@nodellmcache/pinecone

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

Install

npm install @nodellmcache/pinecone @nodellmcache/core @pinecone-database/pinecone

Quick start

Create an index in the Pinecone console (or API) with the cosine metric and your embedding dimension, then:

import { PineconeAdapter } from '@nodellmcache/pinecone'

const store = new PineconeAdapter<{ source: string }>({
  apiKey: process.env.PINECONE_API_KEY,
  index: 'my-index',
  namespace: 'docs', // optional
})

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')

Options

| Option | Default | Description | |--------|---------|-------------| | index | — (required) | Existing index name | | apiKey | PINECONE_API_KEY env | API key | | namespace | — | Optional namespace within the index | | maxRetries | 3 | Attempts per op on transient failure | | client | constructed | Inject an existing Pinecone index handle (or compatible) |

Notes

• The index must already exist (index creation is provisioned out-of-band). Use the cosine metric so score is cosine similarity (what SemanticCache expects).
• String ids are native. Metadata filters use Pinecone's $eq form (keys AND-ed). Pinecone metadata values must be string/number/boolean/string[].
• Pinecone is eventually consistent — a freshly upserted vector may take a moment to appear in queries.

Testing

Unit tests use an in-memory fake. Integration tests require credentials and an existing index:

PINECONE_API_KEY=... PINECONE_INDEX=my-index pnpm --filter @nodellmcache/pinecone test

License

MIT