deep-search

name	deep-search
description	Deep multi-round web research with parallel fetching. Triggers: deep search, research, 深度搜索, 调研, investigate, deep research.

Deep Search

Overview

The search tool performs deep multi-round web research. It iteratively searches from multiple angles, fetches pages in parallel, chases the most-referenced external sources, and produces a structured research report.

Usage

Call search with a query and optional depth. The tool will:

1. Round 1: Search the web using Perplexity Sonar (preferred) or other available engines
2. Rounds 2+: Generate follow-up queries from different angles (time-qualified, subtopic-focused, controversy/analysis)
3. Parallel fetch: Crawl all discovered URLs concurrently (8 connections)
4. Reference chasing: Extract outbound links from crawled pages, fetch the most-cited external sources
5. Report: Build a structured report with overview, source previews, and search query log
6. Save: Everything saved under ./research/<query-slug>/

Parameters

• query (required, string): The research topic or question to investigate.
• depth (optional, integer, default: 2): Research depth:
  • 1 = Quick: single search round + crawl (~1 min, up to 10 pages)
  • 2 = Standard: 3 search rounds + reference chasing (~3 min, up to 30 pages)
  • 3 = Thorough: 5 search rounds + aggressive link chasing (~5 min, up to 50 pages)
• max_results (optional, integer, default: 8): Number of search results per round (1-10).
• search_engine (optional, string): Preferred search engine. Options: perplexity, duckduckgo, brave, you. Defaults to auto-detection (prefers Perplexity).

Example

{
  "query": "AI regulations worldwide 2026",
  "depth": 2
}

Output

Returns a structured research report including:

• Overview (initial search answer)
• Source details with inline previews (first 2000 chars of each page)
• List of all search queries used
• Summary with page count and save location

Use read_file on specific source files for full content when you need detailed synthesis.

Saved Files

Results are saved to ./research/<query-slug>/:

• _<query-slug>.md -- structured research report (topic-named; leading _ keeps it out of read_sources' source-ingestion path on subsequent synthesis runs; see issue #897)
• _search_results.md -- combined raw search results from all rounds
• 01_<domain>.md -- full page content from first source
• 02_<domain>.md -- full page content from second source
• etc.

Environment Variables

• PERPLEXITY_API_KEY -- enables Perplexity Sonar (recommended, best for deep research)
• BRAVE_API_KEY -- enables Brave Search
• YDC_API_KEY -- enables You.com search

Without API keys, DuckDuckGo HTML search is used as fallback.