Local Rag Search

**Quick judgment**: A strong, well-documented skill for local, RAG-style web search and multi-engine deep research. It’s especially useful when you want rich web context, care about privacy, or don’t have external API keys. Best suited as a general-purpose web search layer for agents that favor semantic relevance over raw keyword results. **What it does well** - Wraps the `mcp-local-rag` server into several focused tools (`rag_search_ddgs`, `rag_search_google`, `deep_research`, and Google/DDG-only deep research shortcuts). - Uses semantic similarity ranking (RAG-like scoring) so returned results are typically more on-topic than naive keyword search. - Supports multi-engine deep research (DuckDuckGo, Google, Bing, Brave, Wikipedia, and other engines) with tunable parameters for breadth and depth. - Provides clear, concrete usage patterns and example workflows (current events, technical deep dives, multi-perspective research). **Key risks / limitations** - **Content quality & safety rely on search engines**: It can surface biased, incorrect, or harmful content from the open web; there’s no built-in filtering or safety layer described beyond general verification and cross-referencing advice. - **Recency and coverage limited by search backends**: If engines index poorly or have region-specific gaps, results will reflect that; the skill itself doesn’t fix such issues. - **Performance vs completeness tradeoff**: Deep research with many backends and high `num_results` can be slow; the docs mention this but agents must tune parameters carefully. **Recommended scenarios** - Default web search layer when you want semantic ranking and don’t want to manage API keys. - Privacy-conscious general queries (use `rag_search_ddgs` as recommended). - Technical or scientific lookup and documentation discovery (use `rag_search_google` or Google-based deep research). - In-depth, multi-angle research tasks where synthesizing across multiple engines and search terms is important. **Overall**: High usability and functionality, with solid guidance on tool choice and parameter tuning. Main concerns are standard open-web risks (misinformation, unsafe content) rather than issues intrinsic to the skill itself.