To mitigate the high computational cost of using a large k-NN datastore sourced from an entire training dataset, an index for the datastore's vectors can be built and optimized offline before the LLM is run. Because the training data is static, this pre-processing step allows for highly efficient retrieval of similar vectors during inference, making the use of extensive, generalized context computationally feasible. This method of pre-indexing for fast lookups is a standard practice in vector databases.