OpenSearch Adds Agentic and Conversational RAG
What happened
The OpenSearch platform now offers native support for building conversational and agentic RAG systems using models from OpenAI and Anthropic. New documentation details how to configure AI agents that coordinate LLM reasoning and retrieval for dynamic, multi-step workflows. This mainstreaming of agentic search allows for the creation of more sophisticated enterprise assistants directly within the OpenSearch stack.
Why it matters
- The new agent and conversational RAG features were introduced as experimental in OpenSearch version 2.12 and require enabling the `plugins.ml_commons.memory_feature_enabled` setting. - OpenSearch offers four main types of agents: flow agents for sequential, fixed workflows; conversational flow agents that add memory to fixed workflows; conversational agents that use LLM reasoning to dynamically choose tools; and plan-execute-reflect agents for complex, multi-step tasks. - Beyond OpenAI and Anthropic models, the agent framework also supports Google Gemini models and models available through the Amazon Bedrock Converse API. - The functionality is delivered through the ML Commons plugin, which manages machine learning tasks, resource allocation, and provides APIs for interacting with models. - For RAG pipelines, agents typically utilize a `VectorDBTool` for semantic search within OpenSearch indices and an `MLModelTool` to pass the retrieved context to an external LLM. - Persistent agentic memory became generally available in version 3.3, allowing agents to learn and reason across multiple conversations by implementing strategies like semantic fact extraction and conversation summarization. - The framework supports hybrid search, combining traditional keyword-based (BM25) search with semantic vector search to improve the relevance of retrieved documents. - OpenSearch is extending agent capabilities through Model Context Protocol (MCP) connectors, which allow agents to interact with external tools and data sources beyond what is stored in OpenSearch itself.
Key numbers
- - The new agent and conversational RAG features were introduced as experimental in OpenSearch version 2.12 and require enabling the plugins.ml_commons.memory_feature_enabled setting.
- Persistent agentic memory became generally available in version 3.3, allowing agents to learn and reason across multiple conversations by implementing strategies like semantic fact extraction and conversation summarization.
- The framework supports hybrid search, combining traditional keyword-based (BM25) search with semantic vector search to improve the relevance of retrieved documents.
Quick answers
What happened in OpenSearch Adds Agentic and Conversational RAG?
The OpenSearch platform now offers native support for building conversational and agentic RAG systems using models from OpenAI and Anthropic. New documentation details how to configure AI agents that coordinate LLM reasoning and retrieval for dynamic, multi-step workflows. This mainstreaming of agentic search allows for the creation of more sophisticated enterprise assistants directly within the OpenSearch stack.
Why does OpenSearch Adds Agentic and Conversational RAG matter?
The new agent and conversational RAG features were introduced as experimental in OpenSearch version 2.12 and require enabling the plugins.ml_commons.memory_feature_enabled setting. OpenSearch offers four main types of agents: flow agents for sequential, fixed workflows; conversational flow agents that add memory to fixed workflows; conversational agents that use LLM reasoning to dynamically choose tools; and plan-execute-reflect agents for complex, multi-step tasks. Beyond OpenAI and Anthropic models, the agent framework also supports Google Gemini models and models available through the Amazon Bedrock Converse API. The functionality is delivered through the ML Commons plugin, which manages machine learning tasks, resource allocation, and provides APIs for interacting with models. For RAG pipelines, agents typically utilize a VectorDBTool for semantic search within OpenSearch indices and an MLModelTool to pass the retrieved context to an external LLM. Persistent agentic memory became generally available in version 3.3, allowing agents to learn and reason across multiple conversations by implementing strategies like semantic fact extraction and conversation summarization. The framework supports hybrid search, combining traditional keyword-based (BM25) search with semantic vector search to improve the relevance of retrieved documents. OpenSearch is extending agent capabilities through Model Context Protocol (MCP) connectors, which allow agents to interact with external tools and data sources beyond what is stored in OpenSearch itself.
