Helpster one of the entries that won Inquiry Bot Challange by IEEE CS KS. It is an LLM-based chat bot that can answer questions related to AICSSYC 23, previous AICSSYC's, IEEE CS KS, IEEE CS, IEEE, and other related questions.

Built using:

• Python
• LangChain
• OpenAI
• ChainLit
• Pinecone

Clone the repository

git clone https://github.com/RohitEdathil/helpster

Install dependencies

pip install -r requirements.txt

Setup environment variables

If you want to enable LangChain tracing (Optional)

Create an index in Pinecone with the name helpster and dimension 1536

python3 load.py

Running this command will load documents from data folder to Pinecone index. All files must be plain text files with .txt extension. One file is considered as one document in the index.

Note: The index helpster must be created in Pinecone before running this command.

chainlit run main.py

This will start the server at http://localhost:8000 where you can ask questions similar to ChatGPT.

This bot uses RAG(Retrieval Augmented Queries). It uses a Pinecone index to retrieve the most similar document to the question. Then it uses the retrieved document as a context to the GPT-3 model to generate the answer.

The bot uses a memory buffer to store the last 3 questions and answers. You can thus ask follow up questions to the bot. It was limited to 3 because of the token consumption limitations.

The bot is explicitly instructed to not answer questions which it is not trained to answer. This prevents hallucination and makes the bot more robust. Also the bot will stay close to its purpose.

Here are the data sources used to train the bot:

• Official website of AICSSYC 23
• Official instagram page of AISSYC 23
• Instagram pages of previous AICSSYC's
• Facebook page of AICSSYC 23
• Archived website of previous AICSSYC's
• IEEE CS KS website
• IEEE CS website
• Wikipedia

The data obtained to train the data is publically available. It was obtained mostly manually and a tool called instaloader was used to download the instagram posts. The data was then cleaned and formatted to be used for training. Even though the data is publically available, it is still securely stored only in the development system and secure Pinecone servers.

All the chat information is monitored through LangSmith, a sister project of LangChain. It is a tool to diagnose and monitor LLM apps. The chat information is stored in a secure database and is only used for debugging purposes. Developers may look into the chat information to improve the bot. The chat information is not shared with any third party.

The LLM used by the app is provided by OpenAI. Hence, data is visible to OpenAI.

Read more about OpenAI's API privacy policy here.

From OpenAI's privacy policy we can see that data from API Platform is not used to train the models. Also, the data is not shared with any third party and is SOC 2 Type 2 compliant.

Read more about SOC 2 Type 2 here.

Hence we can conclude there won't be any data leaks (user information presented to other users).