🧠 IFC Autonomous AI Agent

A self-refining, multi-tool AI agent built with LangGraph, Streamlit, and Google Vertex AI that intelligently answers complex questions using a Planner → Executor → Synthesizer → Critique loop.

---

📸 Overview

This project is an Agentic RAG (Retrieval-Augmented Generation) system designed to answer questions about the IFC Annual Report while also leveraging live web search when needed. The agent autonomously decides which tool to use, collects evidence, drafts an answer, and self-critiques — retrying until it is satisfied or hits a maximum iteration limit.

---

🏗️ Architecture

User Query
    │
    ▼
┌─────────┐     ┌──────────┐     ┌─────────────┐     ┌─────────┐
│ Planner │ ──► │ Executor │ ──► │ Synthesizer │ ──► │ Critique│
└─────────┘     └──────────┘     └─────────────┘     └────┬────┘
     ▲                                                      │
     │              SATISFIED / MAX ITERATIONS              │
     └──────────────────────────────────────────────────────┘

Node	Role
Planner	Selects the right tool (document_search or web_search) and rewrites the query
Executor	Runs the chosen tool and collects raw evidence
Synthesizer	Drafts a complete answer from all collected evidence
Critique	Evaluates the draft and decides to continue refining or finish

---

✨ Features

• 🔁 Self-Refining Loop — Agent retries with improved queries if the answer is incomplete
• 📄 Document Search — Semantic search over IFC Annual Report using FAISS + Vertex AI Embeddings
• 🌐 Web Search — Falls back to DuckDuckGo for general or current knowledge
• 🧠 Gemini 2.0 Flash — Powered by Google's latest LLM via Vertex AI
• 💬 Streamlit Chat UI — Clean chat interface with live agent thinking steps
• 🔍 Transparent Reasoning — Real-time display of plan, execution, and critique steps

---

🛠️ Tech Stack

Layer	Technology
LLM	Google Gemini 2.0 Flash (Vertex AI)
Embeddings	text-embedding-004 (Vertex AI)
Vector Store	FAISS
Agent Framework	LangGraph
Web Search	DuckDuckGo (langchain-community)
Frontend	Streamlit
Cloud	Google Cloud Platform (GCP)

---

📂 Project Structure

├── app.py                  # Main Streamlit app + LangGraph agent
├── faiss_index/            # Pre-built FAISS vector store (from IFC PDF)
│   ├── index.faiss
│   └── index.pkl
├── requirements.txt        # Python dependencies
└── README.md

---

🚀 Getting Started

1. Prerequisites

• Python 3.9+
• A Google Cloud Project with Vertex AI API enabled
• Application Default Credentials (ADC) configured:

  gcloud auth application-default login

2. Install Dependencies

pip install -r requirements.txt

requirements.txt should include:

streamlit
langchain
langchain-google-vertexai
langchain-community
langgraph
faiss-cpu
duckduckgo-search

3. Build the FAISS Index

Before running the app, ingest your IFC PDF to create the vector store:

from langchain_google_vertexai import VertexAIEmbeddings
from langchain_community.vectorstores import FAISS
from langchain_community.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter

loader = PyPDFLoader("ifc_annual_report.pdf")
docs = loader.load()
splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=100)
chunks = splitter.split_documents(docs)

embeddings = VertexAIEmbeddings(model_name="text-embedding-004", project="YOUR_PROJECT_ID", location="us-central1")
vector_store = FAISS.from_documents(chunks, embeddings)
vector_store.save_local("faiss_index")

4. Configure Project Settings

In app.py, update these constants:

PROJECT_ID = "your-gcp-project-id"
LOCATION = "us-central1"          # or your preferred region
MAX_ITERATIONS = 3                # max self-refinement loops

5. Run the App

streamlit run app.py

Open your browser at http://localhost:8501

---

💡 Example Queries

• "What was IFC's total investment commitment in FY2023?"
• "Compare IFC's climate finance strategy with current World Bank trends."
• "What sectors did IFC focus on in emerging markets?"

---

⚙️ Configuration

Variable	Default	Description
PROJECT_ID	gd-gcp-gridu-genai	GCP Project ID
LOCATION	us-central1	Vertex AI region
MAX_ITERATIONS	3	Max agent refinement loops
VECTOR_STORE_PATH	faiss_index	Path to FAISS index folder

---

🔒 Notes

• The FAISS index is loaded with allow_dangerous_deserialization=True — only use indexes you have built yourself.
• Ensure your GCP service account has the Vertex AI User role.
• DuckDuckGo search is rate-limited; for production use consider a paid search API.

---

📄 License

This project is for educational and internal demonstration purposes.

---

🙏 Acknowledgements

• IFC (International Finance Corporation)
• LangChain
• LangGraph
• Google Vertex AI
• Streamlit