Pinecone Backend

The Pinecone backend stores embeddings in Pinecone, a fully managed vector database. Pinecone remains fully supported, but it is the optional managed-cloud path rather than the default production recommendation. If you want the recommended open-source production backend, use Qdrant.

Prerequisites

Requirement	Notes
Pinecone account	Free tier available at pinecone.io
API key	Found in the Pinecone console under "API Keys"
Index created	Create an index in the console or via the Pinecone API
Node.js 18+	Uses native fetch (no SDK dependency)

Configuration

import { PineconeVectorStore } from '@framers/agentos/cognition/rag/implementations/vector_stores/PineconeVectorStore';

const store = new PineconeVectorStore({
  id: 'my-pinecone',
  type: 'pinecone',
  apiKey: process.env.PINECONE_API_KEY!,
  indexHost: 'https://my-index-abc123.svc.aped-1234.pinecone.io',
  namespace: 'agent-default',
  defaultDimension: 1536,
});

await store.initialize();

Configuration options

Option	Type	Default	Description
apiKey	string	required	Pinecone API key
indexHost	string	required	Data Plane URL for your index (from Pinecone console)
apiVersion	string	'2026-04'	Explicit Pinecone API version header used on data-plane requests
namespace	string	''	Default namespace; collections map to namespaces
defaultDimension	number	1536	Embedding dimensions (must match the index)

The indexHost is the Data Plane endpoint for a specific index — not the control plane URL. Find it in the Pinecone console under your index details. It looks like https://my-index-abc123.svc.aped-1234.pinecone.io.

Namespace-based collection isolation

Pinecone namespaces are used as "collections". Each namespace is fully isolated within the same index:

// Agent A's memories
await store.upsert('agent-alice', documents);

// Agent B's memories — completely separate namespace
await store.upsert('agent-bob', documents);

// Query only Agent A's namespace
await store.query('agent-alice', embedding, { topK: 10 });

Namespaces are created implicitly on first upsert. createCollection() is a no-op.

Metadata filtering

Pinecone supports MongoDB-style metadata filter operators. AgentOS translates its unified MetadataFilter format to Pinecone's native syntax:

const results = await store.query('my-namespace', embedding, {
  topK: 10,
  filter: {
    type: { $eq: 'semantic' },        // Equality
    importance: { $gte: 0.5 },        // Range
    tags: { $in: ['project', 'decision'] }, // Set membership
  },
});

Supported operators

Operator	Description	Example
$eq	Equal to	{ status: { $eq: 'active' } }
$ne	Not equal to	{ status: { $ne: 'deleted' } }
$gt, $gte	Greater than (or equal)	{ score: { $gt: 0.8 } }
$lt, $lte	Less than (or equal)	{ age: { $lt: 30 } }
$in	In set	{ type: { $in: ['a', 'b'] } }
$nin	Not in set	{ type: { $nin: ['x'] } }
$exists	Field exists	{ tags: { $exists: true } }

Metadata values must be string, number, boolean, or string arrays. Complex objects are JSON-stringified before storage.

Lifecycle-sensitive metadata operations are also supported:

• scanByMetadata() uses Pinecone's fetch_by_metadata endpoint for filtered iteration
• delete(..., { filter }) uses Pinecone's metadata-delete path
• transient 429 / 5xx failures on query, upsert, scan, and delete are retried with capped exponential backoff

Limitations

No hybrid search

Pinecone requires a separate sparse encoder (e.g., SPLADE) for hybrid search. The AgentOS hybridSearch() method falls back to dense-only search on Pinecone. For true hybrid search, use the Postgres or Qdrant backends.

No knowledge graph

There is no sidecar storage for knowledge graph data. If you enable graph: true in MemoryConfig with the Pinecone backend, graph data is not persisted.

Not self-hostable

Pinecone is a managed service only. You cannot run it on your own infrastructure. If self-hosting is a requirement, use Qdrant or Postgres.

Batch size limit

Pinecone limits upserts to 100 vectors per request. AgentOS handles this automatically by splitting batches, but large ingestion jobs will make many sequential API calls.

No deleteAll count

delete({ deleteAll: true }) returns deletedCount: -1 because Pinecone's API does not report how many vectors were deleted in a bulk operation.

delete(..., { filter }) also returns deletedCount: -1 for the same reason: Pinecone accepts the delete-by-filter request but does not return an exact deleted-row count.

Migration FROM Pinecone to self-hosted backends

Use the AgentOS migration engine to move data from Pinecone to Postgres or Qdrant:

import { MigrationEngine } from '@framers/agentos/cognition/rag/migration/MigrationEngine';

await MigrationEngine.migrate({
  from: {
    type: 'pinecone',
    // PineconeSourceAdapter uses indexHost + apiKey + namespace
    url: 'https://my-index-abc123.svc.aped-1234.pinecone.io',
    apiKey: process.env.PINECONE_API_KEY!,
  },
  to: {
    type: 'postgres',
    connectionString: 'postgresql://postgres:password@localhost:5432/agent_memory',
  },
  batchSize: 100,
  onProgress: (done, total, table) => {
    console.log(`[${table}] ${done}/${total}`);
  },
});

The migration reads vectors via Pinecone's list + fetch APIs and writes them to the target backend. Non-vector data (knowledge graph, conversations) is not stored in Pinecone and will not be migrated.

Cost comparison

Tier	Vectors	Monthly cost	Notes
Starter (free)	100K	$0	1 index, 1 project, community support
Standard	1M+	~$70+	Multiple indexes, backup, 99.95% SLA
Enterprise	10M+	Custom	Dedicated infra, HIPAA, SOC2

Self-hosted alternatives for comparison:

Backend	Vectors	Monthly cost	Notes
Postgres + pgvector	10M+	~$15 (Neon free) to ~$50 (RDS)	Full SQL, hybrid search included
Qdrant (Docker)	10M+	Cost of your VM (~$5-20)	Built-in BM25, quantization
Qdrant Cloud	1M+	~$25+	Managed Qdrant, auto-scaling

Pinecone is the easiest managed path to start with but becomes expensive at scale. For production agents processing large knowledge bases, Postgres or Qdrant usually offer better cost efficiency and more features, and Qdrant is the default OSS production recommendation.