# /geekbench/ai_results/search

`POST /api/geekbench/ai_results/search`

Price: 5 credits

Search uploaded Geekbench AI runs by device or chip name, or walk the newest and the top-scoring runs when no query is given. Each row carries the run id and URL, the device and chip name, the inference framework, and the single-precision, half-precision and quantized scores; the run list also carries the upload time. Feed an id to the Geekbench AI run endpoint for the workload breakdown, the backend and the host machine.

## How to use it

Two upstream surfaces sit behind one call and they are exclusive: with `query` you get text matching in a fixed order, without it you get the run list, which is the only place `sort` and `order` apply. Use the sorted list to find the fastest hardware at a precision, and the query to find a specific device. The query matches the device and chip strings the machine reported, so a desktop often appears under its board name and not under any product name.

## Parameters

- `access-token` (string, required)

## Request body

- `timeout` (integer) — Max scrapping execution timeout (in seconds) (default: 300; min: 20; max: 1500)
- `query` (string, nullable) — Free text matched against the uploaded device and chip name. Leave it out to walk the newest or the top-scoring runs instead (examples: "iPhone 17", "Apple M4", "RTX 5090"; minLength: 1)
- `sort` (string) — Ordering of the run list. Only available without a query — the text search always returns its own ordering (default: "created_at"; one of: "created_at", "sp_score", "hp_score", "q_score")
- `order` (string) — Ascending or descending, applied to sort. Only available without a query (default: "desc"; one of: "desc", "asc")
- `count` (integer, required) — Max number of results to return (min: 1; max: 1500)

## Response

### 200 — Successful Response

- `@type` (string) (default: "GeekbenchAiResultRow")
- `id` (string, required)
- `url` (string, required)
- `version` (string, required)
- `system` (string, nullable)
- `model` (string, nullable)
- `platform` (string, nullable)
- `user` (string, nullable)
- `uploaded_at` (integer, nullable)
- `frequency_mhz` (integer, nullable)
- `core_count` (integer, nullable)
- `details` (object) (default: {})
  - `*` (string)
- `framework` (string, nullable)
- `single_precision_score` (integer, nullable)
- `half_precision_score` (integer, nullable)
- `quantized_score` (integer, nullable)

## Errors

### 422 — Validation Error

sort or order was passed together with a query, which the source cannot combine

What to do: Check the fields against this schema. A URN with the wrong prefix is the most common cause.

- `detail` (array)
  - `loc` (array, required)
  - `msg` (string, required)
  - `type` (string, required)
  - `input` (any)
  - `ctx` (object)

### 408

The request ran past its time limit

What to do: Raise `timeout` in the request body, up to the maximum this endpoint documents. Lowering `count` or turning off the `with_*` flags also helps, because less work finishes sooner.

### 412

The entity was not found, or a precondition failed

What to do: Retrying will not help: either the entity does not exist, or the input points at a different one.

### 429

Too many requests: a rate limit or a usage window is exhausted

What to do: When the response carries an X-Retry-After header, wait that many seconds and retry: the same number is in the body as `detail.retry_after`, and the limit clears once that window passes. The message in the body names the limit that was hit.

### 500

Something broke on our side

What to do: Retrying will not help. If it keeps happening, send us the X-Request-ID from the response headers.

### 529

Rate limit reached, or the endpoint is overloaded

What to do: Wait at least 30 seconds, then retry.

## Response envelope

Success: Array of objects (may be empty if no results)

Error: Error may coexist with partial results if it occurs mid-execution. Check X-Error header and status code.

Every response carries these headers:

- `X-Error` — Error message text (present only on error)
- `X-Request-ID` — Unique request identifier
- `X-Execution-Time` — Execution time in seconds
- `X-Result-Count` — How many records the body carries. 0 means an empty result, which is a normal answer and not by itself an error. A non-zero count can come back together with X-Error when the failure happened partway through — read this header and X-Error independently.
- `X-Total-Available-Results` — How many records exist for this query, when the endpoint can say. On a `dry_run` request this is the answer and the body is empty. It saturates: the endpoint's documented maximum means 'at least that many', any smaller number is exact.
- `X-Warning` — Present when the request body carried keys this endpoint does not document. They were ignored, so any filter you meant to apply through them did not apply. Check the spelling against this schema and retry.
- `X-Retry-After` — Seconds to wait before retrying. Present only on 429.