Skip to main content

Building RAG Agent using LangChain

Reference Notebook: 0_create_agent.ipynb

In this lesson, we will focus on initializing a ReAct (Reason + Act) Agent. This is a model that helps LLM not only think but also be capable of performing specific actions through tools:

  • Operating Mechanism: Agent performs a loop: Reasoning -> Action -> Observation.
  • Flexibility: This is an important foundation for building complex Agent systems (Deep Agents), allowing smart processing of multi-step tasks.

Practice Project: RAG Agent System for FPT Policy​

Goal: Build an automated question-answering system about FPT's internal regulations and policies.

Implementation Process:​

  1. Input Data: Gather FPT's policy documents, internal guide documents.
  2. Vector Database: Convert text data into embeddings and store in VectorDB for retrieval.
  3. Retrieval System: Use semantic search mechanism to retrieve most relevant text segments (Context).
  4. Integrate ReAct Agent:
  • Agent will receive questions from users.
  • Use retriever tool to search information in VectorDB.
  • Synthesize information and answer based on actual data from company policy.

Processing Flow (Pipeline): > FPT internal docs → Embeddings → VectorDB → ReAct Agent (Retrieval Tool) → Final Answer