Get an OpenAI API key for embeddings

Create the account

Go to https://platform.openai.com and sign up (Google/GitHub/email).
The API platform is billed separately from ChatGPT Plus. A Plus subscription gives you zero API credits.
Verify your phone number — required before a key can be created.

Add credit

Settings → Billing → Add payment method.
Add credits (minimum is a few dollars; $5 lasts a very long time for embeddings).
Optional but recommended: Usage limits → set a hard monthly cap so a runaway loop cannot drain the card.

Without credit, every request comes back 429 insufficient_quota. It is not a rate limit, it is an empty balance.

Create the key

Settings → API keys → Create new secret key.
Name it after the project (todo-ai-search) so you can revoke it in isolation.
Optional: restrict permissions to Model capabilities only.
Copy it now — sk-... is shown exactly once.

# .env.local — never commit this
OPENAI_API_KEY="sk-proj-..."

Make sure .env*.local is in .gitignore. If a key leaks, revoke it in the dashboard; it dies instantly.

Pick an embedding model

Model	Dimensions	Relative cost	Use when
`text-embedding-3-small`	1536	cheapest	default, good enough for most search
`text-embedding-3-large`	3072	~6.5x	you have measured that `small` is not good enough

The dimension count becomes part of your DB schema (vector(1536)).
dimensions in the request body can shorten the output (e.g. 512) — but shortened vectors are only comparable to other shortened vectors.

Smoke test the key

curl https://api.openai.com/v1/embeddings \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"text-embedding-3-small","input":"buy ice"}' \
  | head -c 300

You should get {"object":"list","data":[{"object":"embedding","index":0,"embedding":[0.02....

Batch form — one round trip, many inputs:

-d '{"model":"text-embedding-3-small","input":["buy ice","order the cake"]}'

Results come back in the same order as input, each with its index.

Notes / gotchas

No SDK needed. Embeddings are one POST; fetch is enough.
Errors worth recognising:
- 401 — bad or revoked key, or key from the wrong org/project.
- 429 insufficient_quota — no credit.
- 429 rate_limit_exceeded — actual rate limit, back off and retry.
Always set a timeout (AbortSignal.timeout(10_000)), a hung provider should not pin a request open.
Embeddings are deterministic-ish and cacheable — same text + same model = same vector. Do not re-embed unchanged text.
Cost scale: embedding a few thousand short strings with 3-small is cents, not dollars.

CarloGiNotes

Explorer