crow-hq/llm-wiki-v2 released v0.5.0 on October 16, 2026. This repository is a project that reads documents and turns them into a Markdown wiki with citations. It collects documents into a wiki and uses that wiki as the basis for attaching citations to answers. The main change in this version is that Word, presentation, spreadsheet, and OpenDocument files can now be read without new dependencies. The release notes are in the v0.5.0 tag on GitHub (https://github.com/crow-hq/llm-wiki-v2/releases/tag/v0.5.0).
Spreadsheets are handled sheet by sheet, with the documents and the original source materials processed separately. The repository description says this wiki works with any LLM provider, including Ollama, which runs local models. The same description claims that CROW classifier routing raises "right note" judgments from 62% to 89%, makes ingest (the process of reading documents and moving them into the wiki) 2.1 times faster, and cuts token usage by 68%. The repository has 225 stars.
Another change in this version is that even if a document's response is wrong, only that document is isolated as an error and the rest of the sync proceeds. However, the comparison target behind 89% versus 62% and the evaluation conditions are not given in the repository description. It is also not stated whether the per-document error isolation retries failed documents or records the cause separately. Considering that the figures are self-reported by the repository, the benefit actually gained in real workloads has to be re-checked for each environment.