Zero-retention privacy architecture for engineering AI

Engineering organizations do not have the luxury of treating artificial intelligence as a clever autocomplete toy. When the corpus includes foundation load reports, unreleased turbine specifications, client contracts under NDA, and internal interpretations of IEC or Eurocode clauses, every architectural decision about where data lives, how long it persists, and who can retrieve what becomes a board-level risk question — not an IT convenience question.

Consumer AI products optimized for growth taught the market a dangerous habit: drag files into a chat window, ask a question, copy the answer into an email. That workflow embeds three failures engineering buyers should reject on first principles: ephemeral context without audit trail, vague or missing citations, and **data handling policies written for marketers rather than for teams who sign structural calculations.

Vevesh is built on a different premise: zero retention as a design constraint, not a policy footnote; military-grade boundary separation between marketing surfaces, operational workers, and customer knowledge planes; and local-first indexing with optional cloud paths that never confuse “hosted retrieval” with “we own your source folder.”

This article explains what that means in practice — for security reviewers, engineering managers, and the technical leads who must defend tool choices in front of clients and regulators.

Why retention is the wrong default for engineering AI

“Retention” in AI products usually hides several distinct behaviors:

1. Session retention — uploaded files stay available until the session ends.
2. Account retention — files persist across sessions for convenience.
3. Operational retention — logs, embeddings, and derived artifacts persist for product functionality.
4. Training retention — user content influences future model behavior for other customers.

Engineering teams care about all four, but consumer terms of service often blur them. A product that “does not train on your data” may still store uploads indefinitely for re-embedding, support debugging, or abuse monitoring. That may be acceptable for marketing copy. It is not acceptable when the upload is a geotechnical survey tied to a confidential site location.

Zero retention, as Vevesh uses the term, means:

• Your source folder remains authoritative on your machine or your infrastructure — not silently mirrored into a vendor-owned “workspace” that becomes the de facto system of record.
• Derived index artifacts (chunks, vectors, metadata) live in explicit data planes with known isolation — separate Cloudflare accounts for user global RAG storage versus operational workers versus marketing sites.
• Query content is processed to produce answers and usage accounting — not harvested into analytics pipelines that treat engineering prompts like engagement metrics.

This is not paranoia. It is the minimum posture for consultancies, OEM engineering departments, and owner-operators who already classify load case spreadsheets at the same sensitivity tier as financial forecasts.

Military-grade boundaries: what we mean (without the theater)

“Military-grade” is abused in SaaS marketing. In Vevesh documentation it refers to concrete boundary patterns, not a sticker on a slide:

Account and plane separation

Customer knowledge base content — chunks, vectors, raw uploads in server extraction mode — runs on dedicated Cloudflare infrastructure (account 2 in our deployment model), isolated from:

• Account 1 operational workers (embeddings gateway, admin tooling)
• Account 3 marketing (this website on Pages)

The point is not branding. It is blast-radius reduction: a compromise or misconfiguration in marketing infrastructure must not expose customer corpora; a rotation of bridge tokens for one worker class must not invalidate unrelated security domains.

Self-hosted server path

Teams with air-gap or on-prem requirements can run the Dell/server stack locally while still using optional edge components only where policy allows. Hybrid is not “cloud-only with a VPN”; it is explicit toggles for local extraction versus server pipeline, per knowledge base.

Access graphs, not single gates

Membership in a global knowledge base is not binary “can use AI.” Project-level overrides (allowedGlobalRagIds) restrict which shared corpora appear in which project context. Redis-cached access checks bust on membership writes so stale privilege windows do not survive personnel changes.

Disclosure modes as policy engines

Ghost mode and open mode are not UI themes. They change retrieval and citation behavior immediately — without re-ingestion — because policy belongs in the access layer, not in re-chunking PDFs.

Reviewers should ask vendors: show me the diagram of accounts, tokens, and storage paths. If the answer is a single multi-tenant bucket and a hand wave about encryption at rest, that is not military-grade anything.

Local AND cloud: privacy follows the deployment choice

Vevesh supports local extraction on the desktop and server extraction through the edge pipeline. Privacy implications differ; the product makes that explicit instead of hiding it behind “sync.”

Local mode:

• Files are read from your filesystem; extraction and chunking run on your hardware.
• Upload artifacts are processed chunks and metadata required for retrieval — not a mandate to exfiltrate raw source trees.
• Ideal for IT policies that restrict egress, field laptops with intermittent connectivity, or preliminary reviews before promoting a corpus to a team knowledge base.

Server mode:

• Raw files upload to isolated R2 storage with content-hash metadata; extraction runs on workers; embed and ingest follow atomic batch rules.
• Billing meters server extraction by engine tier — transparency is part of privacy governance because surprise data processing is a privacy failure too.

The wrong question is “cloud vs on-prem religion.” The right question is which corpus, which mode, which audit requirement — and whether the tool lets you choose per knowledge base without forking your workflow.

Phone, desktop, browser: multi-surface without multi-copy

Engineering work happens in the site trailer, the home office, and the certification lab. Vevesh targets:

• Desktop-native filesystem access for heavy indexing and large folder trees.
• Web entry for lighter workflows and reviewer access.
• Mobile-friendly surfaces for approvals, evidence review, and usage visibility on site.

Multi-surface does not mean “copy your entire project to four devices.” It means one index graph, consistent citations, role-aware sector pickers — so the answer on phone matches the answer on desktop because they query the same grounded index, not because they re-uploaded the same PDF twice.

Zero retention versus “we need history”

Teams hear “zero retention” and ask: but we need conversation history, revert, and audit logs.

Correct. Zero retention is not amnesia for governance features. Vevesh separates:

Data class	Typical handling
Source folders	Stay client-side or customer-controlled; never rewritten by the indexer
Index generations	Hash-based incremental sync; revert paths when bad sync lands
Chat history	Account-scoped for workflow; not training fuel for other tenants
Usage metering	Billing and allowance tracking with explicit categories
Evidence exports	Deliberate, user-initiated packs for design review

The distinction matters for procurement: you are not buying a chat log mined for ads. You are buying traceable engineering queries tied to indexed artifacts you already controlled.

Evidence reporting and verification loops

Privacy without accuracy is useless. Vevesh pairs boundary architecture with:

• Page-level citations on grounded answers
• Verification loops that retrieve, reason, cite, and flag conflicts when sources disagree
• Evidence reporting suitable for design review — exports that chain claims to file paths and sections

When Rev C contradicts Rev B, advanced conflict logic surfaces the fight instead of blending contradictions into fluent prose. That is a privacy feature too: undetected wrong answers leak intellectual property confidence — teams act on mistakes they would not publish if they saw the primary sources side by side.

Team privileges: privacy is a graph, not a checkbox

Shared knowledge bases introduce social attack surfaces: the intern who inherits the whole standards library, the contractor who retains sector access after offboarding, the client project that must not see internal methodology corpora.

Vevesh models:

• Owners and members on global KBs with Stripe-backed subscriptions
• Project-level allowedGlobalRagIds restrictions
• Role-aware sector lists in chat pickers
• Ghost disclosure when citation exposure itself is sensitive

Customizable privileges are not enterprise upsell theater — they are how consultancies stay client-clean while reusing internal playbooks.

How consumer AI fails the engineering privacy review

When you evaluate ChatGPT file upload, Copilot on SharePoint-only corpora, or Notion AI over wikis, ask:

1. Where does the canonical copy live after upload?
2. Can I get page-level citations on every grounded claim?
3. Can I restrict retrieval by project and membership without re-uploading?
4. Can I run local indexing for NDAs and server indexing for shared KBs in one product?
5. Show me account isolation for embeddings, raw storage, and admin tools.

Consumer tools optimize time-to-first-answer. Engineering platforms optimize time-to-defensible-answer — and that requires architecture consumer vendors did not build because their buyer is not signing fatigue load combinations.

See our platform comparison for an honest pros/cons matrix.

A buyer’s checklist: questions to ask any engineering AI vendor

Use this verbatim in RFPs:

1. Retention: Are prompts, uploads, or embeddings retained for product improvement or model training? For how long? Opt-out or contractually zero?
2. Isolation: Are customer corpora commingled in shared indices or buckets? Draw the account diagram.
3. Local path: Can indexing run fully on customer hardware without mandatory cloud upload of raw files?
4. Citations: Are citations path-stable and page-specific? Can reviewers open primary sources in one click?
5. Policy modes: Can citation exposure be restricted per corpus (ghost vs open) without re-ingestion?
6. Access control: Are project-scoped restrictions available for shared corpora?
7. Metering transparency: Are extraction and inference costs attributable per engine and per KB?
8. Revert: What happens when a bad sync or wrong revision enters the index?
9. Evidence export: Can answers be exported as audit bundles?
10. Subprocessors: List every region and provider that touches raw or derived content.

Vevesh publishes architecture aligned with these questions because we sell to teams who already run checklists — on structures, not just on software.

Implementation patterns we see in private preview

OEM engineering department: local Express indexing on project laptops; team standards KB in server Structure mode; ghost disclosure for pre-release methodology.

Consultancy: open citations on client deliverable folders; ghost internal KB; consultant mode workflows with evidence packs for submission packages.

Research group: Premium vision on scanned annexes; PhD-literature-style deep search across PDF and supplementary data; history management for long-running thesis threads.

Owner-operator: self-hosted server with optional edge embed; strict allowedGlobalRagIds per JV project so partners never see each other’s sectors.

None of these patterns require “trust us, it’s encrypted.” They require controls you can explain to a client auditor in plain language.

The cost of getting privacy wrong

Privacy failures in engineering AI are not abstract GDPR fines alone. They include:

• Loss of client trust when a methodology leak surfaces in a competitor bid
• Regulatory exposure when undocumented AI paraphrases replace cited standards clauses
• Rework cost when undetected conflicts between revisions propagate into drawings
• Tool ban backlash when IT disables AI entirely after one reckless upload workflow

Zero retention and boundary-first design are how you avoid the ban hammer: give engineers AI that respects the same document control instincts they learned from ISO 9001 and client QA gates.

Conclusion: privacy is the feature surface

Vevesh shouts super private, zero retention, military-grade boundaries because those are not marketing adjectives — they are the constraints that make every other feature trustworthy: turbo extraction engines, global sector knowledge bases, consultant mode, verification loops, and deep cross-format search.

If your AI vendor cannot draw the data plane on a whiteboard in five minutes, you are not buying engineering infrastructure. You are renting a chat window.

Next steps:

• Read the full features breakdown
• Compare Vevesh vs ChatGPT, Copilot, Notion, and Glean
• Request early access for design-partner onboarding

We are publishing twenty deep-dive articles — each 2,000+ words — on every major capability in this stack. This is article one. The series continues with military-grade deployment boundaries, hybrid local–cloud guides, and verification loop design.

Your folders already contain the truth. Vevesh’s job is to make AI query that truth without becoming another place secrets leak.