Huffman coding assumes your corpus is a string of discrete elements (symbol strings) without any continuous structure (eg. topology/geometry). With that fairly mild assumption, it gives a recipe to reorganize (transform/encode) your data as a prefix-tree, to minimize the bits of information needed to communicate the contents of your corpus i.e. reducing (on average) the bits of information you need to identify a specific item. Eg. To go back to the analogy from my previous comment above... if the function you are inverting via search has long plateaus then you could simply front-load those as guesses; that's roughly the spirit of Huffman coding, except it eschews monotonicity.