v.0.0.4
What's Changed
The previous SBERT all-MiniLM-L6-v2
was not giving great results, so doing a couple of large changes to improve results of the model:
- Using asymmetric search much more like a search engine would with
msmarco-distilroberta-base-v3
- Including more training data: book author and book review from the original goodreads dataset to pull out semantic meaning in the search results
- Reconfiguring Redis indexer to write everything as a hash versus having multiple hash lookups
- Remove Grafana for now - more noise than it's worth
- Including the author in the resultset
- A number of fixes: by @veekaybee in #31
- Use PyTorch CPU only dependencies by @veekaybee in #33
- Retrain embeddings using asymmetric semantic search by @veekaybee in #39
- Fixing Docker Compose for prod by @veekaybee in #41
- Indexing speed by @veekaybee in #42
Full Changelog: v.0.0.3...v.0.0.4