Skip to module content
๐Ÿง AI Academy
0/91
Sign in
Module 09 ยท ~8 min

The Research System

Deep research, search APIs, and the doctrine that beats stuffing.

Reading progress
0/8 ยท 0%

The big idea

๐Ÿ’กKey idea
Search APIs give you the territory โ€” the raw web content that exists. But synthesizing that territory into a coherent, citable answer belongs to your model, where you control the reasoning and the citation behavior. For anything client-deliverable, prefer raw-retrieval APIs (Exa, Tavily, Firecrawl, Brave, Parallel, Linkup) and do the synthesis yourself. Use synthesized-answer products like Perplexity Sonar for your own internal speed, where you don't need to stand behind every citation. As the web fills with AI-generated content, extraction quality โ€” getting clean, complete, de-noised text out of a page โ€” is increasingly the bottleneck, not search ranking. That's the argument for Firecrawl-class tooling in every research pipeline.
Quick check
1 question ยท instant feedback
0/1
  1. For client-deliverable research, prefer:

Numbers that matter

130kโ˜…
Firecrawl open-source core โ€” the extraction leader in the 2026 stack.
~670ms
Brave Search API latency โ€” the fast/cheap breadth option.
9
Parallel Task API tiers ($0.12โ€“$2.40/request) for multi-hop research budgets.
Quick check
1 question ยท instant feedback
0/1
  1. The doctrine 'retrieve wide, re-rank hard' means:

Deep dive

3/3 open

Each API has a distinct job to be hired for.

Exa is best for semantic research โ€” its 'Find Similar' and Websets features make it the go-to for finding conceptually related content, not just keyword matches. Tavily is agent-focused and works well in tool-calling loops. Firecrawl leads on extraction quality: it can search, scrape, interact with dynamic pages, and crawl entire sites; the open-source core has 130k GitHub stars and integrates natively with Lovable.

Perplexity Sonar combines search and synthesis with inline citations, priced on both tokens and a flat $5 per 1,000 searches. Parallel handles multi-hop research (questions that require chaining several searches) across 9 cost tiers from $0.12 to $2.40 per request, and its Monitor API handles continuous tracking. Linkup ranks #1 on the SimpleQA factuality benchmark, making it the right choice for accuracy-critical work in finance, legal, or health-adjacent domains. Brave offers an independent index with strong privacy defaults, roughly 670ms latency, and 2,000 free requests per month.

The rule is simple: raw-retrieval APIs plus your model equals client-deliverable research. Synthesized-answer APIs are for your own internal speed.

The reason matters: when you control the synthesis step, you control the reasoning transparency, the hedge language, and the completeness of citations. When a black-box API does the synthesis, you can't guarantee any of those things โ€” and you can't explain them to a client who asks.

Owning the synthesis is a professional posture, not just a technical preference.

Three offers map cleanly onto this module's capabilities.

The pre-meeting brief skill produces two pages in under five minutes: company profile, recent news, tech stack, key people, likely pain points, suggested angle. Fully automatable.

The competitor delta monitor runs a weekly crawl and diff, surfacing only what changed โ€” pricing, messaging, job postings, product updates. Clients pay for the signal, not the crawl.

The market map sprint uses multi-agent orchestration with a citation-verification pass to produce a landscape document that used to cost โ‚ฌ15k from an analyst firm and take three weeks. It's now a repeatable, scalable deliverable.

Quick check
1 question ยท instant feedback
0/1
  1. Firecrawl's position is:

How to run it

  1. Retrieve wide, re-rank hard
    High recall + precise top-5 is where RAG pipelines start or stop working.
  2. Own the synthesis
    Raw retrieval + your model for client work.
  3. Extraction over ranking
    Firecrawl-class extraction beats raw search as the web fills with slop.
  4. Subagent-per-section
    Multi-agent orchestration for market maps and briefs.
Quick check
1 question ยท instant feedback
0/1
  1. Linkup's use case is:

In the field

๐Ÿ”ฌWorked example
Pre-meeting brief skill: domain in โ†’ Exa semantic search + Firecrawl clean extraction โ†’ subagent per section (company profile, recent news, tech stack, key people, likely pain, suggested angle) โ†’ orchestrator synthesizes โ†’ 2 pages, 5 minutes. Multi-agent research: orchestrator + subagents on the Anthropic 'How We Built Our Multi-Agent Research System' pattern.
๐ŸšซWhen not to reach for it
Don't use synthesized-answer APIs (Perplexity Sonar) for client-deliverable research โ€” you lose control of reasoning and citation. Use them for personal speed; return raw retrieval + your own synthesis to the client.
Quick check
1 question ยท instant feedback
0/1
  1. The productized market-map sprint replaces:

Pitfalls & takeaways

Failure modes

  • One search API for everything โ€” different jobs need different retrievers.
  • Naive dump of 50 chunks into prompt โ€” the fix is retrieve-wide + re-rank-hard.
  • Client-deliverable research using synthesized-answer APIs (loses citation control).
  • No extraction step โ€” raw search results are increasingly polluted by AI slop.

Durable takeaways

  • Different jobs, different retrievers โ€” no one-size-fits-all.
  • Your model owns synthesis for anything client-deliverable.
  • Extraction quality beats search ranking as the web fills with slop.
  • Multi-agent orchestration is the design pattern for market maps.

Do the work

๐Ÿ‹๏ธProve you learned it

Build the pre-meeting brief skill. Wire Firecrawl for extraction + Exa for semantic search. Use a subagent per brief section. Bench: 5 real prospect domains, human graders on accuracy + usefulness, must beat manual-research baseline on time and match on quality.

0 chars
๐Ÿ“ฆArtifact to produce
Pre-meeting brief skill (domain in โ†’ 2-page brief out in <5 min) + a competitor-delta monitor.

Sources

  • ยท GMS Field Manual ยง9 (Research System)
  • ยท Anthropic 'How We Built Our Multi-Agent Research System'