ONNX (Open Neural Network Exchange) is a portable model format that lets RankGear’s development semantic enrichment evaluator run a pinned embedding model on its own, independent of the framework the model was originally trained in.
| Term | ONNX |
|---|---|
| Category | APIs, Reliability and Publishing |
| Also known as | Open Neural Network Exchange |
| Where it appears | Semantic enrichment evaluator |
What it means in RankGear
ONNX is the on-disk format the semantic enrichment evaluator loads when it turns page text into embeddings. Exporting the model to ONNX lets RankGear pin one exact version of the embedding model and run it through a standard runtime instead of carrying the full training stack it came from. That is what keeps the evaluator’s numbers stable and reproducible from one machine and one run to the next.
How to interpret it
ONNX is infrastructure, not a score, so you read it through the evaluator’s operational signals rather than as a value on a page. When the semantic evaluator misbehaves, the visible status, the diagnostics, the retry decision, and the credential test are what separate a transient failure — a slow model load or a timed-out call — from a real configuration problem such as a missing file or wrong credentials. A one-off stall usually clears on the next attempt; a genuine config error repeats identically every time.
Example
A run across a batch of pages stalls on a single URL while the evaluator is loading embeddings. Because the fault is transient, that page receives one bounded longer retry and then completes. Had the failure instead been an authorization error against a remote target, the evaluator would have stopped there rather than firing repeated paid calls.
Important considerations
- ONNX is a plumbing detail of the development evaluator, not a ranking factor or a score — the format itself has no bearing on how a page ranks.
- Because the model is pinned, the embeddings it produces are comparative indicators meant for like-with-like comparison, not Google scores, and correlation with rank is not causation.
- Where the evaluator touches credentials or remote publishing targets, tests should stay read-only where possible and secrets must remain redacted in logs and diagnostics.
Related terms
Part of the RankGear glossary · how RankGear measures · the 870 factors.