Deep Research: Multi-Step AI Research Reports with Sources

How Deep Research Works

Deep Research takes a high-level question and breaks it into sub-questions, researches each one using web search and AI models, cross-references findings, and produces a structured report. The process takes 2-5 minutes but produces reports that would take a human researcher hours.

The heart of the system is a planning loop. Before any searching happens, the research agent decomposes your prompt into a tree of sub-questions — for a query like "how mature is WebAssembly for production backends," it might spin up separate threads for runtime performance, security sandboxing, tooling support, and real-world case studies. Each sub-question becomes its own mini-investigation, and the agent decides how deep to go based on what it finds. This planning step is what separates a deep-research report from the single-shot answer you'd get from a web search agent, which retrieves and summarizes but does not re-plan as evidence accumulates.

Planning, Searching & Synthesizing

Once the plan exists, the agent searches broadly — often issuing dozens of queries per sub-question and pulling in academic papers, official documentation, changelogs, forum threads, and news. It reads far more than it will ever cite, discarding low-signal pages and following promising leads deeper. As findings come in, the agent cross-references them: where three independent sources agree, confidence rises; where they conflict, it flags the disagreement rather than silently picking a side.

Synthesis is the final and hardest step. Reading many sources is cheap; weaving them into a coherent, non-repetitive narrative that answers your actual question is where the AI earns its keep. Vincony routes this stage to models chosen for long-context reasoning via its Smart Model Router, then hands the assembled evidence to a synthesis pass that produces an executive summary, structured findings, and a full citation list. If you need to merge reports across several runs or model perspectives, the dedicated research synthesizer handles the cross-report merge.

Generate a Deep Research Report

import vincony

client = vincony.Client(api_key="YOUR_KEY")

report = client.research.deep(
    question="What are the security implications of WebAssembly in production?",
    depth="comprehensive",  # "quick", "standard", "comprehensive"
    max_sources=30,
    require_citations=True,   # every claim must trace to a source
    output_format="markdown"
)

# Deep research runs async
result = report.wait()
print(f"Report: {len(result.sections)} sections")
print(f"Sources: {len(result.citations)} citations")

# Inspect the cited findings before trusting them
for finding in result.findings:
    print(f"[{finding.confidence}] {finding.claim}")
    for c in finding.citations:
        print(f"   -> {c.title} ({c.url})")

result.save("wasm-security-report.md")

Report Quality

• • Multi-step methodology: Automatically generates and pursues follow-up questions
• • Source diversity: Academic papers, documentation, blog posts, news
• • Confidence ratings: Each finding scored by source reliability
• • Structured output: Executive summary, findings, recommendations, sources

When Deep Research Beats a Quick Answer

Not every question deserves a five-minute investigation. If you need a fact, a definition, or a single current data point, a fast search agent call is faster and cheaper. Deep Research pays off when the answer is contested, spread across many sources, or requires you to reconcile trade-offs — comparing frameworks, evaluating a vendor, scoping a migration, or writing a literature-style overview of a fast-moving topic.

The rule of thumb: reach for deep research when a wrong or shallow answer would be expensive. A one-off lookup that's slightly stale costs you nothing; a strategy memo built on an unverified claim costs you a project. Because Vincony runs every model behind one key, you can escalate from a quick search to a comprehensive report without switching providers or juggling API accounts — the same client call, a different depth setting.

Verifying & Fact-Checking Findings

A cited report is only as trustworthy as its sources, and AI research agents can still misread a page or over-generalize from a single blog post. That's why every finding ships with its supporting citations and a confidence rating — so you can spot-check the claims that matter most before acting on them. Treat the report as a well-organized starting point, not a verdict.

For claims that carry real risk, pipe the report's key findings through a second-pass verification step. Vincony's Fact Checker (see the fact-checker guide) re-checks individual assertions against fresh sources and returns a verdict with evidence, which catches the cases where a single unreliable citation slipped into the synthesis. Running research and verification as separate stages — generation then adversarial checking — is far more reliable than trusting one model to police its own output.

Limits to Keep in Mind

• • Time & cost: A comprehensive run takes minutes and burns 20-60 credits — not something to fire off casually in a tight loop
• • Source quality ceiling: The report can only be as good as what's on the open web; behind-paywall or unpublished data won't appear
• • Recency gaps: Very recent events may be under-covered until multiple sources catch up
• • Synthesis bias: The narrative reflects the balance of sources found, so a lopsided web can produce a lopsided summary

Pricing

Deep Research costs 20-60 credits per report depending on depth. Available on Power ($54.99/mo) and Business ($199/mo) plans.

FAQ

How is deep research different from a normal search agent? A search agent answers a single query in one pass — retrieve, summarize, cite. Deep research plans a tree of sub-questions, runs many searches, follows leads, and synthesizes everything into a structured multi-section report. Use the search agent for a fast fact; use deep research for a decision.

Can I trust the citations without checking them? Trust but verify. Every claim traces to a source with a confidence score, which makes spot-checking easy — but for high-stakes findings, run them through the Fact Checker as a separate verification pass.

Do I need multiple API keys for the models it uses? No. Deep Research runs across 800+ models through Vincony's single unified key, so the planning, searching, and synthesis stages each use the best-fit model without you managing separate provider accounts. See the Developer API docs or start a free account to try it.