limit: 返す結果の最大数（デフォルト: 5）

Implementation Reference

• src/mcp_server_docs/server.py:54-62 (handler)

  MCP tool handler for semantic_search. Registers the tool with @mcp.tool() and delegates execution to DocumentManager.semantic_search.

  @mcp.tool()
  async def semantic_search(query: str, limit: int = 5) -> str:
      """意味的に関連する内容を検索
  
      Args:
          query: 検索クエリ
          limit: 返す結果の最大数（デフォルト: 5）
      """
      return doc_manager.semantic_search(query, limit)

• src/mcp_server_docs/document_manager.py:298-339 (helper)

  Core implementation of semantic search using OpenAI embeddings, cosine similarity, and preview extraction from cached document embeddings.

  def semantic_search(self, query: str, limit: int = 5) -> str:
      """意味的に関連する内容を検索"""
      if not self.client:
          return "Error: OpenAI API key not configured"
  
      if not self.embeddings_cache:
          return "Error: No embeddings available. Run 'python scripts/generate_metadata.py' first."
  
      try:
          # クエリのembeddingを取得
          query_embedding = self._get_embedding(query)
  
          # 各ドキュメントとの類似度を計算
          similarities = []
          for doc_path, doc_embedding in self.embeddings_cache.items():
              # embeddingがリストとして保存されているので、そのまま使用
              similarity = self._cosine_similarity(query_embedding, doc_embedding)
              similarities.append((doc_path, similarity))
  
          # 類似度でソート
          similarities.sort(key=lambda x: x[1], reverse=True)
  
          # 結果を構築
          results = []
          for doc_path, similarity in similarities[:limit]:
              description = self.docs_metadata.get(doc_path, "")
              result_line = f"{doc_path} (相似度: {similarity:.3f})"
              if description:
                  result_line += f" - {description}"
              results.append(result_line)
  
              # 関連する内容を一部抽出
              if doc_path in self.docs_content:
                  content = self.docs_content[doc_path]
                  preview = self._extract_preview(content, query)
                  if preview:
                      results.append(f"  → {preview}")
  
          return "\n\n".join(results)
  
      except Exception as e:
          return f"Error during semantic search: {e}"