Anvia

Completions ​

Use completionModel(...) with an Agent when a workflow needs instructions, tools, memory, context, or multiple turns.

ts
import { Agent } from '@anvia/core'
import { MistralClient } from '@anvia/mistral'

const mistral = new MistralClient({
  apiKey: process.env.MISTRAL_API_KEY!,
})

const model = mistral.completionModel({
    modelId: 'mistral-large-latest'
})

export const supportAgent = new Agent({
  id: 'support',
  model,
  instructions: 'Answer support questions clearly and concisely.',
})

Generate one answer ​

ts
const result = await supportAgent.generate({
    prompt: 'Explain why a refund can take two business days.'
})

if (result.type === 'response') {
  console.log(result.output)
}

generate(...) runs the agent and returns a normalized response, interaction, or blocked outcome. Use a direct completion when the application owns a single model call and does not need the agent runtime:

ts
import { generateCompletion } from '@anvia/core'

const result = await generateCompletion({
    prompt: 'Checkout requests timed out for 12 minutes.',
    model,
    instructions: 'Write a concise internal incident summary.',
    maxTokens: 180
})

console.log(result.text)
console.log(result.usage.totalTokens)

The input, configured model, and request options belong in the same options object.

Stream an agent run ​

agent.stream(...) returns an async iterable. The run advances as the application consumes it:

ts
for await (const event of supportAgent.stream({
    prompt: 'Draft a response to this support ticket.'
})) {
  if (event.type === 'text_delta') {
    process.stdout.write(event.delta)
  }
}

Consume the stream through its terminal event so usage, tool results, observers, and completion state can settle. Use @anvia/server for browser transport instead of exposing provider credentials.

Supported requests ​

The adapter maps text instructions and message history, temperature, maximum output tokens, streaming text and tool-call deltas, tools, tool choice, output schemas, and provider-specific parameters.

It rejects chat image inputs and data or URL document inputs before making the provider request. A file part carrying { type: 'text', text } data is the exception: the adapter silently flattens it into the prompt text. Use Mistral OCR as a separate extraction step for scanned documents and images.

Provider-specific parameters ​

Pass a Mistral-specific value through providerOptions only at the narrow boundary that needs it:

ts
const result = await generateCompletion({
    prompt: 'Give this release note a short title.',
    model,
    providerOptions: {
        randomSeed: 42,
    }
})

The adapter preserves the normalized model and messages fields even if those keys appear in providerOptions. Verify provider option names against the Mistral API version used by your deployment.

Built for Anvia.