huridocs / uwazi

Uwazi is a web-based, open-source solution for building and sharing document collections
http://www.uwazi.io
MIT License
237 stars 79 forks source link

[IX] Training is returning "no context" #6815

Closed aphilop closed 3 months ago

aphilop commented 4 months ago

There are cases when no context information is provided and the column "Context" displays the text "no context", while in other cases there seems to be context information.

No context image


Context provided image

txau commented 3 months ago

Moreover, there seems to be the same context for all suggestions, ie:

image

Same context is provided for both suggestions, while in the document it is clearly stated that "Item 2" is a correct suggestion, but that particular context is not provided:

image

txau commented 3 months ago

This has been discussed with @gabriel-piles and also diagnosed.

The cases with "no context" seem to be errors, in the example, because the files are missing.

Returning the same context is a matter of how the models select the context. This depends on the ML model, so we can close the issue here.