refine_prompt

Instructions

Implementation Reference

• src/index.ts:138-161 (handler)

  The CallTool handler for 'refine_prompt'. It takes the prompt argument, optionally retrieves relevant semantic memories via LanceDB vector search, and returns a refined prompt with context.

  if (name === "refine_prompt") {
    const prompt = args?.prompt as string;
    let contextExtra = "";
  
    // Try to fetch relevant memories
    if (table) {
      const queryVector = await getEmbedding(prompt);
      const results = await table.vectorSearch(queryVector).limit(3).toArray();
      
      if (results.length > 0) {
        contextExtra = "\n[Context Retrieved from Memory]:\n" + 
          results.map((r: any) => `- ${r.text}`).join("\n");
      }
    }
  
    return {
      content: [
        {
          type: "text",
          text: `[Refined Prompt]: ${prompt}\n${contextExtra}\n\n(Antigravity can now use the information above to generate a more precise response)`,
        },
      ],
    };
  }

• src/index.ts:76-88 (schema)

  Registration of 'refine_prompt' in the ListTools handler, including its description and inputSchema (requires a 'prompt' string).

    name: "refine_prompt",
    description: "Refines a prompt using semantic memory to make it more contextual and efficient.",
    inputSchema: {
      type: "object",
      properties: {
        prompt: {
          type: "string",
          description: "The original prompt that needs refinement.",
        },
      },
      required: ["prompt"],
    },
  },

• src/index.ts:111-164 (registration)

  The CallToolRequestSchema handler that dispatches to 'refine_prompt' (alongside 'learn_context') based on the tool name.

  server.setRequestHandler(CallToolRequestSchema, async (request) => {
    const { name, arguments: args } = request.params;
  
    if (name === "learn_context") {
      const info = args?.information as string;
      const category = (args?.category as string) || "general";
      
      const vector = await getEmbedding(info);
      
      const data = [{
        vector,
        text: info,
        category,
        timestamp: new Date().toISOString()
      }];
  
      if (!table) {
        table = await db.createTable(TABLE_NAME, data);
      } else {
        await table.add(data);
      }
  
      return {
        content: [{ type: "text", text: `Learned and stored in semantic memory: "${info}"` }],
      };
    }
  
    if (name === "refine_prompt") {
      const prompt = args?.prompt as string;
      let contextExtra = "";
  
      // Try to fetch relevant memories
      if (table) {
        const queryVector = await getEmbedding(prompt);
        const results = await table.vectorSearch(queryVector).limit(3).toArray();
        
        if (results.length > 0) {
          contextExtra = "\n[Context Retrieved from Memory]:\n" + 
            results.map((r: any) => `- ${r.text}`).join("\n");
        }
      }
  
      return {
        content: [
          {
            type: "text",
            text: `[Refined Prompt]: ${prompt}\n${contextExtra}\n\n(Antigravity can now use the information above to generate a more precise response)`,
          },
        ],
      };
    }
  
    throw new Error(`Tool not found: ${name}`);
  });

• src/index.ts:72-107 (registration)

  The ListToolsRequestSchema handler that registers 'refine_prompt' as an available tool with its input schema.

  server.setRequestHandler(ListToolsRequestSchema, async () => {
    return {
      tools: [
        {
          name: "refine_prompt",
          description: "Refines a prompt using semantic memory to make it more contextual and efficient.",
          inputSchema: {
            type: "object",
            properties: {
              prompt: {
                type: "string",
                description: "The original prompt that needs refinement.",
              },
            },
            required: ["prompt"],
          },
        },
        {
          name: "learn_context",
          description: "Memorizes important information (preference, technical rule, context) for future use.",
          inputSchema: {
            type: "object",
            properties: {
              information: {
                type: "string",
                description: "The information to be remembered.",
              },
              category: {
                type: "string",
                description: "Information category (e.g., 'preference', 'architecture', 'style').",
              },
            },
            required: ["information"],
          },
        },
      ],

• src/index.ts:52-58 (helper)

  getEmbedding is a helper function that uses Ollama to generate embeddings for text, used by refine_prompt to perform vector similarity search.

  async function getEmbedding(text: string): Promise<number[]> {
    const response = await ollama.embed({
      model: EMBEDDING_MODEL,
      input: text,
    });
    return response.embeddings[0]!;
  }