OpenSearch has released version 3.8 of its open-source search and analytics platform, with updates across AI, vector search, and observability tools.
The release adds broader support for Model Context Protocol integrations, faster vector ingestion, revised radial search, and new query and dashboard functions for log and metric analysis. It also expands automated search relevance evaluation and introduces a gRPC transport for streaming machine learning predictions.
One of the main changes focuses on vector workloads, which have grown in importance as companies use search systems for AI applications and retrieval tasks. Version 3.8 supports Base64-encoded vector ingestion for knn_vector fields, removing JSON array serialisation overhead on both the client and server.
According to the project, a 768-dimensional float vector that takes about 16 KB as a JSON array falls to 4 KB in Base64 form. It says the change can deliver up to 4.16 times higher bulk ingestion throughput and 83% lower median latency, while working with supported k-NN engines without mapping changes.
Search performance has also been updated for radial queries. A redesigned graph traversal method bounds graph exploration for expensive queries, reducing the latency spread previously seen in radial workloads.
In benchmark tests on a dataset of 10 million vectors with 768 dimensions, the project said radial queries delivered up to 2.1 times higher throughput and 45% lower median latency than version 3.7. Mean recall also improved from 0.85 to 0.97.
Agent tools
Another focus is support for AI agents that connect to external systems. Version 3.8 extends MCP support to flow and conversational flow agents, so all four supported agent types can now use MCP integrations through a consistent connector configuration.
The release also adds a list tools API, allowing users to discover tools available on a connected MCP server, including name, type, description, and input schema. Connector-level tool description overrides let users change how tools are described to language models without modifying the external server itself.
Storage management for agent memory has been updated with configurable retention policies for agentic memory containers. These policies can delete expired sessions, long-term memories, and history entries based on time or count limits, with operators able to set defaults across a cluster or apply settings per container.
For streaming inference, ML Commons now supports gRPC transport. This lets users stream responses from externally hosted models such as OpenAI Chat Completions and Amazon Bedrock Converse over protocol buffers on HTTP/2, rather than relying only on REST with Server-Sent Events.
The project says the binary transport reduces latency and CPU overhead for applications that need token-by-token responses. It also expanded LLM-as-a-Judge in Search Relevance Workbench, so users can generate automated relevance judgments through any ML Commons connector rather than only through OpenAI.
Ready-made blueprints are available for providers including Azure OpenAI, DeepSeek, Ollama, Google Gemini, Anthropic Claude on Amazon Bedrock, and the Amazon Bedrock Converse API. The update also adds clearer reporting of timeouts and rate-limit errors, along with metadata summaries and per-document status indicators in the Dashboards judgment view.
Search Relevance Workbench also now supports Mustache template variables in search configurations. This allows users to build experiments with multiple parameters, such as category, brand, or status filters, without duplicating configurations for each variation.
Observability shift
On the observability side, OpenSearch 3.8 introduces a visual Piped Processing Language query builder in Explore logs. It lets users create queries by selecting filters, aggregations, and sort criteria from menus, while still allowing a switch to a raw editor when needed.
Discover logs now includes experimental SQL query support, with integration across logs, visualisation, and statistics views. The release also adds an onboarding canvas with per-index cards showing severity histograms and log-line previews to help users inspect data before writing queries.
OpenSearch has added four new commands to the Calcite implementation of PPL: makeresults, foreach, xyseries, and timewrap. These are intended to help users generate in-memory rows, apply expressions across multiple fields, pivot grouped results into wider tables, and compare time periods in time-series analysis.
A new PPL lint engine validates queries as users type in the editor. It can identify issues such as unsupported functions, unknown fields, misspelt commands, and non-deterministic operations, while offering inline suggestions and quick fixes. The feature is optional, and rules can be enabled, disabled, or assigned different severity levels.
The release also introduces a Grok Debugger in Dev Tools, giving users a way to test Grok patterns directly in OpenSearch Dashboards. Users can paste in a sample log line, write a pattern, and simulate field extraction without leaving the interface.
Metric monitoring has also been tightened. Users working with a Prometheus data source in Discover Metrics can now create an alert rule from the same page where they identify a metric pattern, instead of moving to a separate workflow.
The new toolbar action lets users enter a PromQL expression, set thresholds, select comparison operators, and define evaluation intervals before saving the rule to the configured Prometheus data source.
The release reflects how search platforms are increasingly expected to do more than return results. OpenSearch 3.8 combines changes aimed at handling AI-driven vector workloads with features intended to reduce the effort needed to query logs, inspect metrics, and manage external agent tools.
Among the headline performance claims, the project says vector ingestion can be up to 4.16 times faster, radial search throughput can increase by up to 2.1 times, and the system now supports a wider range of agent architectures and model providers in a single release.