SearchOS - An Open-Source Multi-Agent Search Collaboration Framework by Ant Group and Others

What Is SearchOS

SearchOS is an open-source multi-agent search collaboration framework jointly launched by the Gaoling School of Artificial Intelligence at Renmin University of China and Ant Group. The framework draws on the design philosophy of relational databases, abstracting open-domain information retrieval into dynamic relational schemas. Through search-oriented context management, pipeline parallel scheduling, and a search tool middleware layer, it addresses system-level issues in long-horizon tasks such as agent amnesia, repeated queries, and source decoupling, supporting multi-agent shared state, efficient collaboration, and evidence traceability.

SearchOS's Main Features

• Relational Search Schema Construction: Automatically transforms natural language queries into dynamic relational schemas containing table structures, primary keys, and foreign keys, discovering entities while completing attributes.
• Multi-Agent Collaborative Search: The Orchestrator performs 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): Maintains four types of shared state: Frontier Tasks, Evidence Graph, Coverage Map, and Failure Memory, enabling progress visualization and cross-agent synchronization.
• Search Tool Middleware Layer: Includes Context Middleware, Sensor Middleware, and Evidence Extraction Middleware, responsible for state injection, loop detection, and structured evidence extraction.
• Hierarchical Skill Library: Presets about 280 skills, covering strategic skills such as ranked retrieval, multi-hop search, and entity disambiguation, as well as access skills for anti-crawling, login walls, and dynamic pages.
• Real-Time Progress Tracking: Supports CLI, full-screen TUI, and Web workbench, allowing real-time viewing of schema completion progress, agent task flows, and cell-by-cell evidence, with support for resuming after interruption.

SearchOS's Technical Principles

• Relational Search Abstraction: Defines information retrieval tasks as a relational schema S = ({T_m}, R), where T_m is a table schema (including column attributes A_m and primary key P_m) and R represents relationships between tables, unifying various retrieval tasks under a structured framework.
• Shared State Infrastructure: The core judgment is that search state should exist in the system. SOCM externalizes task state into four types of shared data structures. Agents collaborate by querying the shared state, avoiding fact loss caused by context compression.
• Pipeline Parallel Scheduling: Adopts a micro-batch approach, treating search→open→find as a search pipeline, with multiple Search Agents entering in staggered phases and advancing in an overlapping manner; the continuous dispatch mechanism immediately supplements new tasks after an agent completes, improving slot utilization and wall-clock efficiency.
• Middleware Intervention Mechanism: Inserts three layers of middleware between the model and tools. Context Middleware controls context size, Sensor Middleware identifies loops and stagnation and triggers strategy switching, and Evidence Extraction Middleware performs structured extraction, citation anchoring, and evidence storage.
• Dynamic Schema Evolution: The system dynamically adjusts table structures as entities are discovered. Experiments show that even when the optimal fixed structure (Oracle) is pre-specified, the results are still lower than SearchOS's dynamic schema.

How to Use SearchOS

• Environment Preparation: Visit the GitHub repository to clone the project, and complete model provider and API Key setup through the configuration wizard.
• Launch Interaction: Choose one of three interfaces to start the system: CLI, full-screen TUI, or Web research workbench.
• Enter a query: Enter a natural language question in the search box, and the system will automatically identify the task type.
• View progress: Observe in real time the construction process of the relational schema, the completion status of the Coverage Map, and the task assignment and execution flow of each Agent.
• Review results: After the task is completed, view the structured output; each piece of data comes with a citation that can be traced back to the original webpage and specific location.
• Interruption and resumption: Long-horizon tasks support resuming after exiting midway, and the historical search trajectory and evidence chain can be reviewed at any time.

Core advantages of SearchOS

• System-level state externalization: Migrate the search state from the dialogue context to the system infrastructure, fundamentally solving the problems of amnesia and repetition in long-horizon tasks.
• Dynamic relational schema: Adjust the table structure in real time as entities are discovered, which fits the diversity of real retrieval scenarios better than a fixed schema; Oracle experiments verify its effectiveness.
• High-throughput scheduling: Pipeline parallelism and continuous dispatching reduce task completion time by about 24% and the number of LLM calls by about 13%, while also improving the F1 score.
• Full-chain evidence traceability: Every value preserves a citation, supporting tracing from conclusions back to the specific location in the original webpage text, satisfying the rigorous requirement of "find as much as possible, and every item has a source".
• Rich preset skills: About 280 skills cover search strategies and website access; after using skills, search time is shortened by 36.6%, and the number of API calls is reduced by 39%-43%.

Project addresses of SearchOS

• 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

Comparison of SearchOS with similar competing products

Application scenarios of SearchOS

• Competitor and market research: Systematically collect information such as competitor companies' funding rounds, product features, pricing strategies, etc., and generate large-scale comparison tables with sources.
• Academic literature review: For a specific research direction, enumerate related papers, authors, institutions, and publication venues, and fill in key conclusions and method summaries.
• Investment research and due diligence: Track the shareholder structure, executive changes, litigation records, and financial data of target companies to achieve multi-source cross-verification.
• Product Comparison and Selection: Collect technical parameters, user reviews, and price ranges of similar products, and generate structured comparison reports to assist decision-making.
• Open-Domain Knowledge Enumeration: Information collection tasks requiring high recall, such as "list all Fields Medal winners since 2000 and their affiliated institutions and countries."