Skip to content

design_logic.md

Latest commit

 

History

History
92 lines (74 loc) · 3.4 KB

design_logic.md

File metadata and controls

92 lines (74 loc) · 3.4 KB
Raw

Treg Research Assistant - Google ADK & Vertex AI Design

1. System Overview

The system is a Multi-Agent System (MAS) built using Google Agent Development Kit (ADK) patterns and deployed on Google Vertex AI Platform.

It leverages Vertex AI Reasoning Engine for the agent runtime and Gemini 1.5 Pro/Flash for intelligence.

2. Architecture Diagram

graph TD
    subgraph "Client Layer"
        UI[React Frontend]
        API[FastAPI Gateway / Cloud Run]
    end

    subgraph "Vertex AI Platform"
        subgraph "Reasoning Engine - Runtime"
            Orchestrator[Orchestrator Agent - Gemini 1.5 Pro]
            Researcher[Researcher Agent - Gemini 1.5 Flash]
            Analyst[Analyst Agent - Gemini 1.5 Flash]
        end
        
        subgraph "Managed Services"
            V_Search[Vertex AI Search - PubMed/CT Index]
            V_Model[Vertex AI Model Garden - Gemini APIs]
        end
    end

    subgraph "Memory & State"
        Firestore[Firestore - Session History]
        VectorDB[Vertex AI Vector Search - Long-Term Memory]
    end

    %% Flows
    UI <--> API
    API --"Vertex AI SDK"--> Orchestrator
    
    Orchestrator <--> Firestore
    Orchestrator <--> VectorDB
    
    Orchestrator --"Delegate"--> Researcher
    Orchestrator --"Delegate"--> Analyst
    
    Researcher <--> V_Search
    Researcher <--> V_Model
    Analyst <--> V_Model
Loading

3. Component Mapping (ADK & Vertex AI)

A. Multi-Agent System (ADK Pattern)

  • Framework: Python (using google-cloud-aiplatform and ADK best practices).
  • Orchestrator:
    • Implemented as a Reasoning Engine application.
    • Uses Gemini 1.5 Pro for planning and delegation.
  • Sub-Agents:
    • Researcher: Specialized in information retrieval using Vertex AI Search.
    • Analyst: Specialized in data processing using Vertex AI Code Interpreter (if available) or local Python sandbox.

B. Tools (Vertex AI)

  • Knowledge Retrieval:
    • Instead of a local ChromaDB, we use Vertex AI Search (Agent Builder) to index PubMed/ClinicalTrials data.
    • Benefit: Managed, scalable, and semantic search out-of-the-box.
  • Google Search:
    • Use Vertex AI Grounding with Google Search.
  • Code Execution:
    • Use Gemini's Code Execution capability (built-in tool).

C. Sessions & Memory (ADK)

  • Session State:
    • Stored in Google Firestore (NoSQL), following ADK's "Session" schema.
  • Long-Term Memory:
    • Vertex AI Vector Search for storing user preferences and past successful experiments.

D. Observability (Vertex AI)

  • Tracing: Vertex AI TensorBoard or Cloud Trace for agent steps.
  • Evaluation: Vertex AI Evaluation Service (AutoSxS) to compare agent answers against a golden set.

E. Deployment

  • Runtime: Vertex AI Reasoning Engine.
    • Allows deploying the Python agent code as a managed service.
  • Frontend: Deployed on Google Cloud Run.

4. Implementation Steps (Revised)

  1. Setup: Enable Vertex AI, Firestore, and Cloud Run APIs.
  2. Data: Upload PubMed/CT JSON to Vertex AI Search datastore.
  3. Agent:
    • Define Orchestrator class using Vertex AI SDK.
    • Register tools (Search, Code).
    • Deploy to Reasoning Engine.
  4. Frontend: Update React app to call the Reasoning Engine endpoint (via a lightweight proxy if needed).