Prompt conversion apparatus

Abstract

To provide a prompt conversion device for converting an input prompt before transmitting it to a generation AI in order to reduce the risk of information leak due to the use of the generation AI.SOLUTION: The query conversion device 1 includes a word / phrase extraction module 11 that extracts a plurality of words / phrases from an original query, a meta information generation module 12 that generates a first combination of a first code and first meta information corresponding to each of the words / phrases, a replacement module 13 that replaces a word / phrase in the original query with the first code to generate a coded query, and a transmission / reception module 14 that transmits a prompt including the coded query and the first combination to a generation AI3 and receives a response from the generation AI3.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a prompt conversion device that prevents information leakage to a generation AI by replacing a part of the prompt sent to the generation AI. [Background technology]

[0002] In recent years, generative AI has made rapid progress in handling various tasks such as answering questions, translating, and generating text. It can quickly organize and analyze large amounts of data, automate routine tasks, and even provide initial suggestions for advertising copy. Utilizing generative AI in business operations can be expected to improve operational efficiency and reduce human error, allowing human capital to be allocated to more creative tasks. Summary of the Invention [Problem to be solved by the invention]

[0003] Generative AI is built by learning from large amounts of data, and useful learning targets include the data contained in the prompts sent to the generative AI and further questions posed in response to the generative AI's responses to those prompts. The information learned and accumulated in the generative AI can potentially be used to provide answers to other users of the same generative AI service.

[0004] To mitigate concerns about information leaks, some services claim that they do not learn the information entered, but this does not immediately mean that confidential information can be entered with peace of mind.On the other hand, if generation AI is run in a local environment, such as an in-house server, rather than provided via the web, the risk of information leaks can be avoided, but local generation AI is often not sufficient in terms of accuracy or convenience.

[0005] One aspect of the present invention relates to a prompt conversion device that performs data conversion on input prompts before sending them to a generation AI in order to reduce the risk of information leakage due to the use of a generation AI. [Means for solving the problem]

[0006] A prompt conversion device according to one aspect of the present invention comprises: a phrase extraction module that extracts a plurality of phrases from the query source; a meta information generation module that generates a first combination of a first code and first meta information corresponding to each of the phrases; a substitution module that substitutes the first code for the phrase in the original query to generate a coded query; a transceiver module that transmits a prompt including the coded query and the first combination to a generation AI and receives a response from the generation AI; Includes: [Brief explanation of the drawings]

[0007] [Figure 1] FIG. 1 shows a prompt conversion device 1 according to the first embodiment and an external device connected to the prompt conversion device 1. As shown in FIG. [Figure 2] FIG. 2 shows the functions and operations of the prompt conversion device 1 according to the first embodiment. [Figure 3] FIG. 3 shows an example of an original query received from the input terminal 2 in the first embodiment. [Figure 4] FIG. 4 shows examples of phrases extracted from the original query text by the phrase extraction module 11 in the first embodiment. [Figure 5] FIG. 5 shows an example of a first combination of a first code and first meta information generated by the meta information generating module 12 in the first embodiment. [Figure 6] FIG. 6 shows an example of a prompt including the coded query generated by the substitution module 13 and the first combination generated by the meta information generation module 12 in the first embodiment. [Figure 7] FIG. 7 shows an example of a response obtained from generation AI3 in the first embodiment. [Figure 8] FIG. 8 shows the functions and operations of a prompt conversion device 1a according to the second embodiment. [Figure 9] FIG. 9 shows an example of an original query received from the input terminal 2 in the second embodiment. [Figure 10] FIG. 10 shows an example of a design drawing attached as a data file to a query text in the second embodiment. [Figure 11] FIG. 11 shows an example of a specification attached as a data file to the query text in the second embodiment. [Figure 12] FIG. 12 shows an example of a sales plan attached as a data file to the query text in the second embodiment. [Figure 13] FIG. 13 shows examples of phrases extracted from the original query text by the phrase extraction module 11 in the second embodiment. [Figure 14] FIG. 14 shows an example of a first combination of a first code and first meta information generated by the meta information generating module 12 in the second embodiment. [Figure 15] FIG. 15 shows an example of text acquired by the text acquisition module 10 based on the data file of the design drawing in the second embodiment. [Figure 16] FIG. 16 shows examples of words and numerical values ​​extracted by the word extraction module 11 from the text-converted design drawing in the second embodiment. [Figure 17] FIG. 17 shows an example of a second combination of a second code and second meta information generated by the meta information generating module 12 in the second embodiment. [Figure 18] FIG. 18 shows an example of a converted numerical value generated by the meta information generating module 12 in the second embodiment. [Figure 19] FIG. 19 shows examples of words and numerical values ​​extracted from specifications by the word extraction module 11 in the second embodiment. [Figure 20] FIG. 20 shows an example of a second combination of a second code and second meta information generated by the meta information generating module 12 in the second embodiment. [Figure 21] FIG. 21 shows an example of a converted numerical value generated by the meta information generating module 12 in the second embodiment. [Figure 22] FIG. 22 shows examples of words and numerical values ​​extracted from a sales plan by the word extraction module 11 in the second embodiment. [Figure 23] FIG. 23 shows an example of a second combination of a second code and second meta information generated by the meta information generating module 12 in the second embodiment. [Figure 24] FIG. 24 shows an example of a converted numerical value generated by the meta information generating module 12 in the second embodiment. [Figure 25] FIG. 25, together with FIGS. 26 to 29, shows an example of a prompt including a coded query generated by the replacement module 13 and the first and second combinations generated by the meta information generation module 12 in the second embodiment. [Figure 26] FIG. 26, along with FIG. 25 and FIGS. 27 to 29, shows examples of prompts. [Figure 27] Figure 27, along with Figures 25, 26, 28 and 29, shows examples of prompts. [Figure 28] FIG. 28, along with FIGS. 25 to 27 and 29, shows examples of prompts. [Figure 29] FIG. 29, along with FIGS. 25 to 28, shows examples of prompts. [Figure 30] FIG. 30 shows an example of a response obtained from generation AI3 in the second embodiment. [Figure 31] FIG. 31 shows an example of an original query received from the input terminal 2 in the modified example. [Figure 32] FIG. 32 shows an example of a phrase extracted from the original query by the phrase extraction module 11 in the modified example, and a first combination of a first code and first meta information generated by the meta information generation module 12 in the modified example. [Figure 33] FIG. 33, together with FIG. 34, shows an example of a prompt including the coded query generated by the replacement module 13 and the first combination generated by the meta information generation module 12 in the modified example. [Figure 34] Figure 34, together with Figure 33, shows an example of a prompt. [Figure 35] FIG. 35 shows an example of a response obtained from generation AI3 in the modified example. DETAILED DESCRIPTION OF THE INVENTION

[0008] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. Each embodiment described below shows an example of the present invention and does not limit the content of the present invention. Furthermore, not all of the configurations and operations described in each embodiment are necessarily essential as the configurations and operations of the present invention. Note that the same components are given the same reference numerals, and redundant explanations will be omitted.

[0009] <1. First embodiment> <1-1.Configuration> FIG. 1 shows a prompt conversion device 1 according to the first embodiment and an external device connected to the prompt conversion device 1. The prompt conversion device 1 is a computer system including a CPU, memory, and the like (not shown). The prompt conversion device 1 may be configured as a single computer or may be configured as multiple computers connected via a network. The prompt conversion device 1 is connected to external devices such as an input terminal 2 and a generation AI 3. The prompt conversion device 1 sends a prompt to the generation AI 3 and receives a response from the generation AI 3.

[0010] The input terminal 2 is a terminal operated by a user such as an employee who uses the generation AI 3, and is a computer system equipped with an input device, an output device, a CPU, a memory, etc. (not shown). The input terminal 2 can access the prompt conversion device 1, transmits a query input by the user to the prompt conversion device 1, and receives a response from the generation AI 3 from the prompt conversion device 1. The query input by the user is hereinafter referred to as the original query.

[0011] The generation AI 3 includes a large-scale language model (LLM). The large-scale language model is a language model constructed using large amounts of text data and deep learning technology, and processes tasks such as answering questions, translation, text generation, text summarization, and sentiment analysis in response to prompts sent from the prompt conversion device 1. It is desirable for the large-scale language model to have an attention mechanism. The attention mechanism is a mechanism for extracting important parts from the input, and greatly contributes to improving the processing speed and accuracy of the large-scale language model.

[0012] <1-2. Functions and operations> 2 shows the functions and operations of the prompt conversion device 1 according to the first embodiment. The prompt conversion device 1 includes a word / phrase extraction module 11, a meta-information generation module 12, a replacement module 13, and a transmission / reception module 14. These modules are implemented by loading programs into memory included in the prompt conversion device 1 and executing them by a CPU.

[0013] The phrase extraction module 11 extracts multiple terms from the original query received from the input terminal 2. The multiple terms are, for example, highly important terms from the original query. The terms do not need to be limited to terms that require confidentiality, and the phrase extraction module 11 does not need to determine the level of confidentiality. The highly important terms may be, for example, terms with a score higher than a threshold, such as TFIDF (term frequency-inverse document frequency), or may be terms extracted by a local generation AI (not shown). The local generation AI is a generation AI separate from the generation AI 3 and includes a large-scale language model or a small-scale language model (SLM) that operates in a local environment. The local generation AI does not need to be capable of advanced analysis or highly accurate answer generation like the generation AI 3; it is sufficient if it can extract terms and generate meta-information (described below). Examples of the original query and extracted terms will be described later with reference to Figures 3 and 4.

[0014] The meta information generation module 12 generates a first combination of a first code and first meta information for each of the words extracted from the original query by the word extraction module 11. The first code is a unique code assigned to each extracted word. The first meta information is information describing each extracted word, including, for example, a "part of speech," "type," and "meaning." The "part of speech" may indicate whether the word is a common noun or a proper noun, the "type" may indicate the contextual position of the word, and the "meaning" may indicate a word or phrase that expresses a higher-level concept of the word. The "type" may be selected from pre-prepared options. A single word or phrase may have multiple "meanings." It is desirable for the first meta information to be generated taking into account the context of the original query, and it may be generated by the local generation AI described above. An example of a first combination of a first code and first meta information will be described later with reference to FIG. 5.

[0015] The substitution module 13 generates a coded query by substituting the phrase extracted by the phrase extraction module 11 with the first code generated by the meta information generation module 12 in the original query. The transmission / reception module 14 transmits a prompt to the generation AI 3, including the coded query generated by the substitution module 13 and the first combination generated by the meta information generation module 12, and receives a response from the generation AI 3. By including a coded query in which important phrases in the original query are substituted with the first code in the prompt, the risk of confidential information leakage can be reduced while maintaining minimum readability. The first meta information can include multifaceted information. By transmitting the first meta information in association with the first code, the generation AI 3 can understand the context of the coded query while excluding specific information, thereby eliciting a response from the generation AI 3. An example of a prompt will be described below with reference to FIG. 6, and an example of a response from the generation AI 3 will be described below with reference to FIG. 7.

[0016] <1-3. Specific examples> FIG. 3 shows an example of a query text received from the input terminal 2 in the first embodiment. The query text entered by the user is for a request for code name ideas for a new product development project. It includes specific names such as "Tanaka Coffee Shop" and "Chronoir," as well as information about the new product development project, the existence of which should be confidential. If this query text were sent as is to the generation AI 3, there is a risk of confidential information being leaked.

[0017] 4 shows examples of phrases extracted from the original query text by the phrase extraction module 11 in the first embodiment. Phrases such as "Tanaka Coffee Shop" and "Chronoir" are extracted.

[0018] FIG. 5 shows an example of a first combination of a first code and first meta information generated by the meta information generation module 12 in the first embodiment. A first code and first meta information are generated for each of the phrases shown in FIG. 4. Some of the meta information may be left blank, such as the "type" field for "gift from God." The prompt conversion device 1 stores, in a memory (not shown), the correspondence between the phrases extracted by the phrase extraction module 11 and the first codes generated by the meta information generation module 12.

[0019] FIG. 6 shows an example of a prompt including the coded query generated by the replacement module 13 and the first combination generated by the meta information generation module 12 in the first embodiment. The phrase extracted by the phrase extraction module 11 is not included in the prompt. By reading the coded query while referring to the first combination of the first code and the first meta information, it is possible to roughly understand the content required of the generation AI 3, but no specific information is available. Furthermore, it is desirable that the first code generated by the meta information generation module 12 is randomly generated each time a query text is processed. A change in the correspondence between the first code and the first meta information each time can cause noise in the learning by the generation AI 3, making learning more difficult.

[0020] FIG. 7 shows an example of a response obtained by the generation AI 3 in the first embodiment. The response shown in FIG. 7 is a general response that can be generated even without detailed information such as "Tanaka Coffee Shop" or "Chronoir." However, it proposes three generally positive, future-oriented code names suitable for new product development. Even if the prompt includes specific terms such as "Tanaka Coffee Shop" or "Chronoir," the response that the generation AI 3 can provide will likely be general unless the generation AI 3 has detailed information about those terms. Therefore, the value of the response obtained by the generation AI 3 is not significantly reduced by concealing terms such as "Tanaka Coffee Shop" or "Chronoir." Thus, the present invention is highly useful when a general response is sufficient for the response requested from the generation AI 3. For example, the present invention is highly effective when a non-expert wants to obtain general information, or when an expert wants to abstract a problem and consider a wide range of ideas from a general perspective.

[0021] The prompt conversion device 1 may reverse-convert the first code included in the response from the generation AI 3 back into the original phrase. By transmitting the reverse-converted response to the input terminal 2, readability for the user is improved.

[0022] Alternatively, the prompt shown in Figure 6 may request the generation AI 3 to "use appropriate words without using codes when answering." Because the generation AI 3 does not have information on the correspondence between the words extracted by the word extraction module 11 and the first code generated by the meta information generation module 12, it cannot use the original words. However, because it has the first meta information, it can generate an answer that replaces the original words with a superordinate concept, for example. This also improves readability for the user.

[0023] <1-4. Effects> According to the first embodiment, the prompt conversion device 1: a phrase extraction module 11 that extracts a plurality of phrases from the query source text; a meta information generation module 12 that generates a first combination of a first code and first meta information corresponding to each of the phrases; a substitution module 13 for substituting a first code for a phrase in the original query to generate a coded query; a transceiver module 14 for sending a prompt including the coded query and the first combination to the generation AI 3 and receiving a response from the generation AI 3; Includes:

[0024] According to this, by sending a coded query in which a phrase in a query text is replaced with a first code and a first combination of the first code and first meta information to the generation AI 3, an answer can be elicited from the generation AI 3 without providing specific information to the generation AI 3, thereby reducing the risk of information leakage when using the generation AI 3. The meta information assigned to a single code can include multifaceted information, such as part of speech, type, and meaning, thereby more accurately conveying the intent of the query text than simply providing the generation AI 3 with information that is a superordinate concept of the original phrase. Furthermore, since the meta information generation module 12 generates a first code and first meta information each time a query text is processed, the correspondence between the first code and the first meta information changes each time, which can cause noise in the learning process by the generation AI 3 and make learning more difficult. The prompt conversion device 1 can convert the first code back to the original phrase in the answer from the generation AI 3, thereby generating an answer that is easy for the user to read.

[0025] 2. Second embodiment <2-1. Functions and operations> 8 shows the functions and operations of a prompt conversion device 1a according to the second embodiment. In addition to the various modules included in the prompt conversion device 1 described with reference to FIG. 2, the prompt conversion device 1a further includes a text acquisition module 10 and an inverse conversion module 15. These modules are implemented by loading programs into memory included in the prompt conversion device 1a and executing them by a CPU.

[0026] <2-1-1. Processing data files> The input terminal 2 transmits the query text input by the user as well as a data file attached to the query text to the prompt conversion device 1a. The data file is a file specified by the user. Examples of the query text and the data file will be described later with reference to FIGS. 9 to 12.

[0027] The text acquisition module 10 acquires text information based on a data file received from the input terminal 2. If the data file is a text file, it is possible to simply read the text directly from the data file. If the data file is a file containing an image, it is possible to recognize characters contained in the image using optical character recognition (OCR), or to acquire text describing the image from the image using a local generative AI using a vision language model (VLM). Acquisition of text information from a data file will be described later with reference to FIGS. 10 and 15.

[0028] The phrase extraction module 11 extracts a plurality of phrases from each of the original query received from the input terminal 2 and the text information acquired by the text acquisition module 10 based on the data file. Examples of phrase extraction will be described later with reference to FIGS. 13, 16, 19, and 22.

[0029] The meta information generation module 12 generates a first combination of a first code and first meta information corresponding to each of the terms extracted from the query text, as well as a second combination of a second code and second meta information corresponding to each of the terms extracted from the text information. The second code is different from the first code. By assigning different codes to terms commonly included in the query text and the text information, meta information can be generated that takes into account the context of each of the query text and the text information. However, the different correspondence between codes and meta information between the query text and the text information may make learning by the generation AI 3 more difficult. Examples of the first and second combinations will be described later with reference to FIGS. 14, 17, 20, and 23.

[0030] The substitution module 13 generates a coded query by substituting the first code for the phrase in the original query and the second code for the phrase in the text information obtained based on the data file. The transmission / reception module 14 transmits a prompt including the coded query and the first and second combinations to the generation AI 3 and receives a response from the generation AI 3. Examples of the prompt will be described later with reference to Figures 25 to 29, and an example of the response from the generation AI 3 will be described later with reference to Figure 30.

[0031] <2-1-2. Numerical Processing> The query text received from the input terminal 2 or the text information acquired based on the data file may contain numerical values. For example, there are cases where it is desired to prevent numerical values, such as those contained in detailed product designs or accounting information, from being leaked to the outside as they are.

[0032] Therefore, the phrase extraction module 11 may extract multiple numerical values ​​in addition to multiple phrases. For example, if the text information acquired by the text acquisition module 10 includes multiple numerical values, the numerical values ​​may be extracted from the text information. Examples of extracted numerical values ​​will be described later with reference to Figures 16, 19, and 22.

[0033] The meta information generation module 12 not only generates a first combination of the first code and the first meta information, but also generates a converted numerical value by converting the numerical value extracted by the phrase extraction module 11 using a first function, which is a monotonically increasing function. For example, if the extracted numerical value is X, A is a positive number, and B is an arbitrary real number, the first function may be a linear function expressed as AX+B. A is, for example, between 0.68 and 1.32 and is set each time using a random number. By using a monotonically increasing function, a converted numerical value that maintains the magnitude relationship between multiple numerical values ​​and allows for inverse conversion can be generated, thereby reducing the risk of leakage of the original numerical value. Examples of converted numerical values ​​will be described later with reference to Figures 18, 21, and 24. Note that instead of converting numerical values ​​using a monotonically increasing function, the risk of leakage of numerical values ​​may be reduced by abstracting the numerical value (e.g., replacing "3 mm" with "dimension" or "3,500 million" with "large amount") or by converting the numerical value using a function that is not a monotonically increasing function. However, a monotonically increasing function is preferable when the generation AI 3 performs the calculation. Also, if date information is included as a numerical value, for example, the year or month of the date may be shifted, or a higher-level concept such as the 2000s may be used.

[0034] The substitution module 13 generates a coded query by substituting a converted numerical value for a numerical value in the original query or text information in addition to substituting a first code for a word. In order to prevent the influence of the numerical value modification from affecting the meta information, it is preferable that the substitution module 13 performs the substitution with the converted numerical value after the meta information generation module 12 generates the meta information.

[0035] The inverse transformation module 15 corrects the answer by inversely transforming the numerical value included in the answer received by the sending / receiving module 14 from the generating AI 3 using a second function that is an inverse function of the first function.

[0036] <2-2. Specific examples> 9 shows an example of a query text received from the input terminal 2 in the second embodiment. The query text input by the user requests the generation AI 3 to analyze the cause of an accident involving a product sold by the company and to simulate the profit and loss associated with a recall. The query text also requests reference to blueprints and specifications for the analysis of the cause of the accident, and to reference sales plan documents for the profit and loss simulation.

[0037] FIG. 10 shows an example of a design drawing attached as a data file to a query text in the second embodiment. FIG. 11 shows an example of a specification attached as a data file to a query text in the second embodiment. FIG. 12 shows an example of a sales plan attached as a data file to a query text in the second embodiment. All of FIGS. 10 to 12 are intended to explain the configuration of the embodiments, and are therefore shown as simply as possible.

[0038] The query text shown in Figure 9 includes the specific name "ZX-1000" as well as information that could be legally unfavorable, such as the recognition of an accident or the need for a recall. Furthermore, the design drawings shown in Figure 10 and the specifications shown in Figure 11 may contain technical confidentiality, and the sales plan shown in Figure 12 may contain business confidentiality. If such query text, design drawings, specifications, and sales plans are sent to the generation AI 3 as is, there is a risk of confidential information being leaked.

[0039] 13 shows examples of phrases extracted from the original query text by the phrase extraction module 11 in the second embodiment. Phrases such as "main product" and "ZX-1000" are extracted.

[0040] Fig. 14 shows an example of a first combination of a first code and first meta information generated by the meta information generation module 12 in the second embodiment. A first code and first meta information are generated for each of the words and phrases shown in Fig. 13. The prompt conversion device 1a stores in a memory (not shown) the correspondence between the words and phrases extracted by the word and phrase extraction module 11 and the first codes generated by the meta information generation module 12.

[0041] FIG. 15 shows an example of text acquired by the text acquisition module 10 based on the data file of the design drawing in the second embodiment. is the line feed code.

[0042] 16 shows examples of words and numerical values ​​extracted from the text-converted blueprint by the word extraction module 11 in the second embodiment. Words such as "brake lever" and "cable" and numerical values ​​such as "3" and "180" are extracted.

[0043] 17 shows an example of a second combination of second codes and second meta information generated by the meta information generation module 12 in the second embodiment. A second code and second meta information are generated for each of the phrases shown in FIG.

[0044] Fig. 18 shows an example of converted numerical values ​​generated by the meta information generation module 12 in the second embodiment. A converted numerical value is generated for each of the original numerical values ​​shown in Fig. 16. When the original numerical value is Xa, Xa × 0.837 is used as the first function for conversion.

[0045] 19 shows examples of words and numerical values ​​extracted from specifications by the word extraction module 11 in the second embodiment. Words such as "brake cable" and "wire material" and numerical values ​​such as "3.0" and "1800" are extracted.

[0046] 20 shows an example of a second combination of second codes and second meta information generated by the meta information generation module 12 in the second embodiment. A second code and second meta information are generated for each of the phrases shown in FIG.

[0047] FIG. 21 shows an example of a converted numerical value generated by the meta information generation module 12 in the second embodiment. A converted numerical value is generated for each of the original numerical values ​​shown in FIG. 19. When the original numerical value is Xb, Xb×1.32 is used as the first function for conversion. However, for the number of pistons "4," converting it to a value including a fraction less than 1 could cause misunderstanding in the generation AI3, so the converted numerical value is rounded to "5."

[0048] 22 shows examples of words and numerical values ​​extracted from a sales plan by the word extraction module 11 in the second embodiment. Words such as "scooter" and "model" and numerical values ​​such as "20,000" and "35,000" are extracted.

[0049] 23 shows an example of a second combination of second codes and second meta information generated by the meta information generation module 12 in the second embodiment. A second code and second meta information are generated for each of the phrases shown in FIG.

[0050] Fig. 24 shows an example of converted numerical values ​​generated by the meta information generation module 12 in the second embodiment. A converted numerical value is generated for each of the original numerical values ​​shown in Fig. 22. When the original numerical value is Xc, Xc × 0.792 is used as the first function for conversion.

[0051] 25 to 29 show examples of prompts including a coded query generated by the replacement module 13 and first and second combinations generated by the meta information generation module 12 in the second embodiment. Although FIGS. 25 to 29 show one prompt as a whole, they are shown separately in FIGS. 25 to 29 because they cannot fit in one figure. In FIG. 25, document 1 corresponds to the design drawing, and in FIG. 26, document 2 corresponds to the specifications, and document 3 corresponds to the sales plan. In FIG. 27, the combination of meta information and code of the query corresponds to the first combination. In FIGS. 27 to 29, the combination of meta information and code of documents 1 to 3 corresponds to the second combination.

[0052] The words extracted by the phrase extraction module 11 are not included in the prompt. In addition, the numerical values ​​extracted by the phrase extraction module 11 are replaced with converted numerical values. By reading the coded query while referring to the first combination and reading materials 1 to 3 while referring to the second combination, it is possible to roughly understand the content required of the generation AI 3, but no specific information about accidents, recalls, etc. is known. In addition, the numerical values ​​have been altered, so it is not possible to associate the content from the numerical values.

[0053] FIG. 30 shows an example of an answer obtained by the generation AI 3 in the second embodiment. However, this example shows the first and second codes included in the answer from the generation AI 3 reverse-converted to the original phrase by the prompt conversion device 1a. The answer shown in FIG. 30 is a general answer that can be generated even without detailed information such as "ZX-1000." However, even if a specific phrase such as "ZX-1000" is included in the prompt, the answer that the generation AI 3 can provide will often be a general answer unless the generation AI 3 has detailed information about "ZX-1000." Therefore, the value of the proposal obtained by the generation AI 3 is not significantly reduced by hiding phrases such as "ZX-1000" from the generation AI 3. For example, the present invention is highly effective when a non-expert wants to obtain general information, or when an expert wants to abstract a problem and consider a wide range of ideas from a general perspective.

[0054] In other respects, the second embodiment is similar to the first embodiment.

[0055] <2-3. Modifications> FIG. 31 shows an example of a query text received from the input terminal 2 in a modified example. In the second embodiment, the query text described with reference to FIG. 9 was used to request the generation AI 3 to simulate profits and losses associated with a recall. However, the response from the generation AI 3 described with reference to FIG. 30 did not provide sufficient simulation results. This is presumably because the prompts provided to the generation AI 3 in the second embodiment were vague and lacked the information necessary to perform a profit and loss simulation. Therefore, in FIG. 31, the method for the profit and loss simulation is specified, and specific information such as conditions and numerical values ​​is provided to the generation AI 3. The query text shown in FIG. 31 differs from the query text shown in FIG. 9 in that it does not include an analysis of the cause of the accident or a request to refer to blueprints and specifications, but instead includes detailed conditions for the profit and loss simulation and instructions for the calculation procedure. The sales plan may be attached as a data file, as in the second embodiment, but is included in the query text in FIG. 31.

[0056] 32 shows an example of a first combination of a first code and first meta information generated by the meta information generation module 12 in the modified example and a phrase extracted from the original query by the phrase extraction module 11 in the modified example. The numerical value extracted from the original query by the phrase extraction module 11 in the modified example and the converted numerical value generated by the meta information generation module 12 in the modified example are the same as those in FIG.

[0057] 33 and 34 show examples of prompts in a modified example that include a coded query generated by the replacement module 13 and a first combination generated by the meta information generation module 12. Although Figures 33 and 34 show one prompt as a whole, they are shown separately in Figures 33 and 34 because they cannot fit in one figure.

[0058] The words extracted by the phrase extraction module 11 are not included in the prompt. In addition, the numerical values ​​extracted by the phrase extraction module 11 have been replaced with converted numerical values. By reading the coded query while referring to the first combination, it is possible to roughly understand what is required of the generation AI 3, but no specific information such as recall is available. In addition, the numerical values ​​have been altered, so it is not possible to associate the content with the numerical values.

[0059] Figure 35 shows an example of a response obtained from the generation AI3 in the modified example. Here, the prompt conversion device 1a reverse-converts the first code included in the response from the generation AI3 back into the original phrase. In the modified example, the prompt specifies the detailed conditions and calculation procedures for the profit and loss simulation, so the generation AI3 performs specific calculations in accordance with the instructions. The reverse conversion module 15 of the prompt conversion device 1a reverse-converts the numerical values ​​included in the response from the generation AI3 using a second function, which is the inverse function of the first function, to correct the response. In the modified example, the first function used to convert the numerical values ​​included in the sales plan into converted numerical values ​​is Xc × 0.792. Therefore, when the numerical value included in the response from the generation AI3 is Yc, the second function can be Yc / 0.792.

[0060] In other respects, the modified example is similar to the second embodiment.

[0061] <2-4. Effects> According to the second embodiment, the prompt conversion device 1a: a text acquisition module 10 for acquiring text information based on a data file attached to the query text; a phrase extraction module 11 that extracts a plurality of phrases from each of the query original and the text information; a meta information generation module (12) that generates a first combination of a first code and first meta information corresponding to each of the terms extracted from the query text, and a second combination of a second code, different from the first code, and second meta information corresponding to each of the terms extracted from the text information; a substitution module 13 for substituting a first code for a phrase in the original query and a second code for a phrase in the text information to generate a coded query; a transceiver module 14 for sending a prompt including the coded query and the first and second combination to the generation AI 3 and receiving a response from the generation AI 3; Includes:

[0062] This reduces the risk of information leakage via data files even when the query text is accompanied by a data file. By assigning different codes to phrases contained in the query text and phrases contained in the text information obtained based on the data file, meta-information that takes into account the respective contexts of the query text and the data file can be assigned, which may make learning by the generation AI 3 more difficult.

[0063] According to the second embodiment, the prompt conversion device 1a: a text acquisition module 10 for acquiring text information based on a data file attached to the query text; a phrase extraction module 11 that extracts a plurality of phrases and a plurality of numerical values ​​from the query original and the text information; a meta information generation module 12 that generates a first combination of a first code and first meta information corresponding to each of the words and phrases, and that converts each of the numerical values ​​by a first function that is a monotonically increasing function to generate a converted numerical value; a substitution module 13 for substituting a first code for a word and a converted number for a number in the original query and the text information to generate a coded query; a transceiver module 14 for sending a prompt including the coded query and the first combination to the generation AI 3 and receiving a response from the generation AI 3; an inverse transformation module 15 that inversely transforms the numerical values ​​included in the answer using a second function that is an inverse function of the first function to correct the answer; Includes:

[0064] This reduces the risk of information leakage through a data file even when the query text is accompanied by a data file. Furthermore, when the query text or the data file contains numerical values, the generation AI 3 can perform calculations while hiding the specific numerical values ​​from the generation AI 3. The prompt conversion device 1a inversely converts the numerical values ​​included in the response from the generation AI 3 using an inverse function, thereby providing the user with useful calculation results.

[0065] According to a modification, the prompt conversion device 1a a phrase extraction module 11 that extracts a plurality of phrases and a plurality of numerical values ​​from the query original; a meta information generation module 12 that generates a first combination of a first code and first meta information corresponding to each of the words and phrases, and that converts each of the numerical values ​​by a first function that is a monotonically increasing function to generate a converted numerical value; a substitution module 13 for substituting a first code for a word in the original query and a converted number for a number to generate a coded query; a transceiver module 14 for sending a prompt including the coded query and the first combination to the generation AI 3 and receiving a response from the generation AI 3; an inverse transformation module 15 that inversely transforms the numerical values ​​included in the answer using a second function that is an inverse function of the first function to correct the answer; Includes:

[0066] By converting the multiple numerical values ​​contained in the original query using a monotonically increasing function, the generation AI 3 can perform calculations while maintaining the relative magnitude of the numerical values ​​while hiding the specific numerical values ​​from the generation AI 3. The prompt conversion device 1a inversely converts the numerical values ​​contained in the response from the generation AI 3 using an inverse function, thereby providing the user with a calculation result that can be used as reference. [Explanation of symbols]

[0067] 1, 1a...prompt conversion device, 2...input terminal, 3...generation AI, 10...text acquisition module, 11...phrase extraction module, 12...meta information generation module, 13...substitution module, 14...transmission and reception module, 15...inverse conversion module

Claims

1. a phrase extraction module that extracts a plurality of phrases from the query source; a meta information generation module that generates a first combination of a first code and first meta information corresponding to each of the phrases; a substitution module that substitutes the first code for the phrase in the original query to generate a coded query; a transceiver module that transmits a prompt including the coded query and the first combination to a generating AI and receives a response from the generating AI; A prompt conversion device including:

2. a text acquisition module for acquiring text information based on a data file attached to the query source text; a phrase extraction module that extracts a plurality of phrases from each of the original query and the text information; a meta-information generation module that generates a first combination of a first code and first meta-information corresponding to each of the phrases extracted from the original query text, and a second combination of a second code and second meta-information corresponding to each of the phrases extracted from the text information, the second code being different from the first code; a substitution module that substitutes the first code for the phrase in the original query and the second code for the phrase in the text information to generate a coded query; a transceiver module that transmits a prompt including the coded query and the first and second combinations to a generating AI and receives a response from the generating AI; A prompt conversion device including:

3. a phrase extraction module that extracts a plurality of phrases and a plurality of numerical values ​​from the original query; a meta information generation module that generates a first combination of a first code and first meta information corresponding to each of the words and phrases, and that converts each of the numerical values ​​by a first function that is a monotonically increasing function to generate a converted numerical value; a substitution module that substitutes the first code for the phrase and the converted numeric value for the numeric value in the original query to generate a coded query; a transceiver module that transmits a prompt including the coded query and the first combination to a generating AI and receives a response from the generating AI; an inverse transformation module that inversely transforms the numerical value included in the answer using a second function that is an inverse function of the first function to correct the answer; A prompt conversion device including:

4. a text acquisition module for acquiring text information based on a data file attached to the query source text; a phrase extraction module that extracts a plurality of phrases and a plurality of numerical values ​​from the original query and the text information; a meta information generation module that generates a first combination of a first code and first meta information corresponding to each of the words and phrases, and that converts each of the numerical values ​​by a first function that is a monotonically increasing function to generate a converted numerical value; a substitution module that substitutes the phrase with the first code and the numeric value with the converted numeric value in the original query and the text information to generate a coded query; a transceiver module that transmits a prompt including the coded query and the first combination to a generating AI and receives a response from the generating AI; an inverse transformation module that inversely transforms the numerical value included in the answer using a second function that is an inverse function of the first function to correct the answer; A prompt conversion device including:

Images