Compared to v0.64, multi-model integration has been implemented. When encountering knowledge-based questions, the database will call another general model to provide answers based on recorded information through RAG, enhancing the ability to handle knowledge-based queries. The RAG database is still being improved and will integrate more in the future. Additionally, the ability to say "no" has been doubly enhanced. Furthermore, by consulting experts, comprehensive functional annotations for the maize genome have been identified and integrated, enhancing the retrieval capability of functional genes in the B73 version of maize.
The update preliminarily adds support for alfalfa, sweet potato, and Brachypodium distachyon. It further attempts to organize the non-fine-tuning process modules for adding species, enabling quick local deployment. The tutorial for local deployment is also initially launched. Current deployment methods include lightweight overall deployment (only deploying the data server; calling our model API) and full deployment (local deployment of model + data server).
The update adds the ability to search and extract variation information by gene. As this is a new feature and the process is complex, many errors may occur in the real world. Currently, it only supports the rice species, covering entire gene intervals and CDS.
The update implements a feature inspired by previous suggestions: motif prediction for extracted promoters. Based on the integrated JASPAR2024_CORE_non-redundant library from fimoh's jarpar, after extracting promoters, motifs can be directly predicted using the summarize module. Additionally, the gene-based variation detection implemented in v0.67 has been stabilized.
The update achieves broader and more coherent context while maintaining accurate output, making it more like a worker than an operational response machine [smile]. It preliminarily supports continuous context questioning with coherent content (however, due to architectural and computational limitations, operations based on backend program responses still require the summarize module to complete). Additionally, due to the added context, there may be occasional confusion in intent understanding after multiple rounds of questioning (it can still be a bit silly). Taking feedback, we have added a refresh button next to the answer box for one-click initialization.