Wiki Retrieve

Before you start: read this whole install prompt, tell me in a few lines what this skill does and what installing it involves, and wait for my go-ahead before you do anything else. Prerequisites — REINFORCED FRAMING FOR THE OPTIONAL EGRESS, DISCLOSED FIRST: this skill's default and local paths send nothing off your machine: synthetic contextual prefixes are built locally without any LLM (--no-llm), the BM25 index is a local standard-library index, and the default reranker talks only to a localhost Ollama service. ONE optional mode is real egress and is fenced by the skill itself: the explicit Anthropic API profile for contextual prefixes CAN SEND PAGE BODIES OFF-MACHINE to Anthropic, and it requires, all three together, your explicit consent, the --allow-egress flag, and your own configured Anthropic API key; the skill states that consent is never inferred from the presence of an API key or an installed binary, that the scope must be previewed first and the receiving provider stated, and that there is no Claude CLI generation tier and no ambient subprocess fallback. A second optional remote path is fenced the same way: remote Ollama endpoints require your explicit approval plus --allow-remote-ollama (that flag approves the endpoint only, not cloud delegation by the server). Installing this listing does NONE of this: installation copies only this skill's folder, runs nothing, builds no index, and makes no egress of any kind. USE REQUIREMENTS: this skill works on your own Obsidian vault and runs through FOUR scripts that live at the repository root OUTSIDE this skill's folder: scripts/contextual-prefix.py, scripts/bm25-index.py, scripts/retrieve.py and scripts/rerank.py. Installing this listing copies only this skill's folder: without a full claude-obsidian installation (those scripts, python3 to run them, and a selected user vault) the skill cannot run as designed. The family orchestrator is the separate wiki skill of the same repository. No account, key or payment is needed to copy the files; an Anthropic API key is needed only if you later choose the optional Anthropic prefix profile yourself. Real writes at use time, disclosed here: building prefixes and the index writes derived, disposable caches under your vault's .vault-meta/ folder only; the skill never changes canonical notes in wiki/. Install "Wiki Retrieve" for me. It builds and queries a vault-local contextual BM25 retrieval index for your Obsidian wiki, with optional multilingual Nomic cosine reranking, and it never changes canonical notes. Its pipeline has four stages, all run by the product-root scripts named in the prerequisites: contextual-prefix.py splits pages on paragraph boundaries and stores each 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 (retrieval output is not itself evidence; the caller reads the returned pages and synthesizes); rerank.py provides the reranking and diagnostics. Provisioning is preview-first, then local: synthetic prefixes are built without network egress (--no-llm --peek, then --no-llm), the BM25 index is built, and a read-only smoke query runs with --no-rerank. Chunk and index files under .vault-meta/ are disposable runtime state; incremental prefixing skips records whose chunk hashes, page hashes and provider/model profile still match, a complete scan removes surplus records for deleted pages, the prefixer invalidates the BM25 index before changing its chunk set, and prefix and BM25 builds share the vault-wide mutation lock with every other writer, failing closed on a busy vault instead of publishing a partial index. Beyond synthetic prefixes, two explicit local options exist: local Ollama model prefixes, using an already installed exact name:tag only (no server is started and no model is installed or pulled; the adapter verifies cloud-disabled server status and local model identity before sending page text, and unsupported or unavailable preflight falls back to synthetic prefixes for new or changed chunks only), and the fenced remote options described in the prerequisites (the Anthropic API prefix profile and remote Ollama endpoints). Querying is read-only against the prebuilt index: queries are bounded at 8,000 normalized characters and result counts between 1 and 1,000; reranking happens only when explicitly requested by omitting --no-rerank, using Ollama's multilingual nomic-embed-text-v2-moe model by default (approximately 958 MB; the product never pulls it automatically; the smaller English-oriented nomic-embed-text v1.5 model can be selected explicitly). Embeddings are cached by exact model, verified digest, input scheme and hash of the exact prefixed input, and any embedding failure (missing local Ollama service, malformed reply, missing model, unusable vector) falls back for the complete result set to the original BM25 order, never mixing cosine and BM25 score scales. Its integrity rules accept only relative chunk and page paths that resolve under .vault-meta/chunks/ and wiki/ respectively, require matching chunk-body, page and index hashes, and reject absolute paths, symlink escapes, missing pages, mismatched chunk IDs and stale pairs; an empty index is an honest no-result state, and a missing or corrupt index makes retrieve.py exit 10 with a stable rebuild command while callers fall back to the standard vault query/text-search path without fabricating matches. A bench mode measures a prebuilt index against explicit labels read-only and never silently rebuilds an index or calls a model. Repository: https://github.com/AgriciDaniel/claude-obsidian (the skill is the whole skills/wiki-retrieve/ folder at the pinned commit: SKILL.md only, 1 file) Pinned version: install the code at commit 2e545563172f97f02ecaaaab376ceea1246a0583 exactly. Do not substitute a newer version. If this commit is unreachable, stop and tell me. 1. Clone https://github.com/AgriciDaniel/claude-obsidian into a temporary folder, check out commit 2e545563172f97f02ecaaaab376ceea1246a0583, and copy the skills/wiki-retrieve/ folder from that checkout. Then compare every file in the folder with the fingerprints recorded for this listing on 2026-10-22: [FINGERPRINTS] - SKILL.md: SHA-256 654136eec3577929e457abb73444ff88ca9a753e6079b1e881daf59b70519427 If a file does not match, it changed since the listing was prepared: do not install it yet, tell me what you found and let me decide. 2. Read SKILL.md in full, and tell me in plain words: what the skill does when used, what it connects to, whether it reads or sends any personal information or credentials, and anything that seems unusual. Stop on any red flag and tell me before going further. Confirm these points from the files: - The skill folder holds 1 file only: SKILL.md. The four scripts it runs (scripts/contextual-prefix.py, scripts/bm25-index.py, scripts/retrieve.py, scripts/rerank.py) live outside this folder, in the full claude-obsidian product root. - Nothing is sent anywhere during installation itself, and no consent is needed or inferred for installing: the only download at install time is the skill folder from GitHub at the pinned commit, and no vault is read, no prefix is generated, no index is built and no model is called. Do not supply any API key during this install. - What stays local without egress at use time: synthetic prefixes built with --no-llm, the local BM25 index and its caches under my vault's .vault-meta/, read-only queries against the prebuilt index, local Ollama prefixes and reranking against localhost, and all diagnostics and bench measurements. The skill never changes canonical notes in wiki/. - What would leave the machine, and only under the skill's own fence: page bodies sent to Anthropic ONLY under the explicit Anthropic API prefix profile, with my explicit consent AND the --allow-egress flag AND my own configured API key, after a preview that states which provider receives what; and page text sent to a remote Ollama endpoint ONLY with my explicit approval AND --allow-remote-ollama. Consent is never inferred from an API key or an installed binary, and there is no Claude CLI generation tier or ambient subprocess fallback. - Do not build any prefix or index, do not run any of the four scripts, do not call Ollama or Anthropic, and do not use the skill in any way during this install: using it is a separate step that only starts when I ask for it later. 3. When I say "yes, install", install it as a skill: copy the whole wiki-retrieve folder into the agent's skills directory, in a folder named exactly "wiki-retrieve", keeping its 1 file in its original relative layout (SKILL.md at its root). 4. Verify with no network calls: frontmatter valid (name: wiki-retrieve), the 1 file in place with matching fingerprints. Do not run anything from the folder and do not use the skill to verify. 5. Report what was installed, where, and what I still need to do myself (for example: that a full claude-obsidian installation (the four retrieval scripts at the product root) and my own Obsidian vault are required to use it, that its default local path (synthetic prefixes, local BM25 index, localhost reranking) sends nothing off-machine, that building it writes only disposable caches under my vault's .vault-meta/, and that the optional Anthropic egress stays off unless I give explicit consent with --allow-egress and my own API key). GitHub is optional: if I have a GitHub account or the gh CLI, you may use it; otherwise public access is fine. Never require it unless it's in the prerequisites above. Rules: don't touch anything outside the temp folder and the install target. If anything looks off, stop and ask me.