IHI Assistant

Table of Contents

• Introduction
• Features
• Technologies
• Getting started
• Configuration
• Project Structure
• License Copies
• Disclaimer
• Author

Introduction

The IHI Assistant is a Retrieval-Augmented Generation (RAG) AI application for internal use. It uses both local document search and optional web augmentation to generate precise context-aware answers.

---

Features

• Ingests documents from shared drives (PDF, DOCX, PPTX, TXT, CSV)
• Hybrid retrieval using BM25 + FAISS with small and large chunking strategies
• Frontend configuration for model, temperature, tone, reranking
• Optional web search via Bing-compatible API
• Local inference using mistralai/Mistral-7B-Instruct-v0.1
• Frontend built with Next.js, Tailwind CSS, and Lucide Icons

---

Technologies

Backend

• Core: fastapi, uvicorn, pydantic, pydantic-settings
• RAG stack: langchain, langchain-core, langchain-community, langchain-huggingface, langchain-text-splitters
• Transformers + embeddings: transformers, torch, sentence-transformers, faiss-cpu, rank-bm25
• Parsing: pdfplumber, pypdf, docx2txt, python-pptx, pandas, pillow
• Utilities: tqdm, requests, psutil, PyYAML, python-dotenv, regex, scikit-learn, scipy, sentence-transformers, threadpoolctl

Frontend

• next, react, tailwindcss
• lucide-react (icon set)
• Google Fonts: Geist, Geist Mono

---

Getting Started

Note: This project is not hosted online. It is intended for local deployment on internal servers only.

1. Backend

Requires Python 3.11

# Navigate to the project directory
cd rag_model

# Set up virtual environment
python -m venv env
env\Scripts\activate.ps1

# Navigate to the backend directory
cd backend

# Upgrade pip
python -m pip install --upgrade pip

# Install dependencies
python -m pip install -r requirements.txt

# Start the FastAPI server for local use
uvicorn scripts.main:app

# Start the FastAPI server for development
uvicorn scripts.main:app --reload

# Start the FastAPI server for LAN hosting
uvicorn scripts.main:app --host 0.0.0.0 --port 8000

# Access backend at localhost:8000/docs

2. Frontend

# Navigate to frontend directory

cd rag_model/frontend
npm install
npm run dev

# Access the app at localhost:3000

NOTE

The app also contains start.bat, frontend.bat, and backend.bat files for development convenience. These can be run simply:

# Navigate to project directory
cd rag_model

# Example usage (start.bat will run both front- and backend)
start.bat

Configuration

The assistant supports real-time adjustment of LLM parameters via the floating SettingsMenu:

• Model: Switch between supported Hugging Face models
• Temperature: Control generation creativity
• Tone: Choose between formal, neutral, casual These are sent via POST /set-config to the FastAPI backend.

Project Structure

/backend
├── scripts/
│   ├── chunk_documents.py          # Document chunking and processing
│   ├── config.py                   # Prompt templates & constants
│   ├── file_readers.py             # File reading functions
│   ├── handler.py                  # Routes queries by intent (math, code, general inquiry, etc)
│   ├── hybrid_retrievers.py        # Retrieves content based on hybrid retriever
│   ├── llm_utils.py                # Model loading and prompt handling
│   ├── load_utils.py               # Network share and document ingestion
│   ├── main.py                     # FastAPI app
│   ├── rag.py                      # Main pipeline logic
│   ├── retriever_builder.py        # Constructs and stores retrievers with persistance
│   └── utils.py                    # Models

/frontend
├── app/
│   ├── layout.tsx              # Global layout
│   ├── page.tsx                # Main chat page
│   ├──components/
│   │   ├── chat.tsx                # Message list + chat input
│   │   ├── ContextWindow.tsx       # Displays retrieved content
│   │   ├── DocumentUpload.tsx      # Document uploading window
│   │   ├── Dropdown.tsx            # Selection dropdown menu
│   │   ├── LoginForm.tsx           # Login form for LDAP authentication
│   │   ├── SettingsMenu.tsx        # Model configuration menu
│   │   ├── Sidebar.tsx             # Chat list + new chat
│   │   ├── Slider.tsx              # Sidebar slider component
│   │   └── Toggle.tsx              # Sidebar option toggle component

Development Configuration

config.yaml

Create file "config.yaml" inside of backend/

DOCUMENTS (recommended) will be all files the RAG Pipeline will have access to.

BING_API_KEY (optional) is the user or organization API Key for the Bing search API.

MODEL_TOKEN (required) is the user or organization HuggingFace token which enables model usage.

IGNORE_FOLDERS (optional) will be ignored paths inside of DOCUMENTS that the RAG Pipeline will not have access to.

IGNORE_KEYWORDS (optional) are a list of keywords that will be ignored if contained in the path or filename.

ldap (required, but will be made optional) is the LDAP configuration for your organization. NOTE: ldap contains server, user, password, search_filter, and base_dn parameters.

Parameter	Description
server	LDAP server hostname or URI (e.g. ldap://host)
user	DN or user principal for bind (cn=John,... )
password	Bind password
search_filter	RFC 4515 filter for selecting entries
base_dn	Starting DN for search (e.g. dc=example,dc=com)

Consult LDAP documentation for more information: https://ldap3.readthedocs.io/en/latest/

mongo_uri (required) is the mongo port for database access

sharepoint (required, but will be made optional) is the on-prem SharePoint credentials for organization file access. NOTE: sharepoint contains domain, user, and password parameters.

config.yaml example:

DOCUMENTS:
  - /Users/User/SampleDocuments/

BING_API_KEY: None

MODEL_TOKEN: my_model_token

IGNORE_FOLDERS:
  - /Users/User/SkippedDocuments/

IGNORE_KEYWORDS:
  - skip
  - archive

ldap:
  server: "my_ldap_server"
  user: "my_ldap_user"
  password: "password"
  search_filter: "my_ldap_filter"
  base_dn: "my_ldap_base_dn"

mongo_uri: mongodb://localhost:12345

sharepoint:
  domain: "sp_domain"
  user: "sp_user"
  password: "password"

.env

Create file .env in frontend/

NEXT_PUBLIC_HOST_IP (required) is the IP of the hosting server.

This will look like:

NEXT_PUBLIC_HOST_IP: my_host_ip
NEXT_PUBLIC_DOMAIN=my_email_domain
NEXT_PUBLIC_SERVER=my_email_server

License Copies

A consolidated list of licenses for third-party libraries is provided in NOTICE.txt.

You must preserve these notices and licenses if redistributing this project or using it in a commercial product.

Disclaimer

This software is provided "as-is" and intended for internal use only. External distribution or commercial deployment requires verification of all third-party licensing terms.

Author

Developed by Jacob Ma SWE/ML Intern @ IHI, Summer 2025 Contact: jma443@gatech.edu