Exa Semantic Search

Overview

Exa.ai provides neural semantic search optimized for AI consumption. Use when meaning matters more than keywords.

Decision Flowchart

digraph exa_decision {
    rankdir=TB;
    node [shape=box];

    start [label="Need to search the web?" shape=diamond];
    known_url [label="Do you have\na specific URL?" shape=diamond];
    semantic [label="Is this semantic/conceptual?\n(meaning > keywords)" shape=diamond];
    recent [label="Need very recent\nnews/events?" shape=diamond];
    code [label="Is this a coding/\nAPI question?" shape=diamond];

    webfetch [label="WebFetch" shape=box style=filled fillcolor=lightblue];
    websearch [label="WebSearch" shape=box style=filled fillcolor=lightgreen];
    exa_web [label="mcp__exa__web_search_exa" shape=box style=filled fillcolor=lightyellow];
    exa_code [label="mcp__exa__get_code_context_exa" shape=box style=filled fillcolor=lightyellow];

    start -> known_url [label="yes"];
    start -> known_url [label="no" style=invis];
    known_url -> webfetch [label="yes"];
    known_url -> semantic [label="no"];
    semantic -> code [label="yes"];
    semantic -> recent [label="no"];
    code -> exa_code [label="yes"];
    code -> exa_web [label="no"];
    recent -> websearch [label="yes"];
    recent -> websearch [label="no"];
}

Quick Reference

Tool Usage

mcp__exa__web_search_exa

query: "semantic query describing concepts"
numResults: 8 (default, adjust as needed)
type: "auto" | "fast" | "deep"

mcp__exa__get_code_context_exa

query: "React useState hook examples" | "Express middleware patterns"
tokensNum: 5000 (default, 1000-50000 range)

Integration Patterns

Discovery + Extraction:

1. Exa finds relevant sources semantically
2. WebFetch extracts full content from best URLs

Multi-perspective research:

1. Exa: "academic perspectives on X"
2. Exa: "industry implementation of X"
3. Exa: "critiques of X"
4. Synthesize

Fallback:

1. Try Exa for semantic search
2. If results poor, fall back to WebSearch with keywords

Anti-Patterns

When Results Are Poor

1. Switch search type: auto vs fast vs deep
2. Rephrase: more semantic/descriptive
3. Add domain filters via allowed_domains
4. Fall back to WebSearch for keyword matching