Jtoken – lossless JSON compression for LLM prompts (github.com)

The newly released jtoken is a library designed to optimize JSON compression specifically for large language model (LLM) prompts, significantly reducing the number of tokens required without losing any data. Developed by Hermann Samimi, jtoken achieves this by eliminating unnecessary JSON syntax, condensing repeated boolean and null values, and flattening nested dictionaries using dot notation. It is especially beneficial for formats like Elasticsearch and MongoDB, thereby streamlining workflows that rely on these databases. This advancement holds substantial significance for the AI/ML community, as token usage can directly affect usage costs in API calls. By implementing jtoken, developers can experience up to 19.1% savings in token counts for complex document types, leading to reduced operational costs when using models such as OpenAI's GPT. The package is easy to integrate, requiring no extra runtime dependencies for standard usage, and allows the optional installation of the "tiktoken" module for precise token counting compatible with OpenAI models. In summary, jtoken not only facilitates more efficient data handling but also enhances cost-effectiveness in AI applications reliant on LLMs.