@nodellmcache/milvus

Milvus vector-store adapter for NodeLLMCache. Implements VectorStoreAdapter, so it plugs into @nodellmcache/semantic-cache or works standalone.

Install

@zilliz/milvus2-sdk-node is an optional peer dependency (lazily loaded), so install it alongside:

npm install @nodellmcache/milvus @nodellmcache/core @zilliz/milvus2-sdk-node

Run Milvus (standalone) via its docker-compose.

Quick start

import { MilvusAdapter } from '@nodellmcache/milvus'

const store = new MilvusAdapter<{ source: string }>({
  address: 'localhost:19530',
  collection: 'docs', // created on first use
})

await store.upsert('doc-1', embedding, { source: 'wiki' })
const matches = await store.query(queryEmbedding, 5, { source: 'wiki' })
// → [{ id: 'doc-1', score: 0.94, metadata: { source: 'wiki' } }]
await store.delete('doc-1')
await store.disconnect()

Options

| Option | Default | Description | |--------|---------|-------------| | collection | — (required) | Collection name (created on demand) | | address | localhost:19530 | Endpoint (or Zilliz Cloud URI) | | token / username / password / ssl | — | Auth | | vectorSize | inferred | Dimensionality | | maxRetries | 3 | Attempts per op on transient failure | | store | constructed | Inject a MilvusStore (e.g. to avoid the optional SDK) |

Notes

• The collection is created on demand: id (VarChar PK), vector (FloatVector), metadata (JSON), with an AUTOINDEX using the cosine metric so score is cosine similarity.
• String ids are native. Metadata filters are flat equality over the JSON field (metadata["key"] == value, AND-combined).
• The SDK is loaded lazily; if it isn't installed, construction throws a clear error (or inject your own MilvusStore).

Testing

Unit tests use an in-memory fake store. Integration tests require a running Milvus and the SDK installed:

MILVUS_ADDRESS=localhost:19530 pnpm --filter @nodellmcache/milvus test

License

MIT