Uploading data

RINK accepts comma-separated values (.csv) files up to 10 MB. There's no minimum row count for upload, but you'll need at least 30 rows of valid numeric data to train a model.

File requirements

Requirement	Value
Format	.csv (UTF-8 recommended)
Size	≤ 10 MB
Header row	Required — first row is treated as column names
At least one	Numeric column
Recommended	A date / timestamp column

If a non-CSV file is selected, the upload zone rejects it client-side before any bytes leave your browser. The same check is repeated on the gateway and the ML service.

🔒 Your file is encrypted and scanned. Before anything is stored, the ML service scans the upload (format guard + VirusTotal) and rejects binaries, archives, executables, malware, and anything that isn't plain-text CSV. Accepted files are then encrypted at rest — the plaintext CSV never reaches storage. Full details on the Security page.

Date column detection

When you upload, RINK looks for a date column in this order:

1. Named columns: date, timestamp, time, datetime, ds, period — case-insensitive variants like Date, DATE are also matched.
2. Parseable strings: any object-typed column where ≥60% of the first 20 sampled values parse as a date.

If a date column is found, RINK:

• Sorts the rows ascending by date. Reverse-chronological CSVs are reordered, so "most recent" always means the last row after upload.
• Infers the cadence from the median delta between consecutive timestamps. Buckets: daily / weekly / monthly / quarterly / yearly, or every X days for irregular series.

If no date column is found, rows are processed in the order they appear in the file and the cadence shows as unknown. You can still train and forecast, but step semantics fall back to "+1, +2, +3, …" instead of real future dates.

Numeric column detection

After excluding the date column, RINK collects all numeric columns. The first match wins from a preference list:

value, y, target, close, price, pmms30

If none of these are present, the first numeric column in CSV order is used.

You can override the auto-detection at any time using the column picker in the Model card — see Switching columns below.

Panel / grouped data

Some datasets carry many rows per date — for example temperature per city per day, or sales per store per week. This is called panel (or grouped) data, and forecasting it as one blended series gives poor results because several independent series are interleaved.

RINK detects this automatically. When it sees repeating dates plus a low-cardinality categorical column (an ID / group column) that makes each (date, group) pair unique, it:

• Flags the dataset as panel data and suggests the group column (e.g. city).
• Lists the available groups (e.g. Detroit, Austin, Seattle).
• Shows a hint: "Multiple rows share the same date — pick one group to forecast a single, clean series."

A Training Scope card then appears in the workspace with a group picker. Choose a group and click Apply & Re-train to forecast that one series. Leaving it on All groups trains on the combined data.

This profiling is exposed via the /api/analyze endpoint, which also reports date_min / date_max and an encryption_at_rest flag.

What happens after upload

1. The file is forwarded to the ML service, which virus-scans it (VirusTotal), encrypts it, and stores it under users/<your_uuid>/uploaded.csv in object storage (Cloudflare R2, or a local-disk fallback).
2. The model auto-trains. Status toast: "Trained on N rows · weekly cadence · RMSE … · MAE …"
3. The chart and KPIs refresh. The prediction input pre-fills with the most recent N values.

Your file library

Every file you upload is kept in your encrypted file library in the bucket — uploading a new file no longer discards the previous one. The newest upload becomes the active dataset that the tools analyze.

From the My data panel in the workspace (and in your Profile) you can:

• See every file you've uploaded (name, size, row count, date) and which is active.
• Use for analysis — switch the active dataset to any stored file (re-trains automatically).
• Delete a single file, all files, or everything (files + reports) at once.

Files are retained for up to 90 days, then auto-deleted; you can delete sooner anytime.

Switching columns

If your CSV has multiple numeric columns, the Model card shows them all:

• 2–6 columns — emerald pill buttons.
• 7+ columns — a <select> dropdown.

Picking a different column atomically:

1. Refetches /api/data?column=NAME.
2. Calls /api/train with { "column": "NAME" }.
3. Clears stale predictions.

The selection is saved to localStorage so it persists across reloads. Uploading a new CSV clears the saved choice.

Storage and cleanup

• Your CSV is virus-scanned then encrypted before it reaches storage (when an encryption key is configured — always on in production). See Security → encryption at rest.
• Each user's data is isolated under users/<user_id>/ — other users cannot see or access your data.
• Files are retained for up to 90 days and then deleted automatically. You can delete everything sooner from your profile. Signing out does not delete your data.
• Forecast values can be exported to CSV from the Forecast Detail card — see the FAQ.

Limits

Property	Limit
File size	10 MB
Server-stored data per user	One CSV at a time (replaces on re-upload)
Min rows to train	30 (after feature dropping)
Concurrent users on Render Starter	~50 lightly active