Skip to content

Question answering using embeddings error #96

@stephansturges

Description

@stephansturges

I just wanted to point this out:

image

It's obviously no big deal, but given the context window of >8000 tokens with davinci-003 it would be a great opportunity to explain that the semantic proximity may not be a perfect metric for prepending context and in this case we solve the issue by prepending multiple articles since the second-closest one is the correct input.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions