A model cache is a local store of downloaded embedding-model files that RankGear keeps on disk, so the semantic evaluator can reuse them instead of re-downloading and re-initializing the model on every run.
| Term | Model Cache |
|---|---|
| Category | APIs, Reliability and Publishing |
| Also known as | Embedding Model Cache |
| Where it appears | Semantic enrichment evaluator |
What it means in RankGear
The semantic enrichment evaluator depends on an embedding model to turn page text into vectors it can compare. Fetching and loading that model is the expensive part of the work, so RankGear writes the retained model files to a local cache. The first evaluation pays the download-and-initialize cost once; subsequent runs read the same files straight from disk. In practice this is a development-side performance mechanism that keeps repeated semantic checks fast rather than re-paying setup work each time.
How to interpret it
Treat the cache state as your first clue when an evaluation is slow or fails to start. A cold cache (missing or incomplete files) triggers a download and a one-time delay; a warm cache loads almost immediately. Use the visible status, diagnostics, and any credential test to tell three situations apart: a cache miss that is simply still downloading, a transient network failure that warrants one bounded retry, and an invalid configuration (wrong path or bad credentials) that no amount of retrying will fix.
| Cache state | What it tells you |
|---|---|
| Cold | Files absent or partial; first run downloads and initializes, so expect a one-time delay. |
| Warm | Files present locally; the model loads quickly and evaluation proceeds without a fetch. |
| Stale or corrupt | Files present but unusable; clearing the cache forces a clean re-download. |
Example
The first time you open a workbook and run the semantic enrichment evaluator, RankGear downloads the embedding model and writes it to the cache, adding a noticeable one-time delay. When you run the evaluator on the next workbook, the same model loads from disk in a fraction of the time. If a download times out partway through, the evaluator makes a single bounded longer retry instead of repeatedly re-fetching the same files.
Important considerations
- The cache is local and per-machine. Clearing or moving it forces a fresh download on the next run.
- Where the evaluator touches credentials or remote publishing targets, handle them carefully: keep tests read-only where possible and ensure secrets stay redacted in logs and diagnostics.
- Distinguish a transient failure (worth one bounded retry) from invalid configuration (which will keep failing) before re-running, so you avoid wasted work.
- A model cache is an operational and performance mechanism. It speeds up semantic evaluation; it does not change any score and has no bearing on how pages rank in Google.
Related terms
Part of the RankGear glossary · how RankGear measures · the 870 factors.