LongLoom Library Aims to Process Long Documents Through LLMs Without Losing Facts

An open-source library called LongLoom has been introduced to handle documents too large for an LLM's context window without dropping information. Its author reports that a naive chunk-and-summarize approach lost 2–10% of planted facts across runs, which is unacceptable for compliance, legal, and audit uses. To address this, LongLoom offers a verbatim extraction mode that copies fact-bearing lines directly, with the author reporting 100% fact preservation in tests. The pipeline splits documents into chunks, maps each through the LLM, merges results hierarchically, and repairs only when needed. The library is provider-agnostic and supports multiple storage backends, and its benchmark script is publicly available for reproduction.