RAG-GPT

Quickly launch an intelligent customer service system with Flask, LLM, RAG, including frontend, backend, and admin console.
Live Demo

Contents

• Features
• Online Retrieval Architecture
• Deploy the RAG-GPT Service
  • Step 1: Download repository code
  • Step 2: Configure variables of .env
    • Using OpenAI as the LLM base
    • Using ZhipuAI as the LLM base
    • Using local LLMs
  • Step 3: Deploy RAG-GPT
    • Deploy RAG-GPT using Docker
    • Deploy RAG-GPT from source code
      • Set up the Python running environment
        • Create and activate a virtual environment
        • Install dependencies with pip
      • Create SQLite Database
      • Start the service
• Configure the admin console
  • Login to the admin console
  • Import your data
    • import websites
    • import isolated urls
    • import local files
  • Test the chatbot
  • Embed on your website
  • Dashboard of user's historical request
• The frontend of admin console and chatbot
  • admin console
  • chatbot

Features

• Built-in LLM Support: Support cloud-based LLMs and local LLMs.
• Quick Setup: Enables deployment of production-level conversational service robots within just five minutes.
• Diverse Knowledge Base Integration: Supports multiple types of knowledge bases, including websites, isolated URLs, and local files.
• Flexible Configuration: Offers a user-friendly backend equipped with customizable settings for streamlined management.
• Attractive UI: Features a customizable and visually appealing user interface.

Online Retrieval Architecture

Deploy the RAG-GPT Service

Step 1: Download repository code

Clone the repository:

git clone https://github.com/open-kf/rag-gpt.git && cd rag-gpt

Step 2: Configure variables of .env

Before starting the RAG-GPT service, you need to modify the related configurations for the program to initialize correctly.

Using OpenAI as the LLM base

cp env_of_openai .env

The variables in .env

LLM_NAME="OpenAI"
OPENAI_API_KEY="xxxx"
GPT_MODEL_NAME="gpt-3.5-turbo"
MIN_RELEVANCE_SCORE=0.3
BOT_TOPIC="xxxx"
URL_PREFIX="http://127.0.0.1:7000/"
USE_PREPROCESS_QUERY=0
USE_RERANKING=1
USE_DEBUG=0

• Don't modify LLM_NAME
• Modify the OPENAI_API_KEY with your own key. Please log in to the OpenAI website to view your API Key.
• Update the GPT_MODEL_NAME setting, replacing gpt-3.5-turbo with gpt-4-turbo or gpt-4o if you want to use GPT-4.
• Change BOT_TOPIC to reflect your Bot's name. This is very important, as it will be used in query rewriting. Please try to use a concise and clear word, such as OpenIM, LangChain.
• Adjust URL_PREFIX to match your website's domain.
• For more information about the meanings and usages of constants, you can check under the server/constant directory.

Using ZhipuAI as the LLM base

If you cannot use OpenAI's API services, consider using ZhipuAI as an alternative.

cp env_of_zhipuai .env

The variables in .env

LLM_NAME="ZhipuAI"
ZHIPUAI_API_KEY="xxxx"
GLM_MODEL_NAME="glm-3-turbo"
MIN_RELEVANCE_SCORE=0.3
BOT_TOPIC="xxxx"
URL_PREFIX="http://127.0.0.1:7000/"
USE_PREPROCESS_QUERY=0
USE_RERANKING=1
USE_DEBUG=0

• Don't modify LLM_NAME
• Modify the ZHIPUAI_API_KEY with your own key. Please log in to the ZhipuAI website to view your API Key.
• Update the GLM_MODEL_NAME setting, replacing glm-3-turbo with glm-4 if you want to use GLM-4.
• Change BOT_TOPIC to reflect your Bot's name. This is very important, as it will be used in query rewriting. Please try to use a concise and clear word, such as OpenIM, LangChain.
• Adjust URL_PREFIX to match your website's domain.
• For more information about the meanings and usages of constants, you can check under the server/constant directory.

Using local LLMs

If your knowledge base involves sensitive information and you prefer not to use cloud-based LLMs, consider using Ollama to deploy large models locally.

Note

First, refer to ollama to Install Ollama, and download the embedding model mxbai-embed-large and the LLM model such as llama3.

cp env_of_ollama .env

The variables in .env

LLM_NAME="Ollama"
OLLAMA_MODEL_NAME="llama3"
OLLAMA_BASE_URL="http://127.0.0.1:11434"
MIN_RELEVANCE_SCORE=0.3
BOT_TOPIC="xxxx"
URL_PREFIX="http://127.0.0.1:7000/"
USE_PREPROCESS_QUERY=0
USE_RERANKING=1
USE_DEBUG=0

• Don't modify LLM_NAME
• Update the OLLAMA_MODEL_NAME setting, select an appropriate model from ollama library.
• If you have changed the default IP:PORT when starting Ollama, please update OLLAMA_BASE_URL.
• Change BOT_TOPIC to reflect your Bot's name. This is very important, as it will be used in query rewriting. Please try to use a concise and clear word, such as OpenIM, LangChain.
• Adjust URL_PREFIX to match your website's domain.
• For more information about the meanings and usages of constants, you can check under the server/constant directory.

Step 3: Deploy RAG-GPT

Deploy RAG-GPT using Docker

Note

When deploying with Docker, pay special attention to the host of URL_PREFIX in the .env file. If using Ollama, also pay special attention to the host of OLLAMA_BASE_URL in the .env file. They need to use the actual IP address of the host machine.

docker-compose up --build

Deploy RAG-GPT from source code

Note

Please use Python version 3.10.x or above.

Set up the Python running environment

It is recommended to install Python-related dependencies in a Python virtual environment to avoid affecting dependencies of other projects.

Create and activate a virtual environment

If you have not yet created a virtual environment, you can create one with the following command:

python3 -m venv myenv

After creation, activate the virtual environment:

source myenv/bin/activate

Install dependencies with pip

Once the virtual environment is activated, you can use pip to install the required dependencies.

pip install -r requirements.txt

Create SQLite Database

The RAG-GPT service uses SQLite as its storage DB. Before starting the RAG-GPT service, you need to execute the following command to initialize the database and add the default configuration for admin console.

python3 create_sqlite_db.py

Start the service

If you have completed the steps above, you can try to start the RAG-GPT service by executing the following command.

• Start single process:

python3 rag_gpt_app.py

• Start multiple processes:

sh start.sh

Note

• The service port for RAG-GPT is 7000. During the first test, please try not to change the port so that you can quickly experience the entire product process.
• We recommend starting the RAG-GPT service using start.sh in multi-process mode for a smoother user experience.

Configure the admin console

Login to the admin console

Access the admin console through the link http://your-server-ip:7000/open-kf-admin/ to reach the login page. The default username and password are admin and open_kf_AIGC@2024 (can be checked in create_sqlite_db.py).

After logging in successfully, you will be able to see the configuration page of the admin console.

On the page http://your-server-ip:7000/open-kf-admin/#/, you can set the following configurations:

• Choose the LLM base, currently only the gpt-3.5-turbo option is available, which will be gradually expanded.
• Initial Messages
• Suggested Messages
• Message Placeholder
• Profile Picture (upload a picture)
• Display name
• Chat icon (upload a picture)

Import your data

Import websites

After submitting the website URL, once the server retrieves the list of all web page URLs via crawling, you can select the web page URLs you need as the knowledge base (all selected by default). The initial Status is Recorded.

You can actively refresh the page http://your-server-ip:7000/open-kf-admin/#/source in your browser to get the progress of web page URL processing. After the content of the web page URL has been crawled, and the Embedding calculation and storage are completed, you can see the corresponding Size in the admin console, and the Status will also be updated to Trained.

Clicking on a webpage's URL reveals how many sub-pages the webpage is divided into, and the text size of each sub-page.

Clicking on a sub-page allows you to view its full text content. This will be very helpful for verifying the effects during the experience testing process.

Import isolated urls

Collect the URLs of the required web pages. You can submit up to 10 web page URLs at a time, and these pages can be from different domains.

Import local files

Upload the required local files. You can upload up to 10 files at a time, and each file cannot exceed 30MB. The following file types are currently supported: [".txt", ".md", ".pdf", ".epub", ".mobi", ".html", ".docx", ".pptx", ".xlsx", ".csv"].

Test the chatbot

After importing website data in the admin console, you can experience the chatbot service through the link http://your-server-ip:7000/open-kf-chatbot/.

Embed on your website

Through the admin console link http://your-server-ip:7000/open-kf-admin/#/embed, you can see the detailed tutorial for configuring the iframe in your website.

OpenIM chatbot

Dashboard of user's historical request

Through the admin console link http://your-server-ip:7000/open-kf-admin/#/dashboard, you can view the historical request records of all users within a specified time range.

The frontend of admin console and chatbot

The RAG-GPT service integrates 2 frontend modules, and their source code information is as follows:

admin console

Code Repository

An intuitive web-based admin interface for Smart QA Service, offering comprehensive control over content, configuration, and user interactions. Enables effortless management of the knowledge base, real-time monitoring of queries and feedback, and continuous improvement based on user insights.

chatbot

Code Repository

An HTML5 interface for Smart QA Service designed for easy integration into websites via iframe, providing users direct access to a tailored knowledge base without leaving the site, enhancing functionality and immediate query resolution.