Skip to content
haystak
Request a demo
Trust

Nothing leaves unless you say so.

haystak indexes in place and reads only, inside your environment. Where it runs, what it can reach, and how it's tested, laid out plainly.

Deployment

Stay in. Or reach out. Your call.

haystak can run with nothing leaving your network: your data, your models, your hardware. Reaching out to cloud sources or hosted models is a business decision you make, source by source, never a dependency we bake in.

Stay in

Your models, your hardware. Nothing has to leave. Index at the edge and run inference on-prem with Ollama or any OpenAI-compatible server you run (vLLM, LocalAI). No upload, no egress, no per-call bills. The whole pipeline can run air-gapped.

Reach out

Going out is a business decision, not a requirement. When it's the right call, index cloud sources (Box today, more on the same seam) and use hosted models like OpenAI or Anthropic. Your choice, per source.

Read-only, either way

Local or cloud, originals are never modified; derived data and annotations are append-only, under the same RBAC and audit.

Data handling

Your data stays where it lives.

Sources are read in place and never modified. The only data haystak creates (extracted text, metadata, tags, embeddings) is stored in the backend you run, on your infrastructure. There is no haystak cloud in the path unless you connect your cloud sources or your hosted models on purpose.

Index in place

Source bytes never leave the source. haystak reads, extracts, and indexes without moving files.

Your backend, your infra

Derived data lives in the store you run (Elasticsearch, OpenSearch, PostgreSQL, MariaDB, or MongoDB) on hardware you control.

Egress is opt-in

Reaching a cloud source or a hosted model is a choice made source by source. Fully air-gapped is a supported end state.

Obligations

Controls that speak to the rule you are audited against.

Records management, privacy and discovery obligations ask for specific, demonstrable behaviors, not a promise that a vendor is careful. These are the controls haystak implements and the obligations each one answers. They are controls we build and test, not certifications we hold; see what is proven and what is in progress below.

Retention and disposition, on a schedule

Records carry a retention class and are disposed of when it expires, rather than whenever someone remembers. A legal hold refuses the disposal and records the refusal, so a held record cannot quietly age out. Reports show what is due, what is held, and what is overdue. This is the behavior SEC 17a-4(f) and FDA 21 CFR 11.10(e) ask for, and the retention and disposition leg of ISO 15489.

Access limited to a declared purpose

Being entitled to a record for one reason no longer serves it for any reason. Each read declares its purpose and is limited to it. An emergency override exists because refusing outright is worse, but it demands a written reason and lands in the audit as an override rather than an ordinary read. This is HIPAA's minimum-necessary standard and the purpose-limitation requirement in GDPR Articles 5 and 6.

Erasure that reaches the derived data

Erasing a person suppresses the identity first, then removes the tags and facts derived from it across every path that produced them (the part that is usually missed, because derived data does not look like personal data until someone asks). Where a legal hold applies, the erasure is refused and the refusal is recorded. This is GDPR Article 17, including the derived-data problem it creates.

Search completeness you can show

A validation run samples what a search did not return, captures the reviewer's outcome, and reports a confidence interval. That turns "we searched reasonably" into a measurement someone else can check. This is the ground FRCP 26(g) puts a signature on. It measures the set the requesting account was permitted to see, and says so on the result.

An audit trail that shows tampering

Audit records are chained to each other, so a removed or altered entry breaks the chain and verification says so. It is on by default rather than a setting someone has to find, and it survives a restart. This is the integrity property the records-management standards above depend on: a retention claim is only as good as the log that evidences it.

Every one of these is refusable

Each control above can decline. A count that cannot be made exactly refuses instead of estimating; a breakdown too large to be accurate names the limit; a validation run for an unrestricted account is refused rather than guessed. A control that silently degrades under load is not a control, and an auditor will find the case where it did.

These are product controls, enforced inside your environment on every surface haystak exposes.

Data lifecycle

Deletion is a path we built and tested.

Immutability here is an audit property, not an inability to honor a deletion request. The audit trail cannot be quietly rewritten; the content a deletion covers still stops being served. Those are two different things, and haystak does both.

Suppressed everywhere at once

When a file is removed at the source, its entity is marked missing, and every serving surface filters missing entities out: keyword search, the vector retrieval behind RAG, the MCP tools, and the SQL endpoint. A deletion that does not reach all of them is not a deletion, so it reaches all of them.

Dedup-safe and reversible

Identical content is stored once and shared. A deletion suppresses shared content only when no present file still references it, so removing one of two identical files leaves the other intact. If the file returns, the next pass makes it live again. A move undoes itself.

Hide without destroying

An administrator can suppress a record from results without deleting it, applied at the same gate every surface reads through, including fetching a record directly by name. Where the system cannot determine whether something should be hidden, it hides it. A deletion reply says which happened: the rows are gone, or the record is retained and hidden.

Suppression, then purge

Suppression makes content unretrievable across every surface. Reclaiming the stored bytes is a separate, schedulable step. We are exact about the difference: we do not claim the bytes are gone the instant a file is deleted. If your requirement is that the bytes are gone, that is the purge step's job.

Propagation is bounded by how often each source is reconciled, a per-source setting you tune tight on a fast-changing share or looser on a cold archive. An unreachable source is never read as a mass delete: if a volume drops offline, reconciliation is skipped and every entity stays active. And the deletion itself is a recorded, attributable event in the append-only audit.

Assurance

Secured in your environment. Tested before every release.

haystak runs on your infrastructure, behind your identity provider and your network, and nothing leaves unless you configure it to. Here's what the product enforces, and how each release is tested.

In the product today

  • Read-only, fail-closed architecture
  • Tamper-evident, cryptographically chained audit
  • Directory-backed identity and single sign-on: SAML 2.0, LDAP, AD, Open Directory, FreeIPA
  • Database and model-service credentials encrypted at rest
  • Services run under dedicated, low-privilege accounts
  • One ACL enforced across every surface
  • Purpose-based access, with justified break-glass override
  • Retention schedule execution, with a self-verifying disposition certificate
  • Search completeness validation: reproducible, seeded elusion sampling
  • Signed release binaries: Developer ID on macOS, Authenticode on Windows, with SHA256SUMS and an SBOM per artifact
  • Secure development: code review, dependency management

How it's tested

  • Every API operation checked for the authentication and authorization it requires
  • Injection, path-traversal and abuse testing across the API, including hostile files
  • Every release scanned for known vulnerabilities in its dependencies
  • Release gates that fail the build rather than warn

Our API security assessment and its evidence are available to evaluators under NDA.

Disclosure

Found something? Tell us.

Report a suspected vulnerability to security@haystak.app. We acknowledge and remediate. Security questionnaires, architecture and data-flow diagrams, and the current assurance status are available to evaluators on request, under NDA where needed.

  • Read-only: sources are never modified.
  • Fail-closed: deny by default across every surface.
  • On your terms: on-prem, your models, your hardware.
Talk to us

Deploy it your way.

A walkthrough of the deployment posture, data handling, and the audit trail, on a corpus that looks like yours.