A method, device, equipment and medium for solidifying tacit knowledge

Abstract

This invention discloses a method, apparatus, device, and medium for solidifying tacit knowledge, relating to the field of artificial intelligence technology. The method includes: acquiring the current task context; using a generative agent to determine unprocessed items; and having the generative agent search for these unprocessed items in a preset factual knowledge base and / or normative knowledge base to generate valid search records; assembling first prompt words and invoking a large language model to reason about the first prompt words, obtaining preliminary results; comparing the preliminary results with standard examples in the current task context to generate a difference report; assembling second prompt words and invoking a large language model to reason about the second prompt words; extracting tacit knowledge based on the difference report; and solidifying the tacit knowledge contained in the standard examples. This invention mines and extracts the tacit knowledge of experts, achieving the solidification of tacit knowledge and realizing the accumulation, reuse, and structured management of enterprise knowledge assets.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method, apparatus, device, and medium for solidifying tacit knowledge. Background Technology

[0002] In knowledge-intensive organizations, experts' tacit knowledge is their most valuable core asset. Tacit knowledge typically refers to knowledge that is difficult to express explicitly in words and relies on personal experience and intuition, such as design experience, debugging techniques, and review standards. This type of knowledge often exists in the minds of experts and is difficult to record and pass on systematically, leading to numerous challenges for enterprises, including knowledge silos, inconsistent work quality, and long training cycles for new employees.

[0003] Traditional knowledge management methods, such as writing operation manuals, recording training videos, and organizing apprenticeship training, can transmit knowledge content to a certain extent, but the transmission efficiency is low and it is difficult to capture the core content of experts' practice.

[0004] In recent years, Large Language Models (LLMs) have demonstrated powerful capabilities in knowledge generation and question answering. LLMs can generate content based on existing explicit knowledge, and compared to directly asking questions, the generated content is often more structured and relevant. However, the content generated by LLMs still lags significantly behind expert output in terms of professionalism, consistency, writing style, and standardization. This gap directly reflects the lack of tacit knowledge. How to transform experts' tacit knowledge into structured assets that can be used by models and even reused throughout an organization remains a crucial problem that the industry urgently needs to solve.

[0005] Therefore, it is necessary to provide a method that can uncover and extract tacit knowledge. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a method, apparatus, device and medium for solidifying tacit knowledge, thereby solving the problem in the prior art that it is difficult to transform the tacit knowledge of experts into structured assets that can be used by models or even reused by the entire organization.

[0007] According to a first aspect, embodiments of the present invention provide a method for solidifying tacit knowledge, the method comprising:

[0008] The system obtains the user's current task context, uses a generative agent to identify unprocessed items in the current task context, and retrieves them from a preset factual knowledge base and / or normative knowledge base to generate valid retrieval records. The current task context consists of the user's requirement text and product background. The factual knowledge base stores factual information, and the normative knowledge base stores intentional information. Factual information comes from explicit knowledge and is declarative information, while intentional information comes from fixed implicit knowledge and is instructional and rule-based information.

[0009] The effective retrieval records and the current task context are assembled into the first prompt word of the generated agent, and the large language model embedded in the generated agent is called to reason about the first prompt word to obtain preliminary results;

[0010] The preliminary results are compared with standard examples in the current task context to generate a discrepancy report;

[0011] The difference report, standard examples, valid retrieval records, and current task context are assembled into a second prompt word for the solidified agent. The large language model embedded in the solidified agent is then used to reason about the second prompt word. Implicit knowledge is extracted from the difference report, and implicit knowledge contained in the standard examples is solidified.

[0012] In conjunction with the first aspect, in the first embodiment of the first aspect, the step of obtaining the user's current task context, using a generative agent to determine unprocessed items in the current task context, and having the generative agent search for unprocessed items in a preset factual knowledge base and / or normative knowledge base to generate valid search records for unprocessed items specifically includes:

[0013] Get the current task context;

[0014] The generative agent is used to determine at least one question point and / or at least one state point in the current task context; the unprocessed items consist of question points and state points. When it is determined that there are ambiguities in the factual information involved in the current task context, each ambiguous factual information will form a question point. When it is determined that the current task context has different task intentions relative to the historical task intentions, each different task intention will form a state point.

[0015] The agent determines the number of retrieval rounds based on the number of question points and state points, generates search terms for each round of retrieval, and performs retrieval based on the search terms and using a factual knowledge base and / or a normative knowledge base to generate valid retrieval records; in each round of retrieval, one question point and / or one state point are searched.

[0016] Use a generative agent to identify at least one point of doubt and / or at least one state point in valid search records;

[0017] The agent determines the number of retrieval rounds based on the number of question points and state points, generates search terms for each round of retrieval, and performs retrieval based on the search terms and using a factual knowledge base and / or a normative knowledge base, supplementing the valid retrieval records.

[0018] In conjunction with the first implementation method of the first aspect, in the second implementation method of the first aspect, the generating agent determines the number of retrieval rounds based on the number of question points and state points, generates search terms for each round of retrieval, performs retrieval based on the search terms and using a factual knowledge base and / or a normative knowledge base, and generates valid search records, specifically including:

[0019] The generated agent determines the number of retrieval rounds based on the number of question points and state points, and determines the corresponding question points and / or state points for each round of retrieval.

[0020] Based on the points of doubt, factual search terms are generated to query the factual content of the points of doubt, and intent search terms are generated based on the status points to reflect the task intent of the status points.

[0021] Determine the search content for each round of retrieval, identify the valid content from the search content, and add the identified valid content to the end of the existing content;

[0022] Collect all valid content from all rounds of retrieval to obtain valid search records.

[0023] In conjunction with the first aspect, in the third embodiment of the first aspect, the step of comparing the preliminary results with a standard paradigm of the current task context to generate a difference report specifically includes:

[0024] A standard example of obtaining the current task context;

[0025] Determine the preset comparison items and preset ignore items;

[0026] Based on the preset comparison items, the solidified intelligent agent compares the preliminary results with the standard examples item by item to obtain a difference report.

[0027] In conjunction with the first aspect, in the fourth embodiment of the first aspect, the assembly of the difference report, standard examples, valid search records, and the current task context into a second prompt word for the solidified agent, and the invocation of the large language model embedded in the solidified agent to reason about the second prompt word, extracting implicit knowledge based on the difference report, and solidifying the implicit knowledge contained in the standard examples, specifically includes:

[0028] The second prompt word of the solidified agent is assembled from the difference report, standard example, valid search record and current task context; each difference record that makes up the difference report has a corresponding unique knowledge number.

[0029] The large language model embedded in the solidified intelligent agent is invoked to reason about the second prompt word, and the reasoning result is obtained. Each reasoning content that constitutes the reasoning result is summarized into a preset style, and the type of each reasoning content is determined according to the preset style.

[0030] Given that the type of reasoning content is determined to be the first type, implicit knowledge of the deletion class is generated;

[0031] Given that the type of reasoning content is determined to be the second type, implicit knowledge of the modified class is generated;

[0032] Given that the type of reasoning content is determined to be the third type, implicit knowledge of the additional class is generated.

[0033] In conjunction with the fourth embodiment of the first aspect, in the fifth embodiment of the first aspect, the method further includes:

[0034] The tacit knowledge is reviewed and processed, and the approved tacit knowledge is added to the normative knowledge base.

[0035] In conjunction with the fifth implementation of the first aspect, in the sixth implementation of the first aspect, the step of reviewing the implicit knowledge and adding the reviewed implicit knowledge to the standard knowledge base specifically includes:

[0036] Obtain all implicit knowledge output by the fixed intelligent agent;

[0037] Each piece of tacit knowledge is reviewed and processed.

[0038] If it is determined that the tacit knowledge does not require adjustment, add the tacit knowledge to the normative knowledge base;

[0039] If it is determined that tacit knowledge needs to be adjusted, the tacit knowledge is adjusted and added to the normative knowledge base.

[0040] In conjunction with the fifth implementation of the first aspect, in the sixth implementation of the first aspect, the step of having the generator of the standard paradigm review and process each piece of implicit knowledge to obtain the solidified implicit knowledge, and adding the solidified implicit knowledge to the normative knowledge base, specifically includes:

[0041] Each piece of tacit knowledge is reviewed and processed by the generator of the standard paradigm;

[0042] If it is determined that the tacit knowledge does not need to be adjusted, the already solidified tacit knowledge is generated based on the tacit knowledge, and the solidified tacit knowledge is added to the standard knowledge base.

[0043] If it is determined that tacit knowledge needs to be adjusted, the tacit knowledge is adjusted to obtain the solidified tacit knowledge, and the solidified tacit knowledge is added to the normative knowledge base.

[0044] According to a second aspect, embodiments of the present invention also provide a device for solidifying tacit knowledge, the device comprising:

[0045] The hybrid retrieval module is used to obtain the user's current task context, use the generative agent to identify unprocessed items in the current task context, and use the generative agent to search for unprocessed items in a preset factual knowledge base and / or normative knowledge base to generate valid retrieval records for unprocessed items. The current task context consists of the user's requirement text and product background. The factual knowledge base is used to store factual information, and the normative knowledge base is used to store intentional information. The factual information comes from explicit knowledge and is declarative information, while the intentional information comes from fixed implicit knowledge and is instructional and rule-based information.

[0046] The preliminary reasoning module is used to assemble the valid search records and the current task context into the first prompt word of the generated agent, and call the large language model embedded in the generated agent to reason about the first prompt word to obtain preliminary results;

[0047] The information comparison module is used to compare preliminary results with standard examples in the current task context and generate a difference report;

[0048] The knowledge solidification module is used to assemble the difference report, standard examples, valid retrieval records, and current task context into the second prompt word of the solidified agent, and call the large language model embedded in the solidified agent to reason about the second prompt word, extract implicit knowledge based on the difference report, and solidify the implicit knowledge contained in the standard examples.

[0049] According to a third aspect, embodiments of the present invention also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method for solidifying tacit knowledge as described above.

[0050] According to a fourth aspect, embodiments of the present invention also provide a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for solidifying tacit knowledge as described above.

[0051] The present invention discloses a method, apparatus, device, and medium for solidifying tacit knowledge. This involves generating an intelligent agent and producing preliminary results based on the current task context, explicit knowledge, and solidified tacit knowledge. Compared to a standard example generated by an expert, the preliminary results can reflect missing tacit knowledge. By comparing the preliminary results with a standard example in the current task context, a difference report is generated. This difference report contains the missing tacit knowledge, and further refines the tacit knowledge. This solidifies the tacit knowledge existing in the expert's mind and embodied in the standard example, resulting in solidified tacit knowledge. Furthermore, by setting up a system composed of explicit knowledge... This invention utilizes a factual knowledge base and a standardized knowledge base constructed from solidified tacit knowledge. An intelligent agent identifies all unprocessed items in the current task context and performs question-driven fact retrieval and / or intent-driven standardized retrieval based on the type of unprocessed item, thereby improving the effectiveness of information retrieval and collecting as much relevant information as possible. By comparing and refining the data using a chain-of-thought approach, the large language model outputs a series of intermediate, logically coherent reasoning steps before providing the final tacit knowledge, thus improving the accuracy of handling complex problems and making the subsequently solidified tacit knowledge more accurate. This invention systematically mines and extracts the tacit knowledge of experts, solidifies tacit knowledge, and thereby enables efficient accumulation, reuse, and structured management of enterprise knowledge assets. Attached Figure Description

[0052] The features and advantages of the invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the invention in any way. In the drawings:

[0053] Figure 1 One of the flowcharts of the method for solidifying tacit knowledge provided by the present invention is shown;

[0054] Figure 2 The second schematic diagram of the process for solidifying tacit knowledge provided by the present invention is shown.

[0055] Figure 3 A schematic diagram of the structure of the device for solidifying tacit knowledge provided by the present invention is shown;

[0056] Figure 4 This is a schematic diagram of the hardware structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] In knowledge-intensive organizations, experts' tacit knowledge is their most valuable core asset. Tacit knowledge typically refers to knowledge that is difficult to express explicitly in words and relies on personal experience and intuition, such as design experience, debugging techniques, and review standards. This type of knowledge often exists in the minds of experts and is difficult to record and pass on systematically, leading to numerous challenges for enterprises, including knowledge silos, inconsistent work quality, and long training cycles for new employees.

[0059] Traditional knowledge management methods, such as writing operation manuals, recording training videos, and organizing apprenticeship training, can transmit knowledge content to a certain extent, but the transmission efficiency is low and it is difficult to capture the core content of experts' practice.

[0060] LLM refers to a deep learning model trained on massive amounts of text data, possessing powerful expressive and generalization capabilities. It can not only generate natural language text but also deeply understand its meaning, handling various natural language tasks such as text summarization, question answering, and translation. In recent years, LLM has demonstrated powerful capabilities in knowledge generation and question answering. LLM can generate content based on existing explicit knowledge, and compared to directly asking questions, its generated content is often more structured and relevant.

[0061] However, the content generated by LLM still lags significantly behind expert outputs in terms of professionalism, consistency, writing style, and standardization. This gap directly reflects the lack of tacit knowledge. How to transform experts' tacit knowledge into structured assets that can be used by models and even reused throughout the organization remains a crucial problem that the industry urgently needs to solve.

[0062] In conclusion, it is necessary to provide a method that can discover and extract tacit knowledge, solidify tacit knowledge, and thereby achieve efficient accumulation and reuse of enterprise knowledge assets as well as structured management.

[0063] Due to the aforementioned technical problems, this invention provides a method for solidifying tacit knowledge. This method aims to systematically mine and extract the tacit knowledge of experts, solidify this knowledge, and thereby achieve efficient accumulation and reuse of enterprise knowledge assets, as well as structured management. The tacit knowledge solidification method of this invention can be used in electronic devices, including but not limited to computers, mobile terminals, etc. Figure 1 This is a flowchart illustrating a method for solidifying tacit knowledge according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method may include the following steps:

[0064] S101. Obtain the user's current task context, use the generating agent to determine the unprocessed items in the current task context, and use the generating agent to search for the unprocessed items in the preset fact knowledge base and / or normative knowledge base to generate valid search records for the unprocessed items.

[0065] The following example illustrates the implicit knowledge solidification in automotive electronics test cases. The current task context consists of the user's requirement text and product background. The user's current task context can be:

[0066] Please write test cases for automotive electronic products using the Gherkin programming language according to the following requirements.

[0067] [SysRS10186] The climate control module will toggle and report the activation status of the MAX AC function in response to the Max_AC_Pressedvalue.

[0068] Automotive Electronics Product Introduction: This product is an automotive air conditioning controller that communicates with the vehicle's network via a CAN bus. The air conditioning controller has no physical user interface; it receives user input via CAN signals to control functions such as cabin temperature, fan speed, and airflow distribution. The product needs to meet the high reliability and real-time requirements of the automotive industry, ensuring stable operation under various driving conditions.

[0069] The example text states: "Please write test cases for automotive electronic products in Gherkin language according to the following requirements."

[0070] [SysRS10186] The climate control module will toggle and report the activation status of the MAX AC function in response to the Max_AC_Pressedvalue.”

[0071] Accordingly, the product background in the example is: "Automotive Electronics Product Introduction: This product is an automotive air conditioning controller that communicates with the vehicle network via a CAN bus. The air conditioning controller has no physical user interface; it receives button input from the user via CAN signals to control functions such as in-vehicle temperature, fan speed, and airflow distribution. The product needs to meet the high reliability and real-time requirements of the automotive industry to ensure stable operation under various driving conditions."

[0072] It should be noted that the current task context can be structured differently in different application scenarios.

[0073] In this embodiment of the invention, the generated intelligent entity determines the unprocessed items based on the current task context.

[0074] In this embodiment of the invention, the generative agent is one of the agents in a group of agents. The generative agent is not the LLM itself, but a complete application system oriented towards achieving a specific goal. The generative agent consists of three core components: the LLM as the reasoning center, callable external tools, and an orchestration layer responsible for task planning and execution loops. The prompt words are the key carriers of the orchestration layer. The prompt words specify the format in which the LLM embedded in the generative agent should think (such as the Thought / Action / Observation loop in ReAct). Through structured instructions, tool descriptions, and examples, the task decomposition logic, tool invocation strategy, and termination conditions are defined.

[0075] Considering that there may still be ambiguous content in the retrieval information returned by the factual knowledge base and / or normative knowledge base, in this embodiment of the invention, the generating agent will determine the current task context and all unprocessed items in the currently obtained valid retrieval records, and perform question-driven factual retrieval and / or intent-driven normative retrieval according to the type of unprocessed item. Specifically, when the generating agent determines that factual information is needed to resolve an unprocessed item, the generating agent will directly generate a question and search in the preset factual knowledge base; when the generating agent determines that intent information is needed to resolve an unprocessed item, the generating agent will directly generate an intent and search in the preset normative knowledge base.

[0076] It should be noted that unprocessed items can also be determined by the user and provided to the generating agent. When the user believes that there is content that is not clearly explained in the current task context, they can also determine unprocessed items themselves.

[0077] In this embodiment of the invention, the knowledge information that users and intelligent agents can use and understand is divided into two categories and managed using different storage structures. Specifically, the knowledge information includes factual information and intentional information. Factual information is stored in a factual knowledge base, and intentional information is stored in a canonical knowledge base.

[0078] Factual information is declarative information of knowledge information, that is, factual information is knowledge information about "what". Factual information can include: current task, background information, historical project data (such as meeting minutes, project weekly reports), data manuals, etc. By segmenting this information and using a text embedding model to vectorize the content obtained from the segmentation, an index for question-driven fact retrieval can be constructed.

[0079] Intent-based information refers to both instructive and rule-based information within knowledge information. Specifically, it's knowledge about "how to do it," and can include lessons learned, design guidelines, and pre-built tacit knowledge bases. By vectorizing the descriptions of the applicability of this information, indexes for intent-driven, prescriptive retrieval can be constructed.

[0080] Therefore, both the factual knowledge base and the normative knowledge base are vector knowledge bases, but the index objects of the two knowledge bases are different.

[0081] It should be noted that each item in the intent information, which is the tacit knowledge that has been solidified, is stored in a structured form of (scope of application description, normative content description). The scope of application description uses natural language to describe the application scenario and intent of the rule, while the normative content description refers to the specific rule or convention.

[0082] Understandably, there can be multiple descriptions of the scope of application.

[0083] Intent information can be ```(“Gherkin test case Feature item”, “Feature item format is <requirement ID>-<test case description>”)```, or it can be ```(“NVM write”, “Unless otherwise specified by the customer, all reserved bits of the data to be written must also be written”)```.

[0084] By constructing factual and normative knowledge bases for storing factual and intentional information respectively, when the generative agent identifies at least one unprocessed item, it can perform multi-round hybrid retrieval based on these bases. Each round of retrieval processes the unprocessed item, and the form of each round is determined based on the type of the unprocessed item, such as question-driven factual retrieval or intention-driven normative retrieval. The generative agent determines whether the knowledge information retrieved in each round is relevant to the unprocessed item; if relevant, it adds it to the existing information and proceeds to the next round of retrieval.

[0085] In this embodiment of the invention, the knowledge information comes from explicit knowledge and solidified implicit knowledge. Solidified implicit knowledge can be understood as explicit knowledge obtained after solidification. Explicit knowledge is factual information and is stored in a factual knowledge base. Correspondingly, solidified implicit knowledge is intentional information and is stored in a normative knowledge base.

[0086] S102. Assemble the valid search records and the current task context into the first prompt word of the generated agent, and call the large language model embedded in the generated agent to reason about the first prompt word to obtain preliminary results.

[0087] The following is a partial excerpt of the preliminary results obtained from LLM inference:

[0088] Okay, here are the Gherkin test cases written for you >>>

[0089] Feature:MAX AC Function Control

[0090] Scenario: Activate MAX AC function

[0091] …

[0092] In this embodiment of the invention, after multiple rounds of hybrid retrieval by the generating agent, when the generating agent determines that the retrieved knowledge information is sufficient, or when the generating agent determines that neither the factual knowledge base nor the normative knowledge base can return valid information, the generating agent will fuse and assemble the valid retrieval records generated based on the returned factual information ("what it is") and / or normative information ("how to do it") with the current task context into a first prompt word, and call the LLM embedded in the generating agent to perform reasoning to obtain a preliminary result.

[0093] In this embodiment of the invention, a valid retrieval record is a structured cumulative cache used by the user and the generating agent to track information that has been "processed". Based on the number and type of unprocessed items, the number of retrieval rounds required to form a valid retrieval record and the specific content of each round of retrieval can be determined. The multi-round mixed retrieval phase is a cyclical execution phase until the conditions for the preliminary results writing phase are met.

[0094] Understandably, regardless of whether the result is valid, once a round of retrieval is completed, it is considered "processed" to avoid generating agents that repeatedly retrieve the same content.

[0095] S103. Compare the preliminary results with the standard examples in the current task context and generate a difference report.

[0096] In this embodiment of the invention, the above-mentioned comparison stage is performed by a solidified intelligent agent. The solidified intelligent agent performs multi-dimensional comparisons of the preliminary results and standard examples to generate a clear and objective difference report.

[0097] In this embodiment of the invention, the solidified agent is another agent in the agent group. The solidified agent is not the LLM itself, but its architecture is basically the same as that of the generated agent.

[0098] A group of agents can share a single LLM, meaning the same LLM is embedded in both the generated and fixed agents. By creating cue words with different structures, the LLM can infer different content. Alternatively, a group of agents can use different LLMs, meaning different LLMs are embedded in both the generated and fixed agents. No specific restrictions are placed on the LLMs embedded in the generated and fixed agents.

[0099] In this embodiment of the invention, the standard examples are obtained from historical project information or expert output. The standard examples contain the implicit knowledge of the experts. For example, the standard examples are written by professional automotive electronics test and development engineers based on enterprise specifications and / or their own accumulated experience.

[0100] Standard examples can be pre-stored in electronic devices, such as process deliverables from historical projects, i.e., historical project information. Historical projects contain a large amount of historical project information, which can be used as standard examples and saved. Standard examples can also be written on-site by experts based on the current task context. There are no restrictions on the specific form of obtaining standard examples, as long as the electronic device can access them.

[0101] The following is a partial excerpt from the standard example:

[0102] Feature: SysRS10186 - MAX AC Activation Control

[0103] Scenario:Toggle MAX AC function

[0104] …

[0105] The following is a partial excerpt from the difference report obtained by comparing the preliminary results with the standard example:

[0106] 1. Preliminary results snippet: "Okay, here is the Gherkin test case written for you >>>\nFeature: MAXAC Function Control" | Standard example snippet: "Feature: SysRS10186 - MAX AC ActivationControl" | Citation note: "The preliminary results include additional information compared to the standard example";

[0107] 2. Preliminary Result Segment: “Feature: MAX AC Function Control” | Standard Example Segment: “Feature: SysRS10186 - MAX AC Activation Control” | Reference Note: “The standard example Feature item contains the requirement number”;

[0108] …

[0109] In this embodiment of the invention, the solidified agent will generate itemized difference reports. Through the itemized difference reports, each difference identified by the solidified agent will form a record. The record format is as follows: Preliminary result fragment: "<fragment>" | Standard example fragment: "<fragment>" | Attribution description: "<Brief description of the tacit knowledge gap reflected by the difference>".

[0110] By using itemized difference reports, key parts of the original context can be preserved through fragments, and the length of the fragments can be limited in this way, for example, the fragment length should not exceed 100 characters, so as to avoid exceeding the context length limit of LLM, reduce the output time of LLM and reduce the generation cost, and facilitate the extraction of tacit knowledge later.

[0111] S104. Assemble the difference report, standard examples, valid retrieval records, and current task context into a second prompt word for the solidified agent, and call the large language model embedded in the solidified agent to reason about the second prompt word. Extract implicit knowledge based on the difference report and solidify the implicit knowledge contained in the standard examples.

[0112] In this embodiment of the invention, the LLM embedded in the solidified agent verifies the identified discrepancy reports with existing tacit knowledge entries to extract new, solidifiable tacit knowledge. When a new case is found to conflict with existing rules, the rule is located, and a "modification" suggestion is generated; or, if the rule is outdated, a "deletion" suggestion is generated. All this tacit knowledge is generated in a structured entry format and submitted to the standard example generator or other experts for final adjudication. Specifically, when the standard example is written by an expert, the expert is also the generator of the standard example, and it is subsequently submitted to the corresponding generator for final adjudication. When the standard example originates from historical project information, considering that the experts who wrote the historical project may not be able to participate in the review process, it can be submitted to other experts for final adjudication.

[0113] For example, tacit knowledge can be:

[0114] 1. Add tacit knowledge: ("Gherkin test case generation", "Keep the output concise, containing only Gherkin test cases, removing all extra explanations and analysis.")

[0115] 2. Add implicit knowledge: (“Gherkin test case Feature item”, “Feature should include requirement number”)

[0116] Established tacit knowledge can be stored in a specification knowledge base. When test cases need to be generated again, the corresponding tacit knowledge can be retrieved from the specification knowledge base to guide the generation of test cases that closely resemble standard examples. The content in the specification knowledge base can also be directly used as a design guide for training new employees, transforming the tacit knowledge and business logic of experts into manageable and reusable standards, effectively preventing knowledge loss due to personnel turnover.

[0117] The implicit knowledge solidification method of this invention generates an intelligent agent and produces preliminary results based on the current task context, explicit knowledge, and solidified implicit knowledge. Compared with the standard examples generated by experts, the preliminary results can reflect the missing implicit knowledge. By comparing the preliminary results with the standard examples of the current task context, a difference report is generated. This difference report contains the missing implicit knowledge and further refines the implicit knowledge, thus solidifying the implicit knowledge that exists in the expert's mind and is reflected in the standard examples, resulting in solidified implicit knowledge. By setting up a factual knowledge base constructed from explicit knowledge and a normative knowledge base constructed from solidified implicit knowledge, the generated intelligent agent identifies all unprocessed items in the current task context and performs problem-driven factual retrieval and / or intent-driven normative retrieval based on the type of unprocessed items, thereby improving the effectiveness of information retrieval and collecting as much relevant information as possible. By comparing and then refining, and using the concept of thought chains, the large language model outputs a series of intermediate, logically coherent reasoning steps before giving the final implicit knowledge, thereby improving the accuracy of handling complex problems and making the subsequently solidified implicit knowledge more accurate. This invention systematically mines and extracts the tacit knowledge of experts, solidifies the tacit knowledge, and thereby enables the efficient accumulation and reuse of enterprise knowledge assets as well as structured management.

[0118] In this embodiment of the invention, step S101 specifically includes:

[0119] S1011. Obtain the current task context.

[0120] S1012. Generate an agent based on at least one question point and / or at least one state point in the current task context. Correspondingly, unprocessed items consist of question points and state points, and each question point or each state point can be an unprocessed item.

[0121] In this embodiment of the invention, the generated agent first performs a necessity judgment on the retrieval. If it is determined that there are ambiguities in the factual information involved in the current task context, each ambiguous factual information will form a point of doubt. If it is determined that the current task context has a different task intent than the historical task intent, each different task intent will form a state point. For example, the generated agent parses the requirement text in conjunction with the user's product background to determine whether there are ambiguities in the factual information involved in the requirement text, such as signal definitions, triggering conditions, expected behaviors, etc., and whether the current task requirements represented in the requirement text, such as the target of the test case writing being focused on, have different task intents than the historical task intents, such as being more specific or involving new task intents. When there are ambiguities in the factual information, each ambiguous factual information will form a point of doubt; correspondingly, when there are different task intents, each different task intent will form a state point.

[0122] For example, the generative agent analyzes the requirement text based on the user's product background and finds that the definition of Max_AC_Pressed is ambiguous. Therefore, the generative agent generates a direct question, "What is the specific definition of the Max_AC_Pressed signal?", and searches the fact knowledge base. After searching, it finds that this requirement explains the definition of Max_AC_Pressed: ```[SysRS10184] The climate control module will receive and process signal Frt_Btn_Status_1st contained in the CAN message ~Clmt_Button_Stat4_MS1~ with value Max_AC_Pressed, as an adjustment to the Max AC button.```

[0123] S1013. Determine the number of retrieval rounds based on the number of question points and status points, generate search terms for each round, and perform retrieval based on the search terms and using a factual knowledge base and / or a normative knowledge base to generate valid search records. Each round of retrieval involves searching for one question point and / or one status point.

[0124] Step S1013 is mainly completed by the generating agent. The generating agent will perform multiple rounds of mixed retrieval of unclear content in the current task context to obtain effective retrieval records. With this setting, each round of retrieval focuses only on the most pressing factual issue and / or intent dimension, avoiding LLM information overload and thus improving retrieval efficiency.

[0125] In this embodiment of the invention, each round of retrieval has a corresponding original search term, which consists of factual search terms and / or intent search terms.

[0126] Fact search terms are used for question-driven fact retrieval, while intent search terms are used for intent-driven normative retrieval. The generated agent uses fact search terms to search in the fact knowledge base and uses intent search terms to search in the normative knowledge base, thereby forming a series of search results. The search results are then merged to obtain effective search records.

[0127] S1014. Use the generated agent to determine at least one question point and / or at least one state point in the valid retrieval records.

[0128] S1015. The generating agent determines the number of retrieval rounds based on the number of question points and state points, generates search terms for each round of retrieval, and performs retrieval based on the search terms and using the factual knowledge base and / or normative knowledge base, supplementing the valid retrieval records.

[0129] Similarly, step S1015 is mainly completed by the generative agent. Considering that there are still unclear contents in the retrieval information returned by the factual knowledge base and / or the normative knowledge base, the generative agent will also determine whether there are still doubt points and / or at least state points in the valid retrieval records generated in each round of retrieval. If it is determined that there are still doubt points and / or at least state points, multiple rounds of mixed retrieval will continue, and the newly returned information will be used to supplement the valid retrieval records.

[0130] More specifically, step S1013 includes the following steps:

[0131] S10131. Generate an intelligent agent to determine the number of retrieval rounds based on the number of question points and state points, and determine the question points and / or state points corresponding to each round of retrieval.

[0132] S10132. Generate factual search terms based on the question points to query factual content of the question points, and generate intent search terms based on the status points to reflect the task intent of the status points.

[0133] In this embodiment of the invention, fact search terms are used to query factual content such as product design documents and interface specifications (e.g., "What is the data type and valid value of the Max_Defrost_Btn_Stt signal?"), and intent search terms are used to reflect the most specific task intent of the state point. They must be very brief state descriptions and are used to match the scope descriptions in the specification knowledge base (e.g., "Write a Gherkin use case for switching the state of the MAX Defrost button").

[0134] Depending on the specific questions and / or status points focused on in each round of retrieval, the output format of the original search terms includes the following three types:

[0135] Fact search: <fact search terms>; Intent search: <intent search terms>

[0136] Fact search: <fact search terms>

[0137] Intent Search: <Intent Search Terms>

[0138] Unlike directly using written test case tasks as search terms, generating an agent to search based on original search terms obtained from factual search terms and / or intent search terms can significantly improve the usefulness of search results.

[0139] For example, the requirement text is: ```[SysRS10186] The climate control module will toggle and report the activation status of the MAX AC function in response to the Max_AC_Pressed value.``` The semantically most similar text is: ```[SysRS10189] The climate control module will toggle and report the activation status of the MAXDefrost function in response to the Max_Defrost_Pressed value.```. It can be seen that, apart from the function changing from AC to Defrost and the adjustment of the requirement number, they are highly similar semantically, describing the same operating logic of different air conditioning functions. However, this most similar text is not helpful for writing test cases for the original requirement text.

[0140] To write test cases for the original requirement text, the user or the generating agent actually needs to know what `Max_AC_Pressed` actually is. Its definition is found in `[SysRS10184] The climate control module will receive and process signal Frt_Btn_Status_1st contained in the CAN message ~Clmt_Button_Stat4_MS1~ with value Max_AC_Pressed, as an adjustment to the MaxAC button.```, which has a much lower semantic similarity than the aforementioned text. This indicates that directly inputting the original text does not yield the most similar search result, even if it is the most useful. Therefore, it is necessary to introduce an agent to generate appropriate factual search terms, such as `What is the definition of Max_AC_Pressed Value?```, to obtain the semantically closest result.

[0141] Specification-related content that is helpful for writing high-quality test cases often has low text similarity to the current task. Therefore, a generative agent needs to generate a description of the current intent, and the applicable scope text of the specification-related content should be used as an embedded index to improve the usefulness of the search results. For example, a specification with the format `<Requirement ID>-<Test Case Description>` is an example, and its applicable scope is `Gherkin Test Case Feature Item`. When the generative agent generates the intent ```Write Gherkin Test Cases`, the aforementioned applicable scope can be matched, thus adding the specification to the original search terms as prompts, thereby enabling the writing of high-quality specification test cases. By using intent search terms and applicable scope descriptions, specification-related content and the current task context are highly correlated.

[0142] S10133. Determine the search content for each round of retrieval, identify the valid content from the search content, and add the identified valid content to the end of the existing content.

[0143] S10134. Gather the valid content from all rounds of retrieval to obtain valid retrieval records.

[0144] Since factual search terms are indexes for question-driven factual retrieval of the factual knowledge base, and intent search terms are indexes for intent-driven normative retrieval of the normative knowledge base, each search result includes its source. In this embodiment of the invention, the generating agent also filters out a maximum of a preset number (e.g., three) of the most relevant and useful valid content from the returned search results for each original search term; if there are no valid results, a record is generated, with the result field marked as "no valid search results" and the source as "N / A"; each filtered result is output independently as a single line, which can be directly used as an entry in a valid search record.

[0145] Factual search terms can retrieve information with high textual relevance, while intent search terms can retrieve specifications that match the current intent. Taking the implicit knowledge ("Gherkin test case Feature item", "Feature item format is <requirement ID>-<use case description>") as an example, the embedding vector similarity between the intent "write Gherkin test cases" and "Gherkin test case Feature item" is higher than that between the embedding vector of "Feature item format is <requirement ID>-<use case description>". That is, the matching degree between intent search terms and scope descriptions is higher than that between content. By introducing intent search with scope descriptions, we can better find specifications applicable to the current task. Through multiple rounds of mixed search and repeated questioning, we can continuously improve the search results.

[0146] In this embodiment of the invention, whenever valid content is selected, it is added to the end of the existing content before proceeding to the next round of retrieval. Each valid piece of content contains four fields: source type (factual knowledge base or normative knowledge base), original search term, original text fragment of the search result (if none, it is "no valid search result"), and source (such as document ID, chapter, etc.). LLM providers often cache suggestion words; placing the varied parts of the suggestion words at the end can fully utilize the preceding suggestion words, achieve cache hits, thereby improving speed and reducing costs.

[0147] Understandably, in order to avoid generating agents that repeatedly search for the same content, a new question point will only be generated if a corresponding factual search term has not yet appeared in the valid search records for a certain question point. Similarly, a new state point will only be generated if a corresponding intent search term has not yet appeared in the valid search records for a certain state point.

[0148] For example, if the requirement text mentions Max_Defrost_Btn_Stt, but the generating agent is unclear about its meaning and there are no relevant factual search terms in the valid search records, the generating agent will identify this factual information as a point of doubt. If the user's requirement text indicates that the current task intent is "to write Gherkin test cases for switching the state of the MAX Defrost button", and the user has only previously searched for "to write Gherkin test cases", the generating agent will identify this factual information as a state point.

[0149] The following example illustrates the solidification of tacit knowledge in automotive electronics test cases. An example of the first cue word obtained through assembly could be:

[0150] You are a senior automotive electronics R&D and testing engineer with over 10 years of experience in automotive electronic systems testing. You are proficient in the testing logic and methods of core systems such as the body domain, chassis domain, powertrain domain, and smart cockpit. You can accurately identify potential system risks and have the ability to control the entire process from requirements analysis to test case implementation.

[0151] Your task is to write high-quality test cases for automotive electronics products in the Gherkin language, based on the given requirements.

[0152] To ensure the professionalism and completeness of your output, you must strictly follow the following two-stage action framework, including the **retrieval stage** and the **writing stage**.

[0153] *Retrieval phase (must be executed repeatedly until the conditions for entering the writing phase are met)

[0154] Step 1: Determining the Necessity of the Search and Generating Search Terms

[0155] *Please carefully analyze the current task input (including requirement text and product background) to determine if any of the following two types of unprocessed items exist:

[0156] *Points of contention: There are ambiguities in the factual information involved in the requirements (such as signal definition, triggering conditions, expected behavior, etc.), and the corresponding factual search terms for these points of contention have not yet appeared in the valid search records.

[0157] *Status Point: The current task intent (i.e., the goal of writing Gherkin test cases that you are focusing on) is more specific than all historical intents, or involves a new task intent that has not yet appeared in any valid search records with corresponding intent search terms.

[0158] *Example of judgment criteria:

[0159] *If the request mentions Max_Defrost_Btn_Stt, but you are unsure of its meaning and there are no relevant factual searches in the valid records, then this is a point of doubt.

[0160] *If your current intent is "to write Gherkin test cases for MAX Defrost button state toggling", and you have only previously searched for "to write Gherkin test cases", then this is a new state point.

[0161] *Output rules:

[0162] *If all questions and status points have been processed (i.e., all have appeared in valid search records), then the output will be:

[0163] * Search completed, now in the writing stage.

[0164] *Otherwise, at most one fact term and / or one intent term will be generated (the two can exist alone or simultaneously):

[0165] *Fact search terms: Used to search for factual content such as product design documents and interface specifications (e.g., "What is the data type and valid value of the Max_Defrost_Btn_Stt signal?").

[0166] *Intent search terms: Must be very brief state descriptions that reflect the most specific task intent at present, and are used to match applicable conditions in the specification knowledge base (e.g., "Write a Gherkin use case for MAX Defrost button state toggling").

[0167] *Output format (choose one of the following as needed):

[0168] * Fact search: <fact search terms>; Intent search: <intent search terms>

[0169] * Fact search: <fact search terms>

[0170] * Intent Search: <Intent Search Terms>

[0171] *Note: Each round should focus on only one current factual issue and / or one intent dimension to avoid information overload.

[0172] Step Two: Filtering Search Results

[0173] You will receive search results returned by an external system (based on the search terms generated in step one, retrieved from the factual knowledge base and / or the normative knowledge base). Each search result will include its source (such as the requirement document ID, specification section number, etc.).

[0174] * Fact knowledge base: Stores original text fragments of "what is" such as product requirements and signal definitions (e.g., [SysRS10185]...), with original text embedded vector indexes.

[0175] *The specification knowledge base stores "how-to" entries such as test rules and format requirements, but is indexed by the embedding vector of the applicable condition text (such as "Gherkin test case feature item") rather than the rules themselves.

[0176] Your task is to filter the search results:

[0177] For each original search term, select up to three of the most relevant and useful results from the returned results;

[0178] If no valid results are found, a record will be generated with the result field marked as "No valid search results" and the source as "N / A".

[0179] Each filtered result is output as a separate line and can be directly used as an entry for a valid search record.

[0180] Output format:

[0181] Each record is on a separate line, formatted as: ```[Source Type] | Search Term: "<Original Search Term>" | Result: "<Original Text Excerpt of Search Result>" | Source: "<Source Identifier>"```;

[0182] *If a search term yields no valid results, a record should still be output: ```[Source Type] | Search Term: "<Original Search Term>" | Result: "No valid search results" | Source: "N / A"```;

[0183] *If step one becomes a fact search term, then process the corresponding fact knowledge base results and output 1 to 3 results (or 1 result "no valid").

[0184] *If an intent search term is generated, the corresponding standard knowledge base results will be processed, and 1 to 3 results will be output (or 1 result "no valid").

[0185] *All output records are grouped by source type: first output all fact records, then output all canonical records;

[0186] Do not merge multiple results into the same row. Each result must be on a separate row so that they can be appended to the valid search record one by one.

[0187] *Note: This step does not generate new search terms; it only processes existing search results.

[0188] *Writing Stage

[0189] Once the retrieval phase determines that you can proceed to the writing phase, you need to write test cases using the Gherkin language (Given / When / Then structure) based on: the original task input (requirements + product introduction) and complete and valid retrieval records.

[0190] The output is the final answer and requires no further explanation.

[0191] The current task context is as follows:

[0192] >>>

[0193] Please write test cases for automotive electronic products using the Gherkin programming language according to the following requirements.

[0194] [SysRS10186] The climate control module will toggle and report the activation status of the MAX AC function in response to the Max_AC_Pressedvalue.

[0195] Automotive Electronics Product Introduction: This product is an automotive air conditioning controller that communicates with the vehicle's network via a CAN bus. The air conditioning controller has no physical user interface; it receives user input via CAN signals to control functions such as cabin temperature, fan speed, and airflow distribution. The product needs to meet the high reliability and real-time requirements of the automotive industry, ensuring stable operation under various driving conditions.

[0196] <<<

[0197] The following are the valid search records:

[0198] >>>

[0199] <<<

[0200] The search results are as follows:

[0201] >>>

[0202] <<<

[0203] You now need to perform: Search Phase → Step 1: Determining the Necessity of the Search and Generating Search Terms

[0204] In this embodiment of the invention, step S103 specifically includes:

[0205] S1031, A standard example of obtaining the current task context.

[0206] S1032. Determine the preset comparison items and preset ignore items.

[0207] In this embodiment of the invention, the above-mentioned processing is performed by a solidified intelligent agent. The preset comparison items may include: structure, content, format, logic, naming, etc. In this way, the preliminary results are compared with the standard example item by item in the subsequent comparison process to obtain a more comprehensive difference report. By setting preset ignore items, non-significant differences can be filtered out, thereby ignoring minor differences that do not affect the consistency of engineering or the validity of testing (such as irrelevant spaces, line breaks, synonym replacements).

[0208] S1033. Based on the preset comparison items, use the fixed intelligent agent to compare the preliminary results with the standard examples item by item to obtain a difference report. The final difference report only retains those differences that could have been avoided if a certain unstated expert convention had been followed.

[0209] In this embodiment of the invention, step S104 specifically includes:

[0210] S1041. Assemble the difference report, standard example, valid retrieval record and current task context into a second prompt word for the solidified agent.

[0211] S1042. Call the large language model embedded in the solidified intelligent agent to reason about the second prompt word, obtain the reasoning result, summarize each reasoning content that constitutes the reasoning result into a preset style, and determine the type of each reasoning content according to the preset style.

[0212] Each piece of tacit knowledge has a corresponding unique knowledge ID. During the extraction phase, the context of the corresponding tacit knowledge will be carried into it through the knowledge ID.

[0213] In this embodiment of the invention, the preset style is the style used for intentional information. The preset style includes a description of the scope of application and a description of the normative content. That is, the preset style is stored in a structured form of (scope of application description, normative content description). The scope of application description uses natural language to describe the application scenario and intent of the rule, and the normative content description refers to the specific rule or convention.

[0214] S1043. If the type of the reasoning content is determined to be Type I, generate implicit knowledge of the deletion class. If a certain reasoning content is directly falsified by an existing implicit knowledge item, and the corresponding type is Type I, generate implicit knowledge of the deletion class.

[0215] S1044. If the type of the reasoning content is determined to be the second type, generate implicit knowledge of the modification type. If a certain reasoning content is judged to need modification (e.g., the scope of application is described as too broad or too narrow), it is correspondingly of the second type, and implicit knowledge of the modification type is generated.

[0216] S1045. If the type of the reasoning content is determined to be the third type, generate implicit knowledge of the augmented class. If a certain reasoning content has no related intentional information (i.e., existing implicit knowledge entry) in the normative knowledge base, and is correspondingly of the third type, generate implicit knowledge of the augmented class.

[0217] The following example illustrates the tacit knowledge solidification in automotive electronics test cases. An example of the assembled second cue word could be:

[0218] You are a senior automotive electronics testing expert and knowledge engineering architect with over 10 years of experience in test case development and team knowledge system building. You excel at accurately identifying implicit knowledge gaps from the discrepancies between expert outputs and model-generated results, and extracting them into reusable, traceable, and scenario-specific structured knowledge items.

[0219] Your task is to provide operational suggestions (including adding, modifying, or deleting) for the tacit knowledge base based on preliminary results and standard examples.

[0220] To ensure the professionalism and completeness of your output, you must strictly follow the following two-stage action framework, including the **comparison stage** and the **refinement stage**.

[0221] *Input instructions: You will receive the following four types of input:

[0222] *Preliminary results: Test cases output by the generating agent

[0223] *Standard Examples: High-quality examples of the same type written and approved by domain experts in past projects.

[0224] *Task context: original requirements text, product background, etc.

[0225] *Valid Retrieval Records: Valid retrieval records (including fact and normative categories) used by the generating agent during the generation process. During the refinement phase, existing implicit knowledge entries in the normative category records will also be appended with normative basis records based on their knowledge IDs.

[0226] *Comparison Phase

[0227] You need to systematically compare the initial generated results with the standard example, identify all significant differences that can be attributed to a lack of tacit knowledge, and output them in a standard format.

[0228] * Compare the differences between the two in terms of structure, content, format, logic, and naming.

[0229] *Filtering insignificant differences: Ignore minor differences that do not affect engineering consistency or test validity (such as irrelevant spaces, line breaks, and synonym substitutions).

[0230] *Attribution judgment: Only retain those differences that could have been avoided if a certain unstated expert practice had been followed.

[0231] Output format:

[0232] *Itemized difference reports, each difference generates one record, formatted as: ```Preliminary Result Fragment: "<Fragment>" | Standard Example Fragment: "<Fragment>" | Attribution Description: "<Brief description of the tacit knowledge gap reflected by this difference>"```

[0233] *If there is no significant difference, output "No significant difference, can be identified".

[0234] * The fragment must retain the key parts of the original context and must not exceed 100 characters in length.

[0235] *Refining Stage

[0236] Based on the difference report from the first phase, and combined with the task context and valid retrieval records (especially existing tacit knowledge entries that may be contained therein), generate executable knowledge base operation suggestions with knowledge IDs.

[0237] For each discrepancy record, attempt to summarize one implicit knowledge item, in the format: (<Scope of application description>, <Specification content description>).

[0238] *Scope of application description: Extracted from task context, standard example features, or search records, clearly defining the applicable scenarios for the rules (e.g., "Gherkin test case feature item", "state verification after button-type signal triggering").

[0239] *Specification description: Use natural language to explain what should and should not be done (positive / negative statements are supported, such as "should include the requirement ID", "no comments allowed").

[0240] *If there are existing implicit knowledge entries with highly similar semantics in the normative records.

[0241] * Determine if modification is needed (e.g., the scope of application is too broad or too narrow), output the operation type as MODIFY, and use the knowledge ID of that entry.

[0242] If the difference directly disproves an existing tacit knowledge entry, the output operation type is DELETE, and the knowledge ID of that entry is used.

[0243] *If there are no relevant existing tacit knowledge entries in the normative record.

[0244] * The output operation type is ADD, and the knowledge ID uses a fixed placeholder ```[NEW_KB_ENTRY]```

[0245] You do not need to worry about generating the standard basis record. The external system will automatically establish a mapping relationship of "Knowledge ID <-> Standard Basis Record (Preliminary Result, Standard Example, Task Context, Retrieval Record)" based on the current task context.

[0246] *Output Format

[0247] *Each suggestion is on a separate line.

[0248] *For the ADD and MODIFY operation types, the format is ```[<operation type>] | [<knowledge ID>] | entry: ("<scope description>", "<specification content description>")```

[0249] *For the DELETE operation type, the format is ```[DELETE] | [<Knowledge ID>]```

[0250] If no effective knowledge can be extracted, the output will be: "No effective tacit knowledge can be solidified."

[0251] Preliminary results are as follows:

[0252] >>>

[0253] <<<

[0254] The standard example is as follows:

[0255] >>>

[0256] <<<

[0257] The current task context is as follows:

[0258] >>>

[0259] <<<

[0260] The following are the valid search records:

[0261] >>>

[0262] <<<

[0263] The differences are reported as follows:

[0264] >>>

[0265] <<<

[0266] You now need to proceed with the comparison phase.

[0267] Please see Figure 2 In this embodiment of the invention, the method may further include the following steps:

[0268] S201. Obtain the user's current task context, use the generative agent to determine the unprocessed items in the current task context, and based on the unprocessed items, the generative agent searches in a preset factual knowledge base and / or normative knowledge base to generate valid retrieval records for the unprocessed items. Specific details are as follows: Figure 1 The steps in step S101 are shown and will not be repeated here.

[0269] S202. Assemble the valid search records and the current task context into the first prompt word for the generated agent, and call the large language model embedded in the generated agent to reason about the first prompt word to obtain preliminary results. Specific details are as follows: Figure 1 Step S102 is shown in the diagram and will not be repeated here.

[0270] S203. Compare the preliminary results with the standard examples in the current task context and generate a discrepancy report. Specific content is as follows: Figure 1 Step S103 is shown in the diagram and will not be repeated here.

[0271] S204. Assemble the difference report, standard examples, valid retrieval records, and current task context into a second prompt word for the solidified agent, and call the large language model embedded in the solidified agent to reason about the second prompt word. Extract implicit knowledge based on the difference report and solidify the implicit knowledge contained in the standard examples. Specific details are as follows: Figure 1 Step S104 is shown in the diagram and will not be repeated here.

[0272] S205. The tacit knowledge is processed by the generator of the standard paradigm to obtain the solidified tacit knowledge.

[0273] The solidified tacit knowledge includes operational suggestions for that tacit knowledge. In this embodiment of the invention, the tacit knowledge can be reviewed by experts, such as professional automotive electronics test and development engineers, through a user interaction interface. For example, if an expert deems the first piece of tacit knowledge reasonable, they can approve it, thus forming solidified tacit knowledge based on that piece. If an expert believes that the second piece of tacit knowledge can be further refined, they can adjust its content to obtain adjusted tacit knowledge, which will also form solidified tacit knowledge, for example ("Gherkin Test Case Feature Item", "Feature Item Format: <Requirement ID>-<Test Case Description>").

[0274] Specifically, by having the experts who create the standard paradigms—the generators—process the tacit knowledge, the tacit knowledge existing in the experts' minds and embodied in the standard paradigms, as well as the experts' review of the tacit knowledge, is solidified, resulting in solidified tacit knowledge. Through this setup, experts are not directly the creators of tacit knowledge items, but rather the generators of standard paradigms and the reviewers of tacit knowledge. Compared to directly writing the text, reviewing and revising the text is much easier, ensuring the quality of tacit knowledge on the one hand, and improving the efficiency of tacit knowledge generation on the other.

[0275] In this embodiment of the invention, step S205 specifically includes:

[0276] S2051. Obtain all implicit knowledge output by the fixed intelligent agent.

[0277] S2052. Review and process each piece of tacit knowledge. For example, the review and processing can be carried out by the creator of the standard example or other professionals.

[0278] In this embodiment, users, such as experts, only need to review the data and do not need to directly generate the implicit knowledge content. This ensures the quality of the implicit knowledge base and improves its generation efficiency. Furthermore, the newly solidified implicit knowledge is added to the standard knowledge base, thereby updating the stored content of the planned knowledge base. After completing steps S201 to S205, the intelligent agent group automates a complete round of implicit knowledge mining and extraction, obtaining the implicit knowledge solidified in this round.

[0279] S2053. If it is determined that the tacit knowledge does not need to be adjusted, add the tacit knowledge to the normative knowledge base.

[0280] S2054. If it is determined that the tacit knowledge needs to be adjusted, the tacit knowledge shall be adjusted and the adjusted tacit knowledge shall be added to the normative knowledge base.

[0281] Because tacit knowledge has different corresponding types, it contains corresponding operational suggestions. When an expert deems the first piece of tacit knowledge reasonable, it can be approved, thus forming solidified tacit knowledge. If the expert believes that the second piece of tacit knowledge can be further refined, its content can be adjusted to obtain adjusted tacit knowledge, which also becomes solidified tacit knowledge. By reviewing and processing tacit knowledge to solidify it, information compression can be achieved.

[0282] In this embodiment of the invention, the implicit knowledge is generated by assembling difference reports, standard examples, valid search records, and the current task context into second prompt words for a fixed agent, and then using the fixed agent for reasoning. This differs from directly feeding expert-provided standard examples as part of the current task context into a large language model for reasoning. The assembly method of the second prompt words transforms a large amount of example text into clear, explicit implicit knowledge that can be directly used by LLM.

[0283] By adding tacit knowledge that has been solidified to the standard knowledge base, experts' tacit knowledge and business logic can be transformed into manageable and reusable standards, effectively preventing knowledge loss caused by personnel turnover.

[0284] It should be noted that the amount of tacit knowledge obtained from a single round of solidification is often limited, and its content is also often restricted. For comparisons of a single long standard example, a single solidification process may not be able to cover all the differences, i.e., the tacit content. Considering this, in this embodiment of the invention, multiple rounds of solidification can be executed. This involves first solidifying some of the obvious tacit knowledge, adding the solidified tacit knowledge to the specification knowledge base, and then using the updated specification knowledge base for further generation. At this point, the previously obvious differences converge, and the previously indistinct differences become obvious, facilitating further solidification. For comparisons of a large number of standard examples, a single standard example may not be able to determine whether a difference is universal or specific. Through multiple rounds of iteration, the universality and specificity of the differences can be identified.

[0285] That is, standard examples can be repeatedly extracted for the current task context, or multiple standard examples can be used alternately for optimization, thereby continuously enriching and updating the specification database.

[0286] The apparatus for solidifying tacit knowledge provided in the embodiments of the present invention will be described below. The apparatus for solidifying tacit knowledge described below and the method for solidifying tacit knowledge described above can be referred to in correspondence with each other.

[0287] Due to the aforementioned technical problems, this invention also provides a device for solidifying tacit knowledge, which aims to systematically mine and extract the tacit knowledge of experts, solidify the tacit knowledge, and thereby achieve efficient accumulation and reuse of enterprise knowledge assets and structured management. Figure 3 This is a structural schematic diagram of a method for solidifying tacit knowledge according to an embodiment of the present invention, as shown below. Figure 3 As shown, the device may include:

[0288] The hybrid retrieval module 10 is used to obtain the user's current task context, use the generating agent to determine the unprocessed items in the current task context, and use the generating agent to search for the unprocessed items in a preset fact knowledge base and / or normative knowledge base to generate valid retrieval records for the unprocessed items.

[0289] The composition of the current task context can differ across different application scenarios. In this embodiment of the invention, the generating intelligence identifies unprocessed items based on the current task context.

[0290] In this embodiment of the invention, the generative agent is one of the agents in a group of agents. The generative agent is not the LLM itself, but a complete application system oriented towards achieving a specific goal. The generative agent consists of three core components: the LLM as the reasoning center, callable external tools, and an orchestration layer responsible for task planning and execution loops. The prompt words are the key carriers of the orchestration layer. The prompt words specify the format in which the LLM embedded in the generative agent should think (such as the Thought / Action / Observation loop in ReAct). Through structured instructions, tool descriptions, and examples, the task decomposition logic, tool invocation strategy, and termination conditions are defined.

[0291] Considering that there may still be ambiguous content in the retrieval information returned by the factual knowledge base and / or normative knowledge base, in this embodiment of the invention, the generating agent will determine the current task context and all unprocessed items in the currently obtained valid retrieval records, and perform question-driven factual retrieval and / or intent-driven normative retrieval according to the type of unprocessed item. Specifically, when the generating agent determines that factual information is needed to resolve an unprocessed item, the generating agent will directly generate a question and search in the preset factual knowledge base; when the generating agent determines that intent information is needed to resolve an unprocessed item, the generating agent will directly generate an intent and search in the preset normative knowledge base.

[0292] It should be noted that unprocessed items can also be determined by the user and provided to the generating agent. When the user believes that there is content that is not clearly explained in the current task context, they can also determine unprocessed items themselves.

[0293] In this embodiment of the invention, the knowledge information that users and intelligent agents can use and understand is divided into two categories and managed using different storage structures. Specifically, the knowledge information includes factual information and intentional information. Factual information is stored in a factual knowledge base, and intentional information is stored in a canonical knowledge base.

[0294] Factual information is declarative information of knowledge information, that is, factual information is knowledge information about "what". Factual information can include: current task, background information, historical project data (such as meeting minutes, project weekly reports), data manuals, etc. By segmenting this information and using a text embedding model to vectorize the content obtained from the segmentation, an index for question-driven fact retrieval can be constructed.

[0295] Intent-based information refers to both instructive and rule-based information within knowledge information. Specifically, it's knowledge about "how to do it," and can include lessons learned, design guidelines, and pre-built tacit knowledge bases. By vectorizing the descriptions of the applicability of this information, indexes for intent-driven, prescriptive retrieval can be constructed.

[0296] Therefore, both the factual knowledge base and the normative knowledge base are vector knowledge bases, but the index objects of the two knowledge bases are different.

[0297] It should be noted that each item in the intent information, which is the tacit knowledge that has been solidified, is stored in a structured form of (scope of application description, normative content description). The scope of application description uses natural language to describe the application scenario and intent of the rule, while the normative content description refers to the specific rule or convention.

[0298] Understandably, there can be multiple descriptions of the scope of application.

[0299] The scope description is extracted from the task context, standard example features, or search records to clarify the applicable scenarios of the rules (such as "Gherkin test case feature item" or "state verification after button signal triggering"). The specification content description uses natural language to explain how to do or not to do something (it can be positive or negative content, such as "should include requirement ID" or "no comments should be added").

[0300] Intent information can be ```(“Gherkin test case Feature item”, “Feature item format is <requirement ID>-<test case description>”)```, or it can be ```(“NVM write”, “Unless otherwise specified by the customer, all reserved bits of the data to be written must also be written”)```.

[0301] By constructing factual and normative knowledge bases for storing factual and intentional information respectively, when the generative agent identifies at least one unprocessed item, it can perform multi-round hybrid retrieval based on these bases. Each round of retrieval processes the unprocessed item, and the form of each round is determined based on the type of the unprocessed item, such as question-driven factual retrieval or intention-driven normative retrieval. The generative agent determines whether the knowledge information retrieved in each round is relevant to the unprocessed item; if relevant, it adds it to the existing information and proceeds to the next round of retrieval.

[0302] In this embodiment of the invention, the knowledge information comes from explicit knowledge and solidified implicit knowledge. Solidified implicit knowledge can be understood as explicit knowledge obtained after solidification. Explicit knowledge is factual information and is stored in a factual knowledge base. Correspondingly, solidified implicit knowledge is intentional information and is stored in a normative knowledge base.

[0303] The preliminary reasoning module 20 is used to assemble the valid search records and the current task context into the first prompt word of the generated agent, and call the large language model embedded in the generated agent to reason about the first prompt word to obtain preliminary results.

[0304] In this embodiment of the invention, after multiple rounds of hybrid retrieval by the generating agent, when the generating agent determines that the retrieved knowledge information is sufficient, or when the generating agent determines that neither the factual knowledge base nor the normative knowledge base can return valid information, the generating agent will fuse and assemble the valid retrieval records generated based on the returned factual information ("what it is") and / or normative information ("how to do it") with the current task context into a first prompt word, and call the LLM embedded in the generating agent to perform reasoning to obtain a preliminary result.

[0305] In this embodiment of the invention, a valid retrieval record is a structured cumulative cache used by the user and the generating agent to track information that has been "processed". Based on the number and type of unprocessed items, the number of retrieval rounds required to form a valid retrieval record and the specific content of each round of retrieval can be determined. The multi-round mixed retrieval phase is a cyclical execution phase until the conditions for the preliminary results writing phase are met.

[0306] Understandably, regardless of whether the result is valid, once a round of retrieval is completed, it is considered "processed" to avoid generating agents that repeatedly retrieve the same content.

[0307] The information comparison module 30 is used to compare the preliminary results with the standard examples in the current task context and generate a difference report.

[0308] In this embodiment of the invention, the above-mentioned comparison stage is performed by a solidified intelligent agent. The solidified intelligent agent performs multi-dimensional comparisons of the preliminary results and standard examples to generate a clear and objective difference report.

[0309] In this embodiment of the invention, the solidified agent is another agent in the agent group. The solidified agent is not the LLM itself, but its architecture is basically the same as that of the generated agent.

[0310] A group of agents can share a single LLM, meaning the same LLM is embedded in both the generated and fixed agents. By creating cue words with different structures, the LLM can infer different content. Alternatively, a group of agents can use different LLMs, meaning different LLMs are embedded in both the generated and fixed agents. No specific restrictions are placed on the LLMs embedded in the generated and fixed agents.

[0311] In this embodiment of the invention, the standard examples are obtained from historical project information or expert output. The standard examples contain the implicit knowledge of the experts. For example, the standard examples are written by professional automotive electronics test and development engineers based on enterprise specifications and / or their own accumulated experience.

[0312] In this embodiment of the invention, the solidified agent will generate itemized difference reports. Through the itemized difference reports, each difference identified by the solidified agent will form a record. The record format is as follows: Preliminary result fragment: "<fragment>" | Standard example fragment: "<fragment>" | Attribution description: "<Brief description of the tacit knowledge gap reflected by the difference>".

[0313] By using itemized difference reports, key parts of the original context can be preserved through fragments, and the length of the fragments can be limited in this way, for example, the fragment length should not exceed 100 characters, so as to avoid exceeding the context length limit of LLM, reduce the output time of LLM and reduce the generation cost, and facilitate the extraction of tacit knowledge later.

[0314] The knowledge solidification module 40 is used to assemble the difference report, standard examples, valid retrieval records and the current task context into the second prompt word of the solidified agent, and call the large language model embedded in the solidified agent to reason about the second prompt word, extract implicit knowledge based on the difference report, and solidify the implicit knowledge contained in the standard examples.

[0315] In this embodiment of the invention, the LLM embedded in the solidified agent verifies the identified discrepancy reports with existing tacit knowledge entries to extract new, solidifiable tacit knowledge. When a new case is found to conflict with existing rules, the rule is located, and a "modification" suggestion is generated; or, if the rule is outdated, a "deletion" suggestion is generated. All this tacit knowledge is generated in a structured entry format and submitted to the standard example generator or other experts for final adjudication. Specifically, when the standard example is written by an expert, the expert is also the generator of the standard example, and it is subsequently submitted to the corresponding generator for final adjudication. When the standard example originates from historical project information, considering that the experts who wrote the historical project may not be able to participate in the review process, it can be submitted to other experts for final adjudication.

[0316] Established tacit knowledge can be stored in a specification knowledge base. When test cases need to be generated again, the corresponding tacit knowledge can be retrieved from the specification knowledge base to guide the generation of test cases that closely resemble standard examples. The content in the specification knowledge base can also be directly used as a design guide for training new employees, transforming the tacit knowledge and business logic of experts into manageable and reusable standards, effectively preventing knowledge loss due to personnel turnover.

[0317] The implicit knowledge solidification device of this invention generates an intelligent agent and produces preliminary results based on the current task context, explicit knowledge, and solidified implicit knowledge. Compared with the standard example generated by the expert, the preliminary results can reflect the missing implicit knowledge. By comparing the preliminary results with the standard example of the current task context, a difference report is generated. This difference report contains the missing implicit knowledge and further refines the implicit knowledge, thus solidifying the implicit knowledge that exists in the expert's mind and is reflected in the standard example, resulting in solidified implicit knowledge. By setting up a factual knowledge base constructed from explicit knowledge and a normative knowledge base constructed from solidified implicit knowledge, the generated intelligent agent identifies all unprocessed items in the current task context and performs problem-driven factual retrieval and / or intent-driven normative retrieval based on the type of unprocessed items, thereby improving the effectiveness of information retrieval and collecting as much relevant information as possible. By comparing and then refining, and using the concept of thought chains, the large language model outputs a series of intermediate, logically coherent reasoning steps before giving the final implicit knowledge, thereby improving the accuracy of handling complex problems and making the subsequently solidified implicit knowledge more accurate. This invention systematically mines and extracts the tacit knowledge of experts, solidifies the tacit knowledge, and thereby enables the efficient accumulation and reuse of enterprise knowledge assets as well as structured management.

[0318] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4As shown, the electronic device may include: a processor 410, a communication interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communication interface 420, and the memory 430 communicate with each other through the communication bus 440. The processor 410 can call logical commands in the memory 430 to execute a method for solidifying implicit knowledge, which includes:

[0319] The system obtains the user's current task context, uses a generative agent to identify unprocessed items in the current task context, and retrieves them from a preset factual knowledge base and / or normative knowledge base to generate valid retrieval records. The factual knowledge base stores factual information, and the normative knowledge base stores intentional information. Factual information comes from explicit knowledge and is declarative information, while intentional information comes from fixed implicit knowledge and is both instructional and rule-based information.

[0320] The effective retrieval records and the current task context are assembled into the first prompt word of the generated agent, and the large language model embedded in the generated agent is called to reason about the first prompt word to obtain preliminary results;

[0321] The preliminary results are compared with standard examples in the current task context to generate a discrepancy report;

[0322] The difference report, standard examples, valid retrieval records, and current task context are assembled into a second prompt word for the solidified agent. The large language model embedded in the solidified agent is then used to reason about the second prompt word. Implicit knowledge is extracted from the difference report, and implicit knowledge contained in the standard examples is solidified.

[0323] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0324] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is capable of executing the method for solidifying implicit knowledge provided by the above methods, the method comprising:

[0325] The system obtains the user's current task context, uses a generative agent to identify unprocessed items in the current task context, and retrieves them from a preset factual knowledge base and / or normative knowledge base to generate valid retrieval records. The factual knowledge base stores factual information, and the normative knowledge base stores intentional information. Factual information comes from explicit knowledge and is declarative information, while intentional information comes from fixed implicit knowledge and is both instructional and rule-based information.

[0326] The effective retrieval records and the current task context are assembled into the first prompt word of the generated agent, and the large language model embedded in the generated agent is called to reason about the first prompt word to obtain preliminary results;

[0327] The preliminary results are compared with standard examples in the current task context to generate a discrepancy report;

[0328] The difference report, standard examples, valid retrieval records, and current task context are assembled into a second prompt word for the solidified agent. The large language model embedded in the solidified agent is then used to reason about the second prompt word. Implicit knowledge is extracted from the difference report, and implicit knowledge contained in the standard examples is solidified.

[0329] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the aforementioned methods for solidifying implicit knowledge, the methods comprising:

[0330] The system obtains the user's current task context, uses a generative agent to identify unprocessed items in the current task context, and retrieves them from a preset factual knowledge base and / or normative knowledge base to generate valid retrieval records. The factual knowledge base stores factual information, and the normative knowledge base stores intentional information. Factual information comes from explicit knowledge and is declarative information, while intentional information comes from fixed implicit knowledge and is both instructional and rule-based information.

[0331] The effective retrieval records and the current task context are assembled into the first prompt word of the generated agent, and the large language model embedded in the generated agent is called to reason about the first prompt word to obtain preliminary results;

[0332] The preliminary results are compared with standard examples in the current task context to generate a discrepancy report;

[0333] The difference report, standard examples, valid retrieval records, and current task context are assembled into a second prompt word for the solidified agent. The large language model embedded in the solidified agent is then used to reason about the second prompt word. Implicit knowledge is extracted from the difference report, and implicit knowledge contained in the standard examples is solidified.

[0334] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0335] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for solidifying tacit knowledge, characterized in that, The method includes: The system acquires the user's current task context, uses a generative agent to identify unprocessed items within that context, and then retrieves these items from a pre-defined factual knowledge base and / or normative knowledge base to generate valid retrieval records. The factual knowledge base stores factual information, while the normative knowledge base stores intentional information. Factual information originates from explicit knowledge and is declarative, while intentional information originates from fixed implicit knowledge and is both instructional and rule-based. Unprocessed items consist of points of doubt and state points. If there are ambiguities in the factual information within the current task context, each ambiguous factual information constitutes a point of doubt. If the current task context differs from historical task intentions, each different task intention constitutes a state point. The effective retrieval records and the current task context are assembled into the first prompt word of the generated agent, and the large language model embedded in the generated agent is called to reason about the first prompt word to obtain preliminary results; The preliminary results are compared with standard examples in the current task context to generate a discrepancy report; The difference report, standard examples, valid retrieval records, and current task context are assembled into a second prompt word for the solidified agent. The large language model embedded in the solidified agent is then used to reason about the second prompt word. Implicit knowledge is extracted from the difference report, and implicit knowledge contained in the standard examples is solidified.

2. The method for solidifying tacit knowledge according to claim 1, characterized in that, The process of obtaining the user's current task context, using a generative agent to determine unprocessed items in the current task context, and then having the generative agent search for these unprocessed items in a preset factual knowledge base and / or normative knowledge base to generate valid retrieval records for the unprocessed items specifically includes: Get the current task context; Use a generative agent to determine at least one point of doubt and / or at least one state point in the current task context; The agent determines the number of retrieval rounds based on the number of question points and state points, generates search terms for each round of retrieval, and performs retrieval based on the search terms and using a factual knowledge base and / or a normative knowledge base to generate valid retrieval records; in each round of retrieval, one question point and / or one state point are searched. Use a generative agent to identify at least one point of doubt and / or at least one state point in valid search records; The agent determines the number of retrieval rounds based on the number of question points and state points, generates search terms for each round of retrieval, and performs retrieval based on the search terms and using a factual knowledge base and / or a normative knowledge base, supplementing the valid retrieval records.

3. The method for solidifying tacit knowledge according to claim 2, characterized in that, The generating agent determines the retrieval rounds based on the number of question points and state points, generates search terms for each round, and performs retrieval based on the search terms using a factual knowledge base and / or a normative knowledge base to generate valid search records, specifically including: The generated agent determines the number of retrieval rounds based on the number of question points and state points, and determines the corresponding question points and / or state points for each round of retrieval. Based on the points of doubt, factual search terms are generated to query the factual content of the points of doubt, and intent search terms are generated based on the status points to reflect the task intent of the status points. Determine the search content for each round of retrieval, identify the valid content from the search content, and add the identified valid content to the end of the existing content; Collect all valid content from all rounds of retrieval to obtain valid search records.

4. The method for solidifying tacit knowledge according to claim 1, characterized in that, The step of comparing the preliminary results with a standard paradigm in the current task context to generate a difference report specifically includes: A standard example of obtaining the current task context; Determine the preset comparison items and preset ignore items; Based on the preset comparison items, the solidified intelligent agent compares the preliminary results with the standard examples item by item to obtain a difference report.

5. The method for solidifying tacit knowledge according to claim 1, characterized in that, The process involves assembling the difference report, standard examples, valid retrieval records, and the current task context into a second prompt word for the solidified agent, and then inferring the second prompt word using the large language model embedded in the solidified agent. This process extracts implicit knowledge from the difference report and solidifies the implicit knowledge contained in the standard examples. Specifically, this includes: The difference report, standard examples, valid search records, and current task context are assembled into the second prompt word of the solidified agent; The large language model embedded in the solidified intelligent agent is invoked to reason about the second prompt word, and the reasoning result is obtained. Each reasoning content that constitutes the reasoning result is summarized into a preset style, and the type of each reasoning content is determined according to the preset style. Given that the type of reasoning content is determined to be the first type, implicit knowledge of the deletion class is generated; Given that the type of reasoning content is determined to be the second type, implicit knowledge of the modified class is generated; Given that the type of reasoning content is determined to be the third type, implicit knowledge of the additional class is generated.

6. The method for solidifying tacit knowledge according to claim 5, characterized in that, The method further includes: The tacit knowledge is reviewed and processed, and the approved tacit knowledge is added to the normative knowledge base.

7. The method for solidifying tacit knowledge according to claim 6, characterized in that, The process of reviewing tacit knowledge and adding the approved tacit knowledge to the standardized knowledge base specifically includes: Obtain all implicit knowledge output by the fixed intelligent agent; Each piece of tacit knowledge is reviewed and processed. If it is determined that the tacit knowledge does not require adjustment, add the tacit knowledge to the normative knowledge base; If it is determined that tacit knowledge needs to be adjusted, the tacit knowledge is adjusted and added to the normative knowledge base.

8. A device for solidifying tacit knowledge, characterized in that, The device includes: The hybrid retrieval module is used to obtain the user's current task context, utilize the generative agent to identify unprocessed items in the current task context, and then use the generative agent to search for unprocessed items in a preset factual knowledge base and / or normative knowledge base to generate valid retrieval records for the unprocessed items. The current task context consists of the user's requirement text and product background. The factual knowledge base is used to store factual information, and the normative knowledge base is used to store intentional information. Factual information comes from explicit knowledge and is declarative information, while intentional information comes from fixed implicit knowledge and is instructional and rule-based information. Unprocessed items consist of question points and state points. If it is determined that there are ambiguities in the factual information involved in the current task context, each ambiguous factual information will form a question point. If it is determined that the current task context has different task intentions compared to the historical task intentions, each different task intention will form a state point. The preliminary reasoning module is used to assemble the valid search records and the current task context into the first prompt word of the generated agent, and call the large language model embedded in the generated agent to reason about the first prompt word to obtain preliminary results; The information comparison module is used to compare preliminary results with standard examples in the current task context and generate a difference report; The knowledge solidification module is used to assemble the difference report, standard examples, valid retrieval records, and current task context into the second prompt word of the solidified agent, and call the large language model embedded in the solidified agent to reason about the second prompt word, extract implicit knowledge based on the difference report, and solidify the implicit knowledge contained in the standard examples.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method for solidifying tacit knowledge as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method for solidifying tacit knowledge as described in any one of claims 1 to 7.

Images