A client-side implementation of our semantic search endpoint (https://beta.openai.com/docs/api-reference/search).

Our endpoint has a special fast implementation of this logic which makes it very fast for calls involving many documents, so we recommend using our implementation rather than this one for latency-sensitive workloads.

We encourage you to try different variants of this client-side logic -- we don't think our setup is likely optimal at all!

The following usage will run a client-side semantic search. This formats each document into a prompt asking the API for the document's relevance, and then post-processes the logprobs to derive relevance scores:

$ ./semanticsearch.py -q 'positive emotion' -d happy -d sad
[client-side semantic search] {'object': 'list', 'data': [{'object': 'search_result', 'document': 0, 'score': 204.448}, {'object': 'search_result', 'document': 1, 'score': 108.208}], 'model': 'ada:2020-05-03'}

We run the exact same logic server-side:

$ ./semanticsearch.py -q 'positive emotion' -d happy -d sad -s
[server-side semantic search] {'object': 'list', 'data': [{'object': 'search_result', 'document': 0, 'score': 204.448}, {'object': 'search_result', 'document': 1, 'score': 108.208}], 'model': 'ada:2020-05-03'}