Currently, in indexing we are using JSON as a serialization form for the JS content assist model. In order to reduce I/O, we should consider moving to a binary (or possible string-friendly binary) format. Short of compressing the data, this will give us a reduce footprint in the index file. This reduction in size becomes particularly important if an index is being merged to the DiskIndex. Smaller sizes will mean less I/O.
JSON parsing is also showing up during profiling when working with large projects (like the closure-library). This would remove that bottleneck, thus improving read times from the index.