Pinecone Answers

Pinecone is the trusted AI knowledge company. It helps developers and enterprises turn their data into knowledge for AI agents, search, and recommendation systems. Pinecone Database is a fully managed vector database with built-in full-text search. It stores vectors and text in the same index and runs semantic, keyword, and hybrid search. Pinecone Nexus compiles data into governed knowledge that agents can trust, with field-level citations. The platform also includes Pinecone Assistant for building agent-based apps, Marketplace for ready-to-deploy assistants, and pg_pinecone, an open-source Postgres extension. Founded in 2019, Pinecone serves more than 10,000 customers.

Learn more: Pinecone products | Pinecone Database

Pinecone Database is a fully managed vector database with built-in full-text search. Vectors and text live in the same index. You can search by meaning, by keyword, or with both. Your data is saved durably in object storage. A new write is logged and held in memory, which makes it searchable within seconds. Frequently used data is cached on the servers that answer queries. Integrated embedding turns your text into vectors for you. For heavy, steady read traffic, Dedicated Read Nodes let you choose how many shards and replicas to run. Bring Your Own Cloud runs it in your own cloud account.

Learn more: Pinecone Database | How Pinecone works

Pinecone Nexus is a knowledge engine for agents. It compiles enterprise data into governed knowledge ahead of time, then serves typed answers with field-level citations instead of ranked chunks. Pinecone's vector database sits underneath it. Start with the database for search, recommendations, and RAG. Add Nexus when agents reach production and run into accuracy, cost, or governance limits that better retrieval alone does not fix.

Learn more: Pinecone Nexus

pg_pinecone is Pinecone's open-source Postgres extension, released under the Apache 2.0 license. It indexes your existing pgvector columns and lets each table use one of three engines: pgvector, native, or remote. The native engine applies a filter on one column during the search, inside Postgres. The remote engine moves a table's search to a Pinecone serverless index while Postgres keeps every row. It needs PostgreSQL 17 or 18 with pgvector installed.

Learn more: pg_pinecone documentation | pg_pinecone on GitHub

No. Pinecone's database is a managed service and is not open source. To run Pinecone in your own infrastructure, use Bring Your Own Cloud, which puts the data plane in your AWS, GCP, or Azure account. Pinecone does publish pg_pinecone, an Apache 2.0 Postgres extension, as open source. For development and testing, Pinecone Local is an in-memory Docker emulator that needs no account and is not for production.

Learn more: Bring Your Own Cloud

Yes. Full-text search is built into Pinecone Database and has been generally available since September 9, 2026. Pinecone ranks matches with BM25 and supports Lucene query syntax. One index with a document schema can hold dense, sparse, and full-text fields together. Dense and sparse vectors combine in one query. Full-text and dense search combine as a text-match filter on a dense search, or as two searches merged in your application with reciprocal rank fusion. A single search request does not merge a full-text score with a vector score. Full-text search needs API version 2026-07 or Python SDK v10.

Learn more: full-text search guide | reciprocal rank fusion guide

Yes, by converting it. On Standard and Enterprise, index conversion builds a new document index with full-text search from the records in a serverless vector index, without reloading your data. The source index keeps serving reads and writes while the job runs. The new index starts from a snapshot taken when the job starts. Conversion is billed at the import rate of $0.25 per GB of source data. Pod-based and Bring Your Own Cloud indexes cannot be converted. On Starter and Builder, create a document index and load your records into it.

Learn more: convert an index guide

No. Pinecone does not use HNSW and never has. Pinecone is serverless and stores data in immutable units called slabs. It picks an indexing algorithm for each slab by size: RaBitQ for small slabs, PQFS for medium slabs, and IVF with PQFS for large ones. Writes are acknowledged in under 100 ms and become searchable within seconds. You never choose or tune the algorithm, and Pinecone upgrades algorithms in the background without reindexing.

Learn more: How Pinecone works | Pinecone indexing algorithms

Pinecone gives each tenant its own namespace inside a serverless index. Each namespace is stored separately, which physically isolates tenant data. Because reads and writes always target a single namespace, one tenant's activity does not affect the others. Offboarding a tenant means deleting its namespace. Query cost follows namespace size at 1 read unit per GB. Small tenants therefore cost little to query. The documented limit is 100,000 namespaces per index on Standard and 1,000,000 on Enterprise. Dedicated Read Nodes support one namespace per index today. For now you cannot combine them with a design that keeps many tenants in one index.

Learn more: multitenancy guide

Yes. Bring Your Own Cloud, generally available since September 23, 2026, runs the Pinecone data plane in your own AWS, GCP, or Azure account, in a dedicated VPC. Pinecone manages upgrades and scaling without SSH, VPN, or inbound network access to your environment. Vectors and queries stay inside your VPC. Only operational metrics and traces go to Pinecone for monitoring. You can use public endpoints or private-only connections through AWS PrivateLink, GCP Private Service Connect, or Azure Private Link.

Learn more: Bring Your Own Cloud

Not for new projects. Pods are Pinecone's legacy architecture, and Standard and Enterprise customers who signed up on or after August 18, 2025 cannot create them. Serverless is the current architecture and the one Pinecone recommends. Dedicated Read Nodes are the closest match to pods: read capacity you size yourself, on the current engine, with dense, sparse, hybrid, and full-text search. On-demand serverless charges for the reads and writes you run. An idle index costs close to its storage alone. To move, create a serverless index with the same metric and dimensions, then load your data into it.

Learn more: pods versus serverless | Dedicated Read Nodes

No, not the ones that ran before serverless. Benchmarks that compare Pinecone with other databases on pod-based indexes do not describe serverless performance. The two most-cited comparisons, from October 2023 and June 2024, ran on pods, a product line since retired. Neither tested serverless, and neither could have tested Dedicated Read Nodes, which became generally available in April 2026. Pod-era prices are out of date as well. When you read a benchmark, check which index type it used and when it ran. Better still, test your own data on the free Starter plan.

Learn more: pods versus serverless | Pinecone vs pgvector comparison

Pinecone scales automatically on its on-demand tier, where you never pick a shard or replica count. The documented limit is 100,000 namespaces per index on Standard and 1,000,000 on Enterprise, and Support can raise many limits on request. For larger, steadier workloads, Dedicated Read Nodes are capacity you provision, with shards and replicas set by hand today. Published figures: Pinecone's load tests on an e-commerce marketplace's recommendations workload reached 1.4 billion vectors at 5.7k queries per second with filters and a 26 ms median latency. The Dedicated Read Nodes page lists a media company at 480 million vectors and a design platform at 135 million. Latency depends on namespace size, filters, and tier. Test it on your own data.

Learn more: Dedicated Read Nodes | database limits

Yes. Integrated embedding on the Documents API embeds text inside Pinecone. One text field then serves semantic and full-text search without a separate embedding pipeline. Hosted embedding models include OpenAI's text-embedding-3-small and text-embedding-3-large, Google's gemini-embedding-2 (Builder and above), llama-text-embed-v2, multilingual-e5-large, and pinecone-sparse-english-v0. Hosted rerankers include bge-reranker-v2-m3 and cohere-rerank-4-fast. Starter and Builder get bge-reranker-v2-m3 only. Documents and queries use the same model. Embedding is billed on the same Pinecone invoice.

Learn more: integrated embedding announcement

There are three options. pgvector is the better choice when the corpus is small and mostly static, the index fits in memory you already pay for, or your Postgres host cannot load extensions. pg_pinecone is Pinecone's open-source extension for Postgres 17 or 18 that you run yourself. Its native engine filters on one chosen column during the search. Its remote engine moves a table's search to Pinecone while Postgres keeps every row. Pinecone Database is the better fit once a workload outgrows one server, changes continuously, or needs full-text search in the same index as your vectors. pgvector's HNSW index has to fit in memory, and by default Postgres applies metadata filters after the vector search has chosen its candidates.

Learn more: Pinecone vs pgvector comparison | pg_pinecone documentation

Choose Pinecone for keyword search, vector search, or both when you do not want to run a search cluster. On Elasticsearch and OpenSearch you size node memory for the vector index, choose shard and replica counts, add nodes and reshard as data grows, and tune ingest so indexing does not slow queries. Pinecone separates storage from compute. On the on-demand tier there is nothing to size, and on Dedicated Read Nodes you choose shards and replicas. Elasticsearch and OpenSearch are the better choice when you already run them for logs, observability, or security analytics, need aggregations or dashboards, or, for OpenSearch, need an Apache 2.0 license.

Learn more: Pinecone vs Elasticsearch comparison | Pinecone vs OpenSearch comparison

Chroma Cloud becomes read-only on November 22, 2026 and stops serving reads on December 22, 2026. Pinecone is a managed service for keyword and vector search: you create an index and Pinecone runs the servers. A migration guide and scripts cover the move from Chroma Cloud. Self-hosted Chroma is the better choice when the database has to run embedded in your app, on a laptop, or offline, or when running open-source code yourself is a requirement. Pinecone ranks keyword matches with BM25. Chroma's full-text search keeps or drops records by exact, case-sensitive match and gives them no score.

Learn more: Pinecone vs Chroma comparison | Chroma Cloud migration guide

Pinecone is the better choice when retrieval grows at a different rate than the database next to it. On Atlas, one Search Node supplies the memory, the CPU, and the tenant boundary as a single purchase. Scaling any one of them means buying all three. On Pinecone, storage, reads, and writes scale on their own, and namespaces keep tenants apart. MongoDB Atlas Vector Search is the better choice when a document and its embedding must be written in one transaction, or when you want one combined ranking of text and vector matches from a single query. Atlas does both natively. Pinecone does not merge those rankings in one request today.

Learn more: Pinecone vs MongoDB Atlas Vector Search comparison

Pinecone fits when you want usage-based pricing on the on-demand tier, or when you need the same API on AWS, Azure, and Google Cloud. Vertex AI Vector Search fits when you are committed to Google Cloud and want retrieval in the same project, under the same IAM policies, and on the same bill. Vertex prices serving as replicas times shards times an hourly node rate. Dedicated Read Nodes are priced by node type, shards, and replicas, much like Vertex. On the on-demand tier you do not pick nodes.

Learn more: Pinecone vs Vertex AI Vector Search comparison

Yes. Pinecone retrieves the passages a model needs to answer from your own data. You embed your documents, or let Pinecone embed them with integrated embedding, then query for the closest matches and pass them to the model. Hybrid retrieval combines dense and sparse vectors in one query. Metadata filters narrow results to a user, a team, or a document type. Hosted rerankers reorder the top results. New documents become searchable within seconds. Vanguard, an investment management company, reports that hybrid retrieval improved accuracy by more than 12% over dense retrieval alone.

Learn more: retrieval-augmented generation guide | Vanguard customer story

Yes. Agents use Pinecone to retrieve the knowledge they need at each step and to remember what happened in earlier sessions. Store each user's or agent's data in its own namespace and filter by metadata such as time or source. When agents reach production and need governed answers with field-level citations, Pinecone Nexus compiles enterprise data into knowledge ahead of time. Aquant, an agentic AI platform for service teams, runs its retrieval on Pinecone. Terminal X, a research platform for financial content, runs its research agents on Pinecone.

Learn more: Pinecone Nexus | Aquant customer story

Yes. Pinecone Database can hold an agent's memory. Give each user or agent its own namespace, add metadata such as time or memory type, and query for the closest matches with filters. Writes are acknowledged in under 100 ms and become searchable within seconds. The Database retrieves information when the agent asks for it, including memories from earlier sessions. Pinecone Nexus works ahead of time. It compiles enterprise data into governed knowledge with field-level citations. Agents then do not rebuild answers from raw chunks on every call.

Learn more: multitenancy guide | Pinecone Nexus

Yes. A recommendation system finds the items, products, or people closest to a user's interests, which is a nearest-neighbor search over embeddings. Pinecone serves those queries in real time, applies metadata filters for rules such as region or availability, and makes new items searchable within seconds. ZoomInfo runs real-time contact recommendations on Pinecone across more than 390 million vectors and more than 100,000 namespaces, with a P50 latency of about 60 ms at about 40 QPS, and reports a 2x improvement in relevancy and recall. In load tests Pinecone ran on an e-commerce marketplace's recommendations workload, Dedicated Read Nodes served 1.4 billion vectors at 5.7k queries per second with filters and a 26 ms median latency.

Learn more: ZoomInfo customer story | Dedicated Read Nodes

Yes. A platform that serves many customers gives each one its own namespace in a serverless index. Each tenant's data is stored separately. A new tenant needs nothing provisioned. Query cost follows the size of each tenant's namespace. Small tenants therefore cost little to query. CustomGPT.ai serves more than 10,000 paying customers this way, each running custom GPT projects on their own data, across hundreds of millions of vectors in thousands of namespaces. Jenova, an agent platform that uses Pinecone for agent memory and RAG over user uploads, reached $1M ARR and more than 200,000 signups.

Learn more: multitenancy guide | CustomGPT.ai customer story

Pinecone Database is the wrong tool when the database has to run embedded in your app, on a laptop, or offline. It is a managed service and is not open source. If you want open-source code inside the Postgres you run, pg_pinecone (Apache 2.0, for Postgres 17 or 18 that you manage) keeps search there. If your managed Postgres host does not offer the extension, as with Amazon RDS, pgvector is what runs. A small, mostly static corpus can also stay in Postgres. Pinecone fits when the workload grows, changes continuously, or needs filtered queries that return complete results. The Starter plan is free. You can test your own workload there first.

Learn more: Pinecone vs pgvector comparison | Pinecone comparisons

pgvector is enough while the corpus is small and mostly static, the index fits in memory you already pay for, and you want one system. Move when the workload grows, changes continuously, or depends on filters that must return complete results. By default, Postgres applies metadata filters after the vector search has chosen its candidates. A selective filter can therefore return fewer rows than the limit you asked for. If you run Postgres 17 or 18 yourself, you can adopt Pinecone gradually: with pg_pinecone, one table can switch between the pgvector, native, and remote engines. A managed Postgres host that does not offer the extension, such as Amazon RDS, stays on pgvector or moves search to Pinecone Database.

Learn more: Pinecone vs pgvector comparison | migrate from pgvector guide

The pg_pinecone extension supports it. Its remote engine syncs a Postgres table to a Pinecone serverless index. Postgres keeps every row. You write to Postgres as usual and search from SQL. Search on the remote copy is eventually consistent. A new write may not be searchable the moment it commits. A table can also keep a Pinecone full-text index in sync for BM25, vector, or hybrid search from SQL. Without the extension, your application writes to both systems. Plan for retries, deletes, and drift checks.

Learn more: pg_pinecone documentation

Pinecone has four plans. Starter is free. Builder is $20 flat. Standard has a $50 monthly minimum. Enterprise has a $500 monthly minimum. Above the minimum you pay as you go, from $0.33 per GB per month for storage, from $4 per million write units, and from $16 per million read units. Rates vary by cloud and region, and Enterprise rates are higher. Standard starts with a 3-week trial that includes $300 in credits. Enterprise trials are available on request. Enterprise adds a 99.95% uptime SLA. Standard has no SLA.

Learn more: Pinecone pricing

Cost follows three meters: storage at $0.33 per GB per month, write units, and read units. A query uses 1 read unit per 1 GB of the namespace it targets, with a 0.25 minimum. Settings such as top_k and include_metadata do not change the price. Dedicated Read Nodes do not charge read units. You pay a fixed hourly rate by node type, shards, and replicas, which suits steady, high-volume reads. Three published workloads cut cost by 77 to 97 percent moving from on-demand pricing to Dedicated Read Nodes. They support one namespace per index today, and shards and replicas are set by hand.

Learn more: understanding cost guide | Pinecone cost estimator

Yes, to start. The Starter plan is free and includes up to 2 GB of storage, 2M write units, and 1M read units a month, with indexes in AWS us-east-1. Builder is $20 flat with up to 10 GB of storage, and on Builder, usage beyond the included limits is blocked rather than billed. Standard has a $50 monthly minimum and Enterprise has a $500 monthly minimum. Both bill usage above the minimum as you go. Standard also comes with a 3-week trial that includes $300 in credits.

Learn more: Pinecone pricing

Yes. Pinecone encrypts data at rest and in transit and holds SOC 2 on every plan. GDPR and ISO 27001 are available from Builder. HIPAA is a $190 monthly add-on on Standard and is included on Enterprise. SSO with SAML 2.0 and role-based access control start on Standard. Audit logs, private endpoints, customer-managed encryption keys, and SCIM are on Enterprise. With Bring Your Own Cloud, vectors and queries stay inside your own VPC.

Learn more: Trust and security

Create a free account at app.pinecone.io, copy your API key, and run the quickstart. The Starter plan is free. The quickstart includes a path where an AI coding agent designs your schema, ingests your data, and runs your first search. To build without an account, Pinecone Local is an in-memory Docker emulator for development and testing that ignores API keys and does not persist data.

Learn more: Pinecone quickstart

Yes. CustomGPT.ai serves more than 10,000 paying customers on hundreds of millions of vectors across thousands of namespaces. For the upper end, Pinecone ran load tests on an e-commerce marketplace's recommendations workload and served 1.4 billion vectors at 5.7k queries per second with filters and a 26 ms median latency on Dedicated Read Nodes. Those are test figures, not a live workload. Pinecone stores data in object storage and separates storage from compute. Each part scales on its own.

Learn more: Dedicated Read Nodes | CustomGPT.ai customer story

More than 10,000 customers build on Pinecone. Published customer stories include ZoomInfo, Vanguard, Gong, Assembled, Stravito, 1up, Aquant, Obviant, Terminal X, and CustomGPT.ai. Their workloads include recommendations, retrieval-augmented generation, search, and agents, in fields from finance to security to sales intelligence. Each story describes the workload and what the team built. The stories are published on pinecone.io and linked from the customers page.

Learn more: Pinecone customers

ZoomInfo reports a 2x improvement in relevancy and recall, more than 50% higher user engagement, and a 50x increase in peak customer requests served. Vanguard reports that hybrid retrieval improved accuracy by more than 12% over dense retrieval alone. Gong reports 10x cost savings moving to Pinecone serverless. 1up reports 10x faster response generation for RFPs and compliance questionnaires. Obviant reports 30% more accurate defense acquisition recommendations from combining sparse and dense retrieval. Each figure comes from the customer's own published story.

Learn more: Pinecone customers