Autonomous AI Research Agent

Md Jisan Hossen — B.Sc. Artificial Intelligence, NUIST

A ReAct agent that decides on its own whether to answer from your documents (RAG over FAISS) or from the live web — with multi-turn memory and a fully exposed reasoning trace.

---

Motivation

Most "chat with your PDF" demos do exactly one thing: retrieval. Real research questions mix private context ("what did this paper claim?") with live information ("what changed in LangChain this month?"). This project builds an agent that routes between the two sources autonomously using LangChain's ReAct (Reason + Act) loop, and — importantly for trust and debugging — surfaces every intermediate thought/action step in the UI rather than hiding the chain of reasoning.

---

What It Does

An end-to-end autonomous AI agent that can:

• Search the web in real time (DuckDuckGo, no API key needed)
• Answer questions from your documents (PDF, DOCX, TXT) using RAG
• Remember the conversation with multi-turn memory
• Reason step-by-step using LangChain's ReAct agent framework
• Interactive UI via Streamlit + REST API via FastAPI

---

Architecture

User (Streamlit UI)
       │
       ▼
FastAPI Backend  (/chat, /upload, /reset)
       │
       ▼
LangChain ReAct Agent  (gpt-4o-mini)
  ├── Tool 1: DocumentRetriever  →  FAISS Vector Store (local embeddings)
  └── Tool 2: WebSearch          →  DuckDuckGo
       │
       ▼
ConversationBufferMemory  (multi-turn context)

---

Setup Instructions

Step 1 – Clone the repository

git clone https://github.com/j1s4nn/AI_Agent.git
cd AI_Agent

Step 2 – Activate your conda environment

conda activate ml

Step 3 – Install dependencies

pip install -r requirements.txt

First run downloads the MiniLM embedding model (~90 MB). This is a one-time download.

Step 4 – Set your OpenAI API key

Copy the example environment file and add your key:

cp .env.example .env

Edit .env and replace the placeholder:

OPENAI_API_KEY=sk-your-actual-key-here

Get a key at → https://platform.openai.com/api-keys
(gpt-4o-mini costs ~$0.0002 per query – very cheap)

Step 5 – Run the project

python run.py

This starts both servers simultaneously:

Service	URL
Streamlit UI	http://localhost:8501
FastAPI Docs	http://localhost:8000/docs

---

How to Use

1. Open http://localhost:8501 in your browser
2. Upload documents (PDF/DOCX/TXT) using the sidebar → click Ingest Documents
3. Ask questions – the agent will decide whether to search your docs or the web
4. View reasoning – click "View reasoning steps" under any response to see the agent's thought process
5. Reset memory – click "Clear Chat Memory" to start fresh

---

File Structure

AI_Agent/
│
├── app/
│   └── main.py            ← FastAPI backend (API endpoints)
│
├── utils/
│   ├── agent.py           ← LangChain ReAct agent + memory
│   ├── tools.py           ← WebSearch + DocumentRetriever tools
│   ├── vector_store.py    ← FAISS ingestion & retrieval
│   └── config.py          ← Environment config loader
│
├── data/
│   └── vectorstore/       ← Auto-created; stores FAISS index
│
├── docs/
│   └── sample_query.txt   ← Sample document to test with
│
├── streamlit_app.py        ← Streamlit frontend UI
├── run.py                  ← One-command launcher
├── requirements.txt        ← All dependencies
├── .env.example            ← Template for environment variables
├── .env                    ← Your actual API keys (DO NOT commit)
├── .gitignore
└── README.md

---

Example Queries to Try

After uploading docs/sample_query.txt:

"What is RAG and how does it work?"
"What are the benefits of retrieval-augmented generation?"
"Search the web for the latest LangChain updates"
"Compare RAG with fine-tuning an LLM"

---

System Requirements

Component	Minimum	This Project Tested On
RAM	8 GB	16 GB
GPU	Not required	RTX 3060
Python	3.10+	3.10+
OS	Windows/Linux/Mac	Windows 11

---

Technology Stack

Python · LangChain (ReAct agent) · OpenAI API (gpt-4o-mini) · FAISS (vector store) · sentence-transformers (all-MiniLM-L6-v2 embeddings) · DuckDuckGo Search · FastAPI · Uvicorn · Streamlit · pypdf / python-docx (document loaders)

Evaluation & Limitations

This is a service rather than a benchmarked model, so there is no accuracy table. It was validated functionally: the agent correctly routes document questions to the RAG tool and live questions to web search, preserves multi-turn context, and exposes a complete reasoning trace. Known limitations: retrieval quality depends on the MiniLM embedding and chunking strategy; the agent inherits gpt-4o-mini's reasoning limits; and there is no automated end-to-end evaluation harness yet.

Future Improvements

• Add an evaluation harness (e.g., a RAG QA benchmark) to report retrieval and answer quality instead of relying on manual checks
• Cross-encoder re-ranking of retrieved chunks for borderline queries
• Configurable chunk size / overlap and a persistence layer for the vector store
• Swap the OpenAI dependency for a local LLM (llama.cpp) to run fully offline
• Tool-use guardrails and rate limiting for the web-search tool

License

MIT License – free to use, modify, and distribute.