Query Rewriting for Security

The authorization engine that evaluates access policies against the user's context and the requested resource's attributes to issue a permit or deny decision. In query rewriting, the PDP provides the boolean logic that the rewriter injects as metadata filters.

• Evaluates ABAC and RBAC policies in real-time
• Returns a set of scoping constraints (e.g., department=finance)
• Must respond in < 10ms to avoid adding latency to the retrieval pipeline
• Often implemented via Open Policy Agent (OPA) or XACML engines

An authorization technique where the search query is modified or the search space is restricted before the vector similarity search executes. Query rewriting is the primary implementation mechanism for this strategy.

• Injects metadata filters directly into the query payload
• Prevents unauthorized documents from ever entering the candidate set
• More efficient than post-retrieval filtering for large indexes
• Requires tight coupling with the vector database's filter syntax

The secure transmission of the end-user's authenticated identity context through the RAG pipeline layers. Without accurate identity propagation, query rewriting cannot apply user-specific permission scopes.

• Uses JWT tokens or SAML assertions passed via request headers
• Must survive microservice boundaries without stripping claims
• Critical for multi-tenant SaaS applications
• Failure mode: rewriting applies anonymous or default permissions, causing data leakage

A technique used in vector databases to restrict search results by applying boolean conditions to the key-value attributes associated with each document chunk. Query rewriting translates user permissions into these filter predicates.

• Common operators: $eq, $in, $and, $or, $nin
• Filters on fields like classification_level, project_id, region
• Combines with semantic similarity to narrow the search space
• Example: { "$and": [ { "clearance": "secret" }, { "department": "legal" } ] }

The principle of granting a RAG system or user only the minimum necessary data access required to answer a specific query. Query rewriting enforces this by dynamically scoping each retrieval to the user's current authorization boundary.

• Reduces the blast radius of prompt injection attacks
• Prevents lateral movement across document collections
• Requires just-in-time evaluation of permissions, not cached roles
• Aligns with NIST SP 800-207 Zero Trust Architecture principles

The systematic recording of every retrieval event, including the original user query, the rewritten query, and the documents accessed. This creates an immutable record proving that security policies were correctly applied during query rewriting.

• Logs must capture both pre-rewrite and post-rewrite query strings
• Essential for SOC 2 and GDPR compliance demonstrations
• Enables forensic analysis of potential data leakage incidents
• Should be stored in append-only storage with cryptographic integrity checks