An open analytical structure combining depth and breadth with dual engines of thoughts and method thereof

Abstract

The present invention relates to an open analytical structure combining depth and breadth with dual engines of thoughts and method thereof. The open analytical structure involves a basic prompter, configured to clarify and refine a question input by a user (e.g. a user query); a solver agent module, electrically connected to the basic prompter and configured to solve the question and provide an answer; a dual-engine system, electrically connected to the solver agent module and configured to extend a subsequent analysis path and produce more than one analysis result; and a final response agent, electrically connected to the dual-engine system and the solver agent module, and configured to integrate the more than one analysis result and perform final content optimization and structured article output. Therefore, the dual-engine system makes reasoning more comprehensive and focused, and can flexibly structure planning, query and reasoning.

Description

Technical Field

[0001] This invention relates to the technical field of an open analysis framework, and more particularly to an open analysis structure and method that uses a dual-engine approach that combines depth and breadth. Prior Technology

[0002] In the context of global news analysis, news serves as a crucial reflection of social change, directly influencing financial markets, political decisions, and corporate strategies. In the era of big data, the ability to extract valuable insights from news information has become paramount.

[0003] However, analysis by human experts is not only inefficient and biased, but also makes it difficult to quickly extract key points from large amounts of real-time information. Furthermore, some researchers use traditional Large Language Models (LLMs) for analysis, but these models are primarily designed for single tasks and struggle with open-ended problems. For techniques known for using LLMs to analyze problems, please refer to the following published literature: CoT Wei et al., 2023; CoT-SC Wang et al., 2023; ToT Long, 2023; GoT Besta et al., 2024, etc. Summary of the Invention

[0004] The main objective of this invention is to provide an open analytical structure and method that combines depth and breadth thinking. By setting up a dual-engine system, the engine controller dynamically balances the multi-dimensional exploration of the breadth analysis engine and the in-depth analysis of the depth analysis engine, making reasoning more comprehensive and focused. Furthermore, through the setting of a basic prompter and a solver agent module, it is possible to flexibly structure planning, querying, and reasoning, and supports customized tools.

[0005] To achieve the aforementioned objectives, this invention provides an open analysis structure based on dual-engine thinking, comprising: a basic prompter configured to clarify and optimize a user-input question; a solution agent module electrically connected to the basic prompter and configured to solve the question and provide an answer; a dual-engine system electrically connected to the solution agent module and configured to extend a subsequent analysis path and generate a deep analysis result question generated by deep analysis and a breadth analysis question generated by breadth analysis, and return them to the solution agent module for iterative processing until one of the termination conditions is met to produce an analysis result; and a final response agent electrically connected to the dual-engine system and the solution agent module and configured to perform final content optimization and structured article output on the analysis result.

[0006] In some embodiments, the solver module includes a planner, an executor, and an analysis data toolbox. The planner is electrically connected to the base prompter, the executor is electrically connected to the planner and the dual-engine system, and the analysis data toolbox is electrically connected to the planner and the executor.

[0007] In some embodiments, the planner is configured to perform a task decomposition, an allocation tool, and a plan verification; the executor is configured to perform a task execution, an information integration, and an execution verification; and the analytics toolkit is configured to include multiple agent tools for use and support customization requirements.

[0008] In some embodiments, the dual-engine system includes an engine controller, a breadth analytics engine, and a depth analytics engine. The engine controller is electrically connected to the actuator and the final response agent. The breadth analytics engine is electrically connected to the engine controller, and the depth analytics engine is electrically connected to the engine controller.

[0009] In some embodiments, the engine controller is configured to receive information from the actuator to complete the information integration in order to determine whether to transmit to the breadth analysis engine or the depth analysis engine. The breadth analysis engine is configured to perform a multi-dimensional exploration to ensure that the analysis covers a wide range of possibilities and background knowledge, and the depth analysis engine is configured to perform in-depth reasoning analysis to ensure that the analysis results have logical rigor and high resolution.

[0010] This invention also provides an open analysis method based on dual-engine thinking, comprising: an input step: inputting a user query question; an optimization step: optimizing the user query question and transforming it into a structured analysis question; a solution step: decomposing the structured analysis question and ensuring the completeness and accuracy of the reasoning through a rigorous verification mechanism to provide a preliminary answer to the user query question; a strategy selection step: based on the preliminary answer, deciding whether to perform a breadth analysis step, a depth analysis step, or a termination analysis step, thereby producing a depth analysis question generated by completing the depth analysis step and a breadth analysis question generated by completing the breadth analysis step, and then executing the solution step again for iterative processing until a termination condition is met and the termination analysis step is executed to produce an analysis result; and a final response step: performing final content optimization and structured article output on the analysis result and outputting a final answer.

[0011] In some embodiments, the optimization step includes a ambiguity refinement step, a time addition step, a clarification step, a name unification step, and an error correction step. The ambiguity refinement step improves the ambiguity of the user query to make it more resolvable. The time addition step ensures that an execution time includes an appropriate time range to improve timeliness. The clarification step standardizes the terms and concepts in the user query to avoid misinterpretation. The name unification step unifies the names of terms and concepts to ensure consistency and avoid incorrect matching due to name differences. The error correction step checks the correctness of the user query and makes corrections accordingly.

[0012] In some embodiments, the optimization step further includes an error handling step, which analyzes a reason for the failure of the user query and provides a correction suggestion, while preserving the core intent of the user query to ensure that the user query still meets the user's needs.

[0013] In some embodiments, the solution step includes a task planning step and a plan verification step; wherein, the task planning step decomposes the user query into 1 to 3 core sub-tasks through a multi-agent system, matches a suitable analysis tool module, generates executable inputs according to an analysis data toolbox, and ensures that the dependencies and execution order among these core sub-tasks form a cohesive analysis process; the plan verification step includes an integrity check, a redundancy check, and a format check; wherein, the integrity check ensures that all key information in the user query is covered and that the allocation of these core sub-tasks is correct, the redundancy check eliminates duplicate or unnecessary inclusion relationships to improve efficiency, and the format check ensures that these core sub-tasks conform to a standard format to avoid execution errors.

[0014] In some embodiments, the solution step further includes a task execution step and a result verification step; wherein, the task execution step ensures that each core subtask is executed only after a prerequisite is met based on the dependency relationship between the core subtasks, and uses the analysis tool module in the analysis data toolbox to solve the core subtasks, and integrates all output results to form a structured summary to provide a basis for subsequent analysis; the result verification step performs a task verification and a summary verification, the task verification ensures that the output results of each core subtask are consistent with the facts and records the parts to be corrected, and the summary verification ensures that each output result is consistent with the user's query question without logical contradictions or erroneous information.

[0015] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the following detailed description is provided in conjunction with the specific embodiments listed in the figures. Simple Explanation of the Diagram

[0016] Figure 1 is a block diagram of the open analysis structure of the dual-engine thinking of this invention. Figure 2 is a flowchart illustrating the open analysis method based on the dual-engine thinking of this invention. Figure 3 is a schematic diagram of the thinking tree of the dual-engine system in the open analysis method of dual-engine thinking of the present invention. Figures 4 to 9 are schematic diagrams illustrating embodiments of the open analysis method of the dual-engine thinking of this invention. Figure 10 is a schematic diagram showing the comparison results between the open analysis method of the dual-engine thinking of this invention and the five major areas and five major standards of GPT-4o. Figure 11 is a histogram diagram based on the five major areas in Figure 10. Figure 12 is a histogram diagram based on the five standards in Figure 10. Figure 13 is a schematic diagram showing the comparison results between the open analysis method of the dual-engine thinking of this invention and the five domains and five standards of Perplexity AI (llama-3.1-sonar-small-128k-online). Figure 14 is a histogram diagram based on the five major areas in Figure 13. Figure 15 is a histogram diagram based on the five standards in Figure 13. Implementation

[0017] The following description, in conjunction with the accompanying drawings and component symbols, provides a more detailed account of the embodiments of the present invention, so that those skilled in the art can implement them after studying this specification.

[0018] Figure 1 is a block diagram of the open analysis structure of the dual-engine thinking of this invention.

[0019] Referring to Figure 1, the open analysis structure 10 of the dual-engine thinking of this invention includes a basic prompter 100, a solution agent module 200, a dual-engine system 300, and a final response agent 400. The dual-engine thinking is implemented by the dual-engine system 300.

[0020] The Base Prompter 100 can be configured to clarify and optimize a user-input question (i.e., a user query question). In other words, the Base Prompter 100 is responsible for optimizing (or refining) the question input and transforming it into a structured analytical query question, providing accuracy and parsability.

[0021] The Solver Agent module 200 can be electrically connected to the base prompter 100 and configured to solve the problem and provide an answer. In some embodiments, the Solver Agent module 200 may include a Planner 210, an Executor 220, and an Analysis ToolBox 230. The Planner 210 can be electrically connected to the base prompter 100 and can be configured to perform task decomposition, an allocation tool, and a plan verification. The Executor 220 can be electrically connected to the Planner 210 and the dual-engine system 300 and can be configured to perform task execution, information integration, and execution verification. The Analysis ToolBox 230 can be electrically connected to the Planner 210 and the Executor 220 and can be configured to include multiple agent tool modules for use and support customization needs. In some embodiments, the multiple proxy tool modules of the data analysis toolbox 230 may be a news search tool module, an event extraction tool module, a historical analysis tool module, an information searching tool module, and a direct reasoning tool module. The news search module utilizes Perplexity AI (Perplexity's AI search engine, llama-3.1-sonar-small-128k-online) to retrieve and process the latest news articles, playing a crucial role in providing real-time information and data. The event extraction module uses Large Language Modeling (LLM) to analyze time series and causal relationships, constructing detailed event timelines and extracting essential information about key entities and their roles for subsequent analysis. The historical analysis module uses LLM to match current events with historical precedents, identifying key factors influencing past events and helping to better predict and explain potential future outcomes. The information retrieval module is supported by real-time data from Perplexity AI, filling knowledge gaps and improving the comprehensiveness and accuracy of the analysis. The direct reasoning module integrates the outputs of various analysis tools through LLM, transforming raw information into valuable insights through logical reasoning and inductive analysis, and providing directions for further exploration.

[0022] The Dual-Engine System (DEoT) 300 is electrically connected to the solver agent module 200 and configured to extend a subsequent analysis path and produce a depth analysis result problem generated by depth analysis and a breadth analysis problem generated by breadth analysis. These results are then fed back to the solver agent module for iterative processing until one of the termination conditions is met, producing an analysis result. In some embodiments, the Dual-Engine System 300 may include an Engine Controller 310, a Breadth Engine 320, and a Depth Engine 330. Engine controller 310 is electrically connected to actuator 220 and final response agent 400 and can be configured to receive information from actuator 220 after information integration to determine whether to transmit to breadth analysis engine 320 or depth analysis engine 330; that is, engine controller 310 is responsible for deciding when to perform a breadth analysis by breadth analysis engine 320, a depth analysis by depth analysis engine 320, or a termination analysis, while the large language model (LLM) can be dynamically adjusted according to complexity, progress, problem-solving, and resource usage. Breadth analysis engine 320 is electrically connected to engine controller 310 and can be configured to perform multi-dimensional exploration to ensure that the analysis covers a wide range of possibilities and background knowledge; that is, breadth analysis engine 320 can generate multiple key questions from different perspectives (e.g., economic, political, technological, etc.) to help establish a complete background for the problem and generate corresponding analytical queries. The deep analysis engine 330 is electrically connected to the engine controller 310 and can be configured to perform in-depth reasoning analysis to ensure the logical rigor and high resolution of the analysis results. That is, the deep analysis engine 330 can generate a question for further in-depth analysis, delve into key issues, and provide more detailed insights and conclusions through multi-level reasoning. This dual-engine system architecture achieves a balance between comprehensive exploration and detailed analysis, providing efficient and hierarchical analysis results.

[0023] The final response agent 400 can be electrically connected to the dual-engine system 300 (i.e., actuator 220) and the solver agent module 200 and configured to perform final content optimization and structured article output on the analysis results.

[0024] Figure 2 is a flowchart illustrating the open analysis method based on the dual-engine thinking of this invention.

[0025] Please refer to Figure 2. The open analysis method S100 of the dual-engine thinking of the present invention may include steps S110, S120, S130, S140 and S150, which will be described in detail below in correspondence with the open analysis structure 10 shown in Figure 1 above.

[0026] Step S110 is an input step. In step S110, a user query is input.

[0027] Step S120 is an optimization step. In step S120, the user query question can be optimized and transformed into a structured analysis question through the basic prompter 100 shown in FIG1. ​​In some embodiments, the optimization step S120 may include a ambiguity refinement step S121, a time addition step S122, an ambiguity clarification step S123, a name unification step S124, and an error correction step S125. The ambiguity refinement step S121 improves the ambiguity of the user question to make it more resolvable; the time addition step S122 ensures that an execution time includes an appropriate time range to improve timeliness; the ambiguity clarification step S123 standardizes the terminology and concepts in the user query question to avoid misinterpretation; the name unification step S124 unifies the names of terms and concepts to ensure consistency and avoid incorrect matching due to name differences; and the error correction step S125 checks the correctness of the user query question and makes corrections accordingly.

[0028] In some embodiments, the optimization step S120 may further include an error handling step S126. The error handling step S126 analyzes a reason for the failure of the user's query and provides a correction suggestion, while preserving the core intent of the user's query to ensure that the user's query still meets the user's needs.

[0029] Step S130 is a solution step. In step S130, the structured analysis problem can be decomposed using the solution agent module 200 shown in FIG1, and a rigorous verification mechanism can be used to ensure the completeness and accuracy of the reasoning to provide a preliminary answer to the user's query question. In some embodiments, the solution step S130 may include a task planning step S131 and a plan validation step S132. The task planning step S131 can use the planner 210 shown in FIG1 to decompose the user's query question into 1 to 3 core sub-tasks through a multi-agent system (MAS), match a suitable analysis tool module, generate executable inputs according to an analysis data toolbox, and ensure that the dependencies and execution order between these core sub-tasks form a combined analysis process. The plan verification step S132 can be completed by the planner 210, and may include a completeness check, a redundancy check, and a format check. The completeness check ensures that all key information in the user query is covered and that the allocation of core subtasks is correct. The redundancy check eliminates duplicate or unnecessary inclusion relationships to improve efficiency. The format check ensures that the core subtasks conform to a standard format to avoid execution errors. If the plan verification step S132 fails, plan regeneration is initiated (i.e., returning to step S131 and re-executing the planner 210), providing specific feedback to correct the user query while retaining the original analysis objectives, ensuring that the corrected plan still conforms to the expected analysis direction.

[0030] In some embodiments, the solution step S130 may further include a task execution step S133 and a result validation step S134. The task execution step S133 uses the executor 220 shown in FIG1 to ensure that a prerequisite is met before executing each corresponding core subtask based on the dependencies between these core subtasks. It then uses the analysis tool module in the analysis toolbox 230 to solve these core subtasks and integrates all output results to form a structured summary to provide a basis for subsequent analysis. The result validation step S134 performs two levels of validation: task validation and summary validation. The task validation ensures that the output results of each core subtask are consistent with the facts and records the parts to be corrected. The summary validation ensures that each output result is consistent with the user's query question without logical contradictions or errors. If the result validation step S134 fails, plan regeneration is initiated (i.e., returning to step S131 and executing the planner 210 again), providing specific feedback, correcting the problem, and preserving the original analysis objective.

[0031] Step S140 is a strategy selection step. In step S140, the engine controller 310 shown in FIG1 determines, based on the preliminary answer, whether to perform a breadth analysis step S141 of the breadth analysis engine 320 shown in FIG1, a depth analysis step S142 of the depth analysis engine 330 shown in FIG1, or a termination analysis step S143. This produces a depth analysis problem generated by the depth analysis (i.e., depth analysis step S142) and a breadth analysis problem generated by the breadth analysis (i.e., breadth analysis step S141). The solution step (i.e., step S130) is then executed again (e.g., the depth analysis problem and the breadth analysis problem are passed back to the solution agent module 200) for iterative processing until one of the termination conditions (i.e., termination analysis step S143) is met to produce an analysis result.

[0032] Step S150 is a final response step. In step S150, the final response agent 400, as shown in Figure 1, performs final content optimization and structured article output on the analysis results, and outputs a final answer.

[0033] Figure 3 is a schematic diagram of the thinking tree of the dual-engine system in the open analysis method of dual-engine thinking of the present invention. Figures 4 to 9 are schematic diagrams illustrating embodiments of the open analysis method of dual-engine thinking of the present invention.

[0034] Please refer to Figure 3. The thought process tree diagram of the dual-engine system 300 includes: Please refer to Figures 2, 3, and 4. The first level L1 (i.e., Initial Processing) receives the preliminary answer processed by steps S110, S120, S130, and the basic prompter 100 and the solver agent module 200; Please refer to Figures 2, 3, and 5. The second level L2 (i.e., Engine Selection) determines the next analysis direction through the engine controller 310. This next analysis direction includes the breadth analysis step S141 performed by the breadth analysis engine 320, the depth analysis step S142 performed by the depth analysis engine 330, and the direct termination analysis step S143, thereby producing more than one analysis result; and Please refer to Figures 2, 3, and Figures 6 to 8. The third level L3 (i.e., Iterative Processing) The analysis results generated by the processing module 200 are used for the next round of analysis. Each pair of [question Q - answer A] is a node Ni,j, forming an analysis structure. Here, N represents the number of nodes, i represents the number of levels (i.e., L1, L2, L3), and j represents the number of pairs of [question Q - answer A]. The analysis continues until one of the termination conditions is met. The termination conditions include: (1) reaching the maximum level depth, (2) reaching the upper limit of the number of nodes, and (3) all analysis objectives have been met. Please refer to Figure 9 to collect all node information, generate a complete report, and respond to the user's questions in a structured format (i.e., the final output (Final Response)).

[0035] Figure 10 is a schematic diagram comparing the five domains and five criteria of the open analysis method of the dual-engine thinking of this invention with GPT-4o (Generative Pre-trained Transformer 4 Omni). Figure 11 is a histogram diagram based on the five domains in Figure 10. Figure 12 is a histogram diagram based on the five criteria in Figure 10. Among them, GPT-4o is an advanced general-purpose LLM, known for its powerful natural language processing and logical reasoning capabilities.

[0036] In Figure 10, the horizontal content section selects five core areas to test reasoning ability in different scenarios, while the vertical content section selects five evaluation criteria for comparison. The five core areas include Biomedicine (including medical research, pharmaceuticals, biotechnology, health innovation, etc.), Economics (including finance, markets, trade, business, and economic news), Geopolitics (including international relations, political events, conflicts, and the impact of policies on the market, etc.), Industry (including manufacturing, energy, transportation, construction, and other industrial sectors), and Technology (including technological progress, IT industry, hardware and software, artificial intelligence, etc.). 100 questions are selected in each area, totaling 500 open-ended questions, providing complete comparative data for the analytical performance of the open-ended analysis structure 10 and its analysis method S100 of this invention. The five evaluation criteria include: Analytical Depth (measuring the model's deep understanding of the logical relationships, influencing factors, and key issues of the problem); Specific Arguments (assessing whether the model can effectively retrieve and utilize relevant data, cases, or historical analogies to support its arguments); Innovation (measuring whether the model can provide unique insights and innovative solutions in real-world applications, demonstrating creativity); Practicality (verifying the feasibility of the conclusions and considering their operability in real-world applications, rather than merely theoretical value); and Logical Coherence (verifying the structural integrity and internal consistency of the model's reasoning process, ensuring clear logical arguments). To avoid order bias, a dual comparison test is used. Each response undergoes two rounds of independent comparison with the input order swapped, and each round is scored independently. The evaluation encompasses the five evaluation criteria and the overall content, employing a six-item binary decision system. Finally, the win rate is calculated.

[0037] Referring to Figures 10 and 11, the open-ended analytical structure 10 and its analytical method S100 of the dual-engine thinking method of this invention achieve a total win rate of 85.5% across all test domains, with individual scores exceeding 81.5% in each domain. It performs exceptionally well in the geopolitics domain, achieving an 89% win rate, demonstrating its advantages in international relations analysis, policy evaluation, and multi-level reasoning. These evaluation results indicate that the open-ended analytical structure 10 and its analytical method S100 of the dual-engine thinking method of this invention can effectively integrate information and clarify causal relationships, providing more in-depth analysis and reasoning results. Referring again to Figures 10 and 12, the open-ended analytical structure 10 and its analytical method S100 of this invention demonstrate excellent performance in analytical depth, achieving a win rate of 92.5%, showcasing its powerful multi-level reasoning ability. It outperforms GPT-4o in logical coherence (92.5%) and innovation (70.6%), demonstrating its competitive advantage in structured reasoning for complex problems. The lower win rate of 22.9% in practicality may be due to the higher depth of reasoning and the more academic or theoretical answers, rather than providing a decision solution directly in a short time.

[0038] Figure 13 is a schematic diagram comparing the open analysis method of the dual-engine thinking of this invention with the five domains and five standards of Perplexity AI. Figure 14 is a histogram diagram based on the five domains in Figure 13. Figure 15 is a histogram diagram based on the five standards in Figure 13. Perplexity AI (llama-3.1-sonar-small-128k-online) is an LLM that combines real-time retrieval, specializing in information synthesis and rapid response generation.

[0039] In Figure 13, the horizontal content section selects five core areas to test reasoning ability in different scenarios, while the vertical content section selects five evaluation criteria for comparison. The five core areas include Biomedicine (including medical research, pharmaceuticals, biotechnology, health innovation, etc.), Economics (including finance, markets, trade, business, and economic news), Geopolitics (including international relations, political events, conflicts, and the impact of policies on the market, etc.), Industry (including manufacturing, energy, transportation, construction, and other industrial sectors), and Technology (including technological progress, IT industry, hardware and software, artificial intelligence, etc.). 100 questions are selected in each area, totaling 500 open-ended questions, providing complete comparative data for the analytical performance of the open-ended analysis structure 10 and its analysis method S100 of this invention. The five evaluation criteria include: Analytical Depth (measuring the model's deep understanding of the logical relationships, influencing factors, and key issues of the problem); Specific Arguments (assessing whether the model can effectively retrieve and utilize relevant data, cases, or historical analogies to support its arguments); Innovation (measuring whether the model can provide unique insights and innovative solutions in real-world applications, demonstrating creativity); Practicality (verifying the feasibility of the conclusions and considering their operability in real-world applications, rather than merely theoretical value); and Logical Coherence (verifying the structural integrity and internal consistency of the model's reasoning process, ensuring clear logical arguments). To avoid order bias, a dual comparison test is used. Each response undergoes two rounds of independent comparison with the input order swapped, and each round is scored independently. The evaluation encompasses the five evaluation criteria and the overall content, employing a six-item binary decision system. Finally, the win rate is calculated.

[0040] Referring to Figures 13 and 14, the overall win rate of the open analysis structure 10 and its analysis method S100 of the dual-engine thinking method of this invention is 77.2%, slightly lower than the win rate against GPT-4o. This may be attributed to the fact that Perplexity AI (llama-3.1-sonar-small-128k-online) has real-time retrieval capabilities, giving it an advantage in N2Q (News-to-Question Dataset, an open-ended question set automatically generated based on the latest news content). In the economics field, it achieved a win rate of 83.0%, demonstrating excellent performance and showing superiority over Perplexity AI (llama-3.1-sonar-small-128k-online) in market dynamics analysis, policy impact assessment, and economic modeling. The above evaluation results indicate that the open analysis structure 10 and its analysis method S100 of the dual-engine thinking method of this invention can more effectively analyze the fluctuating economic environment and provide structured explanations and inferences. Referring to Figures 13 and 15, the open-ended analytical structure 10 and its analytical method S100 of this invention demonstrate outstanding performance in innovation (91.9%), primarily benefiting from its dual-engine architecture which promotes more creative and diverse reasoning. It surpasses Perplexity AI (llama-3.1-sonar-small-128k-online) in analytical depth (85.7%) and logical coherence (74.5%), further demonstrating its advantages in multi-level reasoning and structured thinking. Its performance in specific arguments (3.5%) is poor, possibly because Perplexity AI relies on instant retrieval, giving it an advantage in evidence support and detail richness.

[0041] In summary, the dual-engine thinking open analysis structure 10 and its analysis method S100 of this invention, through the setting of a dual-engine system, dynamically balance the multi-dimensional exploration of the breadth analysis engine and the in-depth analysis of the depth analysis engine through the engine controller, making reasoning more comprehensive and focused. Furthermore, through the setting of the basic prompter and the solver agent module, it can flexibly structure planning, querying and reasoning, and support customized tools.

[0042] The above description is merely for explaining preferred embodiments of the present invention and is not intended to limit the present invention in any way. Therefore, any modifications or changes made to the present invention under the same inventive spirit should still be included within the scope of protection intended by the present invention.

[0043] 10: Open Analytical Structure of Dual-Engine Thinking

[0044] 100: Basic Prompt

[0045] 200: Solving the proxy module

[0046] 210: Planner

[0047] 220: Actuator

[0048] 230: Data Analysis Toolbox

[0049] 300: Dual-engine system

[0050] 310: Engine Controller

[0051] 320: Breadth-of-Size Analysis Engine

[0052] 330: Deep Analysis Engine

[0053] 400: Final Response Proxy

[0054] L1: First level

[0055] L2: Second level

[0056] L3: Third level

[0057] S100: An Open-Ended Analytical Approach Based on Dual-Engine Thinking

[0058] S110: Steps

[0059] S120: Steps

[0060] S121: Steps

[0061] S122: Steps

[0062] S123: Steps

[0063] S124: Steps

[0064] S125: Steps

[0065] S126: Steps

[0066] S130: Steps

[0067] S131: Steps

[0068] S132: Steps

[0069] S133: Steps

[0070] S134: Steps

[0071] S140: Steps

[0072] S141: Steps

[0073] S142: Steps

[0074] S143: Steps

[0075] S150: Steps

Claims

1. An open analytical structure based on dual-engine thinking, including: A basic prompter, configured to clarify and optimize a user input for a problem; A solver agent module, electrically connected to the basic prompter and configured to solve the problem and provide an answer; a dual-engine system, electrically connected to the solver agent module and configured to extend a subsequent analysis path and produce a depth analysis result problem generated by depth analysis and a breadth analysis problem generated by breadth analysis, and return them to the solver agent module for iterative processing until one of the termination conditions is met to produce an analysis result; and a final response agent, electrically connected to the dual-engine system and the solver agent module and configured to perform final content optimization and structured article output on the analysis result.

2. The open analytical structure of dual-engine thinking as described in Request 1, wherein, The solver module includes a planner, an executor, and an analysis data toolbox. The planner is electrically connected to the basic prompter, the executor is electrically connected to the planner and the dual-engine system, and the analysis data toolbox is electrically connected to the planner and the executor.

3. The open-ended analytical structure of dual-engine thinking as described in claim 2, wherein, The planner is configured to perform task decomposition, allocation, and plan verification; the executor is configured to perform task execution, information integration, and execution verification; and the analytics toolkit is configured to include multiple proxy tools and support customization.

4. The open analytical structure of dual-engine thinking as described in claim 3, wherein, The dual-engine system includes an engine controller, a breadth analysis engine, and a depth analysis engine. The engine controller is electrically connected to the actuator and the final response agent. The breadth analysis engine is electrically connected to the engine controller, and the depth analysis engine is electrically connected to the engine controller.

5. The open analytical structure of dual-engine thinking as described in claim 4, wherein, The engine controller is configured to receive information from the actuator to complete the information integration in order to determine whether to transmit it to the breadth analysis engine or the depth analysis engine. The breadth analysis engine is configured to perform a multi-dimensional exploration to ensure that the analysis covers a wide range of possibilities and background knowledge. The depth analysis engine is configured to perform in-depth reasoning analysis to ensure that the analysis results have logical rigor and high resolution.

6. An open-ended analytical method based on dual-engine thinking, including: Input steps: Enter a user's query question; The process involves: an optimization step, optimizing the user query and transforming it into a structured analysis problem; a solution step, breaking down the structured analysis problem and using a rigorous verification mechanism to ensure the completeness and accuracy of the reasoning to provide a preliminary answer to the user query; a strategy selection step, based on the preliminary answer, deciding whether to perform a breadth analysis step, a depth analysis step, or a termination analysis step, thereby generating a depth analysis problem generated by completing the depth analysis step and a breadth analysis problem generated by completing the breadth analysis step, and then executing the solution step again for iterative processing until a termination condition is met, at which point the termination analysis step is executed to produce an analysis result; and a final response step, performing final content optimization and structured article output on the analysis result, and outputting a final answer.

7. The open-ended analysis method based on dual-engine thinking as described in claim 6, wherein, The optimization process includes a ambiguity refinement step, a time-addition step, a clarification step, a naming standardization step, and an error correction step. The ambiguity refinement step improves the ambiguity of the user query, making it more resolvable. The time-addition step ensures that the execution time includes an appropriate time range to improve timeliness. The clarification step standardizes the terminology and concepts in the user query to avoid misinterpretation. The naming standardization step unifies the names of terms and concepts to ensure consistency and avoid incorrect matching due to name differences. The error correction step checks the correctness of the user query and makes corrections accordingly.

8. The open-ended analysis method based on dual-engine thinking as described in claim 7, wherein, The optimization process also includes an error handling step, which analyzes the reasons for the failure of the user's query and provides a correction suggestion, while preserving the core intent of the user's query to ensure that the user's query still meets the user's needs.

9. The open-ended analysis method based on dual-engine thinking as described in claim 8, wherein, The solution process includes a task planning step and a plan verification step. The task planning step uses a multi-agent system to break down the user query into 1-3 core sub-tasks, matches them with a suitable analysis tool module, and generates executable inputs based on an analysis data toolkit. It also ensures that the dependencies and execution order among these core sub-tasks form a cohesive analysis flow. The plan verification step includes a completeness check, a redundancy check, and a format check. The completeness check ensures that all key information in the user query is covered and that the allocation of core sub-tasks is correct. The redundancy check eliminates duplicate or unnecessary inclusion relationships to improve efficiency. The format check ensures that the core sub-tasks conform to a standard format to avoid execution errors.

10. The open-ended analysis method based on dual-engine thinking as described in claim 9, wherein, The solution process also includes a task execution step and a result verification step. The task execution step ensures that each core subtask is executed only after a prerequisite is met, based on the dependencies between the core subtasks. It uses the analysis tool module in the analysis toolbox to solve the core subtasks and integrates all outputs to form a structured summary to provide a basis for subsequent analysis. The result verification step performs a task verification and a summary verification. The task verification ensures that the outputs of each core subtask are factual and records any parts requiring correction. The summary verification ensures that the outputs are consistent with the user's query and contain no logical contradictions or errors.

Images