Search the site

GETTING STARTED

APP

DATA

BACKEND

OPERATIONS

SHIPPING

REFERENCE

BACKEND

AI

Call Anthropic, OpenAI, Gemini, OpenRouter or a local Ollama model through one API, DV.AI.

Turn a backend function into a tool a model can call, with the JSON Schema written for you.

ON THIS PAGE

Pick a provider

Chat, extract and embed

Let a model call your code

Run a prompt with a budget and a fallback

Status

Pick a provider

void configureAI() {
  DV.AI.configure(AnthropicDVAIAdapter(
    apiKey: DV.Secrets.get('ANTHROPIC_API_KEY'),
  ));
}

Copy code to clipboard

AnthropicDVAIAdapter

Calls: Anthropic. No embeddings or transcription

OpenAIDVAIAdapter

Calls: OpenAI, including Whisper for transcription

OpenRouterDVAIAdapter

Calls: OpenRouter, with the OpenAI request shape

GeminiDVAIAdapter

Calls: Google Gemini

OllamaDVAIAdapter

Calls: A local Ollama server, localhost:11434 by default

LocalDVAIAdapter

Calls: Nothing. A predictable answer for tests and offline work

Nothing picks a provider for you. Call configure once at start-up, or every call throws.

Chat, extract and embed

final String summary = await DV.AI.chat('Summarise this ticket: $ticket');

final DVJsonObject triage = await DV.AI.structuredOutput(
  'Triage this ticket: $ticket',
  const <String, DVJsonValue>{
    'type': DVJsonString('object'),
    'properties': DVJsonMap(<String, DVJsonValue>{
      'urgent': DVJsonMap(<String, DVJsonValue>{'type': DVJsonString('boolean')}),
    }),
  },
);

final List<double> vector = await DV.AI.embed(ticket);

Copy code to clipboard

structuredOutput takes a JSON Schema and returns a map in that shape.

embed returns a vector you can store and search.

transcribe and runAgent are there too. A provider that cannot do one throws UnsupportedError.

Let a model call your code

@DVAITool(description: 'How many orders are waiting in a status')
Future<int> openOrders(String status) async => status == 'paid' ? 3 : 0;

Copy code to clipboard

dartvel routes writes a JSON Schema and a handler for each public @DVAITool function under lib/backend.

The generated backend registers every tool before it serves.

An argument of the wrong type is refused by name, so a tool never runs on a guessed value.

Serve your tools over MCP

DVMcpServer answers tools/list and tools/call for the registered tools, and DVMcpClient adopts the tools of another MCP server. dartvel mcp is a different server: it lets a coding agent read your project's routes, models and jobs.

Run a prompt with a budget and a fallback

final DVPrompts prompts = DVPrompts()
  ..register(
    const DVPrompt(id: 'ticket.summary', version: 4),
    const DVPromptTemplate(
      system: 'Summarise the ticket for a support agent.',
      input: <String, Type>{'body': String},
    ),
  );

final DVAIFeatures features = DVAIFeatures(
  prompts: prompts,
  adapter: const LocalDVAIAdapter(),
  model: 'local',
)..register(const DVAIFeature(
    prompt: 'ticket.summary',
    fallback: <DVAIFallback>[DVAIFallback.degrade],
  ));

final DVAIFeatureResult result = await features.run(
  'ticket.summary',
  input: <String, Object?>{'body': body},
  idempotencyKey: 'summary:$ticketId',
);
final String? summary = switch (result) {
  DVAIAnswered(output: DVJsonString(:final String value)) => value,
  _ => null, // degraded, refused by a budget, or the provider was down
};

Copy code to clipboard

Prompts are versioned. A stored override can be rolled back, and every change is kept in an audit list.

A budget is a usage meter checked before the call. A budgeted run needs an idempotency key, so a retry is counted once.

The result says what happened: answered, degraded, refused by the budget, or unavailable.

DVAIEval replays golden transcripts against a feature and fails when too few match.

Partial

Spec section: AI Operations

Planned work and implementation limits

Prompts and features are registered in code. No generator reads @DVPrompt or @DVAIFeature, and nothing is wired into DV.AI.

The prompt store is in memory, and there is no dartvel ai eval command.

Tokens are estimated at 4 characters each, since adapters report no usage.

Status

Partial

Spec section: AI

Planned work and implementation limits

A tool's schema has a type per parameter and no descriptions or enums.

A tool that returns a type DVJsonCodec cannot encode throws when the result is sent.

FSL-1.1-MIT licensed. Built with Dartvel.

Dartvel is made by

To the bottom