lhh737/LangChain-ReAct-Agent

LangChain ReAct Agent – An AI‑powered chatbot with reasoning, tool use and RAG

What it is

• A Python application that builds a ReAct (Reasoning + Acting) agent on top of LangChain and LangGraph.
• The agent can answer questions from a knowledge base, call external tools (weather, user data, report generation) and switch prompts dynamically, all while showing its internal thought‑action‑observation loop in a Streamlit UI.

Key capabilities

Feature	What it does
ReAct reasoning loop	The agent repeatedly generates a Thought, decides on an Action (e.g., call a tool or run a RAG query), receives an Observation, and loops until it can answer.
RAG‑enhanced QA	Documents (txt/pdf) about robot vacuums are indexed in a Chroma vector store using DashScope embeddings; the agent retrieves relevant chunks and lets the LLM summarise them.
Multi‑tool calling	Built‑in tools for weather lookup, user‑specific data retrieval, and report generation. The middleware automatically selects the right tool based on the user’s intent.
Dynamic prompt switching	Two system‑prompt templates – one for ordinary Q&A, another for report generation – are swapped at runtime by the middleware.
Streamlit front‑end	A web UI shows the conversation, streams the LLM output token‑by‑token, and visualises the agent’s internal reasoning chain.
Modular, YAML‑driven design	Agent logic, RAG settings, vector‑store parameters and prompt paths live in separate YAML files, making the system easy to re‑configure without code changes.

How it works (high‑level flow)

1. User types a query in the Streamlit chat box.
2. The middleware decides which prompt template to use and starts the ReAct loop.
3. The agent produces a Thought → decides on an Action (e.g., RAG search or call weather API).
4. The selected tool runs, returns an Observation.
5. Steps 3‑4 repeat until the agent can emit a final answer, which is streamed back to the UI.

Technology stack

• LLM: Alibaba DashScope ChatTongyi (via the DashScope API).
• Agent framework: LangChain 0.3 + LangGraph 0.2.
• Vector store: Chroma (persistent on‑disk index).
• Embedding model: DashScope embedding service.
• Document processing: PyPDF, RecursiveCharacterTextSplitter.
• Front‑end: Streamlit 1.40.
• Configuration: YAML files for agent, RAG, prompts, Chroma.

Getting started (as described in the README)

# 1. Clone
git clone https://github.com/lhh737/LangChain-ReAct-Agent.git && cd LangChain-ReAct-Agent

# 2. Install
pip install -r requirements.txt

# 3. Set your DashScope API key
export DASHSCOPE_API_KEY="your-key"   # or `set DASHSCOPE_API_KEY=your-key` on Windows

# 4. Build the knowledge base (first run only)
python -c "from rag.vector_store import VectorStoreService; VectorStoreService().load_document()"

# 5. Launch the UI
streamlit run app.py

Open http://localhost:8501 in a browser, then try sample questions such as:

• “扫地机器人有哪些主要功能？” → triggers RAG QA.
• “如果机器人无法正常回充，该如何处理？” → triggers fault‑diagnosis reasoning.
• “请根据用户数据生成一份个性化使用报告” → triggers the report‑generation tool chain.

Project layout

• agent/ – core ReAct logic and tool definitions.
• rag/ – vector‑store handling and RAG service.
• model/ – factory that creates the LLM and embedding client.
• config/ – YAML files controlling behaviour.
• prompts/ – raw prompt templates.
• utils/ – helper modules (config loader, file handling, logging, etc.).
• app.py – Streamlit entry point.

Who might use this

• Developers wanting a reference implementation of a ReAct‑style agent with tool use.
• Teams building AI‑augmented customer‑service bots that need to combine knowledge‑base lookup with external data.
• Anyone exploring LangChain/LangGraph patterns for dynamic prompt management and RAG integration.

License: MIT (see LICENSE).