sqlite-vec requires compilation on some platforms:
uv add sqlite-vec --reinstall
# If that fails on macOS
brew install sqlite
uv add sqlite-vec --reinstall- Check content was indexed:
SELECT COUNT(*) FROM chunks WHERE content LIKE '%yourterm%'
- FTS5 tokenization may differ from expectation. Try simpler queries.
- For code symbols, ensure they weren't stripped during HTML→Markdown conversion.
- Test manually:
docs-mcp --db comsol.db --test "test query" - Check paths in IDE config are absolute
- Restart IDE after changing MCP configuration
- Check IDE's MCP debug logs
Embedding generation is the bottleneck (~5 min for 10k chunks). Options:
- Use
--no-embeddingsfor initial testing, then rebuild with embeddings - Use a smaller embedding model (trades accuracy for speed)
- Run on a machine with GPU (sentence-transformers auto-detects CUDA)
The embedding model loads on first semantic search. This is a one-time cost per session. Use --mode keyword for fast CLI testing:
docs-mcp --db comsol.db --test "query" --mode keywordSQLite WAL mode can cause file locking issues on Windows. The server handles this automatically, but if you see errors:
- Close any other programs accessing the database
- Delete
.db-waland.db-shmfiles if present - Restart the MCP server
On native Windows (not WSL), wrap NPX commands with cmd /c:
# Instead of: npx -y some-package
cmd /c npx -y some-packageAlways use forward slashes or escaped backslashes in config files:
{
"args": ["--db", "C:/Users/name/docs-mcp/comsol.db"]
}Or use environment variables:
{
"args": ["--db", "comsol.db"]
}(Database files in %LOCALAPPDATA%\docs-mcp\ are found automatically.)