# Debug and Fix Plan for MCP Server Issues
## Issue Analysis
### 1. Python File Count Shows 0
**Root Cause**: Multiple issues identified:
- The `get_indexed_count()` method was checking for wrong attribute (`_file_contents` instead of `index`)
- Fixed this in both basic and semantic Python plugins
- However, still showing 0 because the running MCP server instance may be using different plugin instances
**Additional Finding**:
- The plugin manager initialization in debug script shows 0 plugins loaded
- SQLite schema issues preventing plugin initialization ("no such table: repositories")
- This suggests the MCP server running through Claude Code uses a different initialization path
### 2. Complex Semantic Queries Return No Results
**Root Cause Analysis**:
- Basic Python plugin returns empty list for semantic queries: `if opts and opts.get("semantic"): return []`
- Semantic plugin exists but may not be loaded when `SEMANTIC_SEARCH_ENABLED=true`
- Environment variable loading happens after plugin imports in some cases
**Key Issues**:
1. Plugin selection happens at import time in `__init__.py`
2. Environment variables from `.env` may not be loaded early enough
3. Voyage AI integration requires API key which may not be available to the MCP server process
### 3. Semantic Search Partially Implemented
**Implementation Status**:
- ✅ SemanticIndexer class fully implemented
- ✅ Voyage AI client integration complete
- ✅ Qdrant vector storage working (tested with in-memory)
- ✅ Enhanced plugin base class created
- ⚠️ Plugin switching logic may not work in production
- ❌ Semantic indexing not called during regular indexing
- ❌ Complex queries fail because embeddings aren't created
## Comprehensive Fix Plan
### Phase 1: Fix Python File Count (Quick Win)
1. **Update Dispatcher Status Collection**
```python
# In mcp_server_cli.py, line 219
# Current: lang = getattr(plugin, 'language', getattr(plugin, 'lang', 'unknown'))
# Should be: lang = getattr(plugin, 'lang', getattr(plugin, 'language', 'unknown'))
```
2. **Ensure Consistent Plugin Attribute**
- Add `language` property to base plugin class that returns `lang`
- Or update all references to use `lang` consistently
### Phase 2: Fix Semantic Search Loading
1. **Early Environment Loading**
```python
# In mcp_server_cli.py - already fixed
from dotenv import load_dotenv
load_dotenv()
```
2. **Fix Plugin Selection Logic**
```python
# In plugin __init__.py files
# Move environment check to lazy loading or factory method
def get_plugin_class():
if os.getenv("SEMANTIC_SEARCH_ENABLED", "false").lower() == "true":
return PythonPluginSemantic
return Plugin
```
3. **Ensure API Keys Available**
- Verify VOYAGE_API_KEY is in environment
- Add fallback for missing keys
- Better error messages
### Phase 3: Complete Semantic Integration
1. **Fix Semantic Indexing Call**
```python
# In plugin_semantic.py indexFile method
# Currently: self.index_with_embeddings() is called
# Need to verify: _enable_semantic is True
# Add logging to trace execution
```
2. **Debug Embedding Creation**
```python
# Add debug logging
logger.info(f"Semantic enabled: {self._enable_semantic}")
logger.info(f"Creating embeddings for {len(symbols)} symbols")
```
3. **Fix Search Method**
```python
# Ensure semantic search actually queries embeddings
# Currently may be falling back to traditional search
```
## Implementation Steps
### Step 1: Create Test Harness
Create a test script that mimics MCP server initialization to debug issues in isolation.
### Step 2: Add Comprehensive Logging
Add debug logging at key points:
- Plugin initialization
- Environment variable loading
- Semantic indexer creation
- Embedding generation
- Search query routing
### Step 3: Fix Configuration Loading
Ensure all configuration is loaded before plugin initialization:
- Load .env at startup
- Validate required API keys
- Check Qdrant connectivity
### Step 4: Implement Fixes
1. Fix get_indexed_count() ✅ (already done)
2. Fix language/lang attribute usage
3. Fix plugin selection logic
4. Ensure semantic indexing is called
5. Verify search routing
### Step 5: Test End-to-End
1. Index a repository with semantic enabled
2. Verify embeddings are created
3. Test semantic search queries
4. Confirm file counts are accurate
## Testing Strategy
### Unit Tests
```python
def test_python_plugin_count():
plugin = PythonPlugin()
plugin.indexFile("test.py", "def test(): pass")
assert plugin.get_indexed_count() == 1
def test_semantic_plugin_loading():
os.environ["SEMANTIC_SEARCH_ENABLED"] = "true"
from mcp_server.plugins.python_plugin import Plugin
assert isinstance(Plugin(), PythonPluginSemantic)
```
### Integration Tests
1. Start MCP server with semantic enabled
2. Index sample repository
3. Verify status shows correct counts
4. Test semantic search returns results
## Next Actions
1. **Immediate**: Fix language/lang attribute issue in status collection
2. **Short-term**: Improve plugin loading to respect environment variables
3. **Medium-term**: Add comprehensive debug logging
4. **Long-term**: Refactor plugin system for better configurability
## Success Criteria
- [ ] Python file count shows actual number of indexed files
- [ ] Semantic search returns relevant results for natural language queries
- [ ] All language plugins report accurate counts
- [ ] No errors in logs during indexing or searching
- [ ] Performance remains acceptable with semantic search enabled
