RAG Challenge Winner Solution

Read more about this project:

• Russian: https://habr.com/ru/articles/893356/
• English: https://abdullin.com/ilya/how-to-build-best-rag/

This repository contains the winning solution for both prize nominations in the RAG Challenge competition. The system achieved state-of-the-art results in answering questions about company annual reports using a combination of:

• Custom PDF parsing with Docling
• Vector search with parent document retrieval
• LLM reranking for improved context relevance
• Structured output prompting with chain-of-thought reasoning
• Query routing for multi-company comparisons

Disclaimer

This is competition code - it's scrappy but it works. Some notes before you dive in:

• IBM Watson integration won't work (it was competition-specific)
• The code might have rough edges and weird workarounds
• No tests, minimal error handling - you've been warned
• You'll need your own API keys for OpenAI/Gemini
• GPU helps a lot with PDF parsing (I used 4090)

If you're looking for production-ready code, this isn't it. But if you want to explore different RAG techniques and their implementations - check it out!

Quick Start

Clone and setup:

git clone https://github.com/IlyaRice/RAG-Challenge-2.git
cd RAG-Challenge-2
python -m venv venv
venv\Scripts\Activate.ps1  # Windows (PowerShell)
pip install -e . -r requirements.txt

Rename env to .env and add your API keys.

Test Dataset

The repository includes two datasets:

1. A small test set (in data/test_set/) with 5 annual reports and questions
2. The full ERC2 competition dataset (in data/erc2_set/) with all competition questions and reports

Each dataset directory contains its own README with specific setup instructions and available files. You can use either dataset to:

• Study example questions, reports, and system outputs
• Run the pipeline from scratch using provided PDFs
• Use pre-processed data to skip directly to specific pipeline stages

See the respective README files for detailed dataset contents and setup instructions:

• data/test_set/README.md - For the small test dataset
• data/erc2_set/README.md - For the full competition dataset

Usage

You can run any part of pipeline by uncommenting the method you want to run in src/pipeline.py and executing:

python .\src\pipeline.py

You can also run any pipeline stage using main.py, but you need to run it from the directory containing your data:

cd .\data\test_set\
python ..\..\main.py process-questions --config max_nst_o3m

CLI Commands

Get help on available commands:

python main.py --help

Available commands:

• download-models - Download required docling models
• parse-pdfs - Parse PDF reports with parallel processing options
• serialize-tables - Process tables in parsed reports
• process-reports - Run the full pipeline on parsed reports
• process-questions - Process questions using specified config

Each command has its own options. For example:

python main.py parse-pdfs --help
# Shows options like --parallel/--sequential, --chunk-size, --max-workers

python main.py process-reports --config ser_tab
# Process reports with serialized tables config

Some configs

• max_nst_o3m - Best performing config using OpenAI's o3-mini model
• ibm_llama70b - Alternative using IBM's Llama 70B model
• gemini_thinking - Full context answering with using enormous context window of Gemini. It is not RAG, actually

Check pipeline.py for more configs and detils on them.

License

MIT