Exa Neural Search

Exa is the search backend for web_search in Hermes. Firecrawl handles page extraction via web_extract and web_crawl.

When Exa Shines

Exa's neural search is better than keyword search when:

• You need semantic understanding of a query, not just keyword matching
• Searching for research papers, arXiv content, GitHub repos
• Need date-filtered results (last week, this year, date ranges)
• Want results only from specific domains
• Looking for specific content categories (news articles, company info, financial reports)

Advanced Search Parameters

All parameters are available through the standard web_search tool. Unrecognized params are silently ignored by non-Exa backends.

Date Filtering

# Only results from 2024 onwards
start_published_date: "2024-01-01"

# Specific date range
start_published_date: "2024-06-01"
end_published_date: "2024-12-31"

# Recent news (combine with category)
start_published_date: "2025-04-01"
category: "news"

Date params accept ISO date strings (YYYY-MM-DD). Both start and end are optional.

Domain Targeting

# Only academic sources
include_domains: ["arxiv.org", "scholar.google.com", "academic.oup.com"]

# Block low-quality sources
exclude_domains: ["wikipedia.org", "medium.com"]

# GitHub-only search
include_domains: ["github.com"]

Content Categories

The category param scopes results to a specific content type:

• "research paper" — academic papers, preprints
• "news" — news articles
• "pdf" — PDF documents
• "github" — GitHub repositories and discussions
• "company" — company websites and profiles
• "tweet" — tweets (X/Twitter)
• "personal site" — personal websites, blogs
• "linkedin profile" — LinkedIn pages
• "financial report" — financial documents, SEC filings

Effective Query Patterns

Research literature search:

web_search(query="CRISPR delivery nanoparticles lipid", 
           category="research paper", 
           start_published_date="2024-01-01",
           limit=10)

Competitive analysis:

web_search(query="AI coding assistant CLI tools",
           category="company",
           limit=10)

GitHub projects:

web_search(query="websocket proxy relay",
           include_domains=["github.com"],
           limit=5)

Recent news:

web_search(query="LLM pricing changes",
           category="news",
           start_published_date="2025-04-01",
           limit=5)

Backend Architecture

• web.search_backend: exa — all web_search calls go through Exa
• web.extract_backend: firecrawl — web_extract and web_crawl use Firecrawl

This split gives us neural search quality + Firecrawl's superior JS rendering and crawling.