Гибридный поиск по Obsidian-хранилищу

★ 7.6 · knowledge-management

wiki-retrieve is a Claude Code skill that builds and queries a vault-local hybrid BM25 retrieval index with optional multilingual semantic reranking via Nomic embeddings for searching an Obsidian vault. The pipeline comprises four scripts: `contextual-prefix.py` splits pages at paragraph boundaries and attaches page-level prefixes, `bm25-index.py` builds the index using the Python standard library, `retrieve.py` selects candidates and deduplicates by page, and `rerank.py` applies cosine similarity using `nomic-embed-text-v2-moe` (~958 MB) through a local Ollama instance. All derived caches are stored under `.vault-meta/` and no data leaves the machine without explicit user consent — sending content via the Anthropic API or a remote Ollama endpoint requires the `--allow-egress` or `--allow-remote-ollama` flags respectively. When reranking is unavailable, the skill falls back deterministically to BM25 order without mixing score scales. It suits developers and knowledge workers who need chunk search, contextual retrieval, or retrieval diagnostics over a local wiki.