🧪 Deep Research: Advanced Multi-Stage Investigation

Deep Research is an autonomous, multi-stage recursive engine designed for high-stakes technical investigations, market analysis, and exhaustive fact-finding. It identifies missing data points and performs follow-up searches to close information gaps in every investigation.

---

🏗️ The 4-Stage Lifecycle

The tool follows a deterministic state machine to ensure exhaustive coverage of the topic:

1. Search Query Expansion

The engine begins by using a high-fidelity model to expand your initial prompt into ~6 highly specific search queries.

To ensure broad coverage, the engine purposefully targets five distinct data categories:

1. Official Corporate Records: Investor Relations, Annual Reports, and SEC filings.
2. Technical Resources: Specifications, Documentation, and GitHub repositories.
3. Market Analytics: Financial analysis and industry providers.
4. Current Events: Recent press releases and news.
5. Technical Assessment: Expert critiques and whitepapers.

---

2. Search Discovery & Filtering (Phase 1)

The engine detects your configured Search Provider and routes to one of two entirely separate logic paths for URL discovery. The goal of this phase is to rapidly identify candidate links using specialized scouts.

3. Deep Page Analysis & Processing (Phase 2)

Once candidate URLs are discovered, the engine launches a unified, parallelized Chromium engine to analyze each page. Choose between Headless (Background) or Headed (Visible Window) mode to bypass hardware-level fingerprinting or manually resolve traffic challenges. This browser-based scraping allows the tool to process modern, JavaScript-heavy documentation into clean Markdown.

4. Recursive Gap Analysis (Discovery Loops)

After gathering initial findings, the engine identifies critical information gaps and performs follow-up discovery:

• Autonomous Query Generation: The engine automatically writes 3 targeted follow-up queries (Gaps) based on missing data points.
• Iteration Depth: Controlled by the max_depth setting. A depth of 3 means the engine will perform up to 3 recursive loops to find missing facts before synthesis.

---

⚡ Search Discovery: DuckDuckGo Scout

The DuckDuckGo implementation acts as a high-speed discovery scout for initial URL retrieval.

• URL Discovery Only: Targets html.duckduckgo.com (Zero-JS) to retrieve long lists of candidate links without browser overhead.
• Direct Scraping: This phase does not use a browser; it utilizes high-performance Rust-based CSS selectors.
• Massive Parallelism: Search queries are fired in a single parallel burst to populate the candidate pool instantly.

---

🕵️ Search Discovery: Google Search Scout (Stealth Mode)

The Google implementation acts as a Stealth Scout designed to navigate defensive security layers and resolve high-fidelity results.

• Chromium-Driven Search: Uses a Chromium engine even for the initial result discovery. Supported in both background and visible modes.
• Sequential Stealth Mode: Searches are performed one by one with a mandatory 1.2-second jitter delay between queries to avoid traffic flagging.
• Dynamic Redirect Resolution: Natively detects and follows Google's "Enable JS" and "Unusual Traffic" redirects by monitoring tab context switching.
• Broad Bucket Collection: Acts as a pure data harvester, capturing every link and metadata field before delegating relevance ranking to the semantic engine.

---

🚀 The Multi-Tab Browsing Engine (Unified)

Regardless of the search provider, the Deep Page Analysis phase is unified and uses a specialized Chromiumoxide engine.

• Native Challenge Recovery: Auto-detects secondary tabs spawned by JS-challenges (e.g., enablejs redirects), ensuring the scraper always binds to the active result context.
• Profile Isolation: Each research session uses a unique, process-level temporary directory (zed-research-profile-<PID>), preventing SingletonLock conflicts and ensuring private, state-free browsing for every task.
• Content Conversion: Every page is analyzed, scrolled to trigger lazy-loading, and converted from raw HTML into Technical Markdown.

---

🤖 Customizing the System Prompts

Advanced users can refine the internal logic used during each stage by overriding these prompts in their settings:

1. search_system_prompt: Controls the initial search query expansion strategy.
2. gap_analysis_system_prompt: Controls how missing information is identified and follow-up queries are written.
3. condensation_system_prompt: Controls the final report's tone, structure, and technical density.
4. use_headed_browser: Toggles between a hidden background browser and a visible window for manual audit and CAPTCHA resolution.