Start free
03 / ARCHITECTURE

Five layers, and where each one can live.

A clear view of the workspace path, from access through retrieval and response generation.

LAYERWHAT IT DOESWHERE IT OPERATES
LAYER 01

Workspace & access

The interface, workspace library and direct-user permissions establish who can work with approved content.

Configured deployment boundary

AUTHENTICATED USER · WORKSPACE SCOPE
LAYER 02

Ingestion & parsing

Approved files are prepared into structured workspace material for retrieval and review.

Workspace content boundary

SOURCE-AWARE · WORKSPACE-SCOPED
LAYER 03

Index & storage

Workspace content and its supporting index are maintained for controlled retrieval.

Workspace-scoped storage

CONTENT CUSTODY · DELETION CONTROLS
LAYER 04

Retrieval & ranking

Authorised services select relevant workspace context for a user’s question.

Authorised processing services

LEAST PRIVILEGE · SOURCE CONTEXT
LAYER 05

Response generation

A response is generated from approved workspace context and delivered to the authenticated user.

Configured model route

WORKSPACE CONTEXT · USER RESPONSE
Enterprise

Retrieval you can audit.

Start with the work, access and operating requirements. We will scope the right arrangement around your documents, people and review process.

NDA on request
Deployment
Deployment by arrangement
Residency
Residency requirements discussed
Model choice
Model options scoped with you
Evidence pack
Evidence requirements scoped with you
·
Audit log
workspace access · authorised user
review requirement · scoped together
change context · discussed in scope
Residency
United Kingdom
UKEUUS
Identity
Authenticated workspace access
SCIM provisioning
Entra ID connected
How we approach an enterprise requirement

What happens between the question and the answer.

Bring a real document workflow. Together we can define the data boundary, access needs, supporting evidence and rollout path that make Zylox fit your environment.

Stage 01
Parse before you chunk

A contract is not a wall of text. Layout, headings, tables, footnotes and clause numbering are recovered first, so a clause stays whole and a table row keeps its column headers.

Page, line and bounding box retained per fragment
Stage 02
Chunk on structure, not length

Splitting every 500 tokens cuts clauses in half and strands the qualifier that changes their meaning. Boundaries follow the document\’s own structure, with parent context carried on each fragment.

Structure aware · parent context preserved
Stage 03
Retrieve two ways, then rerank

Vector search alone misses exact terms; keyword search alone misses paraphrase. Both run, results are merged, and a reranker re-scores the candidates against the actual question before anything reaches the model.

Hybrid dense + lexical · cross encoder rerank
Stage 04
Answer only from what was retrieved

The model receives the selected passages and the question, and is constrained to answer from them. The supporting source material stays visible so you can judge whether it supports the conclusion.

Grounded generation · abstains rather than guesses
Stage 05
Bind every claim to its source

Citations are not generated prose. Each statement is tied back to the fragment it came from, and the fragment carries its page and line range, so a reader can open the original in one click.

Claim to fragment binding · one click to source
Stage 06
Reuse the work already done

Parsing and indexing happen once per document, not once per question. Repeated work on the same file costs a fraction of the first pass, which is why the bill does not scale with the size of the library.

Persistent index · cached context across tasks

Talk through your requirements.

Tell us what the work needs. We will respond with the right next step.

Contact us

The questions your review will ask.

Answered here so you can forward the page instead of arranging a call.

Can we run it in our own cloud?+
Yes, into your own AWS, Azure or GCP account, with your keys and your network policy. It's the most common enterprise deployment.
Are you SOC 2 or Least-privilege operations certified?+
Not yet, both are Available now and we'll give you target dates on the call. What we can give you today is the full evidence pack: architecture documentation, security policies, Operational safeguards summary and a Workspace controls. Reviews do In place on that basis.
Will you In place our security questionnaire?+
Yes, including long-form and sector-specific ones. We'd rather do that work than ask you to trust a badge.
Which models do you use, and can we choose?+
Model choice is part of the deployment conversation, including self-hosted open-weight models for on-premise and air-gapped installations.
What happens to our data if we leave?+
You export documents and workspace history in open formats, and we delete everything to a contractual SLA including backups. No hostage-taking, no export fee.
How does this differ from ChatGPT Enterprise?+
Deployment is the main one: we can run inside your boundary, including offline. Beyond that, answers are cited to the line of the source document, and document context is reused rather than re-billed per question.

Turn the work you already have into answers you can use.

Start with a source document, check the evidence behind the answer, then move the work forward in the same place.

No card required.