Verify citations programmatically. One endpoint, streamed over SSE.

Base URL

https://api.citetrue.com

Quickstart

  1. Go to Dashboard → API Keys and click New API Key. Copy the key starting with sk_. It is shown once.
  2. Include the key in Authorization: Bearer sk_… on every request.
  3. POST a task to /verify/v2. The response is a text/event-stream; consume SSE events until you see task_completed.

Minimal call:

curl -N -X POST https://api.citetrue.com/verify/v2 \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"text":"[1] Vaswani, A., Shazeer, N., Parmar (2017). Attention is all you need."}'

-N disables curl buffering so SSE events stream line by line. There is no depth to choose — every reference goes through the same flow. Most cost 1 credit; tricky ones that need more data sources and more analysis can cost up to 5.

Client examples

SSE is a long-lived HTTP response where each event ends with a blank line (\n\n). Skip lines that don't start with data: and JSON-parse the rest.

Browser note: the built-in EventSource can't send an Authorization header — use fetch + ReadableStream as shown below.

Node 18+ with native fetch.

const API = 'https://api.citetrue.com'
const KEY = process.env.API_KEY

async function verify(text) {
  const resp = await fetch(`${API}/verify/v2`, {
    method: 'POST',
    headers: {
      'Authorization': `Bearer ${KEY}`,
      'Content-Type': 'application/json',
    },
    body: JSON.stringify({ text }),
  })
  if (!resp.ok) throw new Error(`${resp.status}: ${await resp.text()}`)

  const reader = resp.body.getReader()
  const decoder = new TextDecoder()
  const refs = []
  let buf = ''

  while (true) {
    const { value, done } = await reader.read()
    if (done) break
    buf += decoder.decode(value, { stream: true })
    let idx
    while ((idx = buf.indexOf('\n\n')) !== -1) {
      const chunk = buf.slice(0, idx); buf = buf.slice(idx + 2)
      for (const line of chunk.split('\n')) {
        if (!line.startsWith('data: ')) continue
        const evt = JSON.parse(line.slice(6))
        if (evt.event === 'ref_completed') {
          refs.push(evt.data.ref)
        }
        if (evt.event === 'task_completed') return { refs, balance: evt.data.balance }
        if (evt.event === 'error') throw new Error(evt.error)
      }
    }
  }
}

const out = await verify('[1] Vaswani, A., Shazeer, N., Parmar (2017). Attention is all you need.')
console.log(out)

Authentication

API keys are long-lived bearer tokens in the shape sk_<40-char-token>. Send them on the Authorization header:

Authorization: Bearer sk_AbCd123…
  • Revoke a key from the Dashboard. Revocation is effective within a few seconds.
  • A key can perform anything the owning user can. Treat it like a password — rotate and scope per integration.

Rate limits

Each API key gets a token-bucket limiter:

  • Burst: 10 requests.
  • Sustained: 60 requests / minute (bucket refills at 1 token/s).
  • When exceeded, the request fails with 429 and a Retry-After header (seconds).

The limiter runs per key; different keys on the same account have independent buckets. Credits still apply on top of this.

Request headers

Every endpoint accepts the same authentication and content-type headers; only the request body and URL differ.

HeaderRequiredValue
AuthorizationyesBearer sk_…
Content-Typeyesapplication/json

Responses are 200 OK with Content-Type: text/event-stream.

data: {"event":"<name>","data":{...},"error":""}

POST /verify/v2

Split a text blob into references and verify each. Supply exactly one of text (new task) or taskHash (resume). Every reference goes through the same flow — there is no verification level to choose.

Request body

FieldTypeNotes
textstringNew verify task. Max 10MB. Accepts numbered, bulleted, blank-line-separated, BibTeX, or in-text prose with (Author, Year) citations.
taskHashstringResume an existing task — replays cached state or attaches to live stream. References already charged aren't charged again; any that still need work are finished and billed as usual.
forceboolBypass the task-level dedup cache (forces a fresh run, billed like a new check).
localestringOptional 2-letter language code ("zh", "ja", "de", …). When set, AI translates the per-ref note into that language and surfaces it under noteTranslated[locale]. Default "" / "en" = no translation.

Idempotency: same text + same user returns the cached task on second POST (no re-billing) unless force: true.

Deprecated fields

  • depth — still accepted but ignored (1, 5, 20 or omitted all behave the same). Don't send it.
  • refHash + parentTaskHash — legacy single-reference mode kept for older clients. It ignores depth and simply finishes that one reference, billed under the normal rules. New integrations shouldn't use it.

Example request (text)

{
  "text": "[1] Vaswani, A., Shazeer, N., Parmar (2017). Attention is all you need.\n[2] Piketty, T. (2016). Capital in the twenty-first century. Harvard University Press."
}

Example request (resume)

{
  "taskHash": "f0e1d2c3b4a5..."
}

Typical event sequence

task_created → refs_created (N refs) →
  [ref_processing + ref_completed] × N (parallel) +
  task_progress × several →
task_completed

Per-ref assessment

Each ref_completed event carries an assessment string from this closed set:

  • authentic — paper exists and is the one cited. value contains the matched paper.
  • unsure — a close candidate was found but couldn't be confirmed; value is that candidate and confidence how likely it is the cited work.
  • inauthentic — definitively not found / fabricated / unrelated. value, if present, is only the closest candidate — not a match.
  • invalid — input wasn't a citation; pipeline refused to search before any datasource hit.
  • error — verification failed on this single ref; sibling refs may still succeed. data.errors[] lists tokens (datasource_unavailable, timeout, rate_limited, parse_failed).
  • exceeded — account ran out of credits before this ref could run.

Notices on a result (year mismatch, author mismatch, …) are orthogonal to assessment — even an authentic verdict can carry notices worth showing. data.cost is the total credits this task has charged for the reference (0–5): 1 for most references, up to 5 for tricky ones that needed more data sources and more analysis, 0 if it was already paid for in another task. It covers every connection and resume of the same task, and the same number is sent on every push, replays included — overwrite it, don't add it up across reconnects. data.depth is kept only for compatibility with older clients — ignore it.

Notice values

notices[] on a ref is a set of soft-fail flags. For programmatic checks, match on these rather than on the English note.

ValueMeaning
doi_incorrectThe DOI written in the citation doesn't exist, or it belongs to a different work. Can come with any assessment: on authentic the work is real but the citation carries the wrong DOI; on unsure / inauthentic it is strong evidence that the reference was made up.
yearThe matched record's year differs from the cited year by two years or more. One-year gaps (online-first vs. print), books (editions) and web pages (page dates) are not flagged.
linkOnly a citation record was matched, with no link to the work itself — other papers cite it, but no direct source was found.
authorsIn-text citations: the cited authors don't fully line up with the matched work.
titleIn-text citations: what the sentence describes doesn't fully line up with the matched work.
likely, url_inaccessible, url_inconsistent, url_mismatch, url_unsure, url_incorrectOnly on results of older checks that are replayed from cache; current checks don't produce them.

Notices never change the assessment. New values may appear as more mismatch classes are detected — treat unknown tokens as a generic warning.

SSE events

SSE event names emitted during verification. Listen for task_completed as the terminal success event; error for terminal failure.

  • task_created — task accepted, hash assigned.
  • refs_created — text was split into N refs.
  • split_completed — splitting phase finished.
  • ref_processing / ref_status — per-ref progress signal.
  • ref_completed — per-ref final assessment.
  • task_completed — all refs done; carries cost (credits newly charged during this connection, 0 for a pure replay) and the account balance afterwards (can be negative — see Credits & billing).
  • task_progress — periodic progress %.
  • warn — non-fatal warning.
  • error — terminal task-level failure.

GET /credits

GET /credits — returns the current credit balance (totalBalance) and the active credit records. Auth required. The balance can be negative (see Credits & billing).

Credits & billing

Most references cost 1 credit. Tricky ones that need more data sources and more analysis can cost up to 5. Each reference is charged 1 credit when it's queued; anything extra is added when it finishes. Cached tasks (same text + user) do not re-bill.

Reading the charge: data.cost in ref_completed is the running total this task has charged for that reference — the same number is sent again on reconnects and replays, so overwrite it rather than adding it up. cost in task_completed is what this connection newly charged (0 for a pure replay), so after a reconnect it won't equal the sum of the per-ref costs. balance is the account balance afterwards.

Running out: a reference that hasn't started when the balance can't cover it comes back as exceeded, and the task still ends with task_completed. If you run out of credits mid-check, references already in progress still finish; the overage is deducted at your next credit reset, so balance can be negative.

Error reference

Standard HTTP status codes apply (400 malformed body, 401 missing / invalid / revoked key, 405 wrong method, 429 rate-limited, 5xx upstream). Running out of credits is not an HTTP error: references that can't start come back as exceeded. SSE-level failures arrive as a final error event.

Versioning & stability

Path-versioned (/verify/v2 etc.). Breaking changes bump the version segment; backwards-compatible additions ship in-place.

Maintained by the CiteTrue team · Last updated: