Semantic Search

MidOS uses LanceDB with Gemini embeddings (3072-dimensional vectors) to provide semantic search across 290,000+ knowledge vectors.

Search Tools Comparison

Tool	Tier	Best For
search_knowledge	Free	Quick keyword lookup
semantic_search	Dev	Conceptual similarity
hybrid_search	Dev	Best of both worlds
smart_search	Ops	Auto-detecting best mode

Search Modes

Keyword Mode

Fast grep-based search. Finds exact or partial text matches.

{
  "name": "semantic_search",
  "arguments": {
    "query": "useEffect cleanup",
    "search_mode": "keyword"
  }
}

Best for: Known terms, function names, specific patterns.

Vector Mode

Pure semantic similarity. Finds conceptually related content even with completely different wording.

{
  "name": "semantic_search",
  "arguments": {
    "query": "how to handle side effects when a component unmounts",
    "search_mode": "vector"
  }
}

Best for: Conceptual queries, “how do I…” questions, finding patterns you don’t know the name of.

Hybrid Mode (Default)

Combines keyword and vector search. Keyword results are checked first, then semantic results fill in gaps.

{
  "name": "semantic_search",
  "arguments": {
    "query": "React hooks cleanup patterns",
    "search_mode": "hybrid",
    "rerank": true
  }
}

Best for: General-purpose searching. This is the recommended default.

Filtering by Stack

Narrow results to a specific technology:

{
  "name": "semantic_search",
  "arguments": {
    "query": "authentication patterns",
    "stack": "fastapi",
    "top_k": 10
  }
}

Reranking

When rerank: true (default), results go through a secondary relevance scoring pass. This improves precision but adds slight latency.

Disable reranking when speed matters more than precision:

{
  "name": "semantic_search",
  "arguments": {
    "query": "quick lookup",
    "rerank": false
  }
}

Tips for Better Results

1. Be specific: “FastAPI dependency injection with async database sessions” beats “dependency injection”
2. Use natural language: The vector model understands questions better than keywords
3. Try different modes: If keyword gives no results, hybrid will automatically try semantic
4. Filter by stack: Reduces noise significantly when you know the technology
5. Increase top_k: Set top_k: 20 when exploring a broad topic to see more variety