Yearn Docs: hosted retrieval

Inputs

• •BASE (optional): The docs site origin to query (e.g. a Vercel preview). Defaults to https://docs.yearn.fi.

Example:

BASE=https://docs.yearn.fi

Tip: if you use a bash one-liner like BASE=... python3 ... --base "$BASE" ..., $BASE expands before the assignment takes effect. Prefer either --base 'https://...' directly, or export BASE=... first, or omit --base and rely on the default.

One-approval workflow (recommended)

To avoid repeated network approvals in restricted environments, use the bundled helper. It auto-checks for updates (conditional request) and only downloads docs.jsonl when it has changed, so you typically need only one approval per invocation:

python3 skills/yearn-docs-site-query/scripts/yearn_docs_query.py --base "$BASE" search "yCHAD multisig signers"

If you want absolutely no network calls, force offline mode (no network at all):

python3 skills/yearn-docs-site-query/scripts/yearn_docs_query.py --base "$BASE" --offline search "veYFI gauge"

To fetch a page by route:

python3 skills/yearn-docs-site-query/scripts/yearn_docs_query.py --base "$BASE" get /developers/security/multisig

To clear cached files for that BASE:

python3 skills/yearn-docs-site-query/scripts/yearn_docs_query.py --base "$BASE" cleanup

To see whether your cached data is stale (age/etag/manifest info):

python3 skills/yearn-docs-site-query/scripts/yearn_docs_query.py --base "$BASE" status

To check the remote for updates and refresh the cache if needed (will require network approval in restricted environments):

python3 skills/yearn-docs-site-query/scripts/yearn_docs_query.py --base "$BASE" --check-updates status

To skip update checks and use cached data (if present):

python3 skills/yearn-docs-site-query/scripts/yearn_docs_query.py --base "$BASE" --no-auto-update search "multisig signers"

To force a full re-download (even if unchanged):

python3 skills/yearn-docs-site-query/scripts/yearn_docs_query.py --base "$BASE" --refresh status

Endpoints (relative to BASE)

• •GET /llms.txt — human/agent pointer file.
• •GET /ai/manifest.json — machine-readable pointers (paths, counts, origin).
• •GET /ai/docs.jsonl — plaintext corpus (JSON Lines: 1 JSON object per doc page).
• •GET /ai/raw/... — optional raw .md/.mdx source mirror (if present in corpus records).

Retrieval workflow

1. •Fetch GET {BASE}/ai/manifest.json and GET {BASE}/llms.txt.
2. •Download GET {BASE}/ai/docs.jsonl and parse it as JSONL (stream line-by-line).
3. •Rank candidate records for the user query:
   • •Tokenize the query and score matches.
   • •Boost matches in title and headings[].text over matches in text.
4. •Select the top records (e.g. 3–8), then select the best sections:
   • •Split text by headings and keep the most relevant chunks.
5. •Answer using only the selected chunks (don’t hallucinate fields not present in the docs).

Record shape (docs.jsonl)

Expect at least:

• •title: page title
• •headings: array of { level, id?, text }
• •text: extracted plaintext (may include fenced code blocks)
• •route: the path on the site host (e.g. /developers/addresses)
• •url: canonical URL for the page (often on https://docs.yearn.fi/...)
• •source.htmlPath: where it came from in the static build

Optional/newer:

• •canonicalUrl
• •source.rawPath: path under /ai/raw/... for exact .md/.mdx

Citing and linking

• •Prefer citing the canonical URL in the record:
  • •Use canonicalUrl if present, else url.
• •If url points to the wrong origin for the environment you’re using, cite BASE + route.

Exact-source fallback (for precise quotes/formatting)

When you need exact formatting, tables, or the original MDX:

1. •If source.rawPath exists: GET {BASE}{source.rawPath} and quote from that.
2. •Otherwise fetch the rendered page: GET {BASE}{route} and extract the relevant section from HTML.

Caching

• •Cache GET {BASE}/ai/docs.jsonl per session.
• •If you can, revalidate with ETag/If-None-Match on subsequent runs.