Vercel AI SDK

Agents built with the AI SDK need current facts they can cite. Pasting search pages into the prompt burns tokens and hides where each claim came from. Give the agent Flux as tools and it gets short passages, with their sources and the names and dates they mention.

Build it

1. 1

   Install

   Shell

   npm install ai @ai-sdk/openai zodexport FLUX_API_KEY="<your key>"

2. 2

   Define the tools

   webSearch finds pages. readPage pulls the passages of one page that answer a question.

   flux-tools.ts

   import { tool } from "ai";import { z } from "zod";type Passage = { enriched_text: string };type SearchResult = { url: string; lineage: { results: Passage[] } | null };async function flux<T>(path: string, body: object): Promise<T> {  const response = await fetch(`https://fluxsearch.io/api/v1/${path}`, {    method: "POST",    headers: {      Authorization: `Bearer ${process.env.FLUX_API_KEY}`,      "Content-Type": "application/json",    },    body: JSON.stringify(body),  });  if (!response.ok) throw new Error(`Flux ${path} failed with ${response.status}`);  return (await response.json()) as T;}export const webSearch = tool({  description: "Search the live web. Returns passages with provenance and their source URLs.",  inputSchema: z.object({ query: z.string() }),  execute: async ({ query }) => {    const { results } = await flux<{ results: SearchResult[] }>("search", {      query,      max_results: 5,      enrich: 3,    });    return results.flatMap((result) =>      (result.lineage?.results ?? []).map((passage) => ({        url: result.url,        text: passage.enriched_text,      })),    );  },});export const readPage = tool({  description: "Read one web page and return only the passages that answer a question.",  inputSchema: z.object({ url: z.string(), question: z.string() }),  execute: async ({ url, question }) => {    const { results } = await flux<{ results: Passage[] }>("provenance", {      url,      query: question,    });    return results.map((passage) => passage.enriched_text);  },});

   Show all 48 lines

3. 3

   Let your model use them

   TypeScript

   import { openai } from "@ai-sdk/openai";import { generateText, stepCountIs } from "ai";import { readPage, webSearch } from "./flux-tools";const { text } = await generateText({  model: openai("gpt-5-mini"),  tools: { webSearch, readPage },  stopWhen: stepCountIs(5),  prompt: "Who runs Barclays?",});console.log(text);

Call Flux as a model

For one grounded answer with no loop, add Flux as an OpenAI compatible provider. The reply is the passages Flux read.

import { createOpenAICompatible } from "@ai-sdk/openai-compatible";import { generateText } from "ai";const flux = createOpenAICompatible({  name: "flux",  baseURL: "https://fluxsearch.io/api/v1",  apiKey: process.env.FLUX_API_KEY,});const { text } = await generateText({  model: flux.chatModel("flux-search-1"),  prompt: "Who is the chief executive of Barclays?",});console.log(text);

Next steps