[FEATURE] Smarter handing of runs

[jacobmarks]

Addresses #31

• Added image_similarity, text_similarity, evaluation, and metadata columns to examples spreadsheet
• Changed the ordering of links so example selection comes later
• Added runs filters to vector index so we don't show examples for run types we don't have
• Changed synthetic keys to special tokens that we replace at runtime

Still to do:

• handle case of multiple evaluations
• handle when to use label classes vs text similarity
• classification vs detection and match() / filter_labels()

[jacobmarks]

Added in handling for the following cases:

• when text_similarity exists and is relevant, and no label classes are found, filter out match() or filter_labels() examples
• when identified label fields are classification, filter out detection, and vice versa

Other changes include:

• turning the many filtering $and condition generations into a function
• more examples for sort_by_similarity() now that we are narrowing down what the model sees based on other filters
• more examples for query validation and label class NER, where it was struggling
• reproduced your MATCH issue where it gave you all caps. I think this is because we are telling it to do _MORE_ and _CONFUSED_ in all caps. Put in a post-processing fix for this.

Still need to deal with:

• label field names, likely with a CLASSIF_FIELD and DET_FIELD based approach. The main problem is still that often the LLM thinks multiple fields are relevant, and it is hard to determine if both actually are, or just one will be used in the end.

[jacobmarks]

Added moderation to ensure initial and later queries abide by OpenAI's terms of use