October 2026 Summaries

Neon’s Lakebase Search has introduced the lakebase_tokenizer extension, which lets managed Postgres users define custom synonym and stop-word dictionaries as SQL tables rather than server-side files. Because BM25 ranking depends on the normalized tokens created before indexing, the extension enables applications to map vocabulary variants such as “k8s,” “kube,” and “Kubernetes,” normalize abbreviations like “PG” and “creds,” and strip accents so searches for “zurich” can match “Zürich.” The configuration works with standard PostgreSQL tsvector columns as well as GIN and lakebase_bm25 indexes, and it is durable and branch-aware, allowing dictionary changes to be tested in database branches. An example using support tickets shows that custom tokenization can increase relevant matches and alter BM25 scores by producing more accurate term-frequency and rarity statistics, while custom stop words can prevent conversational filler terms from causing otherwise useful searches to fail.

Atlas is a TanStack Start photo-library application hosted on Vercel that demonstrates how Neon’s backend services can support private uploads, user authentication, and search by text, similar image, or face. It stores original images and face crops in private S3-compatible object storage, uses short-lived presigned URLs for access, and records metadata, ownership, captions, CLIP embeddings, and face descriptors in Lakebase Postgres. Managed Better Auth supplies authenticated user identities, while server-side checks scope every operation to the signed-in owner. Text and image similarity searches use CLIP vectors and Lakebase Search’s ANN indexing, while facial search uses browser-based face detection and descriptors grouped into people through clustering. A central feature is Neon branching, which creates isolated preview environments containing copy-on-write database data, private objects, authentication sessions, and search indexes, allowing developers to test changes against realistic production-like libraries without copying data or affecting production.

Neon has increased Postgres storage on its Free plan from 0.5 GB to 1 GB per project while retaining support for up to 100 projects, with the higher limit applied automatically to existing projects. Designed for coding-agent workflows involving frequent experiments and many mostly idle applications, each free project includes 100 monthly compute-unit hours, autoscaling up to 2 CU, 10 database branches, a six-hour restore window, managed authentication for up to 60,000 monthly active users, 5 GB of S3-compatible object storage, and Node.js Functions with monthly invocation and capacity allowances. Neon promotes creating a separate backend for each idea or demo through its CLI plugin and MCP server, and offers bootstrap templates for applications such as realtime chat, file indexing, and Discord bots. The company also invites users constrained by free-tier limits elsewhere to migrate projects to Neon using its documentation.

Neon has added embedding-model support to its AI Gateway, allowing developers to generate vectors through an OpenAI-compatible `/v1/embeddings` endpoint using the same credentials and SDK setup used for chat models. At launch, the service offers qwen3-embedding-0-6b, a configurable 1,024-dimensional normalized model priced at $0.02 per million input tokens, and gte-large-en, a 1,024-dimensional model priced at $0.13 per million tokens. The feature integrates with Neon’s Lakebase Postgres database and Lakebase Search extensions, enabling vector, keyword, and hybrid retrieval alongside application data. Neon presents this as a complete retrieval-augmented generation workflow in which uploaded files trigger Functions, are chunked and embedded through AI Gateway, stored in Postgres, retrieved through vector or BM25 search, and passed to chat models for answer generation. Because Neon’s database, storage, functions, gateway, and search capabilities branch together, teams can test alternative embedding models, vector dimensions, chunking methods, and ranking strategies against production-like data without duplicating storage.

pg_redact is a demonstration support-inbox application that addresses the limitations of column-level PostgreSQL privacy controls by detecting and masking personally identifiable information embedded within free-text messages. It generates possible PII spans using regex and tokenization rules, then sends those candidates with their surrounding message context to TypeSafe’s Jev classification model, which labels items such as names, emails, addresses, phone numbers, and identification numbers while filtering low-confidence results. The application stores the original message and classified spans in PostgreSQL JSONB fields, where SQL functions assign sensitivity levels to PII types and redact content according to role-based clearance levels: guests see no detected PII, support agents can see lower-sensitivity identifiers such as names and emails, and administrators see all content. By separating classification from database-side enforcement, the system allows clearance and sensitivity policies to change without reclassifying stored messages, while preserving useful non-PII details such as invoice numbers that simple pattern matching might otherwise conceal.

Neon has released @neon/effect, a set of Effect 4 bindings for its TypeScript SDK that brings typed errors, cancellation, timeouts, retries, and Streams for paginated Neon API operations. The package targets developers building automation, CI/CD workflows, or platforms that provision isolated Neon Postgres projects and resources for users, applications, or agents, where asynchronous provisioning, readiness polling, rate limits, and transient failures must be coordinated reliably. It preserves the SDK’s client methods and parameters while returning Effect values for API calls and Streams for lists, allowing applications to handle specific errors, retry network failures, set deadlines, resume timed-out readiness operations, and interrupt active HTTP requests and polling. The announcement argues that Effect’s explicit handling of production concerns is increasingly useful for agent-generated TypeScript code, while noting that Effect 4’s zero runtime dependencies provide a foundation for libraries such as this integration.