The code can extract adresses from raw Turkiye earthquake tweets and classify them for intent via OpenAI GPT Codex API by using few-shot prompting.
Currently the input format is .jsonl
where each line has a json string with "Tweet" field, see an example input file here data/test.jsonl.
Export two environment variables as comma seperated keys:
export OPENAI_API_KEY_POOL=key1,key2,key3...
export GEO_KEY_POOL=key1,key2
optionally for afet org api base urls
export OPENAI_API_BASE_POOL=
To extract the geo location address information:
- Specify your paths in run_addres.sh, then run the script
./run_address.sh
To extract the intent information:
- Specify your paths in run_intent.sh, then run the script.
./run_intent.sh
- To run locally
uvicorn main:app --reload
Running github actions