ClientMind is an intelligent customer service automation agent built for the Elasticsearch Agent Builder Hackathon. It leverages Elasticsearch's advanced search capabilities and LLM reasoning to automate messy e-commerce support workflows.

• Dual-Path RAG (Memory + Knowledge): Unlike standard RAG that only looks at documents, ClientMind queries structured customer history (ES|QL) and unstructured knowledge (ELSER) simultaneously.
• Memory Evolution: After each interaction, the agent extracts new customer insights and updates the Elasticsearch persona index, creating a continuous learning loop.
• Human-in-the-Loop: A professional admin dashboard for reviewing, editing, and approving AI-generated drafts before they are sent.
• High-Performance Retrieval: Leveraging Elasticsearch's unified platform to achieve sub-100ms context retrieval for complex reasoning.

graph TD
    %% Trigger Layer
    Start["1. Webhook (Email JSON)"] --> Extract["2. Extract Fields (Email, Subject, Body)"]

    %% Perception Layer
    Extract --> Classify["3. Intent Classification (LLM)"]
    Classify -- "Non-Sales" --> StandardFlow["4. Standard Template Reply"] --> End1((End))
    Classify -- "Sales Inquiry" --> ParallelStart((5. Dual-Path Retrieval))

    %% Memory Path
    ParallelStart --> MemQuery["5a. Build ES Query (ES|QL)"]
    MemQuery --> ES_Mem["6a. Search clientmind_persona"]
    ES_Mem --> MemResult["7a. Return: Preferences + Interaction History"]

    %% Knowledge Path
    ParallelStart --> KnowQuery["5b. Semantic Search (ELSER)"]
    KnowQuery --> ES_Know["6b. Search clientmind_knowledge"]
    ES_Know --> KnowResult["7b. Return: Relevant Doc Fragments"]

    %% Reasoning Layer
    MemResult --> ContextBuilder["8. Construct Prompt Context"]
    KnowResult --> ContextBuilder
    ContextBuilder --> SystemPrompt["9. Inject System Prompt (Rules & Persona)"]
    SystemPrompt --> LLM_CoT["10. Reasoning (Chain of Thought)"]

    %% Execution Layer
    LLM_CoT --> GenDraft["11. Generate Reply Draft"]
    GenDraft --> ShowUI["12. Push to Admin Dashboard"]

    %% Evolution Layer (Async)
    ShowUI -- "Admin Approval" --> SendMail["13. Gmail API Send"]
    SendMail --> ExtractNew["14. Extract New Insights"]
    ExtractNew --> UpsertES["15. Update ES (Memory Evolution)"] --> End2((Loop Closure))

The project includes a production-ready n8n workflow clientmind-workflow.json that acts as the "nervous system" of the agent:

1. Orchestration: It orchestrates Gmail triggers, intent classification, and the Dual-RAG retrieval API.
2. State Management: It uses n8n's Wait node to handle the human approval process, pausing the execution until the admin clicks "Approve" in the Dashboard.
3. Deployment: Simply import the JSON file into your n8n instance.

To make the workflow operational for judging:

1. Import Workflow: In n8n, go to "Workflows" -> "Import from File" and select clientmind-workflow.json.
2. Configure Credentials:
   • The Gmail Trigger and Reply to a message nodes will show a connection error.
   • Click on these nodes and create/select your own Gmail OAuth2 API credentials.
   • Ensure your Gmail API has https://www.googleapis.com/auth/gmail.modify scope.
3. Local API Connectivity:
   • The workflow points to http://localhost:3000. If running n8n in Docker, ensure it can reach your host machine (e.g., using http://host.docker.internal:3000).
4. Activate: Set the workflow to Active to begin polling for new customer inquiries.

• agent-builder/: Contains the core Agent definitions.
  • system_prompt.md: The "brain" of the agent, defining reasoning steps.
  • tools_dsl.json: Tool definitions using Elastic's ES|QL and Semantic Retrieval.
• clientmind-workflow.json: The complete n8n workflow for email automation.
• src/app/api/: Backend API routes for RAG, Draft Generation, and Classification.
• scripts/: Initialization scripts for Elasticsearch indices and mock data.

• Node.js 20+ & pnpm
• An Elasticsearch Cloud instance (8.15+)
• An OpenAI or Anthropic API Key

1. Clone the repository.
2. Install dependencies:

   pnpm install

3. Copy .env.example to .env and fill in your credentials.

Run the initialization script to create indices and import sample knowledge base data:

pnpm ts-node scripts/init-elastic.ts

pnpm dev

Watch the Demo Video (Please replace this with your actual video link before submission)

To use the Agent Builder with this project:

1. Open your Elastic Cloud Console.
2. Navigate to Agent Builder.
3. Copy the definitions from agent-builder/tools_dsl.json into the Tools section.
4. Paste the content of agent-builder/system_prompt.md into the Instructions.

To help the judging panel evaluate the integration depth, here is a mapping of our implementation to the Elasticsearch Agent Builder contest requirements:

• Dual-Path RAG Engine: Our Core RAG Logic implements a hybrid retrieval strategy. It executes ELSER-powered semantic search for unstructured knowledge and ES|QL pipeline queries for structured customer history simultaneously, achieving sub-100ms retrieval latency.
• Vector & Semantic Search: We utilize the clientmind_knowledge index with ELSER (v2) to ensure high-precision retrieval without the need for external embedding models.
• Structured Analytics (ES|QL): We use ES|QL to perform real-time data processing of customer personas, enabling the agent to reason based on historical order frequency and interaction sentiment.

• Native Agent Configuration: Detailed tool definitions and API integration schemas are located in agent-builder/tools_dsl.json, designed specifically for the Elastic Agent Builder environment.
• Sophisticated System Prompts: Our System Instructions utilize Chain-of-Thought (CoT) prompting, guiding the LLM through intent classification, context synthesis, and draft generation.

• Dynamic Persona Updates: Post-interaction, the system extracts new customer insights and "evolves" the memory stored in Elasticsearch, ensuring the agent becomes smarter with every email.
• n8n Orchestration: The complete "nervous system" of the agent is defined in clientmind-workflow.json, handling everything from Gmail webhooks to human-in-the-loop approvals.

• License: Licensed under Apache 2.0.
• Setup Scripts: Includes init-elastic.ts for one-click index creation and mapping configuration (BM25 + ELSER).

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.