Implementation of a Multi-Agentic Workflow: Travel Advisor

What is Multi-Agentic Travel Advisor?

The Multi-Agentic Travel Advisor establishes a fundamental understanding of multi-agentic workflows within AI-Builder. This pipeline integrates simple API calls, tool invocation, and protobuf definitions to create a well-structured and orchestrated AI workflow for travel planning. This system aims to enable ongoing, dynamic travel interactions—starting from the initial planning phase and extending to post-trip experiences.

This framework can coordinate multiple AI agents to handle tasks such as:

• Destination discovery
• API integrations
• Real-time itinerary adjustments
• Local recommendations
• Post-trip feedback collection and analysis

A initial version of the workflow is as shown below,

System Design

Tasks

• Implement Bidirectional Pipeline (Ref. Sudoku Solver)
• Implement the Chatbot -> Planner -> Agent workflow
• Tool Retrieval Mechanisms
• Investigate and implement retrieval strategies such as: (Top-k Search, Graph-based sampling, Enhanced NLP techniques, Decision trees)
• Toolset Exploration
• Identify relevant tools for the use case
• Explore open-source API hubs for possible integrations
• Prompting Approaches
• Test and refine various complex prompts to improve system responses
• Automated Function Generation
• Develop a mechanism to auto-generate functions based on: The API structure and The required JSON output format
• Planner & Agent Orchestration Enhancements
• Improve multi-agent coordination using advanced techniques like: Chain-of-thought reasoning, Tree-mapping

added Work In Progress label

changed the description

assigned to @swetha1095

changed the description

marked the checklist item Implement the Chatbot -> Planner -> Agent workflow as completed

Update-07-04-2025

Initial Pipeline	Currently Implemented Bidirectional Setting
	In this pipeline, the shared folder is eliminated. We use gRPC communication to flow messages in the pipelines.
	Drawbacks Sometimes the streaming doesn't start automatically – need to check this.

Sequence Diagram

Sample Output

marked the checklist item Identify relevant tools for the use case as completed

marked the checklist item Explore open-source API hubs for possible integrations as completed

Update-28-04-2025

1. Model Context Protocol

• MCP is an open standard developed by Anthropic that enables AI models to interact with external tools and data sources in a structured manner.
• Set up the MCP server and client
• Typically, defining the functions (tools) is sufficient.
• MCP handles the provision of tool schemas to OpenAI or other LLMs

2. gRPC Reflection and gRPC to OpenAI

• The user needs to provide both the function and its corresponding protobuf file
• Reflection discovers the function, and grpc2openai (user provided -what we create) produces a tool schema from the protobuf file

Feature	MCP	gRPC Reflection
Standardization	Open standard by Anthropic	Utilizes gRPC reflection and OpenAI tools
Integration	Requires setting up MCP server and client	Depends on existing gRPC services and reflection
Tool Schema Generation	Defined through MCP specification	Generated from protobuf definitions
Accuracy	Accurate	Depends on how well the protobuf definitions are.

Update-05-05-2025

Update-12-05-2025

MCP-gRPC-LLM Flow:

1. User submits a query from the UI via SendUserQuery.
2. LLM identifies the agent_type and selects the appropriate FuncSchema.
3. In parallel, MCP tools are scanned and their FunctionSchemas are fetched via GetFuncSchemas.
4. As a next step, LLM picks the right tool for the matching agent_type and make the tool call again (ToolRequest).
5. The selected tool processes it and returns context via InvokeTool.
6. LLM uses this augmented context to generate a response, received through RecvCallTool.
7. Final answer is sent back to the UI via RecvResponse.

AI-Builder Pipeline