Skip to content
aptrouter

Search API

POST /v1/search returns ranked, verbatim-cited results for a query. This guide covers what the API guarantees and how to use it economically; the machine-readable contract is the generated API reference.

The verbatim guarantee

Section titled “The verbatim guarantee”

There is no LLM anywhere in the pipeline. Every cited sentence (q) comes verbatim from a fetched source, and every ranking, fusion, dedup, and compression step is deterministic. The same evidence in produces the same results out — nothing is paraphrased, summarised, or hallucinated. Your model (or your reader) does the reasoning over quotable, attributable evidence.

That makes the API a fit where LLM-synthesised search is a dealbreaker: legal and compliance research, medical queries, academic citation, fact-checking, and competitive intel with auditable provenance.

Retrieval lanes

Section titled “Retrieval lanes”

The engine behind the API (aptrouter’s research-mcp) fuses several independent retrieval lanes:

  • Web — Brave Search, an independent ~30B-page index.
  • Academic — OpenAlex, Europe PMC, arXiv, and Semantic Scholar, resolved to open-access full text (OA PDFs and JATS XML) rather than paywalled landing pages.
  • Reference & news — Wikipedia REST and a news RSS lane for recency.
  • Hosted corpus — a scholarly index serving DOI-backed results with dense
    • full-text retrieval.

Lanes are merged with Reciprocal Rank Fusion so each lane’s best results interleave, and a continuous time-decay prior favours fresh sources on recency-relevant queries while leaving evergreen queries unbiased.

Quality & honesty signals

Section titled “Quality & honesty signals”

Ranking is driven by deterministic quality signals, not classifiers:

  • Domain tiers (authoritative / reputable / mid / low / unknown) and Oxford-CEBM evidence levels (systematic review … expert opinion).
  • Cross-source clustering — claims are clustered across sources and labelled Consensus / Corroborated / Singular / Conflict.
  • Spam resistance — listicle/SEO/affiliate pages are demoted by structural signals (no AI-text classifiers, which false-positive on legitimate non-native-English writing).
  • Honest failure — the engine reports a composite confidence score (corroboration + tier + evidence level + diversity − conflict) and a degraded list naming blocked, failed, or empty sources rather than papering over gaps. The hosted /v1/search envelope stays minimal by design (n + results); these signals shape what reaches your results, and the full envelope (confidence, degraded, per-source metadata) is exposed on the engine’s MCP surface.

Token budgets

Section titled “Token budgets”

Long agent runs live or die on token economy, so the API is token-lean by default:

  • Abbreviated keys — d, t, y, a, s, q instead of verbose field names (see the quickstart for the full key table).
  • Opt-in weight — abstracts are omitted unless you ask ("fields": "abstract" adds ab per result).
  • Bounded result counts — limit is clamped to 1–50 (default 8).
  • X-Token-Budget — every response carries an estimated token count (~4 chars/token) so a framework can budget without parsing the body.

Pricing & billing

Section titled “Pricing & billing”

Billing is prepaid, pay-as-you-go — there are no plans, tiers, quotas, or free allowance. You top up a credit balance in the dashboard and each call draws it down:

  • $0.005 per search ($5 per 1,000) — one successful POST /v1/search debits 1 credit. GET /v1/usage and GET /health are free.
  • No charge for failures — server errors (500) and timeouts (504) automatically refund the credit.
  • Empty balance returns 402 — top up to keep serving; nothing throttles you while you have a balance.

Request shape

Section titled “Request shape”
{
  "query": "quantum error correction surface codes",
  "limit": 5,
  "fields": "abstract"
}

query is required (1–2048 characters) and is tokenised full-text + dense retrieval — plain language works best; it is not a boolean DSL.

Every /v1/* request must carry your API key as Authorization: Bearer sk_live_… (issue and rotate keys in the dashboard). A missing or invalid key returns 401.

Reference

Section titled “Reference”