Carbon Footprint
Carbon emissions were measured during the development and inference process using CodeCarbon https://github.com/mlco2/codecarbon
| Phase | Description | Emissions | Source |
|---|---|---|---|
| Development | OCR and parsing all pdf documents with AI | 28gCO2e | CodeCarbon |
| Development | Question Answering development | 114gCO2e | CodeCarbon |
| Inference | Question Answering | ~0.102gCO2e / call | CodeCarbon |
| Inference | API call to turbo-GPT | ~0.38gCO2e / call | https://medium.com/@chrispointon/the-carbon-footprint-of-chatgpt-e1bc14e4cc2a |
Carbon Emissions are relatively low but not negligible compared to other usages: one question asked to ClimateQ&A is around 0.482gCO2e - equivalent to 2.2m by car (https://datagir.ademe.fr/apps/impact-co2/)
Or around 2 to 4 times more than a typical Google search.