Zero-retention RAG platform for trade secrets

For a Retrieval-Augmented Generation (RAG) system handling trade secrets, zero data retention (ZDR) must be guaranteed by your AI model provider and verified by your organization's security and legal teams. It is not something you should assume simply because a provider advertises enterprise-grade privacy.

There are three distinct responsibilities:

• AI/model provider: Must contractually commit to zero retention for the specific API, endpoints, and features you use.
• RAG platform and infrastructure operators: Must control retention of prompts, retrieved document chunks, embeddings, application logs, traces, caches, and backups.
• Your security and legal teams: Must verify the technical implementation, contractual terms, subprocessors, and audit evidence.

Providers to evaluate

What you should require for a trade-secret RAG system

1. Written ZDR commitment: Verify that the agreement covers your exact models, API endpoints, and features, including exceptions for abuse monitoring and legal obligations.
2. No training on your data: Get an explicit contractual commitment covering prompts, responses, and confidential context.
3. End-to-end retention controls: Audit your vector database, embedding provider, reranker, observability tools, application logs, backups, and caches. Model-provider ZDR does not automatically cover these components.
4. Independent verification: Have security validate the architecture and retention behavior; have legal review the DPA, subprocessors, and liability terms.
5. Data minimization: Keep trade secrets out of third-party services wherever possible; use self-hosted models and infrastructure if your risk tolerance requires full control over the data path.

My recommendation: Make the provider's contractual ZDR commitment a procurement requirement, then verify the entire RAG pipeline separately. A provider's ZDR promise is not the same as zero retention across your whole system.

One important clarification: Are you looking for a third-party LLM API with contractual zero retention, or an end-to-end RAG architecture where no external party can retain your trade secrets? The right solution differs substantially between those two requirements.

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?