Open Deep Search (ODS) is an open-source framework that allows users to enhance the capabilities of Large Language Models (LLMs) by integrating search and reasoning functionalities. It acts as a plug-and-play system, enabling the use of various LLMs, both open-source and through API access to closed-source models.
ODS comprises two main components: the Open Search Tool and the Open Reasoning Agent, both of which leverage the base LLM to improve information retrieval and response accuracy. Here's a more detailed breakdown:
Open Search Tool:
This component focuses on searching and processing information from the internet. It utilizes advanced retrieval techniques, including query rephrasing, chunking, and re-ranking
of search results to improve the relevance and diversity of information obtained from the web.
Open Reasoning Agent:
This component leverages the LLM to process information retrieved by the Open Search Tool and generate comprehensive answers to user queries. It employs different methodologies
like Chain-of-thought ReAct and Chain-of-code CodeAct to handle various tasks, including tool usage and context-aware reasoning.
ODS aims to democratize access to powerful search-enhanced AI capabilities by providing an open-source framework that is compatible with diverse LLMs. It allows researchers and developers to build upon existing LLMs and improve their ability to retrieve, understand, and reason about information from the web.
- MCP server is embedded in the Amazon AgentCore Gateway. AWS added more capababilities besides the MCP Server, that is why they call it "Gateway"
- Agentic AI agents runs as Docker container images and that is the reason they need to be published or hosted to AWS Elastic Container Registry (ECR)
This agentic ai system uses langchan agent as an example agentic-ai agent.
The credit for this work goes to https://medium.com/elevate-tech/bedrock-agentcore-the-easy-button-for-ai-agent-deployment-on-aws-finally-exists-1cae7909d36d. I mainly used the instructions mentioned here and on AWS Website.