Frequently asked questions

Sentiqra is a deterministic, multilingual text-analysis platform. You send text to POST /v1/analyze, choose which analyzers to run, and get back labeled, signal-scored results plus a signed receipt. No generative model is used to produce classifications.

12 analyzers are generally available: Trust & Safety, Cybersecurity, Discrimination Analysis, Dating Safety, Fraud & Scam Detection, Crisis Intelligence, Disinformation Analysis, Market Intelligence, Insurance Intelligence, Emotion Analysis, Customer Service Intelligence, Wellness Intelligence. The catalog also lists 1 restricted analyzer (Information Warfare), accessible under an access agreement rather than self-serve.

Determinism. The same input against the same pinned configuration always produces the same output - replayable while the configuration remains supported under the published support policy or your contract. Every successful analysis result includes a signed verification receipt: a configuration ID, fingerprint, and signature you can use to verify a result was produced by a specific version. See the full comparison.

No. Sentiqra returns structured signals - what was detected, at what signal-match strength, under which configuration. Your policy engine, reviewers, or downstream models make the decision. [Trust & Safety specific] Labels map to policy categories, not actions; allow/restrict/remove stays with you.

GA analyzers with published evidence report per-language Top-3 benchmark results tied to a configuration ID - see benchmarks. Where per-language publication is pending, the analyzer and benchmark pages say so instead of quoting numbers. Ranges vary by analyzer and language; the benchmark pages show the honest spread per language rather than a single headline number.

The published performance target is p95 under 500ms for a single text input, including multi-analyzer requests - a target that applies to every plan, not a contractual commitment; contractual SLAs are an Enterprise entitlement. Latency is consistent because detection is non-generative signal evaluation, not token generation. Measured p50/p95/p99 figures are on the Performance Dashboard snapshot.

The detection models capture semantic meaning beyond keywords, including coded and evasive phrasing. [Disinformation specific] Manipulation techniques like deflection and urgency amplification are first-class labels. [Trust & Safety specific] Dogwhistles and evasion patterns have dedicated indicator families.

Coverage varies by analyzer - from 8 languages (Market Intelligence) to 41 (Wellness Intelligence). Every analyzer page lists its exact language set and its benchmark publication status; per-language results are published where available and marked pending otherwise. Do not assume parity across analyzers.

Yes - 5,000 analyzer runs per month, all GA analyzers, and every language each analyzer supports. No credit card required.

All generally-available analyzers are included on every plan, including Free. You pay for volume and features (custom quotas, dedicated support, governance options), not per-analyzer access. See pricing.

One analyzer run = one analyzer applied to one request. A single /v1/analyze call requesting two analyzers counts as two runs.

Free: 2 RPS. Build: 10 RPS. Growth: 25 RPS. Production: 50 RPS. Scale: 200 RPS. Enterprise: custom. Per-key rate limits and daily quotas can be configured from the dashboard.

Yes. Upgrades take effect immediately; downgrades apply at the next billing cycle. Annual billing is available with a discount on all paid plans.

One endpoint, one auth scheme, one response shape. Generate an API key, then POST /v1/analyze with {"text": ..., "language": ..., "analyzers": [...]}. See the quickstart - first call in under five minutes.

Yes - pass an array in analyzers, up to a maximum of 2 analyzers per request (each requested analyzer counts as one billed run). Multi-analyzer pipelines chain calls: a cheap primary screen first, then specialized analyzers only when it warrants. [Customer Service specific] A common pairing is Customer Service Intelligence + Emotion Analysis: intent and urgency plus the emotional register, in one call.

Sentiqra is a REST API that works with any HTTP client; examples are provided in curl, Python, JavaScript, Go, and PHP. See SDKs and client libraries for current status.

Every successful analysis result includes a signed verification receipt: a configuration ID (sqcfg_...), a fingerprint (fp_...), and a signature (sig_...). Store them - you can replay the analysis while the configuration is supported under the published support policy or your contract, and verify the receipt offline at any time via the verification endpoint.

No. Text is held in memory while the request is processed and released when the response completes. Request metadata (timestamps, key identifier, usage units) persists on the plan retention schedule; submitted text is never persisted. See data handling.

No. We do not use submitted API content to train or improve detection models - this is a commitment enforced by processing and retention controls, with evidence available on request. See no training on API inputs.

No. The API receives only the text you send - no user IDs, thread context, or metadata is required or used. [Discrimination specific] This also means the analyzer cannot assess a person's history or intent - only the submitted text.

Today: production compute is EU-hosted (Germany); transient API processing also uses US cloud infrastructure under SCCs, with submitted-content-not-retained handling as standard. An EU residency option covering submitted-content processing is in progress - see the Trust Center roadmap.

Sentiqra is a deterministic, non-generative analysis system - not a generative or general-purpose AI system. Whether it falls under a specific high-risk category depends on your use case - a hiring-screening deployment is assessed differently than a spam filter. Your legal team makes the final determination; we provide the architectural documentation to support it.

Every detection is reproducible while its configuration is supported under the published support policy or your contract: same input + same pinned configuration = same output. When a user contests an automated decision, you can show the exact labels, scores, and configuration that produced it - specific evidence rather than "the model said so."

Yes. [Trust & Safety specific] Structured signal output maps to DSA statement-of-reasons requirements - you can cite the specific indicator and configuration behind each moderation action, and reproduce it for auditors.

[Cybersecurity specific] Structured threat signals feed SOC workflows under NIS2 incident-handling obligations. [Market Intelligence specific] For DORA, reproducible classification supports the record-keeping expectations on ICT risk monitoring. Details are on the compliance architecture page.

No. Sentiqra detects language patterns; legality is a determination for your counsel and your jurisdiction. [Discrimination specific] The analyzer flags exclusionary language - it does not determine whether unlawful discrimination occurred. [Disinformation specific] It flags manipulation techniques - it does not decide which political claim is true.