👷Mutlivectors Improvement Proposals

Current Proposals(non prioritized):

• add minibatch in retriever.store method
• Add agents interfaces, and use multivector docstring as prompt
• Rework signatures to allow retriever "warm start" from non empty vector store
• add metadata table in vector store to avoid storing multiple signature instances and store metadat that could be usefull for advanced vectors (like idf counts ?)
• Rust extensions to accelerate hash based vectors Improve min hash perf
• Add "mutate_query" function to change the queryside during eval
• add code to generate a gradio chat frontend from a well initialised VS (either in memory or codegen - preferably both) to allow superfast projects quickstart once ingestion is taken care of
• Add support for images in RetrieverTool (currently smolagents only allows tools with 1 modality, we will need to find a workaround)
• Add compatibility matrix in signature checks
• Add BM25 approximate vectors for efficient full text search
• Rework litellm batching to extend it to all vectors
• Add complex vectors and fourier derived similarity measures (heaveside, RBF Kernels)
• Add "maximize" vectors to emulate sorting
• Add AG-UI compatibility in codegen for better out of the box UIs and richer interactions
• Add knn classifiers and regressors
• Add preprocessors
• Add reformulation vectors to facilitate Hyde usecases / information extraction usecases
• Add weight optimisation for knn variants instead on learning to rank
• Add LSH vectors to generalise Lp distances with p<2
• Add ProbMinHash vector
• Node in Graph vector types based on graph traversal techniques to generalize hierarchy vectors
• Retreive multivector and multivector
• Add multihash class that operates like a multivectors but on binary arrays where similarity is XNOR pop count / N (potentially usefull to leverage more efficient storage)