chatbot sample

English | 日本語

Sample for building a self-hosted chatbot
You can import source code and document files stored on your local PC.

The article ChatGPT Chatbot for Internal Use introduces the mechanism.

Demo

Frequently Asked Questions.

• Data usage policy
  • Data sent by customers via API Data sent by customers are not used for training in principle (opt-in if you are willing to share data for training).
  • Retain API data for 30 days to monitor for abuse or misuse (after that, delete it. (After that, it will be deleted, except where required by law)
  • A limited number of authorized OpenAI employees and specialized third-party contractors may access this data only to investigate and verify suspected abuse.
  • "Content (input prompts and their responses, uploaded and generated images) in non-API consumer services such as ChatGPT and DALL-E may be used to improve the service
  • Enterprise customers deploying use cases with low potential for abuse could request that no API data be stored at all
• Difference between ChatGPT and ChatGPT API
  • The ChatGPT API is an easier-to-use version of ChatGPT's functionality, allowing programmers to easily use ChatGPT.
  • This sample uses ChatGPT API.
  • To use ChatGPT API, you need to create an OpenAI account and issue an API key.

Installation

Set up API key

export OPENAI_API_KEY=<Open AI API Key>

If you need Google search, refer to Try LangChain's Google custom search integration and set the following.

export GOOGLE_CSE_ID=
export GOOGLE_API_KEY=

Build the container

docker-compose up -d --build chatbot

create vector database

docker-compose exec chatbot python /app/ingest.py --help
ingest.py [OPTIONS] TARGET

TARGET : Specify the path to the file or directory to be processed. (Required field)
Options include.

• -o, --output-file TEXT : specifies the output vector database file name.
• -l, --loader-cls [text|pdf_miner|pymupdf|pdf|html] : Specifies the loader class to use. Default is text.
• -dl, --dir-loader-cls [directory|readthedocs] : Specifies the loader class for directories. Default is directory.
• -e, --file-ext TEXT : Specifies the file extension to process.
• -cs, --chunk-size INTEGER : specifies the size of the text chunks. Default is 1000.
• -co, --chunk-overlap INTEGER : specifies overlap of text chunks. Default is 200.
• -d, --dry-run : do not actually add the document to the vector store.

For example, . /samples/pdf If you want to create a database file by processing all PDF files in the directory .
The -d option allows you to see the number of tokens and fees that will be consumed without building the database.

docker-compose exec chatbot python /app/ingest.py -e pdf -l pdf_miner -o /data/projectname -d /samples/pdf
load 654 documents
use 338462 token
price 0.0169231 USD
Dry run mode enabled. Exiting without adding documents to vectorstore.

Start ChatBOT

Set the environment variable to the database file you just saved and start the chatbot.
Connect to the chatbot via a browser to "http://localhost:3000".

docker-compose up -d

Reference

• https://dev.classmethod.jp/articles/openai-data-usage-policy/
• https://github.com/hwchase17/chat-langchain
• https://github.com/mckaywrigley/chatbot-ui