@anvia/milvus: Examples

Minimal search

import { MilvusVectorStore } from "@anvia/milvus";

const store = await MilvusVectorStore.connect({
  collectionName: "support_docs",
  vectorSize: 1536,
});
const index = store.index(embeddingModel);

const results = await index.search({
  query: "enterprise support",
  topK: 5,
});

console.log(results.map((result) => result.id));

Retrieval inside an agent

import { AgentBuilder } from "@anvia/core";
import { MilvusVectorStore } from "@anvia/milvus";

const store = await MilvusVectorStore.connect({
  collectionName: "support_docs",
  vectorSize: 1536,
});
const index = store.index(embeddingModel);

const agent = new AgentBuilder("support", completionModel)
  .instructions("Answer from retrieved support documentation when it is relevant.")
  .dynamicContext(index, {
    topK: 4,
    threshold: 0.72,
  })
  .build();

Harness shape

import { describe, expect, it } from "vitest";

describe("retrieval index", () => {
  it("returns filtered ids", async () => {
    const index = store.index(embeddingModel);
    const matches = await index.searchIds({
      query: "password reset",
      topK: 3,
      filter: { product: "support" },
    });

    expect(matches.length).toBeLessThanOrEqual(3);
  });
});

Use integration tests with disposable collections, tables, or namespaces. For agent tests, inject a fake VectorSearchIndex so prompt behavior can be tested without a live database.