A full instance reads with an LLM and searches with an embedding model; a lite
instance declares neither. Give a full instance its key, on the page or in the
.env beside revetly.toml.
Open the configuration page. If it is not open, run revetly config and
open the address it prints.
You should see One answer still needed.
Enter your own key in the OpenAI API key box.
Select Save.
You should see Configured and running.
Read the provider line the page carries now.
You should see LLM and embedding: OpenAI on the instance card.
Select Test on the LLM and embedding line. Revetly asks the provider whether it takes this key and can run the model.
You should see 'gpt-4.1-mini' resolves on that line.
Read what is still missing, before anything is spent:
revetly doctor --offline
[FAIL] provider key OPENAI_API_KEY is not set
In .env at the project root, set the empty key line revetly init wrote:
# .env
OPENAI_API_KEY=<YOUR_PROVIDER_KEY>
Replace <YOUR_PROVIDER_KEY> with the key you hold from your provider.
Read the same line back:
revetly doctor --offline
[PASS] provider key OPENAI_API_KEY set …
Revetly reads it from the .env beside the declaration, which revetly init
git-ignores.
Price the re-embedding a promotion off a lite corpus buys:
revetly ingest --dry-run
about <STAND_INS> stand-in vector(s) written while this instance declared no provider are re-embedded first, at roughly 16 token(s) each (the engine's figure) at the `[embedding]` sink's rate — counted on the graph by the mark, the same sweep the warm-up itself runs
If this instance never declared none, the quote names no stand-in vector.
Run the pass that warms the vector space:
revetly ingest
.env beside the declaration. Add it there, or inject it as a secret.[llm] model in revetly.toml.