Deniable FAQ
Learn how synthetic test data, AI training datasets, RAG evaluation cases, EML files, API credits, licensing, and free sample access work.
01
Start with the difference between prepared synthetic cases and tidy placeholder fixtures.
Synthetic test data is deliberately created material that behaves like the input a system expects without copying a real customer record. Deniable organises pre-generated documents around a clear testing job, such as parsing, retrieval, extraction, classification, or workflow routing.
Simple placeholders rarely include the missing context, conflicting details, format changes, or uneven language that cause failures. Synthetic documents let you test those behaviours repeatedly while keeping real production records out of the test loop.
02
Use synthetic material to reduce exposure without treating it as a replacement for every lawful production validation step.
Synthetic data can reduce privacy risk because it is not copied from customer records, but safety still depends on how it is generated, stored, shared, and used. Deniable documents are synthetic; keep real personal data out of uploads and protect downloaded files and API keys.
No. It is a controlled layer for development, regression, QA, and evaluation. Teams should add lawful, privacy-aware production-like validation when the final risk assessment requires it.
03
Choose document categories and labels for NLP training, evaluation, and model adaptation.
AI training datasets are collections of examples used to teach or evaluate a model. Deniable packs can support tasks such as sentiment, intent, entity extraction, text classification, clustering, chatbot behaviour, and fine-tuning where the product licence permits it.
Yes. The Deniable licence allows internal development, evaluation, model training, and fine-tuning for purchased material, subject to the product licence. Resale, public hosting, and redistribution are not permitted.
04
Evaluate retrieval and generation with difficult, traceable document inputs.
Use documents with similar distractors, long context, conflicting details, missing evidence, and questions that require combining sections. Track retrieved passages, citations, groundedness, refusal behaviour, and answer quality separately.
RAG evaluation datasets pair documents and questions with a known testing purpose. They help teams measure whether a system finds the right evidence and responds appropriately when evidence is incomplete or contradictory.
05
Test headers, threading, encoding, body parts, and unusual messages without using a real mailbox.
Use EML files that vary headers, message relationships, dates, character sets, multipart bodies, quoted-printable sections, and incomplete context. Compare parsed fields with expected outcomes and keep the raw input with the test result.
EML can be returned as raw text through the document API. PDF and other binary files are intentionally excluded from the JSON document endpoint; a future binary delivery path would use a separate authenticated download flow.
06
Turn the difficult variations that clean fixtures miss into repeatable regression cases.
Useful families include reordered sections, alternate labels, synonyms, missing optional fields, date and timezone differences, incomplete threads, conflicting values, mixed languages, unusual lengths, and nested or repeated structures.
A case with one clear testing job is easier to debug and maintain than random corruption. Record the expected behaviour, category, format, language, and scenario so a failure can be reproduced and explained.
07
Test context, intent, interruptions, handoffs, and language variation across multi-turn conversations.
A useful conversational set includes incomplete requests, corrections, slang, typos, topic changes, multiple intents, escalation points, and different speaker turns. It should test the workflow around the model as well as the model response.
Yes. Synthetic support tickets and chat logs can provide controlled examples for intent, priority, routing, sentiment, and response evaluation without exposing real customer conversations.
08
Understand prepaid retrieval before you create a key or make a request.
An API credit is a prepaid unit used for an authorised document delivery. One standard case costs 1 credit; long-form documents and ebooks cost 10 credits. Credits are purchased once and there is no automatic renewal.
Choose a top-up amount of at least €5, add credits through Stripe, and use them through authenticated API requests. The dashboard records top-ups and deliveries, and requests stop when the balance is insufficient.
No. Deniable does not use monthly subscriptions or automatic renewals. You choose when to add more prepaid balance.
09
Choose between fixed downloadable packs and authenticated structured retrieval.
Available formats are listed on each catalogue release. The API can deliver JSON, TXT, EML, CSV, and SQL text cases; PDFs and images remain available inside paid ZIP downloads. Product pages remain the source of truth for the exact formats and languages in a pack.
Choose a fixed pack when your team wants a ready-to-download collection. Choose the API when a server-side workflow needs metered JSON retrieval, delivery history, filters, or unseen-document selection.
10
Use purchased material internally while keeping the catalogue protected from redistribution.
Yes, internal commercial development, QA, evaluation, training, and fine-tuning are allowed where the product licence says so. The licence does not permit reselling, sublicensing, public hosting, or building a competing dataset service from the collection.
No. Keep the raw collection and API responses inside authorised internal workflows. Employees, contractors, or advisers may access material only when they need it for the permitted internal purpose and are bound by suitable confidentiality obligations.
11
A free entry point can help teams inspect the delivery format before choosing a paid release.
Yes. The account-gated free sample workflow lets teams choose a published category, inspect a randomly selected pre-generated document, and test the delivery format before buying. Web accounts can claim up to ten samples; an authenticated API key can retrieve up to five free sample documents.
No. A free sample is intended to demonstrate structure and delivery, not to replace a full catalogue release. Paid packs and prepaid API access provide the documented scope, volume, and usage rights shown on their product pages.
Yes. An account is required so access can be limited, replay and abuse controls can work, and every delivery has a clear licence and usage boundary. No credits are charged for the free sample pool.
12
Keep credentials private and design integrations for predictable, controlled delivery.
Store it in a server-side environment variable or secret manager. Never place it in browser JavaScript, a public repository, a prompt, a URL, or a client application. Keys are shown once and can be revoked from account settings.
The API rejects a request before delivery when the balance is insufficient. It does not continue serving documents and silently create a negative balance. Automatic renewal is disabled, so nothing is charged or replenished without your action.
Read the response headers, handle HTTP 429 explicitly, wait for the reset or Retry-After value, and use exponential backoff. Request only the document count the current job needs.
13
Get access, manage keys, and resolve account or delivery questions through the workspace.
Yes. Dataset purchases and API access require an authenticated account so delivery, licensing, payments, credits, and security controls can be associated with the correct person or organisation.
Use the account workspace to view your balance, create or revoke API keys, review usage, manage notifications, request a data export, and access settings.
Use the contact form for support, purchase, API, or account questions. Feature ideas can be sent through the dedicated feature request page.
Ask about a workflow, format, dataset scope, API integration, or future sample access.
Contact Deniable →Send a feature request →
Deniable