Pinecone Node.js SDK ·

This is the official Node.js SDK for Pinecone, written in TypeScript.

Documentation

• Reference Documentation
• If you are upgrading from v0.x, check out the v1 Migration Guide.
• If you are upgrading from v1.x, check out the v2 Migration Guide.

Example code

The snippets shown in this README are intended to be concise. For more realistic examples, explore these examples:

• Semantic search
• Article recommender
• Image search

Upgrading the SDK

Upgrading from 2.x to 3.x

There is a breaking change involving the configureIndex operation in this update. The structure of the object passed when configuring an index has changed to include deletionProtection. The podType and replicas fields can now be updated through the spec.pod object. See Configure pod-based indexes for an example of the code.

Upgrading from older versions

• Upgrading to 2.x : There were many changes made in this release to support Pinecone's new Serverless index offering. The changes are covered in detail in the v2 Migration Guide. Serverless indexes are only available in 2.x release versions or greater.
• Upgrading to 1.x : This release officially moved the SDK out of beta, and there are a number of breaking changes that need to be addressed when upgrading from a 0.x version. See the v1 Migration Guide for details.

Prerequisites

The Pinecone TypeScript SDK is compatible with TypeScript >=4.1 and Node >=18.x.

Installation

npm install @pinecone-database/pinecone

Productionizing

The Pinecone Typescript SDK is intended for server-side use only. Using the SDK within a browser context can expose your API key(s). If you have deployed the SDK to production in a browser, please rotate your API keys.

Usage

Initializing the client

An API key is required to initialize the client. It can be passed using an environment variable or in code through a configuration object. Get an API key in the console.

Using environment variables

The environment variable used to configure the API key for the client is the following:

PINECONE_API_KEY="your_api_key"

PINECONE_API_KEY is the only required variable. When this environment variable is set, the client constructor does not require any additional arguments.

import { Pinecone } from '@pinecone-database/pinecone';

const pc = new Pinecone();

Using a configuration object

If you prefer to pass configuration in code, the constructor accepts a config object containing the apiKey value. This is the object in which you would pass properties like maxRetries (defaults to 3) for retryable operations (upsert, update, and configureIndex).

import { Pinecone } from '@pinecone-database/pinecone';

const pc = new Pinecone({
  apiKey: 'your_api_key',
  maxRetries: 5,
});

Using a proxy server

If your network setup requires you to interact with Pinecone via a proxy, you can pass a custom ProxyAgent from the undici library. Below is an example of how to construct an undici ProxyAgent that routes network traffic through a mitm proxy server while hitting Pinecone's /indexes endpoint.

Note: The following strategy relies on Node's native fetch implementation, released in Node v16 and stabilized in Node v21. If you are running Node versions 18-21, you may experience issues stemming from the instability of the feature. There are currently no known issues related to proxying in Node v18+.

import {
  Pinecone,
  type PineconeConfiguration,
} from '@pinecone-database/pinecone';
import { Dispatcher, ProxyAgent } from 'undici';
import * as fs from 'fs';

const cert = fs.readFileSync('path-to-your-mitm-proxy-cert-pem-file');

const client = new ProxyAgent({
  uri: '<your proxy server URI>',
  requestTls: {
    port: '<your proxy server port>',
    ca: cert,
    host: '<your proxy server host>',
  },
});

const customFetch = (
  input: string | URL | Request,
  init: RequestInit | undefined
) => {
  return fetch(input, {
    ...init,
    dispatcher: client as Dispatcher,
    keepalive: true,  # optional
  });
};

const config: PineconeConfiguration = {
  apiKey:
    '<your Pinecone API key, available in your dashboard at app.pinecone.io>',
  fetchApi: customFetch,
};

const pc = new Pinecone(config);

const indexes = async () => {
  return await pc.listIndexes();
};

indexes().then((response) => {
  console.log('My indexes: ', response);
});

Indexes

Create Index

Create a serverless index with minimal configuration

At a minimum, to create a serverless index you must specify a name, dimension, and spec. The dimension indicates the size of the vectors you intend to store in the index. For example, if your intention was to store and query embeddings (vectors) generated with OpenAI's textembedding-ada-002 model, you would need to create an index with dimension 1536 to match the output of that model.

The spec configures how the index should be deployed. For serverless indexes, you define only the cloud and region where the index should be hosted. For pod-based indexes, you define the environment where the index should be hosted, the pod type and size to use, and other index characteristics. For more information on serverless and regional availability, see Understanding indexes.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

await pc.createIndex({
  name: 'sample-index',
  dimension: 1536,
  spec: {
    serverless: {
      cloud: 'aws',
      region: 'us-west-2',
    },
  },
  tags: { team: 'data-science' },
});

Create a pod-based index with optional configurations

To create a pod-based index, you define pod in the spec object which contains the environment where the index should be hosted, and the podType and pods size to use. Many optional configuration fields allow greater control over hardware resources and availability. To learn more about the purpose of these fields, see Understanding indexes and Scale pod-based indexes.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

await pc.createIndex({
  name: 'sample-index-2',
  dimension: 1536,
  metric: 'dotproduct',
  spec: {
    pod: {
      environment: 'us-east4-gcp',
      pods: 2,
      podType: 'p1.x2',
      metadataConfig: {
        indexed: ['product_type'],
      },
    },
    tags: { team: 'data-science' },
  },

  // This option tells the client not to throw if the index already exists.
  suppressConflicts: true,

  // This option tells the client not to resolve the promise until the
  // index is ready.
  waitUntilReady: true,
});

Checking the status of a newly created index

The createIndex method issues a create request to the API that returns quickly, but the resulting index is not immediately ready for upserting, querying, or performing other data operations. You can use the describeIndex method to find out the status of an index and see whether it is ready for use.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

await pc.describeIndex('serverless-index');
// {
//    name: 'serverless-index',
//    dimension: 1536,
//    metric: 'cosine',
//    host: 'serverless-index-4zo0ijk.svc.us-west2-aws.pinecone.io',
//    deletionProtection: 'disabled',
//    spec: {
//       serverless: {
//          cloud: 'aws',
//          region: 'us-west-2'
//       }
//    },
//    status: {
//       ready: false,
//       state: 'Initializing'
//    }
// }

Waiting until the index is ready

If you pass the waitUntilReady option, the client will handle polling for status updates on a newly created index. The promise returned by createIndex will not be resolved until the index status indicates it is ready to handle data operations. This can be especially useful for integration testing, where index creation in a setup step will be immediately followed by data operations.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

await pc.createIndex({
  name: 'serverless-index',
  dimension: 1536,
  spec: {
    serverless: {
      cloud: 'aws',
      region: 'us-west-2',
    },
  },
  waitUntilReady: true,
});

Create a pod-based index from a Pinecone collection

ℹ️ Note

Serverless and starter indexes do not support collections.

As you use Pinecone for more things, you may wish to explore different index configurations with the same vector data. Collections provide an easy way to do this. See other client methods for working with collections here.

Given that you have an existing collection:

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

await pc.describeCollection('product-description-embeddings');
// {
//   name: 'product-description-embeddings',
//   size: 543427063,
//   status: 'Ready',
//   dimension: 2,
//   vectorCount: 10001498,
//   environment: 'us-east4-gcp'
// }

Note: For pod-based indexes, you can specify a sourceCollection from which to create an index. The collection must be in the same environment as the index.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

await pc.createIndex({
  name: 'product-description-p1x1',
  dimension: 256,
  metric: 'cosine',
  spec: {
    pod: {
      environment: 'us-east4-gcp',
      pods: 1,
      podType: 'p1.x1',
      sourceCollection: 'product-description-embeddings',
    },
  },
});

When the new index is ready, it should contain all the data that was in the collection, ready to be queried.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

await pc.index('product-description-p2x2').describeIndexStats();
// {
//   namespaces: { '': { recordCount: 78000 } },
//   dimension: 256,
//   indexFullness: 0.9,
//   totalRecordCount: 78000
// }

Create or configure an index with deletion protection

You can configure both serverless and pod indexes with deletionProtection. Any index with this property set to 'enabled' will be unable to be deleted. By default, deletionProtection will be set to 'disabled' if not provided as a part of the createIndex request. To enable deletionProtection you can pass the value while calling createIndex.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

await pc.createIndex({
  name: 'deletion-protected-index',
  dimension: 1536,
  metric: 'cosine',
  deletionProtection: 'enabled',
  spec: {
    serverless: {
      cloud: 'aws',
      region: 'us-west-2',
    },
  },
});

To disable deletion protection, you can use the configureIndex operation.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

await pc.configureIndex('deletion-protected-index', {
  deletionProtection: 'disabled',
});

Create or configure an index with index tags

You can create or configure serverless and pod indexes with tags. Indexes can hold an arbitrary number of tags outlining metadata you would like attached to the index object itself, such as team ownership, project, or any other relevant information.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

// Create index with tag
await pc.createIndex({
  name: 'tag-index',
  dimension: 1536,
  metric: 'cosine',
  spec: {
    serverless: {
      cloud: 'aws',
      region: 'us-west-2',
    },
  },
  tags: { team: 'data-science' }, // index tag
});

// Configure index with a new tag
await pc.configureIndex('tag-index', {
  tags: { project: 'recommendation' }, // new index tag
});

// Delete an existing tag
await pc.configureIndex('tag-index', {
  tags: { project: '' }, // Pass an empty string to an existing key to delete a tag; this will delete the `project` tag
});

Describe Index

You can fetch the description of any index by name using describeIndex.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

await pc.describeIndex('serverless-index');
// {
//    name: 'serverless-index',
//    dimension: 1536,
//    metric: 'cosine',
//    host: 'serverless-index-4zo0ijk.svc.us-west2-aws.pinecone.io',
//    deletionProtection: 'disabled',
//    spec: {
//       serverless: {
//          cloud: 'aws',
//          region: 'us-west-2'
//       },
//    },
//    status: {
//       ready: true,
//       state: 'Ready'
//    }
// }

Configure pod-based indexes

ℹ️ Note

This section applies to pod-based indexes only. With serverless indexes, you don't configure any compute or storage resources. Instead, serverless indexes scale automatically based on usage.

You can adjust the number of replicas or scale to a larger pod size (specified with podType). See Scale pod-based indexes. You cannot downgrade pod size or change the base pod type.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();
await pc.configureIndex('pod-index', {
  spec: {
    pod: {
      replicas: 2,
      podType: 'p1.x4',
    },
  },
});
const config = await pc.describeIndex('pod-index');
// {
//    name: 'pod-index',
//    dimension: 1536,
//    metric: 'cosine',
//    host: 'pod-index-4zo0ijk.svc.us-east1-gcp.pinecone.io',
//    deletionProtection: 'disabled',
//    spec: {
//       pod: {
//         environment: 'us-east1-gcp',
//         replicas: 2,
//         shards: 2,
//         podType: 'p1.x4',
//         pods: 4,
//         metadataConfig: [Object],
//         sourceCollection: undefined
//       }
//    },
//    status: {
//       ready: true,
//       state: 'ScalingUpPodSize'
//    }
// }

Delete Index

Indexes are deleted by name.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

await pc.deleteIndex('sample-index');

List Indexes

The listIndexes command returns an object with an array of index models under indexes.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

await pc.listIndexes();
// {
//   indexes: [
//     {
//       name: 'serverless-index',
//       dimension: 1536,
//       metric: 'cosine',
//       host: 'serverless-index-4zo0ijk.svc.us-west2-aws.pinecone.io',
//       deletionProtection: 'disabled',
//       spec: {
//         serverless: {
//           cloud: 'aws',
//           region: 'us-west-2',
//         },
//       },
//       status: {
//         ready: true,
//         state: 'Ready',
//       },
//     },
//     {
//       name: 'pod-index',
//       dimension: 1536,
//       metric: 'cosine',
//       host: 'pod-index-4zo0ijk.svc.us-west2-aws.pinecone.io',
//       deletionProtection: 'disabled',
//       spec: {
//         pod: {
//           environment: 'us-west2-aws',
//           replicas: 1,
//           shards: 1,
//           podType: 'p1.x1',
//           pods: 1,
//         },
//       },
//       status: {
//         ready: true,
//         state: 'Ready',
//       },
//     },
//   ],
// }

Collections

ℹ️ Note

Serverless and starter indexes do not support collections.

A collection is a static copy of a pod-based index that may be used to create backups, to create copies of indexes, or to perform experiments with different index configurations. To learn more about Pinecone collections, see Understanding collections.

Create Collection

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

await pc.createCollection({
  name: 'collection-name',
  source: 'index-name',
});

This API call should return quickly, but the creation of a collection can take from minutes to hours depending on the size of the source index and the index's configuration. Use describeCollection to check the status of a collection.

Delete Collection

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

await pc.deleteCollection('collection-name');

You can use listCollections to confirm the deletion.

Describe Collection

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

const describeCollection = await pc.describeCollection('collection3');
// {
//   name: 'collection3',
//   size: 3126700,
//   status: 'Ready',
//   dimension: 3,
//   vectorCount: 1234,
//   environment: 'us-east1-gcp',
// }

List Collections

The listCollections command returns an object with an array of collection models under collections.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

const list = await pc.listCollections();
// {
//   collections: [
//     {
//       name: 'collection1',
//       size: 3089687,
//       status: 'Ready',
//       dimension: 3,
//       vectorCount: 17378,
//       environment: 'us-west1-gcp',
//     },
//     {
//       name: 'collection2',
//       size: 208309,
//       status: 'Ready',
//       dimension: 3,
//       vectorCount: 1000,
//       environment: 'us-east4-gcp',
//     },
//   ];
// }

Index operations

Pinecone indexes support operations for working with vector data using operations such as upsert, query, fetch, and delete.

Targeting an index

To perform data operations on an index, you target it using the index method.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();
const index = pc.index('test-index');

// Now perform index operations
await index.fetch(['1']);

The first argument is the name of the index you are targeting. There's an optional second argument for providing an index host override. Providing this second argument allows you to bypass the SDK's default behavior of resolving your index host via the provided index name. You can find your index host in the Pinecone console, or by using the describeIndex or listIndexes operations.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();
const index = pc.index('test-index', 'my-index-host-1532-svc.io');

// Now perform index operations against: https://my-index-host-1532-svc.io
await index.fetch(['1']);

Targeting an index, with metadata typing

If you are storing metadata alongside your vector values, you can pass a type parameter to index() in order to get proper TypeScript typechecking.

import { Pinecone, PineconeRecord } from '@pinecone-database/pinecone';
const pc = new Pinecone();

type MovieMetadata = {
    title: string,
    runtime: number,
    genre: 'comedy' | 'horror' | 'drama' | 'action'
}

// Specify a custom metadata type while targeting the index
const index = pc.index<MovieMetadata>('test-index');

// Now you get type errors if upserting malformed metadata
await index.upsert([{
        id: '1234',
        values: [
            .... // embedding values
    ],
    metadata: {
    genre: 'Gone with the Wind',
        runtime: 238,
        genre: 'drama',
        // @ts-expect-error because category property not in MovieMetadata
        category: 'classic'
}
}])

const results = await index.query({
    vector: [
        ... // query embedding
    ],
    filter: { genre: { '$eq': 'drama' }}
})
const movie = results.matches[0];

if (movie.metadata) {
    // Since we passed the MovieMetadata type parameter above,
    // we can interact with metadata fields without having to
    // do any typecasting.
    const { title, runtime, genre } = movie.metadata;
    console.log(`The best match in drama was ${title}`)
}

Targeting a namespace

By default, all data operations take place inside the default namespace of ''. If you are working with other non-default namespaces, you can target the namespace by chaining a call to namespace().

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();
const index = pc.index('test-index').namespace('ns1');

// Now perform index operations in the targeted index and namespace
await index.fetch(['1']);

See Use namespaces for more information.

Upsert vectors

Pinecone expects records inserted into indexes to have the following form:

type PineconeRecord = {
  id: string;
  values: Array<number>;
  sparseValues?: Array<number>;
  metadata?: object;
};

To upsert some vectors, you can use the client like so:

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

// Target an index
const index = pc.index('sample-index');

// Prepare your data. The length of each array
// of vector values must match the dimension of
// the index where you plan to store them.
const vectors = [
  {
    id: '1',
    values: [0.236, 0.971, 0.559],
    sparseValues: { indices: [0, 1], values: [0.236, 0.34] }, // Optional; for hybrid search
  },
  {
    id: '2',
    values: [0.685, 0.111, 0.857],
    sparseValues: { indices: [0, 1], values: [0.345, 0.98] }, // Optional; for hybrid search
  },
];

// Upsert the data into your index
await index.upsert(vectors);

Import vectors from object storage

You can now import vectors en masse from object storage. Import is a long-running, asynchronous operation that imports large numbers of records into a Pinecone serverless index.

In order to import vectors from object storage, they must be stored in Parquet files and adhere to the necessary file format. Your object storage must also adhere to the necessary directory structure.

The following example imports vectors from an Amazon S3 bucket into a Pinecone serverless index:

import { Pinecone } from '@pinecone-database/pinecone';

const pc = new Pinecone();
const indexName = 'sample-index';

await pc.createIndex({
  name: indexName,
  dimension: 10,
  spec: {
    serverless: {
      cloud: 'aws',
      region: 'eu-west-1',
    },
  },
});

const index = pc.Index(indexName);

const storageURI = 's3://my-bucket/my-directory/';

await index.startImport(storageURI, 'continue'); // "Continue" will avoid aborting the operation if errors are encountered.

// {
//   "id": "import-id"
// }

You can start, cancel, and check the status of all or one import operation(s).

Notes:

• Import only works with Serverless indexes
• Import is in public preview
• The only object storage provider currently supported is Amazon S3
• Vectors will take at least 10 minutes to appear in your index upon completion of the import operation, since this operation is optimized for very large workloads
• See limits for further information

Seeing index statistics

When experimenting with data operations, it's sometimes helpful to know how many records/vectors are stored in each namespace. In that case, target the index and use the describeIndexStats() command.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();
const index = pc.index('example-index');

await index.describeIndexStats();
// {
//   namespaces: {
//     '': { recordCount: 10 }
//     foo: { recordCount: 2000 },
//     bar: { recordCount: 2000 }
//   },
//   dimension: 1536,
//   indexFullness: 0,
//   totalRecordCount: 4010
// }

Querying

Querying with vector values

The query method accepts a large number of options. The dimension of the query vector must match the dimension of your index.

type QueryOptions = {
  topK: number; // number of results desired
  vector?: Array<number>; // must match dimension of index
  sparseVector?: {
    indices: Array<integer>; // indices must fall within index dimension
    values: Array<number>; // indices and values arrays must have same length
  };
  id?: string;
  includeMetadata: boolean;
  includeValues: boolean;
};

For example, to query by vector values you would pass the vector param in the options configuration. For brevity sake this example query vector is tiny (dimension 2), but in a more realistic use case this query vector would be an embedding outputted by a model. Look at the Example code to see more realistic examples of how to use query.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();
const index = pc.index('my-index');

await index.query({ topK: 3, vector: [0.22, 0.66] });
// {
//   matches: [
//     {
//       id: '556',
//       score: 1.00000012,
//       values: [],
//       sparseValues: undefined,
//       metadata: undefined
//     },
//     {
//       id: '137',
//       score: 1.00000012,
//       values: [],
//       sparseValues: undefined,
//       metadata: undefined
//     },
//     {
//       id: '129',
//       score: 1.00000012,
//       values: [],
//       sparseValues: undefined,
//       metadata: undefined
//     }
//   ],
//   namespace: '',
//   usage: {
//     readUnits: 5
//   }
// }

You include options to includeMetadata: true or includeValues: true if you need this information. By default, these are not returned to keep the response payload small.

Remember that data operations take place within the context of a namespace, so if you are working with namespaces and do not see expected results you should check that you are targeting the correct namespace with your query.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

// Target the index and namespace
const index = pc.index('my-index').namespace('my-namespace');

const results = await index.query({ topK: 3, vector: [0.22, 0.66] });

Querying by record id

You can query using the vector values of an existing record in the index by passing a record ID. Please note that the record with the specified ID may be in this operation's response.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();
const index = pc.index('my-index');

const results = await index.query({ topK: 10, id: '1' });

Hybrid search with sparse vectors

If you are working with sparse-dense vectors, you can add sparse vector values to perform a hybrid search.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();

await pc.createIndex({
  name: 'hybrid-search-index',
  metric: 'dotproduct', // Note: dot product is the only distance metric supported for hybrid search
  dimension: 2,
  spec: {
    pod: {
      environment: 'us-west4-gcp',
      podType: 'p2.x1',
    },
  },
  waitUntilReady: true,
});

const index = pc.index('hybrid-search-index');

const hybridRecords = [
  {
    id: '1',
    values: [0.236, 0.971], // dense vectors
    sparseValues: { indices: [0, 1], values: [0.236, 0.34] }, // sparse vectors
  },
  {
    id: '2',
    values: [0.685, 0.111],
    sparseValues: { indices: [0, 1], values: [0.887, 0.243] },
  },
];

await index.upsert(hybridRecords);

const query = 'What is the most popular red dress?';
// ... send query to dense vector embedding model and save those values in `denseQueryVector`
// ... send query to sparse vector embedding model and save those values in `sparseQueryVector`
const denseQueryVector = [0.236, 0.971];
const sparseQueryVector = { indices: [0, 1], values: [0.0, 0.34] };

// Execute a hybrid search
await index.query({
  topK: 3,
  vector: denseQueryVector,
  sparseVector: sparseQueryVector,
});

Update a record

You may want to update vector values, sparseValues, or metadata. Specify the id and the attribute value you want to update.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();
const index = pc.index('imdb-movies');

await index.update({
  id: '18593',
  metadata: { genre: 'romance' },
});

List records

The listPaginated method can be used to list record IDs matching a particular ID prefix in a paginated format. With clever assignment of record ids, this can be used to help model hierarchical relationships between different records such as when there are embeddings for multiple chunks or fragments related to the same document.

Notes:

• When you do not specify a prefix, the default prefix is an empty string, which returns all vector IDs in your index
• There is a hard limit of 100 vector IDs if no limit is specified. Consequently, if there are fewer than 100 vector IDs that match a given prefix in your index, and you do not specify a limit, your paginationToken will be undefined

The following example shows how to fetch both pages of vector IDs for vectors whose IDs contain the prefix doc1#, assuming a limit of 3 and doc1 document being chunked into 4 vectors.

const pc = new Pinecone();
const index = pc.index('my-index').namespace('my-namespace');

// Fetch the 1st 3 vector IDs matching prefix 'doc1#'
const results = await index.listPaginated({ limit: 3, prefix: 'doc1#' });
console.log(results);
// {
//   vectors: [
//     { id: 'doc1#01' }
//     { id: 'doc1#02' }
//     { id: 'doc1#03' }
//     ...
//   ],
//   pagination: {
//     next: 'eyJza2lwX3Bhc3QiOiJwcmVUZXN0LS04MCIsInByZWZpeCI6InByZVRlc3QifQ=='
//   },
//   namespace: 'my-namespace',
//   usage: { readUnits: 1 }
// }

// Fetch the final vector ID matching prefix 'doc1#' using the paginationToken returned by the previous call
const nextResults = await index.listPaginated({
  prefix: 'doc1#',
  paginationToken: results.pagination?.next,
});
console.log(nextResults);
// {
//   vectors: [
//     { id: 'doc1#04' }
//   ],
//   pagination: undefined,
//   namespace: 'my-namespace',
//   usage: { readUnits: 1 }
// }

Fetch records by ID(s)

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();
const index = pc.index('my-index');

const fetchResult = await index.fetch(['id-1', 'id-2']);

Delete records

For convenience there are several delete-related methods. You can verify the results of a delete operation by trying to fetch() a record or looking at the index summary with describeIndexStats()

Delete one

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();
const index = pc.index('my-index');

await index.deleteOne('id-to-delete');

Delete many by ID

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();
const index = pc.index('my-index');

await index.deleteMany(['id-1', 'id-2', 'id-3']);

Delete many by metadata filter

Note: deletion by metadata filter only applies to pod-based indexes.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();
const index = pc.index('albums-database');

await index.deleteMany({ genre: 'rock' });

Delete all records in a namespace

ℹ️ NOTE

Indexes in the gcp-starter environment do not support namespaces.

To nuke everything in the targeted namespace, use the deleteAll method.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();
const index = pc.index('my-index');

await index.namespace('foo-namespace').deleteAll();

If you do not specify a namespace, the records in the default namespace '' will be deleted.

Inference

Interact with Pinecone's Inference API (currently in public preview). The Pinecone Inference API is a service that gives you access to inference models hosted on Pinecone's infrastructure.

Notes:

Supported models:

• Embedding: multilingual-e5-large
• Reranking: bge-reranker-v2-m3

Create embeddings

Send text to Pinecone's Inference API to generate embeddings for documents and queries.

import { Pinecone } from '@pinecone-database/pinecone';

const client = new Pinecone({ apiKey: '<Your API key from app.pinecone.io>' });

const embeddingModel = 'multilingual-e5-large';

const documents = [
  'Turkey is a classic meat to eat at American Thanksgiving.',
  'Many people enjoy the beautiful mosques in Turkey.',
];
const docParameters = {
  inputType: 'passage',
  truncate: 'END',
};
async function generateDocEmbeddings() {
  try {
    return await client.inference.embed(
      embeddingModel,
      documents,
      docParameters
    );
  } catch (error) {
    console.error('Error generating embeddings:', error);
  }
}
generateDocEmbeddings().then((embeddingsResponse) => {
  if (embeddingsResponse) {
    console.log(embeddingsResponse);
  }
});

// << Upsert documents into Pinecone >>

const userQuery = ['How should I prepare my turkey?'];
const queryParameters = {
  inputType: 'query',
  truncate: 'END',
};
async function generateQueryEmbeddings() {
  try {
    return await client.inference.embed(
      embeddingModel,
      userQuery,
      queryParameters
    );
  } catch (error) {
    console.error('Error generating embeddings:', error);
  }
}
generateQueryEmbeddings().then((embeddingsResponse) => {
  if (embeddingsResponse) {
    console.log(embeddingsResponse);
  }
});

// << Send query to Pinecone to retrieve similar documents >>

Rerank documents

Rerank documents in descending relevance-order against a query.

Note: The score represents the absolute measure of relevance of a given query and passage pair. Normalized between [0, 1], the score represents how closely relevant a specific item and query are, with scores closer to 1 indicating higher relevance.

import { Pinecone } from '@pinecone-database/pinecone';
const pc = new Pinecone();
const rerankingModel = 'bge-reranker-v2-m3';
const myQuery = 'What are some good Turkey dishes for Thanksgiving?';

// Option 1: Documents as an array of strings
const myDocsStrings = [
  'I love turkey sandwiches with pastrami',
  'A lemon brined Turkey with apple sausage stuffing is a classic Thanksgiving main',
  'My favorite Thanksgiving dish is pumpkin pie',
  'Turkey is a great source of protein',
];

// Option 1 response
const response = await pc.inference.rerank(
  rerankingModel,
  myQuery,
  myDocsStrings
);
console.log(response);
// {
// model: 'bge-reranker-v2-m3',
// data: [
//   { index: 1, score: 0.5633179, document: [Object] },
//   { index: 2, score: 0.02013874, document: [Object] },
//   { index: 3, score: 0.00035419367, document: [Object] },
//   { index: 0, score: 0.00021485926, document: [Object] }
// ],
// usage: { rerankUnits: 1 }
// }

// Option 2: Documents as an array of objects
const myDocsObjs = [
  {
    title: 'Turkey Sandwiches',
    body: 'I love turkey sandwiches with pastrami',
  },
  {
    title: 'Lemon Turkey',
    body: 'A lemon brined Turkey with apple sausage stuffing is a classic Thanksgiving main',
  },
  {
    title: 'Thanksgiving',
    body: 'My favorite Thanksgiving dish is pumpkin pie',
  },
  {
    title: 'Protein Sources',
    body: 'Turkey is a great source of protein',
  },
];

// Option 2: Options object declaring which custom key to rerank on
// Note: If no custom key is passed via `rankFields`, each doc must contain a `text` key, and that will act as the default)
const rerankOptions = {
  topN: 3,
  returnDocuments: false,
  rankFields: ['body'],
  parameters: {
    inputType: 'passage',
    truncate: 'END',
  },
};

// Option 2 response
const response = await pc.inference.rerank(
  rerankingModel,
  myQuery,
  myDocsObjs,
  rerankOptions
);
console.log(response);
// {
// model: 'bge-reranker-v2-m3',
// data: [
//   { index: 1, score: 0.5633179, document: undefined },
//   { index: 2, score: 0.02013874, document: undefined },
//   { index: 3, score: 0.00035419367, document: undefined },
// ],
// usage: { rerankUnits: 1 }
//}

Testing

All testing takes place automatically in CI and is configured using Github actions and workflows, located in the .github directory of this repo.

See CONTRIBUTING.md for more information.