Fast and Accurate EPD Imports with AI
Adding EPDs to the library is a manual and error-prone task, as it involves entering many values. Today, we introduce our production-ready EPD import from PDFs to decrease time for copying the indicator values from the PDF into Real-time LCA with high accuracy.
Explainable neuro-symbolic AI
At Real-time LCA, we follow an explainable AI approach.
A core issue with generative Artificial Intelligence (GenAI) based on multimodal foundation models, is that these models hallucinate: they are wrong but so confidently wrong that it nay be difficult to spot (IBM on hallucinations).
When we want to automate a task like copying hundreds of indicator values from a PDF into Real-time LCA, we want to save time while being confident that the numbers are correct. If we end up checking every single value again, we may just do the entire thing manually. Yet, the hallucinations are real and such GenAI models are inherently making them, there is always a residual risk.
Identify mismatches for human review
Hence, Real-time LCA uses a neuro-symbolic approach that combines the GenAI capabilities of frontier models to understand EPD files with symbolic AI rules that reason over the results and identify hallucinations. GenAI structures the messy data, and conventional symbolic AI checks the now-structured data for errors.
This approach does not remove wrong guesses of the AI. However, we can identify wrong guesses and inform the user that this requires checking.
In practice, this neuro-symbolic pipeline extracts 98% of the values correctly (recall) across 35 test files correctly (app. 10,000 values). The other 2% are missing or highlighted as potential errors.
Out of 10,000 of values, only 11 individual entries were off. For engineering reliability, Real-time LCA aims to also mitigate these remaining hallucinations in future iterations of this tool.
Test the feature
Is this something you are curious about? Comment on this post, and we’ll invite you to our beta group testing this feature.