Legal Center & Policies — Production RAG Agent

🛡️

Enterprise Zero-Training Privacy Pledge

Your documents, uploaded files, vector embeddings, chat queries, and conversation memory belong solely to your organization. We do NOT use, sell, or allow third parties to use customer data to train, retrain, or fine-tune public foundation models (including OpenAI, Anthropic, Google Gemini, or open-source foundation checkpoints).

Scope & Data Controller Information

This Privacy Policy governs the collection, processing, storage, and protection of personal data and proprietary document assets processed by Production RAG Agent (referred to herein as "RAG Agent", "we", "us", or "our"). This policy applies to all users of our multi-tenant document ingestion workspaces, real-time retrieval-augmented generation (RAG) chat interfaces, autonomous multi-agent workflows, Model Context Protocol (MCP) tool executions, and subscription portals.

For the purposes of the European Union General Data Protection Regulation (GDPR) and the UK Data Protection Act:

Data Controller: For account identity, authentication credentials, billing records, and system telemetry, RAG Agent Technologies Inc. acts as the Data Controller.
Data Processor: For proprietary documents, text chunks, vector embeddings, and workspace interactions uploaded or generated by organization members, RAG Agent acts as a Data Processor on behalf of the customer organization (the Data Controller).

Zero Foundation Model Training Pledge

Enterprises rely on RAG Agent to process sensitive internal documentation, confidential source code, legal contracts, and proprietary intellectual property. We provide an unconditional, binding contractual commitment:

Zero Pre-Training & Fine-Tuning: Customer data—including raw documents, chunked text, semantic vectors, user queries, and agent deliberation traces—is strictly excluded from any model training or fine-tuning datasets.
Zero Data Retention with Third-Party Providers: When commercial foundation models (e.g., OpenAI, Google Gemini, Anthropic Claude) are configured as generation or embedding providers, requests are routed via enterprise endpoints governed by strict Zero Data Retention (ZDR) agreements. Third-party providers are prohibited from logging prompt contents beyond transient stateless inference.
Air-Gapped Local Option: For organizations with maximum privacy mandates, RAG Agent provides native support for 100% local, on-premise execution utilizing local Ollama models (such as Qwen2.5 and Nomic-Embed-Text), ensuring zero outbound network transmission.

Categories of Collected Data

We process data strictly to the extent required to deliver grounded document intelligence and agentic task execution:

Data Category	Examples & Formats	Processing Purpose	Retention Standard
Account Credentials	Email address, username, bcrypt salted hash, organization affiliation	User authentication, RBAC authorization, and account security	Duration of active account + 30 days grace
Uploaded Documents	PDF, DOCX, TXT, Markdown, CSV, HTML files uploaded to Rooms	Text extraction, semantic parsing, and chunk generation	Until deleted by user or room owner
Vector Embeddings	Dense high-dimensional vector arrays and chunk metadata	Cosine similarity indexing and nearest-neighbor search via ChromaDB	Purged immediately upon document deletion
Chat Sessions & Citations	User queries, assistant streaming responses, citation coordinates	Session persistence, conversation continuation, and audit trails	Purged on demand via "Clear All Data"
Dual-Tier Memory	Working facts (STM) and persistent cross-room facts (LTM)	Context continuity across multi-step agent reasoning	User-managed via Memory Inspector
Billing Records	Razorpay subscription ID, invoice number, payment timestamp (no card numbers)	Statutory tax compliance, invoice generation, subscription management	7 years (statutory accounting requirement)

Ingestion & Vector Embeddings Processing

When documents are submitted to a Room workspace, they undergo an automated ingestion lifecycle:

Text Extraction & Sanitization: Raw text is extracted, stripped of harmful scripts or embedded macros, and split into chunks based on configurable size and overlap parameters (default 800 characters with 150-character overlap).
Vector Generation: Each chunk is converted into a vector representation by your designated embedding model.
Database Storage: Embeddings and chunk references are saved directly into the isolated ChromaDB vector store mapped to the specific Room ID and Organization ID.

No external service or unauthorized user can search or cross-reference your vector indices.

Dual-Tier Memory & Context Retention

Our architecture incorporates a two-tier memory system designed to provide context-aware responses without compromising privacy:

Short-Term Memory (STM): Transient conversational facts stored in ephemeral RAM cache, scoped strictly to the active session. STM items decay automatically when a session ends.
Long-Term Memory (LTM): Explicitly confirmed user preferences and enterprise domain facts stored in the database. LTM access is governed by strict RBAC (accessible only to authorized Staff and Admin accounts).
Transparent Control: Authorized users can view, prune, revaluate, or delete every individual memory fact in real time through the dedicated Memory Management Panel.

Multi-Tenant Isolation & Security Architecture

We implement comprehensive technical and organizational safeguards:

Encryption Standards: Data in transit is secured using modern TLS 1.3 protocol with strict cipher suites. Data at rest (SQLite databases, ChromaDB vector stores, uploaded files) is encrypted using industry-standard AES-256.
Strict Workspace Partitioning: Every room workspace maintains cryptographic isolation. Document queries enforce strict row-level security and room membership checks at the API middleware layer prior to vector search execution.
Sanitized Execution: Multi-agent autonomous code execution takes place in restricted sandboxes preventing unauthorized system calls, directory traversal, or network escalation.

Subprocessors & LLM Service Providers

To deliver multi-modal capabilities and secure payments, we engage selected sub-processors under rigorous data protection agreements:

Razorpay: Payment processing gateway for subscriptions and recurring billing. Razorpay is certified PCI-DSS Level 1 compliant. RAG Agent never receives, stores, or logs raw payment card credentials.
Cloud Foundation Providers (Optional): When configured by your administrator, queries may be dispatched to Google Cloud Platform (Gemini API), OpenAI Inc., or Anthropic PBC under enterprise zero-data retention agreements.
ChromaDB & SQLite: Localized vector and relational persistence engines operating within your designated environment boundaries.

Data Retention, Purge & Deletion Rights

We adhere strictly to the principle of data minimization:

Instant Self-Service Purge: You may at any time click the "Clear All Data" button in Settings to permanently delete all conversation transcripts, session histories, and contextual memory records.
Document Deletion: Deleting a file from the Document Library immediately unlinks the file, purges its vector embeddings from ChromaDB, and removes the physical file from the storage directory.
Account Termination: Upon account closure, all associated rooms, documents, embeddings, and credentials undergo complete cryptographic erasure within 30 days.

GDPR, CCPA/CPRA & Global Privacy Rights

Regardless of your geographical location, RAG Agent provides comprehensive data rights:

Right to Access & Portability: You may request an export of your account data, document catalog, and chat logs in structured JSON format.
Right to Rectification: You may update erroneous profile information or re-index updated document versions.
Right to Erasure ("Right to be Forgotten"): You have the absolute right to demand complete deletion of your personal records and indexed assets.
Right to Restrict or Object: You may opt out of telemetry logging and disable automated memory extraction.

Data Protection Officer & Inquiries

For privacy inquiries, Data Subject Access Requests (DSAR), or Data Processing Agreements (DPA), please reach our Data Protection Officer:

📧

Office of the Data Protection Officer

Email: privacy@ragagent.io
Address: Legal & Compliance Dept, RAG Agent Technologies Inc.
Response Turnaround: Within 48 business hours

📜

Binding Agreement & Enterprise License

By accessing or using the RAG Agent platform, you confirm your acceptance of these Terms of Service. Customers retain 100% intellectual property ownership of uploaded documents and generated agent outputs.

Acceptance of Terms & Eligibility

These Terms of Service ("Terms") constitute a legally binding agreement between you (individually or on behalf of the entity you represent) and RAG Agent Technologies Inc. ("RAG Agent", "we", "us"). By creating an account, uploading documents, running autonomous agents, or paying for a subscription, you represent that:

You are at least 18 years of age or the legal age of majority in your jurisdiction.
You have full corporate power and legal authority to bind your organization to these Terms.
Your use of the services complies with all applicable export controls, cybersecurity regulations, and data privacy laws.

Platform Architecture & RAG Services

RAG Agent provides an enterprise-grade document ingestion and conversational retrieval platform featuring:

Multi-tenant room workspaces for group-scoped document sharing and contextual inquiry.
Document parsing for PDF, DOCX, TXT, CSV, MD, and HTML files with automated chunking and vector embedding generation.
Real-time Server-Sent Events (SSE) streaming answers with verifiable source chunk citations.
A agent autonomous collaborative system (Decision Maker, Researcher, Task Generator, Task Executor, Validator, Deployer).
Model Context Protocol (MCP) tool calling and sandboxed computational executions.

Accounts, Organizations & RBAC Governance

Users are responsible for safeguarding account credentials, API keys, and session tokens. You agree to notify us immediately of any unauthorized access to your account.

The platform enforces Role-Based Access Control (RBAC):

Admin: Possesses full control over organization settings, member invitations, team billing, audit logs, and memory management.
Staff: Authorized to revaluate document memory, review audit traces, and manage workspace documents.
Member: Permitted to upload documents, participate in assigned room chats, and run multi-agent workflows.
Viewer: Read-only access to room chats and verified citation references.

Intellectual Property & Output Ownership

We respect intellectual property rights and maintain clear delineations:

Customer Materials: You retain 100% sole ownership, copyright, and trade secret rights in all files, documents, text excerpts, and prompts you upload or provide to the platform.
AI Output Ownership: As between you and RAG Agent, you own all rights, title, and interest in and to the synthesis, reports, answers, and artifacts generated by the agent based on your inputs, to the fullest extent permitted by applicable law.
Platform Intellectual Property: RAG Agent retains all rights, title, and interest in our proprietary algorithms, agent coordination graphs, UI designs, codebases, APIs, and documentation.

Acceptable Use & Prohibitions

You agree not to engage in any prohibited activities, including but not limited to:

Reverse engineering, decompiling, or attempting to discover the source code of non-open components or model weights.
Uploading files containing malware, rootkits, viruses, or obfuscated malicious payloads.
Attempting to break out of the Python execution sandbox or executing unauthorized command-line system operations.
Bypassing multi-tenant room isolation or accessing workspaces without explicit invite tokens.
Generating deceptive deepfakes, automated phishing campaigns, hate speech, or malicious exploits.
Automating high-volume query flooding that degrades platform availability for other tenants.

Subscriptions, Mandates & Razorpay Payments

Paid plans (including Pro and Enterprise tiers) are offered on monthly and annual billing cycles:

Recurring Billing Mandates: Paid subscriptions utilize automated recurring billing handled securely via Razorpay. By upgrading to a paid tier, you authorize recurring debits until cancelled.
Taxes: All fees are exclusive of applicable Goods & Services Tax (GST), Value Added Tax (VAT), or sales taxes, which will be charged where required by local statutory law.
Price Modifications: We reserve the right to revise subscription prices upon at least 30 days advance written notice prior to your next renewal cycle.

Rate Limits & Fair Computing

To ensure optimal responsiveness, accounts are subject to tier-based fair usage policies:

File upload caps (default 25MB per document; customizable for Enterprise tiers).
Concurrent vector ingestion limits to prevent queue starvation.
Rate limits on streaming LLM token generation per minute. Excessive programmatic loops may trigger temporary throttles.

Service Availability & SLA

We target a 99.9% uptime Service Level Agreement (SLA) for paid enterprise tiers, excluding scheduled maintenance. Scheduled maintenance is conducted during low-traffic windows with at least 48 hours advance notice posted in the system dashboard.

Disclaimers & Probabilistic AI Outputs

Important Regulatory Notice on Generative AI:

⚠️

Probabilistic Nature of Large Language Models

RAG Agent employs retrieval grounding and citation validation to minimize hallucinations. However, outputs from Large Language Models are inherently probabilistic. The service is provided "AS IS" and "AS AVAILABLE". Outputs should not be solely relied upon for high-risk legal, medical, life-safety, structural engineering, or investment decisions without independent verification by qualified human professionals.

Limitation of Liability

To the maximum extent permitted by applicable law:

Neither party shall be liable for indirect, punitive, special, exemplary, incidental, or consequential damages (including loss of profits, goodwill, data, or business disruption).
RAG Agent's aggregate total liability arising out of or related to these Terms shall not exceed the total fees paid by customer to RAG Agent in the twelve (12) months preceding the event giving rise to liability.

Termination & Data Grace Period

You may terminate your account at any time via the billing portal. Upon termination, customer maintains a 30-day grace period during which all workspace documents, transcripts, and memory data may be exported. Following the grace period, all data is permanently purged.

Governing Law & Arbitration

These Terms shall be governed by and construed in accordance with the laws of Delaware, USA (or the applicable regional entity jurisdiction for international enterprise contracts), without regard to conflict of laws principles. Any unresolved disputes shall be settled through binding individual arbitration under the rules of the American Arbitration Association (AAA).

💳

14-Day Hassle-Free Satisfaction Guarantee

We want you to experience our grounded document retrieval and autonomous multi-agent system with complete confidence. First-time subscribers are entitled to a full 100% refund within 14 calendar days of plan initiation.

14-Day Money-Back Guarantee

If you upgrade to our Pro subscription tier and determine within 14 calendar days that the platform does not meet your business requirements, you may request a complete refund of your initial subscription payment. No interrogation, no convoluted hurdles.

Eligibility Criteria for Refunds

Refund requests will be promptly honored under any of the following circumstances:

First-Time Subscription Window: Request filed within 14 calendar days of the initial subscription charge.
Verified Service Disruptions: Platform downtime or core ingestion failures exceeding our 99.9% SLA threshold during a billing cycle.
Technical Incompatibility: Inability of the platform to index supported file formats (.pdf, .docx, .txt, .csv) due to verified backend bugs.
Billing Anomalies: Duplicate charges, incorrect plan assignment, or billing errors resulting from gateway synchronization anomalies.
Annual Plans: Annual plans cancelled within 30 days of renewal are eligible for a prorated refund for all unused months.

Ineligible Scenarios & Overages

Refunds will not be granted under the following conditions:

Excessive Quota Depletion: Accounts that have consumed more than 80% of allocated monthly LLM token quotas prior to filing a refund claim.
Terms of Service Violations: Accounts suspended or terminated as a result of malicious hacking attempts, automated abuse, or illegal material uploads.
Requests Past the Guarantee Window: Standard monthly renewals requested after the expiration of the 14-day window.
Repeated Cycles: Customers who previously claimed a refund and subsequently subscribed again under the same organization or billing credentials.

Subscription Cancellations & Mandate Halting

You can cancel your recurring subscription at any moment without penalty:

Navigate to the Plans & Pricing Page.
Locate your active subscription banner and select "Cancel Subscription / Stop Recurring Mandate".
Upon cancellation, your card will never be charged again. You retain full access to Pro features until the conclusion of your current paid billing period.

Step-by-Step Refund Claim Workflow

To submit a refund request, follow these straightforward steps:

📩

How to File a Claim

Email refunds@ragagent.io from your registered account email with:
1. Invoice Number (e.g. INV-2026-XXXX or Razorpay Payment ID)
2. Account Username / Organization Name
3. Brief Reason for Refund (optional feedback to help us improve)
Our billing desk reviews and responds to all requests within 24–48 business hours.

Razorpay Settlement Timelines

Once our billing desk approves your refund:

The refund command is dispatched immediately to the Razorpay Payment Gateway.
Funds are routed back exclusively to the original payment instrument (Credit Card, Debit Card, UPI, NetBanking).
Settlement Timing: Most domestic banks credit the funds within 5 to 7 business days. UPI refunds often settle within 24 to 48 hours.

Chargebacks & Dispute Policy

We strongly encourage reaching out to our billing desk prior to filing formal bank chargebacks. Direct refund processing via Razorpay is significantly faster (5–7 days) compared to bank dispute workflows, which can take 60–90 days. We are committed to fair and prompt resolution of all genuine billing inquiries.

🤖

Autonomous AI Governance & Safety Framework

This policy defines the engineering guardrails, ethical boundaries, and human oversight controls governing our agent collaborative team, Model Context Protocol (MCP) tool executions, and retrieval-grounded generation.

Principles & Governance Scope

At RAG Agent, our mission is to empower teams with grounded, verifiable intelligence. Our AI architecture is governed by four fundamental pillars:

Grounded Truth Over Hallucination: Models must justify assertions using verifiable citations from indexed documents. If source documents do not contain the answer, the model must explicitly state insufficient evidence.
Human Agency & Oversight: Autonomous agents must never perform destructive actions without human confirmation.
Zero Data Contamination: Proprietary context is ephemeral and isolated; it is never absorbed into public model weights.
Transparent Tool Execution: Every step taken by autonomous agents—including search queries, script executions, and API calls—must be logged in the inspection audit trail.

The Agent Collaborative Architecture

Our multi-agent system orchestrates a cooperative team of specialized autonomous agents. Each agent operates under explicit behavioral constraints:

Agent Role	Core Responsibility	Safety Constraint & Guardrail
0. Decision Maker	Evaluates user prompt intent, assesses query complexity, determines appropriate execution graph, and assigns sub-tasks.	Enforces organizational scope; rejects prompts violating acceptable use.
1. Researcher	Reformulates natural language queries into dense semantic search vectors to extract relevant document chunks from ChromaDB.	Constrained strictly to documents present in the active room workspace.
2. Task Generator	Deconstructs multi-step objectives into directed acyclic graphs (DAGs) of atomic, verifiable subtasks.	Limits workflow recursion depth to prevent runaway computational cycles.
3. Task Executor	Performs calculations, data transformations, and sandboxed script executions.	Runs in sandboxed Python environments with no raw system privilege.
4. Validator	Cross-references proposed responses against retrieved citation chunks, computing hallucination risk scores.	Mandates citation anchors; rejects ungrounded extrapolations.
5. Deployer	Synthesizes the validated answer, formats markdown and citations, and commits verified context to memory.	Applies privacy filters to strip PII before writing to long-term memory.

Human-in-the-Loop (HITL) Controls

Autonomous execution is subject to human-in-the-loop safeguards:

Confirmation Gates: Actions that modify system state, overwrite stored files, or invoke external paid APIs require explicit confirmation via UI dialog prompts.
Emergency Stop Button: Users can abort running multi-agent workflows at any instant via the real-time cancellation control.
Audit Logs: Every reasoning step, tool invocation, and decision rationale is logged in the Workflow Inspector for post-execution review.

Model Context Protocol (MCP) & Tool Safety

When utilizing Model Context Protocol (MCP) tools for external integrations:

Least Privilege: MCP connectors are configured with minimal required scopes (e.g. read-only file access by default).
Payload Sanitization: All tool inputs and returned data undergo structural validation to mitigate indirect prompt injection attacks.
Execution Isolation: Script executions take place in isolated sandboxes with strict resource caps on memory, CPU, and execution timeout.

Grounding & Hallucination Prevention

We mandate strict grounding standards across all generation pipelines:

Source Attribution: Responses must explicitly link statements to specific document chunk IDs. Users can click any citation pill to inspect the verbatim source text.
Insufficient Evidence Protocol: If the retrieved document context does not contain sufficient factual evidence to substantiate an answer, the model is strictly prompted to declare: "The indexed documents do not provide sufficient information to answer this query."

Dual-Tier Memory Ethical Rules

Our two-tier memory system operates under transparent governance:

Automated PII Scrubbing: Personal identifiers (e.g., social security numbers, credit card numbers, personal health identifiers) are automatically redacted before facts are committed to Long-Term Memory.
Memory Inspection & Purge: Organization members can view, edit, or delete any extracted memory fact at any time in the Memory Tab.
No Cross-Tenant Sharing: Memory items are strictly scoped to the tenant organization and are never shared across organization boundaries.

Strictly Prohibited AI Use Cases

The following activities are strictly prohibited when using RAG Agent:

Developing autonomous cyber weapons, malware, or vulnerability exploitation frameworks.
Deploying models for fully automated legal sentencing, high-stakes medical diagnosis, or credit worthiness decisions without human oversight.
Generating non-consensual deepfakes, synthetic CSAM, or targeted harassment materials.
Conducting biometric mass surveillance or facial identification harvesting.

Model Providers & Audit Telemetry

Users are always informed which foundation model is executing their query (e.g., Ollama Qwen2.5, Google Gemini, OpenAI GPT-4o, or Anthropic Claude). The active model is permanently displayed in the navigation header badge. Detailed telemetry—including token consumption, latency, and step logs—is provided in real time to ensure total transparency.