Agentic AI: Multi-Source RAG with Human-in-the-Loop
Sergiu Nicoara
A professional-grade AI Agent system designed to bridge the gap between private internal knowledge and real-time web data. This project implements a Human-in-the-Loop (HITL) safety mechanism, ensuring that autonomous actions such as web searches are only executed after explicit user approval.
Key Features
Hybrid Knowledge Retrieval — Integrates LlamaIndex for semantic search over private documents and Tavily Search for live web intelligence.
Stateful Orchestration — Built with LangGraph to manage complex reasoning cycles, state transitions, and persistent conversation memory.
Human-in-the-Loop (HITL) — Uses interrupt_before logic to pause the agent and request human permission before accessing external tools.
Interactive Dashboard — Custom Streamlit interface handling real-time chat, tool-call alerts, and manual action approval (Approve / Reject).
Cloud Ready — Includes langgraph.json configuration for seamless deployment to LangGraph Cloud or local servers.
Technical Stack
Component | Technology | Role |
|---|---|---|
Orchestration | LangGraph | State management and HITL interrupt logic |
Retrieval (RAG) | LlamaIndex | Document ingestion, indexing, and query engine |
LLM Engine | GPT-4o-mini | Reasoning core for intent classification and synthesis |
External Search | Tavily API | High-fidelity web search optimized for LLM agents |
Backend API | FastAPI | High-performance REST API |
Frontend | Streamlit | Client-side interface and HITL approvals |
Observability | LangSmith | Trace logging, run inspection, and LLM debugging |
System Workflow
The agent operates on a cyclic graph with a conditional safety gate:
Thinking Phase — The LLM determines whether the request requires internal RAG or an external web search.
Interrupt — If a tool is required, LangGraph pauses execution.
User Approval — Streamlit UI prompts the user to Approve or Reject the tool call.
Final Response — The agent synthesizes all gathered data into a concise answer.
Project Structure
main.py — FastAPI server compiling the LangGraph workflow with a MemorySaver checkpointer.
integrated_agent.py — LlamaIndex VectorStoreIndex and custom tool definitions.
frontend.py — Streamlit application managing session state and HITL UI.
langgraph.json — Deployment manifest for the LangGraph ecosystem.
requirements.txt — Dependency list including FastAPI, LangChain, and LlamaIndex.
Security and Best Practices
HITL Safety — No external web search executes without explicit human authorization, preventing unintended API usage.
Ignored Files — .gitignore excludes .env, venv/, and local vector storage.
Like this project
Posted Jun 15, 2026
A professional-grade AI Agent system bridging private knowledge and real-time web data, with Human-in-the-Loop safety gates ensuring autonomous actions only execute after explicit user approval.
Likes
0
Views
1
