Wiki Retrieve

If your vault has grown past 50 pages and you're tired of Claude reading five irrelevant documents before finding the one paragraph that answers your question, this replaces the old hot→index→drill read order with chunk-level hybrid retrieval. It implements Anthropic's September 2024 Contextual Retrieval paper: contextual prefix generation, BM25 sparse search, and cosine reranking. One-time setup costs about $3.60 per 100 pages with Haiku plus prompt caching, or you can run the fully local tier with synthetic prefixes and Ollama. The smart part is it's opt-in and feature-gated, so wiki-query and autoresearch detect it and upgrade themselves automatically while older skills keep working. If you're running deep research sessions, the quality jump is real.

skills.sh installs: 323
GitHub stars: 16.3K
last updated: May 28, 2026

View on GitHub

Install

$ npx -y skills add agricidaniel/claude-obsidian --skill wiki-retrieve --agent claude-code

Installs into .claude/skills of the current project.

Files

Retrieve relevant passages

This extension derives search data from wiki/ into .vault-meta/. It never changes canonical notes. Always pass the selected vault explicitly.

Resolve the installed product root from this skill's own location, not from the vault or current working directory:

PRODUCT_ROOT=/absolute/path/to/installed/claude-obsidian
PREFIX="$PRODUCT_ROOT/scripts/contextual-prefix.py"
BM25="$PRODUCT_ROOT/scripts/bm25-index.py"
RETRIEVE="$PRODUCT_ROOT/scripts/retrieve.py"
RERANK="$PRODUCT_ROOT/scripts/rerank.py"
test -f "$PREFIX" && test -f "$BM25" && test -f "$RETRIEVE" && test -f "$RERANK"

Pipeline

contextual-prefix.py splits pages on paragraph boundaries and stores the raw chunk plus a short page-level prefix.
bm25-index.py builds a local, standard-library BM25 index over the contextualized text.
retrieve.py selects BM25 candidates, optionally reranks them, rejects invalid records, deduplicates by page, and returns paths and snippets.
The caller reads the returned pages and performs synthesis; retrieval output is not itself evidence.