SearchOS - An open-source multi-agent search collaboration framework developed by Ant Group and others.

What is SearchOS?

SearchOS is an open-source multi-agent search collaboration framework jointly launched by the Renmin University of China's Hillhouse Institute for Artificial Intelligence and Ant Group. Drawing inspiration from relational database design principles, the framework abstracts open-domain information retrieval into a dynamic relational schema. Through search-oriented context management, pipelined parallel scheduling, and a middleware layer for search tools, it addresses system-level issues in long-running tasks such as agent amnesia, duplicate queries, and source decoupling. It supports multi-agent state sharing, efficient collaboration, and evidence traceability.

SearchOS's main functions

• Relationship search pattern constructionIt automatically converts natural language queries into dynamic relationship schemas that include table structures, primary keys, and foreign keys, and completes attributes as entities are discovered.
• Multi-agent cooperative searchThe Orchestrator handles unified planning and scheduling, the Explore Agent discovers candidate entities, the Search Agent executes parallel subtasks, and the Writer Agent generates structured reports.
• Search-Oriented Context Management (SOCM)Maintain four types of shared states: Frontier Tasks, Evidence Graph, Coverage Map, and Failure Memory, to enable progress visualization and cross-Agent synchronization.
• Search tool middlewareIt includes Context Middleware, Sensor Middleware, and Evidence Extraction Middleware, and is responsible for state injection, loop detection, and structured evidence extraction.
• Hierarchical Skills LibraryIt comes pre-loaded with approximately 280 skills, covering strategy skills such as ranking retrieval, multi-hop search, and entity disambiguation, as well as access skills such as anti-scraping, login firewall, and dynamic pages.
• Real-time progress trackingSupports CLI, full-screen TUI and Web workbench, allowing real-time viewing of schema completion progress, agent task flow and cell-by-cell evidence, and supports recovery after interruption.

SearchOS Technical Principles

• Relational Search AbstractionThe information retrieval task is defined as a relation schema S = ({T_m}, R), where T_m is the table schema (including column attributes A_m and primary key P_m), and R is the relationship between tables, so that all kinds of retrieval tasks are unified under a single structured framework.
• Shared state infrastructureThe core judgment is that the search state should exist in the system. SOCM externalizes the task state into four types of shared data structures. Agents achieve collaboration by querying shared states, avoiding the loss of facts caused by context compression.
• pipeline parallel schedulingThe system adopts a micro-batch approach, treating search→open→find as a search pipeline, with multiple Search Agents entering and overlapping at different times. The continuous dispatch mechanism immediately fills in new tasks after an Agent completes its task, improving slot utilization and wall clock efficiency.
• Middleware intervention mechanismThree layers of middleware are inserted between the model and the tool: Context Middleware controls the context size, Sensor Middleware identifies loops and stagnation and triggers policy switching, and Evidence Extraction Middleware completes structured extraction, citation anchoring, and evidence storage.
• Dynamic pattern evolutionThe system dynamically adjusts the table structure as entities are discovered. Experiments show that even if the optimal fixed structure (Oracle) is specified in advance, the performance is still lower than that of SearchOS's dynamic mode.

How to use SearchOS

• Environmental preparationAccess the GitHub repository to clone the project and complete the model service provider and API key settings through the configuration wizard.
• Start interactionChoose one of the three interfaces to start the system: CLI, full-screen TUI, or Web research workbench.
• Enter queryEnter your natural language question in the search box, and the system will automatically identify the task type.
• Check progressReal-time observation of the relationship schema construction process, the completion status of the Coverage Map, and the task allocation and execution flow of each Agent.
• Review resultsAfter the task is completed, you can view the structured output. Each data point is accompanied by a citation that allows you to trace back to the original webpage and specific location.
• Interruption and ResumptionLong-term tasks can be resumed after exiting midway, and historical search trajectories and evidence chains can be reviewed at any time.

SearchOS's core advantages

• System-level state externalizationThis approach migrates the search state from the dialogue context to the system infrastructure, fundamentally solving the problems of amnesia and repetition in long-term tasks.
• Dynamic Relationship PatternThe table structure is adjusted in real time as entities are discovered, which is more in line with the diversity of real retrieval scenarios than a fixed schema. Oracle experiments have verified its effectiveness.
• High-throughput schedulingPipeline parallelism and continuous dispatch reduce task completion time by approximately 24%, LLM call count by approximately 13%, and improve F1 score.
• End-to-end evidence tracingEach value stores a citation, supporting the retrieval of the conclusion back to the specific location in the original webpage, satisfying the rigorous requirement of "finding as many as possible and having a source for every citation".
• Abundant Pre-set SkillsApproximately 280 skills cover search strategies and website access. After using these skills, search time was reduced by 36.6%, and the number of API calls decreased by 39%-43%.

SearchOS project address

• Project official website：https://antins-labs.github.io/SearchOS/
• GitHub repository：https://github.com/antins-labs/SearchOS
• arXiv technical paper：https://arxiv.org/pdf/2607.15257

SearchOS's Competitive Product Comparison

Application scenarios of SearchOS

• Competitive products and market researchSystematically collect information on competitor companies' funding rounds, product features, pricing strategies, etc., and generate a large-scale comparison table with source information.
• Literature ReviewFor a specific research direction, list relevant papers, authors, institutions, and conferences, and complete the abstracts of key conclusions and methods.
• Investment research and due diligenceTrack the target company's shareholder structure, senior management changes, litigation records, and financial data to achieve multi-source cross-verification.
• Product Comparison and SelectionCollect technical parameters, user reviews, and price ranges of similar products to generate structured comparison reports to assist in decision-making.
• Open domain knowledge enumerationTasks requiring high recall include information gathering tasks such as "listing all Fields Medal winners since 2000 and their affiliated institutions and countries".