/google-cloud-solution-rag-enterprise-search-gke-sqldb

Related documentation

Table of contents

• Architecture guides: Lines 12 to 44
• Deployment guides: Lines 45 to 55
• Product-specific best practices guides: Lines 56 to 80
• Product-specific decision-making guides: Lines 81 to 109

<a name="architecture-guides"></a>Architecture guides

Use the following reference architectures and design guides to design infrastructure for a RAG workload on Google Cloud:

• RAG infrastructure for generative AI using GKE and Cloud SQL: A flexible, container-based architecture that provides maximum control to build custom applications with open source tools such as Ray, Hugging Face, and LangChain.
• RAG infrastructure for generative AI using Gemini Enterprise and Agent Platform: An agent-driven architecture that uses Gemini Enterprise as a unified platform to orchestrate an end-to-end RAG dataflow for enterprise applications that require real-time data availability and enriched contextual search.
• RAG infrastructure for generative AI using Agent Platform and Vector Search: A fully managed, serverless architecture that provides optimized, high-performance vector search for large-scale applications.
• RAG infrastructure for generative AI using Agent Platform and AlloyDB for PostgreSQL: An architecture that stores vector embeddings alongside your operational data in a fully managed database like AlloyDB for PostgreSQL.
• Harness CI/CD pipeline for RAG applications: An architecture for a continuous integration (CI) and continuous deployment (CD) pipeline for a RAG application in Google Cloud.
• Private connectivity for RAG-capable generative AI applications: Guidance to configure private endpoints and private service connections for secure RAG data ingestion and processing.
• Networking for AI inference model serving on GKE: Guidance to structure load balancers and private VPC subnets to serve AI models on GKE securely.
• RAG infrastructure for generative AI using Gemini Enterprise and Vector Search: Design considerations and guidelines for leveraging the Gemini Enterprise Agent Platform with Vector Search.

<a name="deployment-guides"></a>Deployment guides

Use the following resources to generate the deployment guidance:

• Example blueprints and Notebooks for RAG on GKE: https://gke-ai-labs.dev/docs/blueprints/rag-on-gke/
• Reference architecture for running inference workloads on GKE: https://github.com/GoogleCloudPlatform/accelerated-platforms/blob/main/docs/platforms/gke/base/use-cases/inference-ref-arch/README.md
• Build a RAG chatbot with GKE and Cloud Storage: https://docs.cloud.google.com/kubernetes-engine/docs/tutorials/build-rag-chatbot.md.txt

<a name="product-specific-best-practices-guides"></a>Product-specific best practices guides

The following guides provide optimization guidelines and operational best practices for the products used in this RAG architecture.

• Optimize GKE for machine learning inference: Best practices for optimizing ML inference workloads, reducing cold-start latency, and sizing compute profiles.
• Best practices for running cost-effective applications on GKE: Best practices for running cost-effective, containerized Kubernetes applications.
• Cloud SQL for PostgreSQL: Example of an embedding workflow Guide with example of embedding workflow using an AI embedding model as a service.
• Generate and manage auto vector embeddings for large tables in AlloyDB: Guide about embedding generation in AlloyDB.
• Build generative AI applications using AlloyDB AI: AI tools and options to build a RAG workflow using AlloyDB.
• Parallel composite uploads strategy: Guide to designing high-throughput object upload pipelines using parallel composite uploads.
• GKE Autopilot security measures: Security best practices and posture configurations for GKE Autopilot clusters.

<a name="product-specific-decision-making-guides"></a>Product-specific decision-making guides

Use the following comparison guides to evaluate design trade-offs and options in the GKE and AlloyDB RAG stack.

• Choose GKE cluster mode: GKE Autopilot versus GKE Standard modes of operation.
• GKE Autopilot compute classes: Predefined ComputeClasses for GKE Autopilot Pods (Balanced, Scale-Out, GPU options like L4).
• Agent Platform RAG Engine: vector database choices:
• Choose a connectivity option for AlloyDB: Compares AlloyDB connectivity methods, including AlloyDB Auth Proxy, Private Service Connect (PSC), and direct VPC peering, to help you select the appropriate method.
• Choose an AlloyDB machine type: Explains how to select the appropriate machine type for your AlloyDB primary and read instances based on vCPUs, memory, and performance.
• Choose a vector index in AlloyDB AI: Compares indexing strategies to help you choose the appropriate model for vector search. When to use Cloud SQL pgvector vs. Agent Platform Vector Search.
• PostgreSQL pgvector indexing strategy: Comparing HNSW and IVFFlat indices in pgvector for approximate nearest neighbor (ANN) search.
• Cloud SQL for PostgreSQL configuration choices: Selecting machine series, storage type, and other configuration options for Cloud SQL for PostgreSQL instances.