START HERE. Returns what this lake actually holds — every dataset with its coverage window, history mode and freshness — plus the query contract that makes an answer correct (adjustment, point-in-time, which series have no real history). Cheap, reads no data. Call it before answering anything factual about A-share data, and use its coverage_start/coverage_end to tell 'this lake has no such data' apart from 'this did not happen'. It also reports live_mode — when that is on, some answers come from the vendor rather than the lake and say so. Then pick a tool: resolve_symbol for names, query_bars for prices, query_fundamentals for financials, query_dataset for everything else, run_sql for aggregation.

Look up a symbol by company name or partial code — '茅台', 'maotai', '600519' all resolve to 600519.SH. Do this before any query when the user named a company rather than a code: a guessed suffix returns an empty result rather than an error. Delisted names are included and flagged with delist_date. Returns the canonical symbol, name and delist_date; several rows mean the fragment is ambiguous, so pick by name before querying.

Daily or intraday price bars. PRICES ARE STORED UNADJUSTED: pass adjust='hfq' for any comparison across time (returns, charts, drawdowns) or a split reads as a crash. Use adjust='qfq' when the levels must match what a quote screen shows today. Pass universe='all_a' for cross-sectional work to drop names that were not tradable that day. datasets: daily_bars, index_bars, minute_bars (1m), minute_bars_5m (5m), and minute_bars_15m / _30m / _60m when the user has derived them (each bar's resampled_from says whether 1m or 5m built it). Every response carries origin: 'lake' is stored and validated; 'live' was fetched just now and cannot be adjusted or universe-filtered — read its warning before using the numbers for anything but a current quote. Not for financial statements (query_fundamentals), other datasets such as fund flow (query_dataset), or ranking and aggregating across many symbols (run_sql).

Financial statement items, point-in-time. as_of is REQUIRED and is what prevents look-ahead: only figures publicly announced on or before that date come back, and a restated figure returns the vintage that was current then. For a present-day view pass today's date; for a historical study pass the date being studied, never today. Results are paged with limit/offset. For prices use query_bars; for valuation ratios read valuation_metrics through query_dataset. Set pit_mode='strict' only to audit what the lake itself held on that date.

Read any other dataset by date window and symbol — fund_flow, margin_trading, dragon_tiger, block_trades, northbound_holdings, share_unlock_schedule, corporate_actions, valuation_metrics, index_constituents and the rest. Call describe_lake for the full list and each one's coverage. Bars go through query_bars and financial statements through query_fundamentals. Results are paged with limit/offset and the response reports the true total and a truncated flag; for cross-sectional ranking use run_sql instead of paging.

One read-only DuckDB SELECT over every dataset, each exposed as a view under its own name. Use this for ranking, aggregation, percentiles and joins across datasets — computing them here is both correct and far cheaper than paging thousands of rows back and adding them up. daily_bars_adj is a ready-made view carrying hfq_open/high/low/close and open-lake qfq_* beside the raw columns; for a bounded historical qfq window use daily_bars_qfq(start_date, end_date) so the anchor matches that window. Prefer these over joining adj_factors yourself. Single SELECT only: no writes, DDL or multiple statements. For one symbol's bars use query_bars, and for plain dataset reads query_dataset.