deepsearch_service_skill

Overview

This skill provides an LLM-driven deep iterative search and reasoning service that breaks down complex questions, gathers multi-source evidence, and delivers structured analytical reports. It is designed to adapt search strategy across multiple iterations to improve completeness and relevance. The output includes a comprehensive report, provenance metadata, and iteration logs for transparency.

How this skill works

The service uses an LLM to decompose the input query into focused subqueries, runs multiple retrieval passes, and aggregates results from diverse sources. After each iteration it evaluates information sufficiency and dynamically adjusts queries and depth. Final output is a cohesive, structured report with ranked sources, iteration history, and timing metrics.

When to use it

• Investigating complex technical or strategic problems requiring multi-angle analysis
• Producing comprehensive research or briefing reports with source attribution
• When initial single-pass search yields incomplete or inconsistent information
• Validating hypotheses by iteratively refining queries and evidence
• Compiling summaries from heterogeneous sources for decision support

Best practices

• Start with a clear, scoped question and provide key constraints or goals
• Adjust max_iterations and queries_per_iteration for depth vs latency trade-offs
• Specify depth_level (quick/normal/deep) to match research needs and time budget
• Review iteration logs and source relevance scores to understand how conclusions were formed
• Combine the generated report with domain expert review for high-stakes decisions

Example use cases

• Designing a high-availability microservices architecture with cross-source justification
• Creating a literature-style review on best practices for Python async programming
• Investigating regulatory impacts across jurisdictions with sourced evidence
• Generating an executive briefing that synthesizes technical, business, and risk factors
• Validating competing claims by tracing evidence across iteration history