API documentation
Reference
Find error codes, input limits, rate limits, and usage details.
Error codes
All errors share a common envelope. The request_id matches the one in usage logs — include it when contacting support.
{
"error": {
"code": "invalid_request",
"message": "horizon: Input should be less than or equal to 512",
"request_id": "req_a1b2c3d4e5f67890a1b2c3d4e5f67890"
}
}| HTTP | Code | Description |
|---|---|---|
| 400 | invalid_request | Schema, value, or limit violation in the request body |
| 401 | unauthorized | Missing, invalid, or revoked API key |
| 402 | payment_required | Insufficient credits |
| 403 | forbidden | Insufficient permissions |
| 404 | not_found | Unknown route |
| 405 | method_not_allowed | HTTP method not allowed for this route |
| 408 | request_timeout | Request body was not received within 15 seconds |
| 413 | payload_too_large | Request body exceeds the size limit |
| 429 | rate_limited | Rate limit exceeded — retry with backoff |
| 500 | internal_error | Unexpected server error |
| 503 | model_unavailable | Model is not available — retry later |
| 504 | timeout | Inference timed out |
Input limits
Requests exceeding these limits are rejected with 400 invalid_request.
| Limit | Value |
|---|---|
| Points per series | 16–2048 |
| Series per request | up to 64 |
| Forecast horizon | 1–512 |
| Total input points per request | up to 131,072 |
Rate limits
Applied per API key with a token bucket: burst of 10 requests, refilling at 1 request/sec (free tier). Exceeding it returns 429 rate_limited.
CSV upload format
The CSV accepted by "Try your data" in the Playground and the result CSV it returns. Same columns as the specimen on that screen.
| Item | Details |
|---|---|
| File | Comma-separated. UTF-8 or Shift_JIS. Up to 10 MB |
| Row 1 | Header row. Column names are free-form (duplicates become _2, _3, …) |
| Column with dates or times | One column (inferred; optional). ISO 8601 date-time or date only. A date-time without an offset is read as the browser's local time. If rows are evenly spaced, future timestamps are filled in on the result |
| Value columns | Numbers only; no empty cells or Infinity. Selecting several forecasts them jointly as one multivariate task (sent with a group) |
| Limits | 16–2048 rows, up to 64 value columns. Same as the API input limits |
| Handling | The file is read in the browser; only the selected numeric columns are sent to the server. Nothing is stored |
timestamp,building_a_kw,building_b_kw,building_c_kw
2014-11-18T01:00:00Z,153.1,98.5,417.7
2014-11-18T02:00:00Z,148.1,97.1,387.1
2014-11-18T03:00:00Z,100.7,99.6,375.3timestamp,type,building_a_kw,building_b_kw,building_c_kw
2015-01-01T00:00:00Z,actual,71.8,72.2,299.8
2015-01-01T01:00:00Z,forecast,127.6,110.6,434.2
2015-01-01T02:00:00Z,forecast,130.7,76.2,347.2The result CSV lists actual rows followed by forecast rows, distinguished by the type column (actual / forecast). Forecast timestamps extend the input interval. Date-only columns stay date-only, and date-times with seconds and a Z keep that format; all other date-times use UTC ISO 8601. Forecast rows also use UTC for offset-less inputs so that times remain distinct across daylight-saving transitions. Actual rows keep their original values. Cells starting with = + - @ are quoted so they are not executed as formulas.
Usage
The usage object in every 200 response shows how many data points the request consumed. The same numbers appear in the request log on the Usage page.
| Field | Description |
|---|---|
| input_points | Total input points across all submitted series |
| output_points | Total points returned. Defined per endpoint: forecast = horizon × series count / anomaly = sum of 2 × (length − test_start) / impute = sum of imputed (null) points / embed = 0 |
| series_count | Number of series in the request |