{
  "mcpServers": {
    "knowledge-base": {
      "command": "python",
      "args": ["/absolute/path/to/knowledge-base-mcp/kb_server.py"],
      "env": {
        "KB_DATA_DIR": "/absolute/path/to/knowledge-base-mcp/kb_data"
      }
    }
  }
}

{
  "mcpServers": {
    "knowledge-base": {
      "command": "python",
      "args": ["/absolute/path/to/knowledge-base-mcp/kb_server.py"],
      "env": {
        "KB_DATA_DIR": "/absolute/path/to/knowledge-base-mcp/kb_data",
        "KB_INITIAL_FILE": "/absolute/path/to/knowledge-base-mcp/initial_knowledge.txt"
      }
    }
  }
}

#kb_save we use splunk on the cloud at https://company.splunkcloud.com
#kb_save when dynatrace fails in tanzu, use DT_DISABLE flag and restart the instance
#kb_save here's our graphql mutation test pattern: ```csharp
[Test]
public async Task TestGraphQLMutation() {
    // test code here
}

### Searching Knowledge

GitHub Copilot will automatically search when you ask questions:

### Manual Search

You can also explicitly search:

## Available Tools

### `kb_save`
Save a memory to the knowledge base.
- **content**: The memory content to save
- **memory_type**: Optional type (environment, code_snippet, operational, architectural)
- **tags**: Optional list of tags for categorization

### `kb_search`
Search for relevant memories.
- **query**: Search query
- **limit**: Maximum results (default: 5)
- **memory_type**: Filter by type
- **include_metadata**: Include detailed metadata

### `kb_list`
List all saved memories.
- **memory_type**: Filter by type
- **limit**: Maximum entries (default: 10)
- **include_content**: Show full content vs summary

### `kb_delete`
Delete a memory by ID.
- **memory_id**: Full or partial memory ID

## Configuration

### Environment Variables

- `KB_DATA_DIR`: Directory for ChromaDB storage (default: `./kb_data`)
- `KB_INITIAL_FILE`: Optional path to initial knowledge file to load on startup
- `KB_UI_PORT`: Port for the Streamlit UI (default: `8501`)

### Initial Knowledge File

You can bootstrap the knowledge base with pre-existing information by providing an initial knowledge file. The file should contain knowledge entries separated by double newlines (`\n\n`).

**Example `initial_knowledge.txt`:**

[Test]
public async Task TestAPI() {
    // test code here
}

**Features:**
- ✅ Automatic metadata extraction (technologies, URLs, memory types)
- ✅ Entries marked with `source: initial_knowledge` 
- ✅ Loads only on first startup (won't duplicate entries)
- ✅ Supports all content types (code, configs, operational knowledge)

### Embedding Model

The server uses `all-MiniLM-L6-v2` by default for local embeddings. This provides:
- Fast inference
- Good semantic understanding
- No external API calls
- Small memory footprint

## Data Storage

All data is stored locally in ChromaDB format:
- **Vector embeddings**: For semantic search
- **Document content**: Raw memory text
- **Metadata**: Extracted technologies, URLs, timestamps, access counts

## Performance

- **Search latency**: < 500ms typical
- **Storage capacity**: 10,000+ memories
- **Memory usage**: ~200MB for model + data
- **Embedding generation**: ~10ms per memory

## Security & Privacy

- ✅ **Local-only storage**: No cloud dependencies
- ✅ **No external APIs**: Embeddings generated locally
- ✅ **File-system permissions**: Standard OS-level access control
- ✅ **Corporate compliant**: Designed for enterprise environments

## Troubleshooting

### Server Won't Start
- Check Python version (3.9+ required)
- Verify all dependencies installed: `pip install -r requirements.txt`
- Check data directory permissions

### Poor Search Results
- Ensure memories are saved with clear, descriptive content
- Use specific technology keywords
- Try different search terms

### Memory Not Found
- Use `kb_list` to see all saved memories
- Check memory type filters
- Verify memory was actually saved (check for success message)

## Development

### Project Structure

### Adding New Features

The server uses FastMCP for easy tool development:

```python
@mcp.tool()
async def new_tool(param: str) -> str:
    """Tool description."""
    # Implementation
    return "Result"