The default retention period is 12 months for all stored data — conversation history, files, and embeddings.

This is fully configurable based on your compliance requirements. Contact your account team to adjust.

All deletions are hard deletions. When an agent or knowledge base is deleted, data is permanently removed with no soft-delete intermediary or backup persistence beyond the configured SLA.

Uploaded files are stored temporarily, encrypted at rest, and follow the same retention policy as all other data.

Supported file types are parsed for text extraction. Document metadata such as comments, tracked changes, and hidden sheets may be processed as part of extraction.

Lyzr's optional PII Guard layer sits in the processing pipeline — before any data is sent to an external LLM provider. It combines two detection approaches:

NLP-based entity recognition

Identifies contextual entities like person names using trained language models.

Pattern & regex matching

Catches structured PII: credit card numbers, SSNs, phone numbers, emails, IP addresses, IBANs. Resultsare merged, deduplicated, and filtered through per-entity confidence thresholds to reduce false positives.

Detected entities are handled in one of three ways, configurable per entity type: Redact — replaced with a numbered placeholder such as PERSON_1; Block — the entire input is rejected; or Disabled — no action taken. A reverse mapping is returned so responses can be de-anonymised on the return path.

85.9% Precision, 83.4% F-Score, 82.8% Recall from 10k samples

The system was evaluated against a synthetic dataset of 10,000 samples spanning 17 PII entity types, including names, addresses, credit cards, SSNs, phone numbers, emails, IBANs, and IP addresses.

The only exception is images: PII embedded visually in scanned documents, screenshots, or photos of IDs will not be detected.

Captured:

• Request identifiers for debugging

• Model and provider info

• Token usage for billing

• Error rates and response times

• Workflow steps — tools and actions run

NEVER Captured:

• Personal information of any kind

• Content of prompts or responses

• Uploaded files or attachments

• Conversation history

• Any organisational identifiers

Lyzr enforces logical tenant separation — each organisation's agents, knowledge bases, conversation history, and configuration are scoped to that organisation's namespace and cannot be accessed by other tenants.

Conversation data is stored per session and isolated per organisation. Message storage can be disabled entirely via a platform setting if your policy requires it.

Yes — all data is encrypted both in transit using TLS and at rest. This applies across all datastores: AWS DocumentDB, Qdrant, ElastiCache, S3, and ClickHouse.

Lyzr uses AWS KMS for key management and supports automatic key rotation. A secrets management layer

handles all provider and API keys securely. Customer-managed keys — BYOK and HYOK — are supported.

Contact your account team for the scope and limitations specific to your deployment tier.

External LLM receives:

• System prompt

• User message or PII-masked version

• Conversation history

• Tool definitions

NEVER receive:

• User IDs or org IDs

• Session identifiers

• Billing or usage data

• Any internal Lyzr metadata

Provider allowlisting, single-provider mode, and regional data residency controls are all available. The platform's fallback routing system can be configured to restrict which providers are eligible, including preventing fallback across provider boundaries entirely.

Discuss your specific provider restriction requirements with Growth Directors to configure the right routing policy for your compliance and data residency environment.

Lyzr enforces responsible AI guardrails as a core part of the agent architecture. These include input inspection, content filtering, and the optional PII Guard layer. Agent traces are available for auditability.

Prompt injection controls and RAG grounding mechanisms are built into the retrieval and processing pipeline.

Log export to external SIEM systems or S3 for independent retention is a supported capability at the enterprise tier. The exact log schema for user requests and admin/config actions, tamper-evidence controls, and export configuration details are available in the enterprise documentation package.

Contact us at Growth Directors to discuss log forwarding and SIEM integration for your compliance needs.