# AI Travel & Policy Assistant

An AI-powered Travel & Policy Assistant for corporate travel and employee operations.

## Features

- Travel and expense policy question answering
- Retrieval-Augmented Generation (RAG)
- Semantic search using Sentence Transformers
- FAISS vector database
- Source attribution for policy answers
- Hallucination prevention
- Employee eligibility checking
- Travel request validation
- Reimbursement calculation
- Agentic routing between RAG, tools, and memory
- Conversation memory
- MCP server
- Flask conversational chatbot interface

## Technologies

- Python
- Pandas
- Sentence Transformers
- FAISS
- Flask
- Ollama
- Llama 3.2 3B
- MCP
- HTML, CSS, JavaScript
- Git and GitHub

## Architecture

User -> Flask Chatbot -> Agent Router

The Agent Router determines whether the request requires:

1. RAG for general policy questions
2. Employee tools for eligibility and travel validation
3. Reimbursement calculation
4. Conversation memory

RAG uses policy documents, document chunking, Sentence Transformer embeddings, FAISS retrieval, and an LLM.

## Policy Documents

The knowledge base contains seven fictional policy documents:

- travel_policy_india.txt
- travel_policy_us.txt
- expense_policy.txt
- employee_eligibility.txt
- cancellation_policy.txt
- approval_policy.txt
- airport_policy.txt

Employee information is stored in:

data/employees.csv

All data is fictional training data.

## RAG Pipeline

Policy Documents
-> Ingestion
-> Preprocessing
-> Chunking
-> Sentence Transformer Embeddings
-> FAISS Vector Index
-> Semantic Retrieval
-> Policy Context
-> LLM
-> Answer with Sources

Embedding model:

sentence-transformers/all-MiniLM-L6-v2

Embedding dimension: 384

## Agent Tools

The project provides three main tools:

### Employee Eligibility

Checks whether an employee is eligible for company travel.

### Trip Validation

Validates a travel request using employee and applicable policy information.

### Reimbursement Calculator

Calculates the reimbursable amount and excess amount.

## Memory

Conversation memory is stored in:

artifacts/memory.json

The memory module supports storing and retrieving relevant conversation information.

## MCP

The MCP server is located at:

mcp/server.py

Available MCP tools:

- employee_eligibility
- trip_validation
- reimbursement_calculator

## Flask Application

Start the application with:

python app/app.py

Then open:

http://127.0.0.1:5000

The interface provides a continuous chatbot conversation where previous questions and answers remain visible.

## Build Knowledge Base

To rebuild the FAISS knowledge base:

python scripts/build_knowledge_base.py

Generated artifacts:

- artifacts/chunks.json
- artifacts/policy.index

## Example Questions

What is the standard trip limit in India?

What happens if I exceed the travel limit?

Is EMP001 eligible for company travel?

Is EMP004 eligible for company travel?

Validate trip for EMP001 to Mumbai costing 1500.

Validate trip for EMP001 to Mumbai costing 2500.

Calculate reimbursement for 2500 with a limit of 2000.

What did we discuss earlier?

## Testing

Test scenarios are documented in:

tests/test_cases.md

Testing covers policy retrieval, semantic search, employee eligibility, trip validation, reimbursement calculation, memory, agent routing, hallucination prevention, Flask API behavior, and error handling.

## Project Structure

ai-travel-policy-assistant/

- app/
- artifacts/
- data/
- mcp/
- scripts/
- src/
- tests/
- README.md
- requirements.txt

## Disclaimer

This is a fictional educational training project. Employee records and company policies are sample data and should not be treated as real corporate policies or employee information.