Product Mapping Guidance for Hybrid Search Solutions

Use the guidance in this file to do the following:

• Recommend appropriate products and features for the components of the hybrid search system
• Explain the rationale and trade-offs for each recommendation.

Database & Vector Search Store

Primary Recommendation: AlloyDB for PostgreSQL (AlloyDB AI)

• Description: Fully managed, PostgreSQL-compatible relational database designed for demanding enterprise workloads, equipped with built-in ScaNN vector indexing and Gemini Enterprise Agent Platform integration using google_ml_integration.
• Advantages:
  • ScaNN index algorithm delivers up to 4x faster vector search queries than standard HNSW/IVFFlat at high recall.
  • In-database vector embeddings (embedding('text-embedding-005', ...)), reranking (ai.rank), and LLM calls (ml_predict_row) eliminate data movement outside the database boundary.
  • Integrated evaluate_query_recall function provides automated, in-database accuracy measurement.
  • Single SQL query seamlessly combines vector similarity ordering with structured relational filters (WHERE category = ANY(...)).

Alternative 1: BigQuery Vector Search

• Description: Enterprise data warehouse and lakehouse platform featuring native hybrid search (AI.SEARCH with mode => 'HYBRID', and VECTOR_SEARCH with lexical_search_columns), batch embedding generation (AI.GENERATE_EMBEDDING), autonomous background embeddings (AI.EMBED), and ScaNN (TREE_AH) / IVF vector indexing (CREATE VECTOR INDEX).
• Advantages:
  • Scales to billions of rows across structured tables, unstructured documents (ObjectRef), and data lakehouses without infrastructure management.
  • Provides native SQL hybrid search functions: AI.SEARCH (TVF combining vector similarity with lexical search on autonomous embedding tables) and VECTOR_SEARCH (with lexical_search_columns).
  • Accelerates hybrid search with CREATE VECTOR INDEX supporting ScaNN (TREE_AH) and Inverted File (IVF) algorithms with STORING clauses for metadata pre-filtering and join elimination.
  • Includes Gemini in BigQuery for AI-assisted data exploration, conversational analytics with custom data agents, automated data insights, and natural language SQL/Python code generation.
  • Supports both batch embedding generation across very large tables via AI.GENERATE_EMBEDDING (TVF) and continuous autonomous maintenance via AI.EMBED stored generated columns.
  • Ideal for gigabyte to petabyte-scale analytical RAG, log analytics, audience segmentation, and batch hybrid search entirely in standard SQL (vs low-latency transactional OLTP in AlloyDB).
• Resources (Only put into context if the user selects this alternative product):
  • https://docs.cloud.google.com/bigquery/docs/vector-search-intro.md.txt
  • https://docs.cloud.google.com/bigquery/docs/gemini-overview.md.txt
  • https://docs.cloud.google.com/bigquery/docs/reference/standard-sql/bigqueryml-syntax-ai-search.md.txt
  • https://docs.cloud.google.com/bigquery/docs/vector-index.md.txt

Alternative 2: Cloud SQL

• Description: Fully managed relational database service offering native vector search (VECTOR data type, ScaNN TREE_SQ indexing, approx_distance and vector_distance functions) for MySQL, and pgvector for PostgreSQL.
• Advantages:
  • Lower entry cost and simpler operational setup for small to medium workloads (less than 10M rows).
  • Fully transactionally consistent, real-time ACID-compliant vector index updates during DML operations.
  • Supports ScaNN algorithm and iterative filtering (cloudsql_vector_iterative_filtering) for ANN similarity queries in Cloud SQL for MySQL.
• Resources (Only put into context if the user selects this alternative product):
  • https://docs.cloud.google.com/sql/docs/mysql/vector-search.md.txt
  • https://docs.cloud.google.com/sql/docs/mysql/integrate-cloud-sql-with-vertex-ai.md.txt

Vector Embeddings Engine

Primary Recommendation: Text embeddings on Gemini Enterprise Agent Platform

• Description: Managed text embedding model producing 768-dimensional dense vector representations optimized for semantic search and retrieval.
• Advantages:
  • Directly invocable from SQL inside AlloyDB, BigQuery, and Cloud SQL.
  • High semantic accuracy across multi-lingual and domain-specific retail search queries.
  • Fully managed scaling and zero model server maintenance.

Alternative 1: Self-hosted embedding model

• Description: Custom-deployed open-source or fine-tuned text embedding model hosted on Cloud Run or Gemini Enterprise Agent Platform.
• Advantages: Complete control over customized fine-tuned open-source models.
• Resources (Only put into context if the user selects this alternative product):
  • https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-garden/self-deployed-models.md.txt
  • https://docs.cloud.google.com/run/docs/deploying.md.txt

Database Abstraction & Agentic Tooling

Primary Recommendation: MCP Toolbox for Databases

• Description: Model Context Protocol (MCP) Toolbox is an open source Model Context Protocol (MCP) server that connects your AI agents, IDEs, and applications directly to your enterprise databases.
• Advantages:
  • Decouples database execution and complex hybrid search SQL logic from application code.
  • Exposes clean REST/MCP tool endpoints for AI agents and application services.
  • Supports agentic tool integration out of the box with custom tools that enforce tenant isolation boundaries.

Alternative 1: Managed Google Cloud Remote MCP Servers

• Description: Fully managed remote MCP servers provided by Google Cloud (managed via Agent Registry) offering built-in IAM governance, Model Armor security scanning, and pre-packaged toolsets for Google Cloud services.
• Advantages:
  • Zero server infrastructure management; built-in IAM authentication, fine-grained access policies, and Model Armor prompt injection protection.
  • Exposes tools, prompts, and resources via standardized MCP discovery (tools/list).
• Resources (Only put into context if the user selects this alternative product):
  • https://docs.cloud.google.com/mcp/overview.md.txt

Alternative 2: Remote Custom MCP Servers Hosted on Cloud Run

• Description: Custom MCP servers built using open-source SDKs (TypeScript, Python, Go, FastMCP) deployed on Cloud Run using streamable HTTP or Server-Sent Events (SSE) transport.
• Advantages:
  • Complete flexibility to implement custom tools, prompts, and resources tailored to hybrid search applications.
  • Supports serverless auto-scaling, sidecar deployment alongside AI agents, and service-to-service IAM authentication (roles/run.invoker).
• Resources (Only put into context if the user selects this alternative product):
  • https://docs.cloud.google.com/run/docs/host-mcp-servers.md.txt

Web Application Hosting

Primary Recommendation: Cloud Run

• Description: Fully managed serverless container platform that automatically scales compute instances based on incoming traffic.
• Advantages:
  • Scales to zero when idle; seamlessly handles traffic bursts.
  • Native integration with Direct VPC Egress for secure private connection to AlloyDB.
  • Supports any containerized web framework (Java Spring Boot, Python FastAPI, Node.js).

Networking & Security Topology

Primary Recommendation: Regional External Application Load Balancer and Direct VPC Egress

• Description: Edge load balancer providing Cloud Armor DDoS/WAF protection, routing traffic to Cloud Run services configured with Direct VPC Egress to communicate with AlloyDB over private IP ranges.
• Advantages:
  • Strict physical and network isolation for database tier.
  • Cloud Armor protection against SQL injection and web threats.