vaaya / docs / reference

Web search

Five search vendors behind one endpoint shape, from 1¢ per query: Exa (semantic, the default), Brave (independent index), Tavily (RAG-tuned), Linkup (cited answers), and Parallel (second source + deep async research). This page is the per-vendor reference; for one call that orchestrates all of them, see SuperSearch.

When to use which

Routing rules that hold up under measurement: start with Exa, always. For time-critical or important queries, fire Exa and Brave concurrently and take the first (or both) — they have different indexes and different cache profiles (Brave is the fastest engine when cold). Use Linkup when you want a cited answer rather than links, and reserve deep research tiers for questions a single search genuinely can’t answer.

exa/search

Price

1¢

Latency

fast (~0.1s warm)

Results

up to 100 per call

Agent-native semantic search — the default “find me pages about X” tool. Semantic recall plus date/domain filters plus inline page content in the same call.

curl -X POST https://vaaya.ai/api/run/exa/search \
  -H "Authorization: Bearer $VAAYA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "latest developments in WebTransport API",
    "type": "auto",
    "numResults": 10,
    "contents": { "text": true },
    "start_published_date": "2026-01-01"
  }'

Optional filters: start_published_date / end_published_date (ISO dates), include_domains / exclude_domains, category (news, research, github, people, company). type accepts auto, neural, keyword, deep-lite, deep, deep-reasoning.

Finding people by ICP

Set category: "people" and phrase an open natural-language query — role + seniority + industry + company size + geography, e.g. "VP Sales at fintech companies with 21-100 employees in India — LinkedIn profiles". This is the primary people-discovery path for GTM work.

Gotchas

• Without a date filter, neural search returns stale results for “recent” queries — always add start_published_date for recent-event queries.
• The people and company categories don’t support date or exclude_domains filters.
• numResults goes up to 100 — use a large value for broad people/lead sweeps instead of paging.
• Already have the URLs? exa/contents (0.1¢/URL per content field) is the cheap extraction path — see Web scraping.

brave/search and brave/news

Price

1¢

Latency

fast — the fastest engine when cold

Index

Brave’s own 20B+ page crawler

The corroboration source outside the semantic engines and the Google ecosystem. Snippets only — chain a scrape for full text.

curl -X POST https://vaaya.ai/api/run/brave/search \
  -H "Authorization: Bearer $VAAYA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"q": "on-device llm inference benchmarks", "count": 10, "freshness": "pm"}'

• count (1–20), offset, country, search_lang.
• freshness — pd/pw/pm/py or an explicit YYYY-MM-DDtoYYYY-MM-DD range.
• result_filter — comma list of web,news,videos,faq,discussions.
• extra_snippets: true — up to 5 extra excerpts per hit.
• brave/news is the news-only vertical with age and breaking flags; count up to 50.

tavily/search

Price

1¢

Latency

~1–2s

RAG-tuned results with per-result relevance scores, built for feeding LLMs. include_answer: true returns a synthesized answer with sources; include_raw_content: "markdown" inlines full page text per hit.

curl -X POST https://vaaya.ai/api/run/tavily/search \
  -H "Authorization: Bearer $VAAYA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query": "vector db comparison for RAG", "max_results": 10, "include_answer": true}'

• Also takes topic (general|news|finance), include_domains/exclude_domains, time_range (day|week|month|year), country.
• include_raw_content makes responses big — prefer it over a chained scrape only when you’ll read most of the hits.
• Measured miss pattern: results favor blog/tutorial pages over official docs domains — use Exa or Brave for authoritative-docs lookups.

linkup/search and linkup/deep-search

Price

1¢ (search) · 5¢ (deep-search)

Latency

~1.5s · tens of seconds

Linkup searches and synthesizes in one call. outputType picks the shape: sourcedAnswer (default — answer plus cited sources), searchResults (raw ranked hits), or structured (pass structuredOutputSchema as a JSON-schema string).

curl -X POST https://vaaya.ai/api/run/linkup/search \
  -H "Authorization: Bearer $VAAYA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"q": "what changed in the EU AI Act final text", "outputType": "sourcedAnswer"}'

• Optional fromDate/toDate (ISO), includeDomains/excludeDomains, includeImages, includeInlineCitations.
• linkup/deep-search (5¢) runs iterative agentic retrieval for multi-hop questions. Escalate here only after one-pass search comes back thin; for a full research report the bigger rung is parallel/task.

parallel/search and parallel/extract

Price

1¢ (search) · 1¢ per URL (extract)

Latency

fast

Parallel.ai’s AI-powered search — use it as the second source in a multi-vendor pipeline (search Exa + Parallel concurrently, dedupe). parallel/extract pulls clean content from URLs with an optional objective to focus the extraction; for cheap bulk extraction, exa/contents at 0.1¢/URL wins — see Web scraping.

parallel/task — deep async research

Price

10¢ (pro) · 30¢ (ultra)

Latency

slow — seconds to minutes, async

Polling

parallel/task-status — free

Submit a complex research question; Parallel runs an LLM-driven search-and-synthesize loop. The call returns { run_id } immediately — poll parallel/task-status (free) until status is completed.

# 1) start the run
curl -X POST https://vaaya.ai/api/run/parallel/task \
  -H "Authorization: Bearer $VAAYA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input": "Compare the top 5 headless CMS platforms for Next.js in 2026", "processor": "pro"}'

# 2) poll until done (free)
curl -X POST https://vaaya.ai/api/run/parallel/task-status \
  -H "Authorization: Bearer $VAAYA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"run_id": "run_..."}'

When not to use it

• Simple “find me information about X” — exa/search is 10–30× cheaper.
• When you need the result fast — runs are multi-second, sometimes multi-minute.
• Tight budgets — an exa/search → parallel/extract chain often reaches similar quality for a few cents.

Measured performance

From a 30-task, 6-engine benchmark (August 2026, steady state):