Philosophy on LLM Usage in an Operational Context

The scope of these DRAFT statements is isolated to ISD Tech Team, not to be taken as reflecting the stance of the American Red Cross in totality.

• Preference for Open Source, therefore HuggingChat preferable to ChatGPT
• Optimistic but want to be judicious around its usage, with clear approaches towards post-interaction, programmatic validation of output to mitigate against hallucinations. In practice this means focusing on purely extractive exercises (eg. not a single word should be present in the output that didn't exist in the input, no mathematical operations performed, etc.)

It is in the near-term roadmap to test the best sequence of prompts to yield very targeted things. Specifically the extraction of key/value fields (where key would be a location for example, and value would be # of families affected) in Q4 2023.