has anyone experienced embedding responses being empty? We have been experiencing a subset of our Embedding responses returning empty, without consistency.
Here are our settings on the call we make to divinci-003 after we we get the embeddings context and after we construct the prompt:
curiously, if we use Davinci-001 instead, we do not experience the empty responses:
but our desire it to call 003 and to not receive empty responses.
Hi @cris
Other folk are getting similar issue with text-davinci-003 as well. If you check your billing, does it show you 0 completion tokens?
Interesting. Most show completion tokens when we hit davinci-003:
1. 2,409 prompt + 903 completion = 3,312 tokens
2. 572 prompt + 1,404 completion = 1,976 tokens
3. 288 prompt + 125 completion = 413 tokens
however there are a few that show 0 completion tokens:
780 prompt + 0 completion = 780 tokens
We discovered the issue in our case was fixed by a combination of removing extra white space and adding a marker, prompting for the bot output.
Q: {{query}}
Use this specific context to answer the question: {{context}}
Output:`
With no newline character after the “Ouput:” and only one line of spacing between the prefix, question, context prefix, and ouput line.
and in the call to get context the additional context strip the output to remove any trailing space
