bmad-distillator

Document distillation for LLM context compresses source files into a lossless, token-efficient distillate that preserves every decision, constraint, and relationship. This avoids lossy summaries while fitting your token budget for downstream workflows.

Semantic splitting divides large documents into self-contained sections to improve retrieval granularity and fit strict token budgets. You need it when a single compressed output is too large for one LLM context window.

To verify document distillation completeness, you can run a completeness check to ensure extracted headings and entities appear in the final distillate. Optional round-trip validation also reconstructs originals from the distillate to generate a validation report.

Yes, Python3 is required as a dependency to run the routing scripts and generate the structured JSON distillates. The workflow uses these scripts to analyze sources and enforce formatting rules.

The best way to process multiple related documents is using fan-out routing compression, which analyzes your sources and generates a unified, structured JSON distillate. This ensures all facts and relationships are preserved for downstream LLM consumption.

A limitation of round-trip validation is that reconstructing original documents from the distillate happens without accessing the originals. This tests semantic completeness but does not produce a byte-for-byte identical copy of the source files.