Decision chain software interaction method and system based on large language model

Abstract

The invention discloses a decision chain software interaction method based on a large language model, which comprises the following steps: S1, decision plan generation: generating a plurality of decision templates by inputting prompts to a user and calling the large language model, and forming a decision path supporting multi-step reasoning; s2, decision path revision: instantiating the decision planning path according to a human-in-loop mode to form a decision path which has operation and editing capabilities and is supported by various data, and revising a sub-path or path branch relationship of the current decision path to form a dynamic decision path application mode; and S3, decision content recommendation: calling and generating a data display sub-model and a module, performing adaptive recommendation to a comprehensive display visual large screen, avoiding the problem of low efficiency caused by calling a large model in each step, solving the problems of low reliability and the like of traditional thinking chain reasoning through a decision chain template and a feedback link, and improving the reasoning efficiency of the thinking chain. And the decision chain software interaction mode based on the large language model, the customization of the system and the systematic design are realized.

Description

Technical Field

[0001] The present invention relates to a large language model thinking chain technology, and in particular to a decision chain software interaction method and system based on a large language model. Background Art

[0002] In the general large model reasoning process, thought chains are used to improve the performance of large models in reasonable reasoning tasks. In actual large language model applications, reasoning and multi-step thinking problems are involved. Usually, a collection of a series of thinking steps is called a thought chain.

[0003] Common thinking chains include input, reasoning chain, and output, which are suitable for some general routine reasoning tasks. At present, thinking chains generally include three forms: 0-sample thinking chains do not need to learn similar problems in advance, and solve problems directly through thinking, but are completely limited by the self-generation and understanding capabilities of large language models; few-sample thinking chains learn by analogy through a few examples, but the quality of samples and large model capabilities are bottlenecks for reasoning performance; self-consistent thinking chains are close to human thinking, and can self-question and correct, but lack knowledge in specific closed fields, and still need to be customized and improved according to actual needs. In addition, in the era of intelligence, the cost of manually constructing thinking chains is high, and it is not easy to dynamically optimize and migrate to multiple closed fields; the quality of automatically constructed thinking chains is relatively poor, overly dependent on given reasoning processes, and inflexible, especially in areas limited by specific knowledge, and the ability of multi-step reasoning and decision-making assistance is insufficient. Summary of the invention

[0004] The purpose of the present invention is to provide a decision chain software interaction method and system based on a large language model, which has strong reasoning ability and high efficiency, does not rely on a given reasoning process, and can be flexibly applied to various knowledge-limited fields.

[0005] The present invention is achieved through the following technical solutions:

[0006] A decision chain software interaction method based on a large language model comprises the following steps:

[0007] S1: Decision planning generation: Based on the text knowledge base of specific field cases, the original decision case data base is realized, a decision chain database is formed, and a decision chain template is constructed. By prompting the user input and calling the large language model, a variety of decision templates are generated to form a decision path that supports multi-step reasoning;

[0008] S2: Decision path revision: Based on the decision path generated by the decision planning generation part, the decision planning path is instantiated in a human-in-the-loop manner to form a decision path supported by multiple data with operation and editing capabilities. If the decision path does not meet the actual needs of human feedback, the sub-path or path branch relationship of the current decision path is revised by calling the background template to form a dynamic decision path application mode;

[0009] S3: Decision content recommendation: Form the final record of the decision path, and call the generated data display sub-model and module based on the data and sub-model process of step-by-step decision reasoning, and adaptively recommend it to the comprehensive display visualization large screen.

[0010] Preferably, the instantiation of the decision planning path in S2 specifically includes the display of geographic information data aggregation, identification, etc. required to complete a specific decision scenario, structured data, chart retrieval display, and unstructured data call display.

[0011] Preferably, S3 also includes feeding back the final visualized data to the big model, optimizing the decision-making link, and forming a user-oriented big model-driven data-decision-recommendation-visualization full-link application mode.

[0012] A decision chain software interaction system based on a large language model, including three parts: large language model operation interaction, large visualization screen and background management;

[0013] The large language model operation interaction part is used to call the large language model to generate a decision template and a decision path through user input prompts, and to generate a final recommended decision path after instantiating and revising the decision path. The layout of the large language model operation interaction part is a left-middle-right three-column structure, with the large language model software interaction generation part on the left, the decision business process editing part of the operator in the middle, and the recommendation display part of the large model on the right;

[0014] The visualization large screen part is used to comprehensively display the recommended data generated by the large language model operation interaction part, the decision sub-function model blocks and the data required for the decision task;

[0015] The background management part uniformly manages the data required for the visualization large screen and the decision chain model template, and realizes the technical route of the decision chain software interaction system based on the large language model.

[0016] Furthermore, the interactive generation part of the large language model software utilizes the inherent planning ability of the large language model to form decision generation supported by a domain-specific decision chain data set driven by the large language model through prompting user input.

[0017] Furthermore, the purpose of the editing part of the decision-making business process of the operator is to complete the revision of the reasoning path, that is, the in-loop correction of the decision chain, specifically to instantiate the decision planning path, including the aggregation of geographic information data required for specific decision-making scenarios, identification and other displays, structured data, chart retrieval display and unstructured data call display, to form a decision path supported by multiple data with operation and editing capabilities.

[0018] Furthermore, the recommendation display part of the large model includes decision content generation and visualization recommendation, and supports being pushed to the display interface of the large-screen visualization part through block combination.

[0019] Furthermore, the background management part includes four layers: data, model, interaction and visualization. The data layer includes the original data of decision cases, decision templates based on the original data, and other relevant precise structured data that can be used to supplement the needs of decision-making tasks. The model layer includes a large language model and a series of small decision models supported by precise data. The interaction layer calls the large model, decision templates and expert knowledge to complete the generation and revision of decision paths. The visualization layer reorganizes and constructs the data and decision module recommendations generated by the interaction layer to form visual data feedback, which feeds back to optimize the large model and strategy chain.

[0020] The present invention has the following beneficial effects:

[0021] (1) The present invention avoids the inefficiency problem of calling a large model at each step. Through the decision chain template and feedback link, it solves the problem of low reliability of traditional thinking chain reasoning, and realizes the decision chain software interaction mode based on the large language model, and the customization and systematic design and implementation of the system.

[0022] (2) The decision chain software interaction system based on a large language model provided by the present invention overcomes the problems of high cost, poor quality, difficulty in optimization and expansion and migration of the artificial construction of the thinking chain in the traditional large language model system, excessive reliance on a given reasoning process, and inflexible reasoning. In a limited field where specific knowledge is closed, the reasoning ability of the decision chain software interaction system of the present application is greatly improved, and through feedback on visualized data, the large model and strategy chain are optimized, so that the reasoning ability of the system is continuously optimized and strengthened, and the reasoning results are better. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a flowchart of the decision chain software interaction method based on the large language model of the present invention;

[0024] Figure 2 It is a system block diagram of the decision chain software interaction system based on the large language model of the present invention;

[0025] Figure 3 It is a schematic diagram of the background management part in the decision chain software interaction system based on the large language model of the present invention. DETAILED DESCRIPTION

[0026] The present invention is further described in detail below in conjunction with specific embodiments, which are intended to explain the present invention rather than to limit it.

[0027] refer to Figure 1As shown, the present invention provides a decision chain software interaction method based on a large language model, comprising:

[0028] S1: Decision planning generation: It includes online and offline parts. The offline part implements the original decision case data base based on the text knowledge base of specific domain cases to form a decision chain database and support the construction of decision chain templates. The online part generates multiple decision templates through user input prompts and calls the large language model to form a decision path, support the decision path of multi-step reasoning, and form a decision generation supported by a domain-specific decision chain data set driven by a large language model.

[0029] S2: Decision path revision: Based on the decision path generated by the decision planning generation part, the decision planning path is instantiated in a human-in-the-loop manner, including the display of geographic information data aggregation, identification, etc. required to complete specific decision scenarios, structured data, chart retrieval display, unstructured data call display, and the ability to operate and edit decision paths supported by multiple data. If the decision path does not meet the actual needs of human feedback, the sub-paths or path branch relationships of the current decision path are revised by calling the background template to form a dynamic decision path application mode;

[0030] S3: Decision content recommendation: Form the final record of the decision path, and call the generated data display sub-model and module according to the data and sub-model process of step-by-step decision reasoning, and adaptively recommend it to the comprehensive display visualization screen. The visualized data feeds back to the big model, optimizes the decision link, and forms a user-oriented big model-driven data-decision-recommendation-visualization full-link application mode.

[0031] refer to Figure 2 A decision chain software interaction system based on a large language model is shown, including three parts: large language model operation interaction, large visualization screen and background management.

[0032] The layout of the large language model operation interaction part is divided into a three-column structure of left-middle-right. The left side is the large language model software interaction generation part, the middle is the operator's decision-making business process editing part, and the right side is the recommendation display part of the large model. It supports pushing to the large-screen visual display interface through block combination.

[0033] The large language model software interactive generation part on the left uses the inherent planning ability of the large language model to generate multiple decision paths through prompts for user input, support multi-step reasoning decision paths, and form a large language model-driven domain-specific decision chain data set supported by decision generation, overcoming the problem of simple stereotyped thinking links in traditional thinking chains that lack domain knowledge.

[0034] The purpose of the decision-making business process editing part of the intermediate operator is to complete the revision of the reasoning path, that is, the decision chain person in the loop correction, using geographic information data as the base map, to realize on-map operations, comprehensive on-map display, specifically to instantiate the decision planning path, including the aggregation of geographic information data required for specific decision-making scenarios, identification and other displays, structured data, chart retrieval display and unstructured data call display, to form a decision path supported by multiple data with operation and editing capabilities.

[0035] The recommendation display part of the large model on the right includes decision content generation and visualization recommendations. It supports pushing to the large-screen visualization display interface through block combination, which is presented as a comprehensive recommendation display of data and models. Finally, the visualization data feeds back to the large model to optimize the decision-making chain.

[0036] For the visualization large screen part, it can comprehensively display the data interactively generated by the large language model for recommendations, the decision sub-function model blocks and the data required for the decision-making task. The decision sub-function model blocks are in the form of a combination of data, charts, etc.

[0037] For the backend management part, the data required for the visualization large screen and the decision chain model template can be managed in a unified manner, realizing the decision chain software interaction mode and system technical route based on the large language model. Figure 3 The system shown includes four layers: data, model, interaction and visualization. The data layer includes the original data of decision cases, decision templates based on the original data, and other relevant precise structured data that can be used to supplement the needs of decision-making tasks. The model layer includes a large language model and a series of small decision models supported by precise data. The interaction layer calls on the large model, decision templates and expert knowledge to complete the generation and revision of decision paths. The visualization layer reorganizes and constructs the data and decision module recommendations generated by the interaction layer to form visual data feedback, which feeds back to optimize the large model and strategy chain.

[0038] An embodiment of the present invention is applied: an example of an interactive method of an intelligent emergency rescue system based on a large language model. For example, at 16:30 on the 3rd of a certain month of a certain year, a magnitude 6.5 earthquake occurred in County B, City A, Province A, and the basic information document of this earthquake is processed.

[0039] The document content is as follows:

[0040] 1) The epicenter was located at 27.1 degrees north latitude and 103.3 degrees east longitude, with a focal depth of 12 kilometers. As of 08:00 on a certain day of a certain month of a certain year, a total of 825 aftershocks were recorded.

[0041] 2) The local weather conditions within five days are as follows:

[0042] 3rd of a certain month: light rain, 17℃~28℃, no sustained wind direction and light breeze;

[0043] 4th of a month: heavy rain, 17℃~27℃, no sustained wind direction and light breeze;

[0044] 5th of a certain month: light rain, 17℃~27℃, no sustained wind direction and light breeze;

[0045] On the 6th of a certain month: The Central Meteorological Observatory predicts that the earthquake-stricken area of ​​County B will mainly experience light rain;

[0046] 7th of a certain month: It is predicted that the earthquake-stricken area of ​​County B will still be mainly cloudy and rainy, with occasional moderate rain.

[0047] 3) Damage

[0048] Casualties: As of 15:00 on a certain day of a certain month of a certain year, 317 people died, 62 people were missing, and 1,143 people were injured;

[0049] Scope of the disaster: The disaster affected 10 counties (districts) in Province A, Province B, and Province C, with 1,088,400 people affected;

[0050] House damage: 80,900 houses collapsed, and a large number of houses were seriously damaged or generally damaged. Economic losses: direct economic losses amounted to 19.849 billion yuan.

[0051] The above cases are used as data. Based on this and previous accumulated cases, the big model extracts the decision chain template and forms the decision chain path:

[0052] 1) Rapid response in the early post-earthquake period (0 to 2 hours)

[0053] 2) Emergency rescue stage (2 to 72 hours)

[0054] 3) Rescue consolidation and expansion phase (72 hours to 7 days)

[0055] 4) Late rescue transition phase (7 days to 1 month)

[0056] For the above case, the system uses the big model to generate decision recommendations based on the above decision chain and perform operations in the middle part of the system: the specific location of the earthquake is displayed on the map according to the longitude and latitude of the epicenter, the range of the epicenter is determined by the earthquake level, and the range of aftershocks may be affected.

[0057] In the initial rapid response period after the earthquake, based on the feedback of real-time data, suggestions and feedback are given to the large model system: how to quickly integrate and coordinate rescue forces in different regions to improve the overall rescue efficiency and ensure that the forces in each rescue hotspot area are sufficient and balanced. At the same time, it recommends the organization of rescue teams, the equipping of professional equipment such as life detectors and search and rescue dogs, the establishment of medical rescue teams, engineering rescue teams and material support teams, etc. The specific locations of the rescue force units in the earthquake zone retrieved from the database are projected onto the map to form optional decision-making sub-function model blocks.

[0058] Based on the above feedback, the big model generates a decision path for the composition and deployment of rescue forces according to the words of the decision-related cases and the relevant locations of the rescue force units in the earthquake-stricken area retrieved from the database. After the expert decision is in line with the actuality and effectiveness of the task, it is stored in the decision-making chain.

[0059] At the same time, the big model generates decision suggestions again based on the above decision chain and performs operations in the middle part of the system:

[0060] 1) Map the rescue force composition and deployment measures on the map;

[0061] 2) Based on the information of the rescue units, the available rescue equipment is gathered to form a sub-model to achieve effective dispatch of rescue supplies.

[0062] During the emergency rescue phase, based on the feedback from real-time data, suggestions and feedback are given to the large model system: In the face of complex geographical conditions, how to effectively allocate different rescue forces to carry out rescue operations and ensure that they arrive at the rescue site at the fastest speed via the optimal path; at the same time, the route planning generated for different rescue forces can be put on the map and form optional decision-making sub-function model blocks.

[0063] After the expert decision is in line with the reality and effectiveness of the task, the rescue force's route and the deployment of the rescue area are stored in the decision chain. At the same time, the big model generates decision suggestions again based on the above decision chain, and performs operations in the middle part of the system: the rescue route planning is mapped to the map according to the rescue force.

[0064] During the rescue consolidation and expansion stage, based on the feedback from real-time data, suggestions and feedback are given to the large model system: expand the rescue scope to form an optimization strategy, and at the same time combine the specific locations of the earthquake-affected areas retrieved from the database to be placed on the map, and form optional decision-making sub-function model blocks.

[0065] After the expert decision is in line with the practicality and effectiveness of the task, it is stored in the decision chain. At the same time, the big model generates decision recommendations again based on the above decision chain, and performs operations in the middle part of the system: mapping the specific locations of secondary disasters generated after the earthquake onto the map.

[0066] In the late transition stage of rescue, in order to reduce the complexity of the command chain caused by the diversity of rescue forces in the disaster area, in order to solve the problem of how to evacuate some rescue forces from the disaster area in an orderly manner without affecting the efficiency of post-disaster rescue, based on the feedback of real-time data, suggestions and feedback are given to the large model system: combined with the real-time location of the rescue forces to be evacuated in the database, the evacuation route is given and placed on the map, and an optional decision-making sub-function model block is formed.

[0067] After the expert decision is in line with the task reality and effectiveness, it is stored in the decision chain. At the same time, the big model generates decision suggestions again based on the above decision chain, and performs operations in the middle part of the system: mapping the evacuation route to the map according to the location of the units to be evacuated after the earthquake.

[0068] At the end of the decision chain, based on the overall process of this rescue mission, the text of the rescue case is formed to provide data for the subsequent decision chain template generation. In addition, based on the optional decision sub-function model blocks formed by the decision chain selection, the relevant data is sent to the visualization large screen to complete the visualization push of data, models, and results.

[0069] The above contents are further detailed descriptions of the present invention in combination with specific embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these inventions. For ordinary technicians in the technical field of the present invention, without departing from the concept of the present invention, they can also make several simple deductions or substitutions, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A decision chain software interaction method based on a large language model, characterized in that: The following steps are involved: S1: Decision planning generation: Based on the text knowledge base of specific field cases, the original decision case data base is realized, a decision chain database is formed, and a decision chain template is constructed. By prompting the user input and calling the large language model, a variety of decision templates are generated to form a decision path that supports multi-step reasoning; S2: Decision path revision: Based on the decision path generated by the decision planning generation part, the decision planning path is instantiated in a human-in-the-loop manner to form a decision path supported by multiple data with operation and editing capabilities. If the decision path does not meet the actual needs of human feedback, the sub-path or path branch relationship of the current decision path is revised by calling the background template to form a dynamic decision path application mode; S3: Decision content recommendation: Form the final record of the decision path, and call the generated data display sub-model and module based on the data and sub-model process of step-by-step decision reasoning, and adaptively recommend it to the comprehensive display visualization large screen.

2. The decision chain software interaction method based on a large language model according to claim 1, characterized in that: The instantiation of the decision-making planning path in S2 specifically includes the display of geographic information data aggregation, identification, etc. required to complete a specific decision-making scenario, structured data, chart retrieval display, and unstructured data call display.

3. The decision chain software interaction method based on a large language model according to claim 1, characterized in that: The S3 also includes the final visualization data feeding back to the big model, optimizing the decision-making link, and forming a user-oriented big model-driven data-decision-recommendation-visualization full-link application mode.

4. A decision chain software interaction system based on a large language model, characterized in that: It includes three parts: large language model operation interaction, large visualization screen and background management; The large language model operation interaction part is used to call the large language model to generate a decision template and a decision path through user input prompts, and to generate a final recommended decision path after instantiating and revising the decision path. The layout of the large language model operation interaction part is a left-middle-right three-column structure, with the large language model software interaction generation part on the left, the decision business process editing part of the operator in the middle, and the recommendation display part of the large model on the right; The visualization large screen part is used to comprehensively display the recommended data generated by the large language model operation interaction part, the decision sub-function model blocks and the data required for the decision task; The background management part uniformly manages the data required for the visualization large screen and the decision chain model template, and realizes the technical route of the decision chain software interaction system based on the large language model.

5. A decision chain software interaction system based on a large language model according to claim 4, characterized in that: The interactive generation part of the large language model software utilizes the inherent planning ability of the large language model to form decision generation supported by a domain-specific decision chain data set driven by the large language model through prompting user input.

6. A decision chain software interaction system based on a large language model according to claim 4, characterized in that: The purpose of the operator's decision-making business process editing part is to complete the revision of the reasoning path, that is, the decision chain human-in-the-loop correction, specifically to instantiate the decision planning path, including the aggregation of geographic information data required for specific decision-making scenarios, identification and other displays, structured data, chart retrieval display and unstructured data call display, to form a decision path supported by multiple data with operation and editing capabilities.

7. The decision chain software interaction system based on a large language model according to claim 4 is characterized in that: The recommendation display part of the large model includes decision content generation and visualization recommendation, and supports being pushed to the display interface of the large-screen visualization part through block combination.

8. The decision chain software interaction system based on a large language model according to claim 4 is characterized in that: The background management part includes four layers: data, model, interaction and visualization. The data layer includes the original data of decision cases, decision templates based on the original data, and other relevant precise structured data that can be used to supplement the needs of decision-making tasks. The model layer includes a large language model and a series of small decision models supported by precise data. The interaction layer calls the large model, decision templates and expert knowledge to complete the generation and revision of decision paths. The visualization layer reorganizes and constructs the data and decision module recommendations generated by the interaction layer to form visual data feedback, which feeds back to optimize the large model and strategy chain.

Images