AI features, no API key needed
Semantic caching, agent memory and search by meaning with models KiviDB runs for you: Amazon Titan Text Embeddings V2 and Amazon Nova Lite, billed per token from your credit. No model provider...
A semantic cache, agent memory and search by meaning all need an embedding model, and agent memory needs a language model too. KiviDB runs both for you. Choose KiviDB as the provider and pay per token from your credit: no account at a model provider, and no key to store, rotate or leak.
Two models, run for you
- Embeddings Amazon Titan Text Embeddings V2 at 256, 512 or 1024 dimensions, normalised for cosine similarity. It turns prompts, memories and queries into vectors, for the semantic cache, agent memory and search by meaning.
- A language model Amazon Nova Lite reads agent memory's conversations in the background, decides what is worth remembering, and summarises long sessions.
- An embeddings API POST up to 64 texts to /instances/{id}/embeddings and get their vectors back, to store beside your documents so search by meaning compares like with like.
Per token, from your credit
| Hosted by KiviDB | Your own provider key | |
|---|---|---|
| Embeddings | $0.05 per million tokens | Your provider's price |
| Language model, read | $0.15 per million tokens | Your provider's price |
| Language model, written | $0.60 per million tokens | Your provider's price |
| A key to store and rotate | no | yes |
| An account at a model provider | no | yes |
| Where the model runs | Amazon Bedrock, AWS us-east-1 | Your provider's region |
A token is roughly four characters of English. A cache answering 10,000 lookups a day of 200-character prompts uses about 15 million tokens a month: $0.75. Usage shows by database on the Usage page.
One setting in each feature
- Semantic cache: Answer a reworded question from the cache instead of the model. Choose KiviDB hosted when you create a namespace.
- Agent memory: Extraction and embeddings on hosted models: a store with no keys at all.
- Search by meaning: Type a query against an index of Titan vectors, at 256, 512 or 1024 dimensions; the Search tab picks the matching model.
What happens to your text
- Not kept, not trained on Text is sent to Amazon Bedrock in AWS's us-east-1 region, whatever region your database is in. Bedrock does not store prompts or use them to train models. If your data must stay in one region, use your own provider.
- Spending stays bounded Hosted models answer an organization only while it has a month of its running databases, plus $1, in credit. Calls are rate-limited per organization, at roughly 20 a second.
- Switch either way Move a feature between hosted models and your own key in its settings. A cache or store that already holds vectors must be emptied first: vectors from two models cannot be compared.
Questions
Do I still need an OpenAI or Anthropic account?
No. With KiviDB hosted chosen, the semantic cache, agent memory and search by meaning need no other account. Your own keys keep working where you prefer them.
Can I mix hosted and my own models?
Yes. An agent memory store can extract with your own model and embed with ours, or the other way round, and each cache chooses its own provider.
Which threshold should a hosted semantic cache use?
Start at the default, 0.45. Titan scored reworded questions 0.29 to 0.78 and unrelated ones at most 0.29 in our measurements; the docs have them by size.
Is it slower than my own key?
A hosted embedding takes about 210 ms, because the model runs in us-east-1. A whole cache lookup took about 250 ms for a database in Frankfurt and about 750 ms for one in us-east-1, as the API is in Frankfurt. Sending your own vectors avoids the embedding call.
Is there a free allowance?
No. Tokens are billed from the first one, from your credit, at the prices above.
Add AI features without a key
Create a database, then choose KiviDB hosted as the provider for a semantic cache or a memory store.