AVAConstructor
import { AVA } from '@antv/ava';
const ava = new AVA({
llm: {
model: 'YOUR_MODEL',
apiKey: process.env.OPENAI_API_KEY,
baseURL: process.env.OPENAI_BASE_URL,
maxRetries: 3,
},
engine: { type: 'duckdb' },
});llm is required. Omitting engine selects DuckDB; frontend applications must explicitly select javascript.
Model options
| Field | Type | Required | Meaning / default |
|---|---|---|---|
provider | 'openai' | 'gateway' | No | Defaults to OpenAI-compatible Chat Completions. Gateway reads AI_GATEWAY_API_KEY. |
model | string | LanguageModel | Yes | Model name or a preconfigured AI SDK model instance. |
apiKey | string | No | API key for the LLM provider |
baseURL | string | No | Optional API base URL |
maxRetries | number | No | Maximum retries for retryable model API failures. Default 3. |
onQueryUsage | (usage: LanguageModelUsage) => void | No | Optional callback for SQL-generation token usage |
String model names use OpenAI-compatible Chat Completions by default. A supplied model object is used directly. Usage callbacks are not a complete accounting stream for every LLM request.
AI Gateway
const ava = new AVA({
llm: {
provider: 'gateway',
model: 'openai/gpt-4o-mini',
apiKey: process.env.AI_GATEWAY_API_KEY,
},
});The gateway provider can also read AI_GATEWAY_API_KEY from the environment when an explicit key is omitted.
Separate retry controls
llm.maxRetries controls model API retries. The direct analysis strategy has a separate maxRetries option for correcting failed executable queries. See analysis strategies.
Credentials in frontend apps
Keep shared provider secrets on the server. For browser-only experiments, obtain credentials at runtime rather than including them in source code or the frontend bundle.