You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I have looked for existing issues (including closed) about this
Feature Request
Loaders could do with some more features to create a more complete experience.
Add a more straightforward chunking experience (including multiple strategies, isolated + overlapping)
Add pipeline integration by implementing a new method or the Op and TryOp traits
Motivation
Loaders first implementations are straightforward, we should continue to grow the aspect of incorporating knowledge for the utility of RAGs by improving our interface.
Proposal
Add .chunk with various options to the loaders
Add Op and/or TryOp trait implementations
Alternatives
Add chunking as apart of the pipeline system instead
This is a fairly interesting thought as chunking is a generic way of breaking of text. This could also apply to length LLM output streams which is natural from a pipeline POV. However, it's mostly used for loaders and the context restraint being only placed on loaders makes for seamless helpers as anyone could just use iterators on their own to create custom chunking if needed.
Create iterator helper methods directly rather than tying it to the context of loaders.
Also another approach which would increase it's flexibility. It would exist more as a loaders helper / utils which would be leveraged in common use-cases for chunking strategies (perhaps the loaders implementation just wraps the internal iterator extenders).
Notes
Loaders share some sense of ideology but the different loaders do not share any traits or implementations so it's very easy to duplicate code. Might want to investigate how some logic could be intelligently shared.
The text was updated successfully, but these errors were encountered:
Feature Request
Loaders could do with some more features to create a more complete experience.
Op
andTryOp
traitsMotivation
Loaders first implementations are straightforward, we should continue to grow the aspect of incorporating knowledge for the utility of RAGs by improving our interface.
Proposal
.chunk
with various options to the loadersOp
and/orTryOp
trait implementationsAlternatives
Notes
Loaders share some sense of ideology but the different loaders do not share any traits or implementations so it's very easy to duplicate code. Might want to investigate how some logic could be intelligently shared.
The text was updated successfully, but these errors were encountered: