October 2026 Summaries

5 posts from Pinecone

Filter
Month: Year:
Post Summaries Back to Blog
Pinecone has introduced pg_pinecone, an Apache 2.0 open-source PostgreSQL extension that adds vector search to existing pgvector data while keeping tables and full vectors in Postgres. Its native engine uses IVF clustering and RaBitQ vector compression rather than an HNSW graph, aiming to reduce index size, accelerate index construction and search, and improve filtered-query accuracy by applying filters during search instead of afterward. The extension preserves standard SQL vector queries and lets users switch among its local native engine, pgvector’s HNSW index, and Pinecone’s managed remote service without changing queries. Reported benchmarks on one million 768-dimensional vectors show smaller indexes and lower latency at comparable recall than HNSW, particularly for narrow filters, although substantial data changes may require reindexing or automatic maintenance. Optional Pinecone-connected features include embedding text through SQL, BM25 full-text search, reranking, and scaling search beyond a single database server, while Pinecone invites community feedback and contributions for the project’s future development.
Oct 27, 2026 1,955 words in the original blog post.
Chroma Cloud will stop accepting writes on November 22, 2026, and cease serving reads on December 22, prompting users to export and validate their data before the deadlines, while open-source Chroma remains unaffected. The guide describes migrating a read-only Chroma Cloud collection containing existing dense embeddings, documents, and metadata to a Pinecone serverless index without re-embedding data, using a six-script toolkit for preflight checks, index creation, export, conversion, loading, and validation. It highlights differences between the platforms, including index and namespace structure, metadata-field rules, ID restrictions, supported filters, and query features that may require redesign. The recommended process includes rehearsing the migration to estimate a write-pause window, blocking or transforming incompatible data, exporting a consistent snapshot, loading into a new namespace, comparing record counts, fields, IDs, filters, and search results, then switching reads and writes together after a decision gate. Following cutover, users are advised to monitor performance and result quality, archive migration mapping files, retire Chroma credentials and integrations before the read-service deadline, and optionally later adopt Pinecone integrated embedding or full-text search. Pinecone also offers eligible Chroma Cloud users a three-week Standard trial with $300 in credits, subject to subsequent plan selection.
Oct 26, 2026 2,442 words in the original blog post.
Pinecone has introduced integrated embedding for its Documents API, allowing users to embed raw text and perform semantic, sparse, and full-text search through a single API and shared text field. The launch includes llama-text-embed-v2 for long multilingual passages, multilingual-e5-large for shorter multilingual inputs, and pinecone-sparse-english-v0 for English keyword-oriented retrieval, while BM25 remains available for conventional text matching. Embeddings are generated automatically when documents are written or updated, ensuring documents and queries use the same model and eliminating separate embedding pipelines, provider credentials, re-embedding jobs, and associated monitoring. Users can combine semantic ranking with exact-phrase full-text filters, optionally avoid storing raw text when only semantic search is needed, and migrate existing content by upserting text into a Documents API index. Pinecone reports that its combined embedding and search request reduced median latency in one test compared with a separate embedding call, and plans future support for OpenAI and Google embedding models, server-side result fusion, reranking, bulk-import embedding, and geographically constrained processing.
Oct 22, 2026 1,111 words in the original blog post.
Pinecone’s Documents API integrates embedding generation with native full-text and vector search, reducing the need for developers to build and operate separate embedding pipelines or manage vectors directly. The evolution began with the Vectors API, which required users to supply embeddings and decide where to store source text, followed by the Records API, which automated embeddings but retained limitations of metadata-based text storage and a single-model structure. Built on Pinecone’s slab-based serverless architecture, the Documents API introduces schema-defined document fields that can store raw text, create a Tantivy-powered full-text index, and generate vectors through a selected embedding model in a single write. Users can search the same field using exact keyword relevance or semantic text queries while applying metadata filters, and integrated embedding eliminates an extra network call, which Pinecone reports made median searches about 3.5 times faster in its tests. Pinecone plans to expand the API with hybrid result fusion, additional models, integrated reranking, bulk import embedding, multimodal search, and data-zone processing, positioning the Documents API as its primary platform for future retrieval workflows.
Oct 22, 2026 3,280 words in the original blog post.
Pinecone has introduced on-demand capacity mode for its Bring Your Own Cloud offering, allowing customers to create indexes without pre-sizing shards individually. The feature uses dynamic sharding across a shared capacity pool, directing new writes to shards with available space and enabling capacity expansion by adding shards without relocating existing data. It is designed to support applications with potentially millions of namespaces, such as multi-tenant services and AI agents, while avoiding overprovisioning or disruptive index rebuilds. Frequently queried namespaces can be moved to dedicated indexes with separate read nodes, while other namespaces remain in the shared pool. The service retains Pinecone BYOC’s zero-access model, keeping vectors, metadata, and queries in customers’ AWS, Google Cloud, or Azure accounts while Pinecone monitors operational metrics without accessing data. On-demand mode is now available to all BYOC customers through the Pinecone console, API, and SDKs.
Oct 15, 2026 508 words in the original blog post.