Pinecone Feature Trails

Pinecone is integrated with Confluent Cloud for Apache Flink, enabling real-time data processing for GenAI applications. This integration allows for filtering, joining, and enriching data streams from various sources and delivering them to Pinecone for use in RAG pipelines. The solution aims to simplify the development of AI applications by providing a unified platform for data streaming and processing, ensuring that AI models have access to up-to-date and trustworthy data.

Pinecone has enhanced its observability capabilities by introducing new Prometheus endpoints and an updated Datadog integration. These features provide deeper insights into index performance and usage, including request frequency and duration, read/write unit consumption, and core latency metrics. The Datadog integration now supports both serverless and pod-based indexes with 25 new metrics, offering granular monitoring, out-of-the-box dashboards, and customizable alerts. The Prometheus endpoint now supports an open-source monitoring stack with pre-built Grafana dashboards and Prometheus metric collection for both Pinecone SaaS (Serverless) and Bring Your Own Cloud (BYOC) deployments, including Kubernetes infrastructure metrics via Node Exporter for BYOC. The stack supports Docker Compose for SaaS-only, Helm charts for BYOC, and a unified instance for mixed environments.

Full Text Search (FTS) is now available in Pinecone, enabling BM25 scoring across multiple text fields per index, Lucene query syntax, and multi-language tokenization. FTS can be combined with vector search and metadata filters within a single query. Text fields, dense vectors, sparse vectors, and filterable metadata are defined in a schema at index creation. Supported operations include upsert, fetch, and delete. Scoring in a single query operates on one modality at a time (BM25, dense, or sparse). The hybrid index allows for keyword-aware semantic search by combining dense and sparse vector representations, offering a unified approach to hybrid search with configurable relevance weighting.

Pinecone has introduced a new Enterprise plan that offers enterprise-grade features and support with self-serve convenience for mission-critical applications. This plan includes multiple availability zones for greater availability when replicas are added to an index, and Prometheus metrics for enhanced observability by allowing users to monitor Pinecone indexes by ingesting performance metrics into their own Prometheus instances or compatible tools. The plan also comes with Premium-level support, including availability SLAs, response-time SLAs, and 24/7 support access. The 'Free' plan has been renamed to 'Starter', and the 'Dedicated' plan to 'Enterprise Dedicated'. The usage estimator has been improved to allow specifying metadata size. This plan is designed for users who need enterprise-grade support and availability but do not require a dedicated, single-tenant cloud environment.

This post introduces a service-aware load balancer for Pinecone Assistant, designed to route requests across embeddings, rerankers, and LLMs. It employs a 'power of two choices' algorithm for adaptive routing without a global coordinator, enabling automatic failover during upstream incidents and reducing latency. Different scoring policies are used for different service types (embeddings/rerankers on latency+reliability, LLMs on reliability+load) to optimize performance and availability.

Pinecone Assistant has evolved into an end-to-end managed knowledge service for AI applications, handling document ingestion, chunking, embeddings, retrieval, query planning, reranking, and answer generation. This includes integration with Amazon Bedrock as a Knowledge Base, allowing developers to leverage Pinecone for RAG applications within the Bedrock ecosystem. The integration involves setting up Pinecone indexes, configuring AWS Secrets Manager for credentials, and defining data sources in S3. Bedrock agents can then interact with the Pinecone knowledge base to retrieve semantically relevant content for GenAI applications, improving performance, cost-effectiveness, and enterprise readiness. The process involves data ingestion into S3, embedding creation by Bedrock, upserting embeddings into Pinecone, and agent interaction with the knowledge base. The post details the setup process for creating a Pinecone index, configuring AWS Secrets Manager, setting up an S3 data source, creating the Bedrock knowledge base, connecting it to Pinecone, and finally setting up and testing an agent. Canopy, an open-source RAG framework, integrates with Pinecone, simplifying the process of building and deploying RAG applications by enabling compatibility with Azure's hosted OpenAI models, enhancing data security, compliance, and offering fine-tuning capabilities. The post details config

Pinecone has implemented 'Janitor', a robust system for managing object lifecycle and safely deleting billions of objects at scale. This system addresses the challenges of immutable storage by identifying and removing stale and orphaned objects through distinct 'Normal', 'Orphan', and 'Customer Deletion' modes. Janitor employs a core protocol of 'Identify -> Verify -> Execute' to ensure safe and auditable deletion, with advanced property-based testing using a mock clock to validate complex, time-dependent behaviors. This infrastructure ensures predictable storage costs and operational reliability by automating garbage collection.

Pinecone has evolved its serverless architecture to optimize for agentic workloads, characterized by millions of namespaces with small, sporadically accessed vector sets. This involves architectural innovations in the write path with adaptive indexing based on log-structured merge trees and in the query path with on-demand data fetching from blob storage and adaptive caching. The architecture also sees improvements for traditional search workloads with enhanced disk-based metadata filtering and performance optimizations in client libraries for faster data ingestion.

Pinecone BYOC allows customers to run the Pinecone data plane within their own AWS, GCP, or Azure accounts, offering a zero-access operating model. This ensures data residency and meets strict security requirements by keeping vectors and query execution within the customer's VPC. The architecture splits into a customer-hosted data plane and a Pinecone-managed control plane. Operations are managed via an outbound-only, pull-based model, and data plane requests are served directly within the customer's cloud. This post announces general availability of Pinecone on Azure, enabling customers to deploy Pinecone within Azure regions, further enhancing performance, simplifying procurement, and meeting enterprise readiness requirements.

Pinecone has introduced new bulk data operations that allow users to update, delete, and fetch records directly using metadata filters. This enhances data management scalability and efficiency by enabling targeted operations on large datasets without needing to collect individual record IDs. The operations leverage existing filter syntax, simplifying integration and common data management tasks like bulk updates, data purges, and selective fetches. Pinecone 2.0 introduces metadata filtering for greater control over search results, allowing users to combine vector similarity search with arbitrary metadata filters to provide more relevant results. This provides fine-grained control over vector search results at ultra-low latencies, accepting arbitrary filters on metadata and retrieving exactly the number of nearest-neighbor results that match the filters. The update also includes hybrid storage for up to 10x lower costs by storing a compressed vector index in memory and the original, full-resolution vector index on disk. The new architecture uses Kafka ingestion and Kubernetes orchestration, separating read and write paths and disassociating storage and compute for fault tolerance, data persistence, and high availability. A new REST API based on the OpenAPI spec is available for ease of use, along with a revamped Python client. Pinecone 2.0 is SOC2 Type II audited.

Pinecone Inference now includes reranking capabilities, allowing users to generate vector embeddings and perform reranking directly within the platform. This simplifies embedding pipelines by enabling automatic embedding and reranking generation as part of the upsert or query process. The reranking feature is in public preview and currently supports the bge-reranker-v2-m3 model, with plans for more models in the future. This integration aims to reduce infrastructure management complexity and cost. This post announces the general availability of Pinecone serverless on Azure and GCP, and the public preview of Pinecone Assistant. It also highlights the addition of backups for Standard and Enterprise users and the early access of API Key Roles for granular access control.

Pinecone Local, a self-hosted, in-memory emulator of the vector database, is now available in public preview. This allows developers to prototype, test, and develop AI applications locally, integrating workflow testing into CI/CD pipelines without incurring usage costs or resource drain. It supports all SDKs and control/data plane endpoints, requiring no API key for local use. The emulator is distributed via a Docker image (`pinecone-local`) and is not intended for production use.

Pinecone has introduced Customer-Managed Encryption Keys (CMEK) and Role-Based Access Control (RBAC) with API key roles. CMEK enhances data security and tenant isolation by allowing customers to manage their own encryption keys, providing greater control and supporting compliance requirements. The system utilizes hierarchical encryption with Key Encryption Keys (KEKs) and Data Encryption Keys (DEKs) to maintain performance and security. RBAC with API key roles provides more granular control over access to Pinecone's control and data planes, complementing existing user roles for organizations and projects.

Pinecone has introduced a date-based API versioning schema managed via a custom header, allowing for incremental changes and a clear, predictable, and manageable API lifecycle. Stable API versions are supported for a minimum of 12 months. The Global Control Plane API has been launched, using a global URL for all control plane operations, supported within Python, Node, and Java clients, and also via Terraform and Pulumi. This update also includes broader availability with a new serverless region (us-east-1), updated SDKs with functionality to list record IDs or those with common ID prefixes (useful for RAG applications), self-serve SSO for enterprise customers, Python proxy support for SDKs, and a refreshed console with improved navigation and a new troubleshooting section in the docs.

Pinecone has introduced Private Endpoints for AWS PrivateLink, allowing users to securely connect to Pinecone serverless from their VPC without exposing traffic to the public internet. This feature enhances data security by keeping traffic within the AWS network and reducing the risk of misconfiguration or unauthorized access. Private Endpoints are set up per region and require authorization of a VPC endpoint ID in Pinecone. Support for GCP and Azure is planned for later in the year.

Pinecone has released an open-source AWS Reference Architecture, defined via Pulumi, to help users deploy production-grade, high-scale AI applications leveraging Pinecone's vector database. This architecture demonstrates best practices for AWS and Pinecone, including distributed systems, queue-based work distribution, network and security separation, ECS microservices, and autoscaling. It has been tested with large data batches and provides a quick start guide and technical walkthrough for deployment. This post provides practical tips for working with Pinecone at scale, including using the gRPC client for high-throughput environments, leveraging the reference architecture for production deployments, scaling indexes vertically or horizontally, and integrating with platforms like Databricks. It also delves into concurrency and parallelism concepts, explaining multithreading for I/O-bound tasks and multiprocessing for CPU-bound tasks, with code examples using ThreadPoolExecutor and ProcessPoolExecutor. Finally, it details batch processing strategies for large datasets, including payload size limits, failure handling with retries, and logging.

Pinecone's vector database architecture is purpose-built for vector search, moving beyond single-algorithm approaches. It employs a per-slab adaptive indexing strategy (Ananas, PQFS, and IVF) on a serverless infrastructure, designed to handle the memory, compute, and scale requirements of real-time AI applications. This approach contrasts with 'bolt-on' solutions that integrate algorithms like HNSW into existing architectures, which Pinecone argues are unable to meet production demands due to me. This release further optimizes performance at scale, significantly lowering search latency for large indexes (e.g., 3.4x faster for 100M vectors on s1 pods). It also improves predictability by flattening the impact of scaling on search latency and reducing variance between p50 and p95 latencies. Data ingestion has been made faster and more reliable, with upserts remaining fast even as the index approaches capacity. Deployment options have been expanded with new regions on GCP and AWS for Standard plan users.

This post details the evolution of Python packaging and dependency management tools, highlighting the challenges and solutions encountered. It specifically discusses the migration of the Pinecone Python client to Poetry, emphasizing simplified local development, better integration with common AI libraries like LangChain, and a mostly backward-compatible user experience. The post also touches upon the broader Python tooling landscape, including pip, virtualenv, conda, pipenv, and PEP 517/518, to provide context for the choice of Poetry.

Pinecone has introduced a new, more efficient architecture for its free tier, running in a new free region called `gcp-starter`. This change enables instant sign-ups, allows new users to keep their free indexes indefinitely, and significantly reduces index provisioning time to seconds. This new architecture also simplifies index creation for new users by removing features like namespaces, collections, and delete by metadata from free indexes. The `gcp-starter` region replaces `us-west1-gcp-free`.