ESG Fact Check

Inspiration

In today’s ESG-focused world, many organizations make bold sustainability claims. But without proper validation, these claims may border on greenwashing. We were inspired to build a system that uses AI to fact-check these claims using real corporate documentation, helping stakeholders make informed decisions.

What it does

ESG Greenwashing Analyzer takes a company’s sustainability claim and cross-verifies it against supporting documents using Retrieval-Augmented Generation (RAG) and LLMs. It rates the credibility of the claim from 1 (likely greenwashing) to 5 (credible) and generates a concise explanation.

How we built it

We built the solution using:

Langflow to design a visual workflow for document ingestion, embedding, and LLM-based analysis. ChromaDB to store and retrieve relevant document chunks. Ollama to run a local lightweight LLM model. Custom components in Python for embedding and analysis logic. A JSON-based Langflow graph to persist and share the entire pipeline.

Challenges we ran into

Dependency conflicts during Langflow installation due to rapid version changes. Local LLM integration via Ollama required adapting existing Langflow components. Debugging issues around environment variable loading and embedding errors. Managing consistent vector storage and retrieval in ChromaDB.

Accomplishments that we're proud of

Successfully integrated local LLMs with Langflow using a custom Ollama endpoint. Built a fully working ESG analyzer that gives interpretable credibility ratings. Overcame complex environment and dependency issues and got everything working end-to-end. Learned Langflow deeply through trial, debugging, and customization.

What we learned

How to build agentic AI workflows using Langflow and RAG. How to integrate third-party LLMs (like Ollama) into a pipeline. Importance of aligning prompt design, chunking strategy, and embedding logic for credible responses. Practical experience with Langchain, vector DBs, and handling model outputs.

What's next for ESG Fact Check

Extend to handle PDFs and URLs as input. Integrate a web-based frontend for user input and claim analysis results. Add multi-language support to analyze claims in local languages. Enable feedback loop to improve credibility scoring using user validations.

Built With

• chaining
• langflow
• ollama
• python