Skip to content

Privacy & Licensing

Knovaryn is a training-data foundry: it ingests documents, cleans and labels them, and exports datasets for model training. Privacy and licensing are design concerns from intake through export. This page describes the built-in controls and the responsibilities that remain with the operator. See hardening.md for the operational checklist and architecture/security.md for the threat model.

The built-in controls reduce risk; they are not a certification. Whether a given corpus may be used for training is a legal determination you must make for your jurisdiction, your contract terms, and the provenance of each document.

General design principles

  • Untrusted-in, controlled-out. Every source document is treated as untrusted at intake (spec §8.6). Nothing in a source is allowed to drive code execution, path traversal, or prompt injection against the operator.
  • Local by default. The control plane and pipeline run on the operator's machine (loopback 127.0.0.1:8765 by default). Data stays on the machine unless the operator explicitly configures external storage or ingestion.
  • Source and provenance. Every document entering the pipeline is tracked with provenance metadata (origin, source path, ingest time, hash) so an operator can audit where a training row came from.

Privacy controls

Intake

  • Restrict the source surface with sources.allowed_roots, disable symlink following, and bound max_file_mb / max_pages so the pipeline cannot be coerced into reading more than the operator intended.
  • Treat all docs as untrusted: parsing, chunking, and extraction never trust content as instructions.
  • URL ingestion is off by default; enabling it extends the intake surface and requires SSRF hardening (see hardening.md).

Storage & secrets

  • Secrets (API keys, tokens, service credentials) are never committed and never stored in document text fields. .gitignore excludes secret material; secret handling is documented with the configured secrets backend.
  • Embedded stores are encrypted at rest by the OS when the machine disk is encrypted; treat *.db files as sensitive and control access to them.

Data minimization

  • Only the fields required for the configured use case are carried through the pipeline; the schema is explicit (spec schema) rather than "load everything."
  • Operators should review any dataset that may contain personal data and apply redaction/anonymization as required before export.

Licensing controls

  • Knovaryn tracks a license field and provenance on each ingested document and surfaces it in the dataset manifest, so downstream consumers can see what terms applied to the source material.
  • The export manifest (manifest.json) records source provenance, hashes, and per-source license metadata so a training corpus can be traced back to its inputs.
  • Knovaryn does not make licensing judgments for you. Decide whether each source corpus permits the intended use (training, redistribution, commercial) under its own terms, and record that decision in the manifest or an operator policy.

Telemetry & analytics policy

  • No telemetry, no analytics. The package collects nothing: no usage metrics, error reporting, or phone-home behavior in any mode. Runtime network calls go only to the provider endpoint and storage backends the operator configures.
  • The documentation site makes zero third-party requests. No webfonts (system font stack), no analytics scripts, no embedded trackers; even the repository star widget was replaced with a plain same-origin link so page views never touch GitHub's API. This property is enforced by automated browser checks that fail the build on any cross-origin request.
  • Adding analytics later requires, before any implementation: a written proposal covering provider, cookies, personal data processed, retention, consent implications, CSP changes, privacy-policy impact, opt-out mechanism, and whether a privacy-preserving alternative would suffice — plus an explicit maintainer decision recorded in the changelog.

Operator responsibilities

  1. Confirm you have the right to ingest and train on each source corpus.
  2. Review datasets for personal data and apply redaction before export where required.
  3. Record a licensing decision per source and keep the audit trail.
  4. Apply the operational controls in hardening.md before any deployment that is not a throwaway local demo.