🤖 IT Purple Hack & Central Bank of Russia

Building LLM-based RAG chat-bot for the Central Bank.

Task Description

🦸‍♂️ Team

We are the finalists of this competition. Get to know us:

• Solomon
• Roman
• Nikita
• Vladislav
• Nikita

Navigation

• dataset_preprocessing - cleaning the corpus of scraped texts
• custom_qa_chain.py - custom question-answer generation chain compatible with LangChain
• qa_generation.ipynb - generating questions for validation and finetuning the solution
• num2text.ipynb - numbers preprocessing
• chroma_db - all files related to launching and populating ChromaDB
• clickhouse - writing data to Clickhouse from a csv file
• simularity.py - script for finding relevant documents
• embeddings.py - obtaining embeddings for the database
• gui.ipynb - the main file with the application front-end

Retrieval pipeline

How to interact with LLM?

Starting the server:

1. Launch the server in LM Studio
2. choco install ngrok
3. ngrok config add-authtoken [authtoken]
4. ngrok http --domain=live-relaxed-oryx.ngrok-free.app 1234

Interacting with LLM https://github.com/Data-Squad-of-Scoofs/cb-purple-hack/blob/db/inference_llm.ipynb

How to interact with the database?

We implemented various options. Eventually, ClickHouse was chosen.

1. ChromaDB

Launching the database in docker:

docker pull chromadb/chroma
docker run -p 8000:8000 chromadb/chroma

Connecting:

pip install chromadb
self.client = chromadb.HttpClient(host=your_ip, port=your_port)

2. ClickHouse

The database was launched via clickhouse.cloud, but it can also be launched via docker. It operates faster than ChromaDB.

Connecting:

pip install clickhouse_connect
client = clickhouse_connect.get_client(host=your_ip, port=your_port, username=your_user_name, password=your_password)

Generating a validation dataset

---