[BUG] Cursor has errors calling search tool

[phernandez]

Bug Description

Cursor cannot figure out how to structure the parameters for that tool call. No matter what Cursor seems to try it gets the errors.

- When I pass an object: "Parameter 'query' must be of type undefined, got object"
- When I pass a string: "Parameter 'query' must be of type undefined, got string"



 and then it reports:  "Invalid type for parameter 'query' in tool search_notes"
Any chance you can give me some guidance with this?

Steps To Reproduce

Steps to reproduce the behavior:

try using search tool in Cursor.

Possible Solution

The tool args should probably be plain text and not json to make it easier to call.

[phernandez]

claude: I tried this also and cursor could not invoke the tool, even though it works via Claude Desktop and via the inspector. I'm wondering if we should refactor search (renamed to search_notes in the latest version) to use simpler args, strings instead of json.

[github-actions]

Claude Code Analysis

Based on my thorough analysis of the issue and codebase, here's my assessment of the problem with Cursor's ability to call the search_notes tool:

Issue Analysis

The bug occurs because Cursor is unable to correctly structure parameters for the search_notes tool, specifically the query parameter. The error messages are:

• When passing an object: "Parameter 'query' must be of type undefined, got object"
• When passing a string: "Parameter 'query' must be of type undefined, got string"

Root Cause

1. The search_notes tool expects its first parameter to be a SearchQuery object, which is a Pydantic model with several optional fields.
2. The tool is defined as:

   async def search_notes(query: SearchQuery, page: int = 1, page_size: int = 10) -> SearchResponse:

3. While this works correctly in Claude Desktop and the MCP Inspector, Cursor's integration with the MCP protocol appears to have difficulty with complex Pydantic models as parameters.
4. Cursor may be trying to pass either a string or an object directly for the query parameter, instead of properly constructing a SearchQuery object.

Potential Solutions

1. Simplify the search_notes interface (Recommended):

   • Modify the tool to accept simple string parameters instead of a complex object:

   @mcp.tool(description="Search across all content in the knowledge base.")
   async def search_notes(
       text: Optional[str] = None,
       title: Optional[str] = None,
       permalink: Optional[str] = None,
       permalink_match: Optional[str] = None,
       page: int = 1,
       page_size: int = 10
   ) -> SearchResponse:
       # Construct SearchQuery internally
       query = SearchQuery(
           text=text,
           title=title,
           permalink=permalink,
           permalink_match=permalink_match
       )
       # Rest of the function...

2. Create a string-only search variant:

   • Add a simpler alternative search tool specifically for clients like Cursor:

   @mcp.tool(description="Simple text search across all content in the knowledge base.")
   async def search_by_text(text: str, page: int = 1, page_size: int = 10) -> SearchResponse:
       query = SearchQuery(text=text)
       return await search_notes(query, page, page_size)

3. Update the schema handling in MCP implementation:

   • Enhance parameter handling to better convert string or object parameters to the expected Pydantic model.

Files to Modify

1. Primary file: /src/basic_memory/mcp/tools/search.py

   • Update the search_notes function signature
   • Internally construct the SearchQuery object

2. If adding a new tool: /src/basic_memory/mcp/tools/search.py

   • Add a new search_by_text function

3. Update tests: /tests/mcp/test_tool_search.py

   • Update tests to match the new function signature

Code Snippet for Solution #1

@mcp.tool(
    description="Search across all content in the knowledge base.",
)
async def search_notes(
    text: Optional[str] = None,
    title: Optional[str] = None,
    permalink: Optional[str] = None,
    permalink_match: Optional[str] = None,
    after_date: Optional[str] = None,
    page: int = 1,
    page_size: int = 10
) -> SearchResponse:
    """Search across all content in the knowledge base.

    This tool searches the knowledge base using full-text search, pattern matching,
    or exact permalink lookup. It supports filtering by content type, entity type,
    and date.

    Args:
        text: Full-text search (e.g., "project planning")
          Supports boolean operators: AND, OR, NOT and parentheses for grouping
        title: Search only in titles (e.g., "Meeting notes")
        permalink: Exact permalink match (e.g., "docs/meeting-notes")
        permalink_match: Pattern matching for permalinks (e.g., "docs/*-notes")
        after_date: Optional date filter for recent content (e.g., "1 week", "2d")
        page: The page number of results to return (default 1)
        page_size: The number of results to return per page (default 10)

    Returns:
        SearchResponse with results and pagination info
    """
    # Construct SearchQuery object from parameters
    query = SearchQuery(
        text=text,
        title=title,
        permalink=permalink,
        permalink_match=permalink_match,
        after_date=after_date,
    )
    
    logger.info(f"Searching for {query}")
    response = await call_post(
        client,
        "/search/",
        json=query.model_dump(),
        params={"page": page, "page_size": page_size},
    )
    return SearchResponse.model_validate(response.json())

This approach maintains backward compatibility with existing code but makes the tool more accessible to clients like Cursor that may struggle with complex object parameters.

---

🤖 Generated with Claude Code GitHub Action

[phernandez]

claude-fix: let make a PR to implement option #1.

[github-actions]

I've analyzed this issue and created a fix in PR: #78

Here's a summary of the changes:

Fixed issue #75 where Cursor was having errors calling the search_notes tool. The problem was that the search_notes tool was expecting a complex Pydantic object (SearchQuery) as input, which was confusing Cursor.

Changes:

1. Modified the search_notes tool to accept primitive types (strings, lists, etc.) as parameters instead of a complex Pydantic object
2. Updated the implementation to create a SearchQuery object internally from these primitive parameters
3. Updated all call sites in the codebase that were using the old function signature
4. Updated tests to use the new function signature

The fix makes it easier for LLMs like Cursor to make proper calls to the search_notes tool, which will resolve the reported error messages:

• "Parameter 'query' must be of type undefined, got object"
• "Parameter 'query' must be of type undefined, got string"
• "Invalid type for parameter 'query' in tool search_notes"

Files modified:

• src/basic_memory/mcp/tools/search.py
• src/basic_memory/mcp/tools/read_note.py
• src/basic_memory/mcp/prompts/search.py
• tests/mcp/test_tool_search.py
• tests/mcp/test_tool_read_note.py

Please review the PR and test the changes to verify they resolve the issue.

---

🤖 Generated with Claude Code GitHub Action