Skip to main content
Instead of making making the search yourself, let the model use Agentset as a search tool. The model decides what to search for, and follow-up with additional queries as necessary. This handles multi-part questions, follow-ups, and ambiguous queries better than simple RAG, which retrieves only once before generating. This guide uses the Vercel AI SDK to wire an Agentset search tool into a tool-calling loop.

​
Prerequisites

  • An Agentset API key and namespace ID.
  • An OpenAI API key (or another AI SDK provider).
Install the dependencies: 
npm install agentset ai @ai-sdk/openai zod
Set your environment variables:
.env.local
AGENTSET_API_KEY=your_agentset_api_key
AGENTSET_NAMESPACE=your_namespace_id
OPENAI_API_KEY=your_openai_api_key

​
System prompt

Use the system prompt below as a starting point and modify it for your use case: 
You are a knowledge base assistant. Answer the user's question using only the
information returned by the search tool.

## Searching

- Always search before answering. Never answer from prior knowledge.
- Write focused queries. For a multi-part question, run a separate search for
  each part rather than one broad query.
- If the first results are weak, rephrase and search again with different terms.

## Answering

- Base every statement on the search results. Do not infer or add outside
  knowledge.
- If the results do not contain the answer, say so plainly. Do not guess.
- Be concise and direct. Do not preface answers with "based on the context."

## Citations

- Cite every factual statement using the chunk id, like this: [<id>].
- Cite multiple sources as [<id1>], [<id2>].
- Only cite ids that appear in your search results. Never invent an id.

​
Define the search tool

Wrap ns.search in an AI SDK tool. The model calls this tool with a query it generates, so describe the tool and its input clearly. 
import { Agentset } from "agentset";
import { tool } from "ai";
import { z } from "zod";

const agentset = new Agentset({
  apiKey: process.env.AGENTSET_API_KEY,
});

const ns = agentset.namespace(process.env.AGENTSET_NAMESPACE!);

const searchTool = tool({
  description: "Search the knowledge base for information to answer the user's question.",
  inputSchema: z.object({
    query: z.string().describe("The search query."),
  }),
  execute: async ({ query }) => {
    const results = await ns.search(query, { topK: 10, rerank: true });

    return results.map((r) => ({
      id: r.id,
      text: r.text,
    }));
  },
});

​
Run the agentic loop

Pass the tool to streamText. Set stopWhen so the model can run several searches before answering—without a limit, a tool-calling loop can run indefinitely. 
import { openai } from "@ai-sdk/openai";
import { streamText, stepCountIs } from "ai";

const result = streamText({
  model: openai("gpt-5.1"),
  system: [
    "Answer the user's question using the search tool.",
    "Search whenever you need information. Run multiple searches for multi-part questions.",
    "Only answer from search results. If they don't contain the answer, say so.",
  ].join("\n"),
  prompt: "How does the billing system handle failed payments and refunds?",
  tools: { search: searchTool },
  stopWhen: stepCountIs(10),
});

for await (const chunk of result.textStream) {
  process.stdout.write(chunk);
}
The model runs the search tool as many times as it needs, then streams the final answer. Multi-part questions—like the billing example above—typically trigger separate searches for “failed payments” and “refunds.”

​
Multi-turn conversations

For a chat application, pass the full message history as messages instead of a single prompt. The model uses earlier turns to interpret follow-up questions and search accordingly. 
import { openai } from "@ai-sdk/openai";
import { streamText, stepCountIs, type ModelMessage } from "ai";

const messages: ModelMessage[] = [
  { role: "user", content: "What regions is the API available in?" },
  { role: "assistant", content: "The API is available in US, EU, and APAC regions." },
  { role: "user", content: "Which one has the lowest latency?" },
];

const result = streamText({
  model: openai("gpt-5.1"),
  system: "Answer the user's questions using the search tool.",
  messages,
  tools: { search: searchTool },
  stopWhen: stepCountIs(10),
});

​
Add citations

Return a stable identifier with each chunk and instruct the model to cite by it. Because the model reads chunks across several searches, citing by ID is more reliable than position-based numbering. 
const searchTool = tool({
  description: "Search the knowledge base.",
  inputSchema: z.object({
    query: z.string().describe("The search query."),
  }),
  execute: async ({ query }) => {
    const results = await ns.search(query, { topK: 10, rerank: true });

    // Expose a short id the model cites by, plus metadata for your UI.
    return results.map((r) => ({
      id: r.id,
      text: r.text,
      source: r.metadata?.filename,
    }));
  },
});
Add the citation contract to your system prompt: 
Cite every factual statement using the chunk id in the form [id].
Only cite ids that appear in the search results.
See Citations for rendering citations in your UI.

​
Tune retrieval

Configure the underlying search per call to balance recall, relevance, and token usage.
OptionRecommendation
topKRaise for broad questions, lower to save tokens. The model can search again if needed.
rerankKeep enabled to surface the most relevant chunks first.
modeUse keyword for exact terms (error codes, IDs); semantic (default) otherwise.
filterScope results by metadata, e.g. per-user or per-document.
You can expose more than one tool—for example, one that fetches a full document by ID—and let the model choose between them.

​
Next steps

  • Search — Configure search parameters and reranking
  • Citations — Add source attribution to responses
  • Filtering — Scope searches with metadata filters
  • Observability — Trace each search and generation step