WattBot 2025: Estimating AI Emissions with RAG

WattBot was a challenge in the 2025 Machine Learning Marathon (MLM25). Teams built retrieval-augmented generation (RAG) systems to extract credible, citation-backed emissions and cost estimates for AI workloads from a corpus of 35+ peer-reviewed papers and 300+ curated Q&A pairs. Systems were expected to return citation-grounded answers or explicitly abstain when evidence was missing – promoting transparency and reproducibility in sustainability reporting.

Challenge design

• Task: Given a natural-language question about AI energy use, water consumption, or carbon emissions, retrieve relevant passages from the provided corpus and generate a citation-backed answer.
• Evaluation: Answers were scored on factual accuracy, proper citation, and appropriate abstention when evidence was insufficient.
• Corpus: 35+ academic papers covering AI sustainability, energy benchmarking, and environmental impact reporting.

Winning approach

The winning solution by KohakuBlueleaf used a RAG pipeline that was later replicated and deployed in both AWS Bedrock and locally with open-source Hugging Face models. See the follow-up talk below for deployment details. The team has since published their approach as an arXiv paper, KohakuRAG: A simple RAG framework with hierarchical document indexing – combining a four-level document tree (document → section → paragraph → sentence) with bottom-up embedding aggregation, an LLM-powered query planner with cross-query reranking, and ensemble inference with abstention-aware voting to take first place on both the public and private leaderboards (final score 0.861).

Links

• Kaggle challenge: WattBot 2025
• Winning solution: KohakuBlueleaf/KohakuRAG
• Winning solution paper: KohakuRAG (arXiv:2603.07612)
• Deployment repo: WattBot in Bedrock and Local

Related resources

• Project: WattBot 2026: Investigating AI’s Environmental Footprint with RAG: The challenge returns for MLM26 – bigger corpus, more questions, and a chatbot bonus track. Open on Kaggle now.
• Talk: Deploying RAG in Bedrock vs. Local: WattBot 2025 Case Study: Follow-up ML+X forum where the winning RAG system was deployed in AWS Bedrock and locally with open-source models.
• Talk: AI’s Environmental Footprint: Insights and Actions: The ML+X forum where WattBot was first introduced alongside LLM energy benchmarking work.
• Notebook: Exploring Fact-Based QA with RAG: Romeo and Juliet: Build an end-to-end RAG pipeline from scratch – a great starting point before tackling WattBot.
• Workshop: Intro to AWS for ML/AI: Covers AWS SageMaker and Bedrock for cloud-based ML/AI workflows, including RAG deployment.
• Notebook: Understanding Quantization and Precision: Learn how quantization (e.g., 4-bit) reduces model size and memory requirements – relevant to the local deployment approach used in this project.
• Paper: KohakuRAG: A simple RAG framework with hierarchical document indexing: The winning team’s writeup of their WattBot 2025 solution, with ablations showing what actually moved the score – prompt ordering, retry mechanisms, and ensemble voting with blank filtering.

Comments