Ashik Sali
AI Agent Chatbot with Tool Use and Streaming (LangChain)

AI Agent Chatbot with Tool Use and Streaming (LangChain)


AI LangChain LLM Agents RAG

Overview

Large Language Models are powerful, but on their own they can hallucinate, struggle with real-time data, and fail on multi-step reasoning tasks.

In this project, I built a production-ready AI agent chatbot using LangChain that goes beyond a simple LLM wrapper. The agent is capable of reasoning, selecting tools dynamically, and streaming both its intermediate steps and final answers to the user in real time.

The result is a transparent, verifiable, and extensible AI system suitable for real-world applications.

System Architecture

At a high level, the system consists of:

  • A frontend chat interface
  • A Python FastAPI backend
  • A LangChain Agent Executor
  • External and internal tools (search, utilities)
  • An LLM providing reasoning and response generation

Step-by-Step Flow

RAG Flow

  1. The user asks a question in natural language
  2. The request is sent to a FastAPI backend
  3. The LangChain agent analyzes intent
  4. The agent decides whether tools are needed
  5. Tools are executed (search, calculations, etc.)
  6. The LLM generates a grounded answer
  7. Output is streamed back to the user in real time

Agent with Tool Calling

The core of the system is an intelligent multi-tool agent that:

  • Chooses between no tool, retrieval, or calculation
  • Chains multiple tools when required
  • Prevents hallucinations by grounding responses in tool outputs

Example

“Find the current temperature in NYC, multiply it by 5, and report the result in Celsius.”

This requires:

  • Web search
  • Numeric computation
  • Unit conversion
    —all handled autonomously by the agent.

Key Features

Intelligent Multi-Tool Reasoning

The agent uses ReAct-style prompting to reason about which tools to invoke and in what sequence.

Dynamic Tool Set

  • External tools: Real-time web search via Serper API
  • Internal tools: Deterministic Python utilities (add, multiply, etc.)

Real-Time Streaming UX

Responses are streamed token-by-token, including:

  • Tool selection
  • Tool inputs
  • Intermediate reasoning
  • Final answer

This significantly improves perceived performance and user trust.

Transparent & Verifiable Output

Final responses include:

  • Clearly structured answers
  • Citations (when web search is used)
  • A list of tools invoked during execution

Technology Stack

  • Language: Python
  • Framework: LangChain (v0.3+)
  • Web Backend: FastAPI
  • LLM Provider: OpenAI API
  • Search Tool: Serper API
  • Frontend: Vue-based chat UI
  • Concepts: Agents, Tool Calling, Streaming, ReAct prompting

Why This Project Matters

This project demonstrates my ability to:

  • Build agentic AI systems, not just LLM wrappers
  • Design transparent, trustworthy AI workflows
  • Implement real-time streaming architectures
  • Apply modern AI engineering patterns used in production systems