Chat History & Effective Queries

[jacobmarks]

• Takes user's query and the history of the chat so far, and decided on an appropriate "effective" query that is then fed in to the rest of the pipeline.
• Uses (and fixes an issue with) semantic NER for classes
• Moving query validation examples to their own file
• Updating formatting for examples to support dictionaries (including map_labels(), switch(), and if_else(). Does so by using LEFTBRACKET and RIGHTBRACKET as special strings. Resolves #19
• NEED MORE EXAMPLES to buff up performance now that we have added some generality.
• Swapped out some view stage descriptions for more examples. (10, 10) --> (5, 20)

Areas for improvement:

• If two label fields have the same classes, then don’t go through the NER twice
• Add examples for getting fields of certain types
• Fix the “not valid query” by providing more examples of valid queries, especially for sort_by_similarity()
• More contains(), is_subset(), and length() examples. If length() in an expression with &, then ( A ) & ( B ) formatting is needed
• Passing previous view stages list into gpt_view_text_generator() so it can just concatenate a view stage if it wants

[ehofesmann]

Potentially resolves https://github.com/voxel51/fiftyone-gpt/issues/19 Need to test