Artificial Intelligence/Generative AI
Creating a foundation service
- Creating a foundation wrapper that could execute GenAI requests, and abstract the model being used (using cloud services at first)
- Ensuring 3rd parties can't abuse the service, while community developpers can use it freely
- Making testing and iteration easy.
- Taking safety measures so that it doesn't cost the Open Food Facts NGO too much
Ensuring capitalization of prompt knowledge
- Listing useful prompts, and creating metrics for them
Avoiding wasteful compute
- Ensuring outputs are stored and shared
- Ensuring we group prompts
- Ensuring prompts are iterated and evaluated before being run at large scale
Enhancing how we work and volunteer
- Ensuring knowledge about Open Food Facts as an org is well disseminated so that LLMs can help us more efficiently (eg: that it knows what Robotoff is, or which database we're using, …)
- Looking for opportunities in the tools we already use (eg GitHub Copilot from GitHub, new LLM features in Odoo, …) and new tools that could save us time or increase our impact
Monitor and document innovative usages outside Open Food Facts
- Document good practises
- Document things we could also implement
- Experiments/GenAI
Think hard about how we can use it
Data quality
Data extraction
Content synthetisation
Data models generation
Code generation
Food intake analysis
Natural search parsing
- https://www.elastic.co/search-labs/blog/large-language-models-elastic-code-langchain
- those are some of the potential themes, but there are many others
- https://www.elastic.co/fr/cloud/generative-ai
- https://eliatra.com/blog/chatgpt-and-llms-to-enhance-opensearch-elasticsearch-usability/
- https://www.elastic.co/search-labs/blog/rag-with-llamaIndex-and-elasticsearch
- https://www.reddit.com/r/elasticsearch/comments/114ckdf/use_natural_language_to_query_elasticsearch_by/
Don't be paralyzed by potential errors or phantasms
- It is possible to mitigate/eliminate those potential errors by keeping humans in the loop, and adopting a copilot approach
- These new techniques can be hidden as a helper in the UI (like a magic button), with the proper warnings