Scraping Nessie v2 /docs for a RAG/LLM index yields one paragraph: the page is a Scalar JS viewer; the full guide is markdown in openapi.json info.description (~33 KB)
Index the spec, not the HTML. curl https://v2.nessieisreal.com/docs returns about 13 KB of HTML whose only static text says the reference is rendered by Scalar, which needs JavaScript. The page calls Scalar.createApiReference('#app', config) and loads /openapi.json or /openapi.yaml in the browser. BeautifulSoup on that page gave our chunker exactly 1 chunk. Everything the docs show is in the spec: jq -r .info.description openapi.json is a ~33,000 character markdown guide with headings like 'Authentication, keys and clearing data', 'Rate limits', 'Accounts, balances and credit cards', 'Reading balances', 'Not enough money?', 'Transactions: how money moves', 'Webhooks: notifications and live decisions', 'Decision hooks', 'The event feed', 'Event types' and 'The polling pattern'. Then there are 49 paths and 59 component schemas with per-field descriptions. A workable chunking: split info.description on ## headings, one chunk per operation (method + path + summary + description + parameters), and one per schema (property names plus descriptions). That gave about 295 chunks and answers questions like 'what is the burst limit' or 'which fields are required for bills'. For keyword-heavy questions (status codes, field names like balance_cents), combine BM25 (rank_bm25) with embeddings, because a small CPU embedder like all-MiniLM-L6-v2 is weak on exact identifiers. The spec is minified single-line JSON (OpenAPI 3.1.0, info.version 2.1.3 at the time), so head openapi.json | jq fails. Fetch it whole.