Vercel AI SDK | Doubleword Inference API

The @doubleword/vercel-ai package provides a Doubleword provider for the Vercel AI SDK, with automatic API key resolution and a pre-configured base URL.

Install

npm install @doubleword/vercel-ai ai

Chat / Text Generation

import { createDoubleword } from "@doubleword/vercel-ai";
import { generateText } from "ai";

const doubleword = createDoubleword({
  apiKey: "{{apiKey}}",
});

const result = await generateText({
  model: doubleword("{{selectedModel.id}}"),
  prompt: "Say hello.",
});

console.log(result.text);

Tool calling

The provider supports multi-step tool use via generateText. The model decides when to call tools, receives the results, and formulates a final answer:

import { createDoubleword } from "@doubleword/vercel-ai";
import { generateText, tool, jsonSchema, stepCountIs } from "ai";

const doubleword = createDoubleword({
  apiKey: "{{apiKey}}",
});

const result = await generateText({
  model: doubleword("{{selectedModel.id}}"),
  tools: {
    calculator: tool({
      description: "Evaluate a basic arithmetic expression",
      inputSchema: jsonSchema({
        type: "object",
        properties: {
          expression: { type: "string", description: "The expression to evaluate" },
        },
        required: ["expression"],
        additionalProperties: false,
      }),
      execute: async ({ expression }: { expression: string }) => {
        return String(new Function(`return (${expression})`)());
      },
    }),
  },
  stopWhen: stepCountIs(5),
  prompt: "What is 137 * 49?",
});

console.log(result.text);

stopWhen: stepCountIs(5) allows up to 5 model→tool→model round-trips before returning. Each step where the model calls a tool automatically feeds the result back for the next step.

Streaming

import { createDoubleword } from "@doubleword/vercel-ai";
import { streamText } from "ai";

const doubleword = createDoubleword({
  apiKey: "{{apiKey}}",
});

const stream = streamText({
  model: doubleword("{{selectedModel.id}}"),
  prompt: "Say hello.",
});

for await (const chunk of stream.textStream) {
  process.stdout.write(chunk);
}

Embeddings

import { createDoubleword } from "@doubleword/vercel-ai";
import { embed } from "ai";

const doubleword = createDoubleword({
  apiKey: "{{apiKey}}",
});

const result = await embed({
  model: doubleword.embeddingModel("Qwen/Qwen3-Embedding-8B"),
  value: "Hello world",
});

console.log(result.embedding.length); // 4096

Doubleword reuses a repeated prompt prefix: the first request stores the processed prefix and later requests read it back cheaply instead of recomputing it. Enable it on the provider to cache the system prefix.

import { createDoubleword } from "@doubleword/vercel-ai";
import { generateText } from "ai";

const doubleword = createDoubleword({
  apiKey: "{{apiKey}}",
  cache: { ttl: "1h" },
});

const { text, usage } = await generateText({
  model: doubleword("{{selectedModel.id}}"),
  system: "…large, stable instructions…", // the ~1024-token floor applies
  prompt: "What is 2 + 2?",
});

console.log(usage.inputTokenDetails); // { cacheReadTokens, noCacheTokens, ... }

ttl is "5m" or "1h", and the cache is left-anchored: a stable prefix followed by a changing tail caches well, a changing prefix caches nothing. Cache activity is reported on usage.inputTokenDetails (cacheReadTokens, noCacheTokens).

To tune a single call, or to cache a different message, pass cacheControl through providerOptions; it overrides the provider default:

const result = await generateText({
  model: doubleword("{{selectedModel.id}}"),
  system: "…large, stable instructions…",
  prompt: "What is 2 + 2?",
  providerOptions: {
    doubleword: { cacheControl: { ttl: "1h", scope: "system" } },
  },
});

scope selects which messages carry the breakpoint: "system" (default), "lastUser", or an array of message indices. Pass cacheControl: false to skip caching for one call.

The same cache option and cacheControl work on createDoublewordAsync and createDoublewordBatch; within one batch the shared prefix is processed once and read by the rest. See the prompt caching guide for the full mechanics.

Default singleton

For convenience, a pre-configured singleton is also exported that reads DOUBLEWORD_API_KEY from the environment:

import { doubleword } from "@doubleword/vercel-ai";
import { generateText } from "ai";

const result = await generateText({
  model: doubleword("{{selectedModel.id}}"),
  prompt: "Say hello.",
});

Batch pricing

For background workloads where latency is not critical, use createDoublewordBatch to transparently route requests through Doubleword's Batch API — cutting inference costs by up to 90% with the Doubleword Inference API. Powered by autobatcher under the hood.

import { createDoublewordBatch } from "@doubleword/vercel-ai";
import { generateText } from "ai";

const doubleword = createDoublewordBatch({
  apiKey: "{{apiKey}}",
  batchWindowSeconds: 2.5,
});

const result = await generateText({
  model: doubleword("{{selectedModel.id}}"),
  prompt: "Summarize this document.",
});

console.log(result.text);
await doubleword.close();

Concurrent generateText calls are automatically collected into batch submissions. The interface is identical to the real-time provider — only streaming is not supported.

Try it end-to-end

A full tool-calling example lives in the repo at examples/tool-calling/. It runs a calculator agent against concurrent arithmetic queries, demonstrating the multi-step agentic loop.