Dynamic Hybrid Search using AlloyDB

This skill provides a workflow to design and implement secure, low-latency, and high-accuracy hybrid search solutions combining structured dataset filtering, vector search indexing, faceted metadata filtering, semantic reranking, recall evaluation, in-database AI validation, database abstraction layers, and serverless application hosting.

Overview of the workflow

The workflow consists of the following phases:

1. Requirements discovery. Gather detailed requirements related to the cloud workload or use case that the user needs assistance for.
2. Solution architecture. Use the requirements that were gathered in Phase 1 to generate a detailed solution architecture for the cloud workload or use case.
3. Solution validation. Create a plan to validate the generated solution, generate validation instructions and scripts, and run the validation.
4. Solution packaging and presentation. Consolidate the generated content and present the solution.

Important notes about the workflow:

• Strict phase separation: During Phase 1 (Requirements discovery), when you ask the user clarifying questions, DON'T recommend, propose, or outline any architectural designs, cloud services, or component mappings. This prevents premature architecture commitments or hallucinations before the full scope is understood.
• Halting for approval: For any step where you are instructed to "obtain approval before proceeding", you MUST stop executing, present the completed tasks to the user, and wait for their explicit approval. You MUST NOT proceed to execute any subsequent tasks or generate any further guidance in that response.
• Ground all generated content: For all tasks across all phases, you MUST first look in the following resources:

  • Product Mapping,

  • Design Recommendations for the required guidance. If the guidance does not provide the required information, you MUST ground the generated content by using the following resources:

  • Google Developer Knowledge MCP server: https://developers.google.com/knowledge/mcp.md.txt

    • Server: https://developerknowledge.googleapis.com/mcp
      • Tools:
        • developerknowledge:search_documents
        • developerknowledge:get_documents
        • developerknowledge:answer_query

  • Relevant skills from https://github.com/google/skills

  • Official Google Cloud documentation in Related Guidance

Product Renaming & Terminology

When generating solution designs, architecture diagrams, and documentation, check the latest Google Cloud documentation for the most up-to-date product names. The table below provides examples of name mappings to be aware of. Note that underlying APIs, Terraform resources, and IAM roles may retain their legacy identifiers.

Legacy Name	Updated Name	Notes
Vertex AI	Gemini Enterprise Agent Platform	Gemini Enterprise Agent Platform can be shortened to Agent Platform after first instance
Vertex AI Embedding	Text embedding on Gemini Enterprise Agent Platform	This refers to the text embedding models available on Gemini Enterprise Agent Platform
Vertex AI Matching Engine	Vector Search

Phase 1: Requirements discovery

In this phase, you must gather detailed requirements related to the hybrid search workload that the user wants to design and deploy in Google Cloud.

Acknowledge provided requirements: If the user's prompt already contains some requirements (functional or non-functional, such as catalog size, search modalities, faceted attributes, or latency targets), you MUST explicitly acknowledge and restate all of these requirements in your response. Do NOT ask the user to describe or re-describe any requirements that they have already provided in the prompt.

Complete the following steps strictly in the specified order:

• Step 1: Ask the user to describe the functional requirements of the workload, including catalog dataset details (e.g., e-commerce apparel, retail products, patent database), search modalities (natural language text, visual search, attribute filters), metadata attributes for faceted filtering (e.g., category, sub_category, color, gender, price), and quality checks (reranking, LLM validation).

• Step 2: You MUST explicitly ask the user to describe ALL of the following six categories of non-functional requirements. You need this information because each category represents a critical architectural pillar, and neglecting any of them can result in a solution that is insecure, unreliable, or inefficient (do NOT omit any of them):

  • Security, privacy, and compliance: E.g., private VPC endpoints, Private Service Connect, Direct VPC Egress, and access control.
  • Reliability: E.g., high availability, failover, disaster recovery goals (RTO/RPO), regional vs multi-region AlloyDB topology.
  • Cost: E.g., budget constraints for compute, database instances, and Gemini Enterprise Agent Platform API calls.
  • Operational excellence: E.g., monitoring, logging, dashboards, and automated deployment.
  • Performance: E.g., target P95 query latency (e.g., < 100ms), vector search recall target (e.g., > 95%), catalog item scale, and QPS expectations.
  • Sustainability: E.g., carbon footprint, low-carbon region selection.

• Step 3: Ask the user whether the workload currently runs on other cloud providers or on-premises.

  • If the user's answer is "yes", then ask the user to describe the architecture of the current deployment.
  • If the user's answer is "no", then proceed to the next step.

• Step 4: Ask the user to describe dependencies, if any, on other workloads, products, or tools (e.g., existing inventory databases, ERP systems, application runtime languages like Java or Python).

• Step 5: Review the input that the user has provided so far, and check whether there are any ambiguities, conflicts, or contradictions in the functional requirements, non-functional requirements, and dependencies. You MUST compare all requirements against each other to identify any conflicts.

Important: STOP, DON'T proceed to generate architecture diagram, architecture description or product recommendations until you have confirmed the generated requirements summary and resolved all ambiguities and contradictions in this phase.

Phase 2: Solution architecture

Task 2.1: Identify Google Cloud products and features required for the workload.

• Step 1: Recommend products and features that are appropriate for each component of the user's workload, prioritizing Google Cloud products.

  Important: The Google Cloud products and features that you recommend MUST be consistent with the guidance in Product Mapping.

• Step 4: Present the generated product recommendations to the user and obtain approval (the user MUST explicitly say "yes" or "I approve") before proceeding to Task 2.2.

  Important: STOP, DON'T proceed to generate architecture diagram until you have confirmed the generated product recommendations with the user.

Task 2.2: Generate an architecture diagram and description

• Step 1: Generate an architecture diagram in the Mermaid format: https://github.com/mermaid-js/mermaid.

  The diagram must show the data flows and request flows across the components of the architecture, based on the gathered requirements and product recommendations. The diagram MUST explicitly show both the ingestion pipeline and serving pipeline.

  The following is an example of the data flows and request flows that the architecture diagram should show:

  • Ingestion pipeline: Catalog Data -> AlloyDB Table (apparels) -> B-Tree Indexes on Facets -> Text embedding (text-embedding-005) -> ScaNN Vector Index.
  • Serving pipeline: User Browser -> Cloud Run Web App -> MCP Toolbox for Databases -> AlloyDB Single-Query Hybrid Search (ScaNN Vector Search + SQL WHERE Filters) -> ai.rank Reranker -> Gemini Pro ai.generate Quality Validation -> Validated Results -> User Browser.

• Step 2: Generate a description that explains the purpose of each component, the relationships between the components, and the task flow or data flow.

• Step 3: Present the generated architecture diagram and description to the user and obtain approval (the user MUST explicitly say "yes" or "I approve") before proceeding to Task 2.3.

  Important: STOP, DON'T proceed to generate design recommendations until you have confirmed the generated architecture description with the user.

Task 2.3: Generate design recommendations.

• Step 1: Generate design recommendations and best practices to optimally configure each component in the architecture based on the workload requirements.

  Important:

  • When you generate design recommendations, consider the following:
    • Functional requirements that were gathered in Phase 1.
    • Non-functional requirements that were gathered in Phase 1.
  • Align the generated design recommendations with the recommendations in Design Recommendations.
  • To generate guidance for the non-functional requirements, use the following skills:
    • google-cloud-waf-security
    • google-cloud-waf-reliability
    • google-cloud-waf-cost-optimization
    • google-cloud-waf-operational-excellence
    • google-cloud-waf-performance-optimization
    • google-cloud-waf-sustainability

• Step 2: Present the generated recommendations to the user and obtain approval (the user MUST explicitly say "yes" or "I approve") before proceeding to Task 2.4.

  Important: STOP, DON'T proceed to generate deployment guidance until you have confirmed the design recommendations with the user.

Task 2.4: Generate deployment guidance.

• Step 1: Generate guidance to deploy the solution, including the following:

  • AlloyDB DDL & SQL setup scripts for extensions (google_ml_integration, alloydb_scan), tables, B-Tree indexes, ScaNN vector indexes, hybrid search SQL, and Gemini validation CTEs.
  • MCP Toolbox deployment configuration on Cloud Run.
  • Python Cloud Run Function shim deployment command.
  • Application deployment command (gcloud run deploy {app_name}).
  • Terraform code or gcloud CLI commands to create required infrastructure.

  Important: The deployment guidance that you generate MUST be consistent with the guidance in the following resources:

  • Related Guidance
  • Relevant skills in https://github.com/google/skills/tree/main/skills/cloud

• Step 2: Present the generated deployment guidance to the user and obtain approval (the user MUST explicitly say "yes" or "I approve") before proceeding to Phase 3.

  Important: STOP, DON'T proceed to generate solution validation until you have confirmed the deployment guidance with the user.

Phase 3: Solution validation

Task 3.1: Pre-deployment validation

• Step 1: Create a pre-deployment plan to statically validate the generated solution and verify that it meets the workload requirements without provisioning live resources:
  • Deployment dry-run: Validate infrastructure syntax and preview the resources that will be provisioned using dry-run commands (e.g., terraform plan or (where supported) gcloud ... --dry-run).
  • Architecture & policy analysis: Perform static verification of network routing topologies, firewall rules, and IAM enforcement against best practices.
• Step 2: Present the static validation plan to the user, obtain approval (the user MUST explicitly say "yes" or "I approve"), and execute the dry-run commands.
• Step 3: Troubleshoot and fix any errors or policy discrepancies identified during dry-run checks until validation succeeds.
• Step 4: Proceed to Task 3.2

Task 3.2: Runtime validation (Post-deployment)

• Step 1: Ask the user whether they choose to deploy the infrastructure now to perform live runtime verification, or skip directly to Phase 4.
• Step 2: If the user chooses to deploy the infrastructure:
  • After the user deploys the infrastructure, generate runtime verification commands (using tools like curl, ping, or gcloud) and provide them to the user to execute, to test live endpoint reachability, networking paths, and load balancer routing.
  • Troubleshoot any deployment or runtime routing issues until checks pass.
• Step 3: Proceed to Phase 4.

Phase 4: Solution packaging and presentation

• Step 1: Consolidate the final text artifacts that were generated in Phase 2 into a single Markdown file named solution-architecture-guide.md, based on the template in Output Template.
• Step 2: Request the user's permission to write the code files in the user's workspace.
• Step 3: After the user gives permission, write the final code files in the user's workspace.