Design recommendations

Important: The recommendations below apply to the primary recommended products in references/product-mapping.md. If an alternative product is selected, use the resources listed under that product's section in references/product-mapping.md to generate appropriate design recommendations.

Use the following guidance to generate design recommendations for a secure, scalable, and high-performance hybrid search solution combining structured relational queries, full-text search, and vector similarity search on Google Cloud.

• Security, privacy, and compliance:

  • Encrypt AlloyDB database data at rest using Customer-Managed Encryption Keys (CMEK) and restrict network access by creating a service perimeter with VPC Service Controls (VPC-SC).
  • Combine IAM least-privilege roles (roles/alloydb.client, roles/alloydb.viewer) with PostgreSQL database-native granular roles (CREATE ROLE) to prevent destructive commands (such as DROP TABLE or TRUNCATE) even if an agent identity is compromised.
  • Use AlloyDB Auth Proxy or Language Connectors for IAM-based connection authorization and TLS 1.3 encrypted data transit.
  • Prevent PostgreSQL search path hijacking by setting ALTER ROLE user_name SET search_path = pg_catalog, pg_temp; for privileged database roles to bypass untrusted public schemas.
  • Disable default run.app ingress URLs for Cloud Run services, routing public traffic exclusively through a regional external Application Load Balancer protected by Cloud Armor WAF rules (sqli-v33-stable and XSS protection).
  • Mitigate multi-tenant prompt injection data leaks by deploying specific custom tools (e.g., lookup_active_order via MCP Toolbox) where tenant identity filters are enforced in backend code outside LLM prompt control, avoiding generic open-ended execute_sql access.
  • Enforce action-selection patterns (hardcoded Action-Selector or Dual-LLM guardrail pre-screening) and strict tool allowlists to prevent unauthorized tool chaining and control-flow manipulation.
  • Enable Model Armor on AI endpoints (gcloud ai endpoints update --enable-model-armor) to sanitize MCP tool inputs and outputs, and configure Sensitive Data Protection (DLP) de-identification templates (model_armor_config.json) to redact PII automatically.
  • Enable Data Access audit logging in Cloud Logging for MCP tools and AlloyDB services, logging raw LLM SQL commands, tool parameters, user IDs, and session IDs to detect exfiltration attempts via log-based alerts.

• Reliability:

  • Provision AlloyDB primary instances with automatic multi-zone high availability (HA) failover and configure microsecond Point-In-Time Recovery (PITR) continuous backups to recover from accidental or malicious data corruption by autonomous agents.
  • Deploy Cloud Run web services and shims across multiple availability zones within a region to ensure automatic load-balancing and zone outage resilience.
  • Use regional or dual-region Cloud Storage buckets combined with Pub/Sub message flow control and retry policies to handle ingestion spikes reliably.

• Operational excellence:

  • To minimize recall degradation and reduce manual interventions, configure ScaNN vector indexes with automatic maintenance enabled.
  • Use AlloyDB Query Insights and System Insights dashboards to monitor query latency, replication lag, and peak database connection counts.
  • Route application logs to Cloud Logging in structured JSON format and trace complete end-to-end execution paths across services using Cloud Trace.

• Cost optimization:

  • Utilize Committed Use Discounts (CUDs) for predictable AlloyDB instance compute workloads, or basic instances for non-production development environments.
  • Configure Cloud Run services to scale to zero instances when idle to eliminate compute costs during off-peak hours.
  • Apply Object Lifecycle Management policies to Cloud Storage buckets to automatically archive or delete temporary ingestion data.
  • Configure Cloud Logging exclusion filters to drop non-critical debug logs and reduce log storage costs.

• Performance efficiency:

  • Configure ScaNN index parameters based on dataset size and tree depth:
    • Two-Level Tree: Set num_leaves = sqrt(rows) for balanced build speed, or rows/100 for optimal quality.
    • Three-Level Tree: Set max_num_levels = 2 and num_leaves =power(rows, 2/3) (balanced) or rows/100 (quality).
    • Four-Level Tree: Set scann.max_allowed_num_levels = 3, max_num_levels = 3, and num_leaves = power(rows, 3/4).
  • Improve search recall on highly selective filters by setting scann.satisfy_limit = 'relaxed_order' (streaming) and capping partition search using scann.max_pct_leaves_to_search = 15.
  • For high-dimensional embeddings (>= 500 dimensions), tune scann.pre_reordering_num_neighbors (default 50 * K) to refine re-ranking precision.
  • Attach Direct VPC Egress to Cloud Run services to establish low-latency, private IP network paths to the AlloyDB instance without internet hops.

• Sustainability:

  • Utilize serverless compute platforms (Cloud Run and Cloud Run functions) to automatically scale compute down when idle, minimizing energy consumption.
  • Execute vector distance calculations, full-text ranking, and embedding generation directly inside AlloyDB to avoid unnecessary cross-service network egress and duplicate compute cycles.