Prodigy-ANN tutorial: semantic search for annotation

Prodigy-ANN helps select text examples for annotation through semantic retrieval. This tutorial, uploaded on October 20, 2023, demonstrates the plugin's workflow at that time; the supplied source gives no software version or separate recording date.

The speaker uses recipe names to explain why exact string matching can miss relevant examples. A search for "meat" may overlook names containing pork, chicken or beef. The proposed approach embeds the text with sentence encoders, stores the vectors in an index and embeds each query to retrieve related examples. The speaker notes that useful results depend on embeddings that suit the task.

The demo creates an index file and then fetches a subset using the original text file, the index and a query. The reported indexing time is about eight seconds for 6,000 examples, a result from this demonstration rather than a general performance guarantee. Retrieved examples include metadata for the query and cosine distance. Queries can also be full sentences.

The final section shows task recipes for named entity recognition, text classification and span classification. These open retrieved examples directly in Prodigy's annotation interface without first saving the subset to a separate file.