Level-Navi Agent: An Open-Source Universal Web Search Agent Framework 192

小编 · 2025-03-02

Level-Navi Agent is an open-source web search agent framework that leverages large language models (LLMs) to break down complex queries and perform precise searches. It works with both open-source and proprietary models, decomposing and understanding complex questions before progressively retrieving information online until delivering an answer.

Level-Navi Agent: An Open-Source Universal Web Search Agent Framework

What is Level-Navi Agent?

Level-Navi Agent is an open-source web search Agent framework that leverages large language models (LLMs) to perform step-by-step analysis and precise searching of complex questions. It can adapt to any open-source or closed-source model, breaking down and understanding complex questions, then progressively searching the web for information until an answer to the user's question is provided.

Level-Navi Agent.webp

Main Features of Level-Navi Agent

Hierarchical Navigation

Problem Decomposition: Breaks down complex questions into multiple sub-questions and solves them one by one through chain-of-thought reasoning.
Parallel Search: After listing the sub-questions, searches for relevant information simultaneously, improving search efficiency.

Intelligent Planning and Problem Decomposition

Chain-of-Thought: Mimics human thinking, progressively clarifying the next piece of information to collect.
Iterative Refinement: Dynamically adjusts operations based on feedback and existing information until sufficient information is gathered to answer the question.

Information Retrieval and Feedback

Online Search: Conducts online searches for sub-questions, collects information, and feeds it back to the planner.
Result Integration: Integrates the search results and provides a clear answer.

Training-Free Universal Framework: Can be used with any open-source or closed-source large language model (LLM), with no need for task-specific training or fine-tuning.

Evaluation and Benchmark: Provides a Chinese web search benchmark dataset (Web24) along with a set of adapted evaluation metrics to comprehensively measure the performance of different LLMs in the community when performing web search tasks.

Application Scenarios of Level-Navi Agent

Complex Query Handling: Can handle complex multi-hop questions, such as "Where is the headquarters of the company that the courier who delivered the 150 billionth package this year works for?"—questions that require understanding complex logical chains and multi-level search reasoning.
Chinese Web Search: Particularly well-suited for Chinese web search tasks, providing a new Chinese web search dataset (Web24) and evaluation metrics, advancing the field of Chinese web search.
Information Retrieval and Knowledge Acquisition: Through hierarchical navigation and parallel search, it can efficiently gather information from the internet, making it suitable for applications that require rapid information acquisition and integration.
aiSearch Agent Evaluation and Optimization: Provides a fair evaluation framework that allows different large language models (LLMs) to be evaluated under the same benchmark, helping researchers and developers optimize model performance in complex search tasks.
Automated Content Generation: Combined with AI content generation technology, it can be used for automated content generation and publishing, especially in scenarios that require collecting and integrating information from multiple sources.