Zero-retention RAG platform for trade secrets

For a RAG system handling trade secrets, the responsibility for zero data retention rests on both your AI model provider and your infrastructure providers. You should not assume that a vendor guarantees zero retention simply because it offers an enterprise plan or says your data is not used for training.

The key distinction is between no training on your data and zero retention of your data. A provider may promise the former while retaining prompts, retrieved documents, outputs, or logs for some period.

Who can provide zero retention?

Official documentation: OpenAI ZDR, Anthropic ZDR, AWS Bedrock retention controls.

Who is accountable in your RAG architecture?

1. Model provider

Obtain a contractual ZDR commitment that explicitly covers your selected models, API endpoints, safety monitoring, and exceptions.

---

2. RAG infrastructure operator

Control retention in vector databases, document stores, caches, tracing, application logs, observability platforms, and backups. Model-provider ZDR does not cover these automatically.

---

3. Your security and legal teams

Verify contractual obligations, access controls, deletion guarantees, subprocessors, incident response, and audit evidence. Your organization remains responsible for the end-to-end system.

My recommendation for trade secrets

Do not select a provider based on a marketing claim alone. Require written confirmation of:

• No persistent storage of prompts, retrieved chunks, or model outputs after processing.
• Explicit treatment of abuse-monitoring logs, safety exceptions, and legal holds.
• No training on confidential inputs without authorization.
• Retention and deletion policies for embeddings, telemetry, backups, and third-party services.
• Documented access restrictions, encryption, and enforceable confidentiality terms.

One important question: Are you building this RAG system using OpenAI, Anthropic, Azure OpenAI, AWS Bedrock, or a self-hosted model? I can narrow this down to the specific provider's ZDR guarantees and the gaps you need to close.

Ensuring zero data retention (ZDR) for trade secrets in a Retrieval-Augmented Generation (RAG) system is a shared responsibility . No single entity handles it alone; rather, it requires a synchronized combination of legal contracts, cloud provider controls, internal infrastructure configuration, and strict governance.reddit.com+2

Accountability for zero retention is distributed across these key pillars:

1. The Cloud / LLM API Provider (Contractual & Technical Enforcement)
   • Who they are: Vendors like OpenAI, Anthropic, or Microsoft Azure (for hosted models).
   • What they do: They provide ZDR-eligible enterprise endpoints . Standard API tiers often log prompts and completions for up to 30 days for safety and abuse monitoring. The vendor ensures zero retention only when operating under an enterprise agreement that explicitly and technically disables persistent logging of inference data.
   • Note on consensus: As noted in discussions on platforms likereddit.com , relying on commercial APIs with strict ZDR agreements is standard, though fully on-premise deployments are often preferred for ultra-sensitive trade secrets.neuraltrust.ai+1
2. Your Organization's Legal & Procurement Team (Contractual Guarantee)
   • What they do: They negotiate the Data Processing Addendum (DPA) and enterprise terms. They ensure the contract explicitly states that the vendor does not retain data, does not use customer inputs/outputs to train or fine-tune foundational models, and provides compliance attestation (e.g., SOC 2, ISO certifications).docs.cloud.google.com+1
3. Your Enterprise Security & Architecture Team (Internal Configuration & Perimeter Control)
   • What they do: They configure the application pipeline so that trade secrets aren't inadvertently leaked or logged internally.
   • Vector Database & Storage: They ensure the vector database (e.g., Pinecone, Milvus, pgvector) and document stores are deployed within a secure internal Virtual Private Cloud (VPC), encrypted at rest and in transit, and governed by strict identity and access management (IAM).
   • Shadow AI Prevention: They block employees from using consumer-grade/free-tier AI subscriptions where ZDR does not apply.airia.com+1
4. Data Governance & Engineering (Data Minimization & Filtering)
   • What they do: They design pre-retrieval filters and data masking policies. They ensure that only the minimum necessary context chunk is sent to the LLM API, and that sensitive metadata or identifiable trade-secret markers are scrubbed or tokenized if they shouldn't transit external APIs.

If you'd like, let me know:

• Are you planning to use external commercial APIs (like OpenAI/Anthropic) or self-hosted/on-premise models?
• What cloud environment or infrastructure stack are you building on?

I can help you outline a specific security checklist for your RAG architecture.

RAG systems for trade secrets: zero data retention

If you're building a retrieval-augmented generation (RAG) system for trade secrets, source code, proprietary formulas, engineering designs, or confidential business documents, the key distinction is between a vendor promising not to train on your data and a system that technically cannot retain or expose your data.

A standard cloud RAG architecture can send sensitive information to several separate providers: the LLM, embedding model, vector database, document parser, observability service, and application logs. Each needs its own data-handling guarantees.

My recommendation: prioritize self-hosted RAG in your own controlled infrastructure if you need the strongest confidentiality guarantees. If you need a managed cloud service, evaluate contractual zero-data-retention (ZDR) terms and the precise scope of each provider's commitment rather than relying on a general “your data is private” statement.

One important caveat: no-retention guarantees are always scoped to particular systems, data types, and exceptions. A vendor's claim does not automatically establish that every component of your RAG pipeline retains zero data.

Shortlist: systems worth evaluating

Based on vendor documentation available as of October 2026.

• 

1. Self-hosted RAG — strongest control over retention

(Best for highly sensitive IP)

Run a local model with Ollama or vLLM, a self-hosted vector database, local embeddings, and your own document-processing pipeline.

• Retention: Can avoid sending content to external AI providers entirely.
• Trade-off: Your team is responsible for logs, backups, access controls, model security, and deletion.
• Best for: Secret formulas, proprietary source code, unreleased designs, and high-value intellectual property.

Project sites: Ollama · vLLM ·Qdrant

---

• 

2. Anthropic Claude API with approved ZDR

(Contract-gated)

Anthropic offers a zero-data-retention arrangement for eligible API customers. Prompts and responses are not stored at rest after the API response, subject to documented exceptions and feature eligibility. ZDR must be enabled for the relevant organization.platform.claude.com+1

• Retention: Zero retention for covered API requests, with exceptions.
• Trade-off: Not all API features qualify; some safety data and legal exceptions remain.
• Best for: Teams that want a managed frontier model but need explicit retention commitments.

Official API retention documentation

---

• 

3. OpenAI API with approved ZDR

(Approval required)

OpenAI offers ZDR for eligible API organizations and endpoints. When enabled, covered requests are excluded from content retention in abuse-monitoring logs, and supported endpoints disable certain application-state storage. Not every endpoint or feature is eligible.developers.openai.com+1

• Retention: Depends on endpoint, model, and approved controls.
• Trade-off: Files, stateful features, third-party services, and specific exceptions need separate review.
• Best for: Existing OpenAI-based RAG applications that need a managed model with a documented ZDR configuration.

Official API data controls

---

• 

4. Amazon Bedrock with an eligible model and retention mode

(Model-specific)

AWS provides configurable data-retention controls, including a none mode for eligible models and approved accounts. However, some models require content retention for safety review, potentially for up to 30 days.docs.aws.amazon.com

• Retention: Depends on the model and account configuration.
• Trade-off: Verify the exact model's retention mode, region, and any cross-region processing.
• Best for: Organizations already operating inside AWS with private networking and centralized cloud governance.

Official Bedrock retention documentation

These are four viable approaches, not four equivalent guarantees. Self-hosting gives you the greatest control over the complete data path; the managed options require more precise contractual and configuration checks.

How I would compare them

For the hosted services, these comparisons concern the model-inference layer, not a complete RAG deployment. They should not be interpreted as a guarantee that the entire system stores no data.

The architecture I would choose for trade secrets

Confidential documents

Encrypted storage under your control

Local parsing and embeddings

No external document-processing APIs

Private vector database

Self-hosted Qdrant or PostgreSQL with pgvector

Local LLM inference

Ollama or vLLM on your own hardware

All document content, embeddings, retrieved passages, and model prompts remain inside your controlled environment, provided the complete pipeline is configured without external data flows.

This architecture can provide a stronger technical guarantee than relying on an external vendor's retention policy. It still requires controls for temporary files, backups, swap, telemetry, crash reports, and administrative access. For particularly sensitive IP, consider an isolated network with outbound traffic blocked and auditable deletion procedures.

What to require before approving any vendor

Use this checklist during procurement and security review.

Security review checklist

of

Bottom line

• If zero external retention is non-negotiable: choose a fully self-hosted RAG stack.
• If you need a managed LLM: investigate Anthropic API ZDR, OpenAI API ZDR, or Amazon Bedrock's none retention mode for your specific model and configuration. Obtain written confirmation before sending trade secrets.
• If you need a legally defensible guarantee: have counsel review the contractual language, exceptions, and the entire data-processing chain. No vendor-level ZDR setting alone guarantees that every component retains absolutely nothing.

One question that would help narrow this to the right systems: Must the RAG system run entirely on your own infrastructure, or is a cloud-hosted LLM acceptable if the provider contractually guarantees zero retention?