1. Discover the folder before scanning it
Use the visible folder-selection action in Open data. Review discovered JSONL, Parquet or Arrow shards and their partition context. A layout such as date=2026-10-01/symbol=BTCUSDT can support useful pruning when recognized by the reader.
Confirm that the folder contains the intended split or dataset version. A filename or directory name alone is not a complete schema contract.
2. Select partitions and fields together
Choose the smallest useful time range, category or symbol partition supported by the reader. Project only the required columns. These choices can avoid unnecessary file access and decoding; they do not guarantee that an arbitrary predicate will prune a file.
Inspect returned types and exact values. When a field appears in one page but not another, investigate schema evolution rather than assuming the reader saw every possible field.
3. Treat the cursor as part of the result
Bounded scans can stop because of a record or byte budget. Continue when nextCursor is present, including when the current page contains no matching rows. An empty page is different from a completed scan.
Keep source-file and source-record provenance when you save findings or ask an assistant to explain them. A page's position is not a substitute for the origin of the record.
Do not infer a global table from a local page
The cross-shard workflow does not claim complete schema reconciliation, global sorting, a global aggregate or an exact row count from bounded discovery. Use a suitable explicit analysis workflow when those guarantees matter.
For database-backed exploration, see the table selection guide. For architectural details, read the performance model.
Try the first page on your own data.
Start with the free Windows Preview: source information and up to 100 records, read-only. Review reader capabilities before planning a full-file Pro workflow.
Pro checkout is being prepared. The initial Windows install enables signed in-app updates for later versions.