What is openJiuwen-DeepSearch

openJiuwen-DeepSearch is a knowledge-augmented, high-performance, high-precision deep retrieval and research engine. It aims to make effective use of structured knowledge and large models, integrate various tools, and provide enterprise-grade Agentic AI search and research capabilities. Built on openJiuwen agent-core, the system implements multi-agent collaboration for query planning, information gathering, understanding and reflection, and report generation, addressing complex reasoning and research tasks.

Use cases

openJiuwen-DeepSearch delivers deep search and deep research for enterprises and consumers. This version focuses on deep research for tasks that need multi-step reasoning, multi-source validation, rigorous logic, and structured output—such as professional or high-stakes decisions.

• Financial analysis reports: Connect to local investment and finance knowledge bases and web-augmented search to plan tasks, gather information, analyze topics (e.g. “Impact of Fed rate cuts in 2025 on A-share tech”), and produce investment and finance reports.
• Academic and policy research: Use local or web-augmented engines for policies and implementation details; plan, collect, analyze, and generate reports (e.g. “Impact of China’s ‘new quality productive forces’ policy on manufacturing SMEs”).

Core capabilities

• Example-based report generation

  • Supports a given report template or extracting a template from a sample report, then generating similar reports from that template.
  • Sample reports may be Markdown, HTML, Word, PDF, etc.; templates can be exported.

• Knowledge-augmented hybrid retrieval

  • Local knowledge bases with keyword, vector, graph, and hybrid retrieval.
  • Hybrid retrieval across local knowledge and the open web.
  • Online dynamic knowledge construction, evaluation, and refinement to improve fused search quality and reduce context cost.

• Collaborative and interactive

  • Natural-language feedback during planning.
  • Collaborative revision based on user feedback.

• Segment-level provenance

  • Outputs and report content include validated citations; citations can be previewed and opened.
  • Segment-level traceability and confidence.
  • Provenance reasoning and visualization for key claims.

• Rich reports with visuals

  • Reports with figures and visualizations; content remains traceable.
  • Markdown output and export to Word, HTML, and other formats.

System architecture

The following diagram illustrates the openJiuwen-DeepSearch architecture. It is mainly built on openJiuwen agent-core and can connect to different LLMs and tools.

DeepSearch consists of a manager, query planning, information gathering, understanding and analysis, and content generation, including:

• Manager: Creates agents, orchestrates workflows, and manages configuration on the openJiuwen agent-core framework so agents coordinate efficiently.
• Query planning: Intent-based routing, structural planning, task decomposition, query rewriting, and related understanding to capture user intent and orchestrate tasks.
• Knowledge retrieval: Offline knowledge construction and online retrieval. Offline: document parsing, chunking, and multiple index types. Online: inverted keyword search, vector search, knowledge-graph search, hybrid modes, and pluggable web search.
• Understanding and analysis: Evaluates, refines, expands, and fuses retrieval results and other context.
• Result generation: Answers, report generation, interactive editing, and provenance.

openJiuwen Studio is an end-to-end AI Agent development platform from build to deploy. openJiuwen-DeepSearch is a reference agent implementation: you can manage models, tools, and knowledge in Studio, submit queries, and experience deep research and reports. openJiuwen Ops supports debugging, evaluation, observability, and tuning for agents including openJiuwen-DeepSearch.

For brevity, later sections use:

• agent-core: openJiuwen agent-core
• Studio: openJiuwen Studio
• DeepSearch: openJiuwen-DeepSearch
• Ops: openJiuwen Ops