Computer program product for outputting text outputs generated using at least one Large Language Model

Abstract

A computer program product for outputting text outputs generated using at least one Large Language Model (LLM), wherein the computer program product includes commands which, when executed by an end device (10), cause the end device to perform the following steps: - Displaying a control area (110) on a user interface (100) of the terminal device (10), - Representing at least one actuable button (S1, S2, Sn) arranged in the control area (110), wherein at least one text instruction (Inst1, Inst2, Instn) is assigned to the at least one button (S1, S2, Sn) in a data-technical manner and is stored in an instruction database (130), - Detect activation of at least one button (S1, S2, Sn), and after activation has been detected: i. Reading user input (Text2) from a text area (120) and / or from the control area (110), ii. Retrieving the text instruction (Inst1, Inst2, Instn) assigned to the pressed button (S1, S2, Sn), iii. Transmitting the input user command and the retrieved text instruction (Inst1, Inst2, Instn) to a Large Language Model (LLM), iv. Receiving a response from the Large Language Model (LLM), and v. Outputting the answer in the text area (120) and / or control area (110).

Description

[0001] The present invention relates to a computer program product for outputting text outputs generated using at least one Large Language Model.

[0002] Furthermore, the present invention relates to a computer-implemented method and a device for generating text outputs using a Large Language Model, as well as a computer-readable data carrier and a communication system. Technical field

[0003] The present invention relates to the technical field of computer-aided text generation using Large Language Models, particularly in the field of text processing programs. State of the art and problem description

[0004] In today's technological landscape, the automated generation of text using Large Language Models (LLM) is already an established state of the art. These LLM models utilize complex algorithms and extensive datasets to create human-like text. Users can interact with LLM systems in interactive chat formats by entering their requests and desired text content into a text input field within a window. By entering so-called "prompts" or text fragments, users can address the LLM to generate text based on this input. This technology enables users to create text more quickly. Text generation using Large Language Models is subsequently referred to as LLM-assisted text generation.

[0005] A problem with the current LLM-based text generation is that the prompts have to be entered anew each time, which is very tedious in the long run, especially when similar texts are generated repeatedly. Resourceful users therefore save their prompts locally on their computers and copy them into the LLM system's input mask as needed. However, this requires manual work.

[0006] Furthermore, a disadvantage of the current LLM-based text generation is that frequent switching between computer tabs is necessary, which wastes time, or the text has to be tediously copied from text field A to text field B. In the process, text formatting is often lost, or the text formatting from the LLM output is adopted.

[0007] Furthermore, a disadvantage of the current LLM-supported text generation is that the text output from LLM is generated as a whole, and as a user, you have no way of recognizing which part is generated by LLM and which originates from the user.

[0008] Furthermore, a disadvantage of the current LLM-based text generation is that, despite good prompts, the LLM outputs are subject to statistical errors and hallucinations, and there is a risk of incorrect outputs.

[0009] In light of this prior art, one objective of the present disclosure is to provide a method or device suitable for enriching the prior art and improving at least the aforementioned aspects of the prior art. In particular, the disclosure aims to provide a solution that makes LLM-supported text generation more resource-efficient, less susceptible to statistical uncertainties, or more controllable for the user.

[0010] The problem is solved by the features of the independent claims. The dependent claims contain further developments of the disclosure. General Revelation

[0011] According to a first aspect of the invention, a device for generating text outputs with at least one Large Language Model is proposed.

[0012] The device comprises a text area and a control area, wherein the text area and the control area can be displayed on a user interface of an end device. This step can also be understood as displaying a text area and a control area on a user interface of an end device with a display device.

[0013] The device also includes at least one operable button arranged in the control area. It is particularly preferred that several buttons are arranged in the control area.

[0014] The device also includes an instruction database in which at least one text instruction is stored, wherein the at least one text instruction is assigned to the at least one button in a data-technical manner.

[0015] The device also includes an event handler that is configured to detect the activation of at least one button.

[0016] The event handler is configured to trigger one of the following steps: reading user input from the text area and / or the control area using a dedicated input device; reading the text instruction assigned to the pressed button using a dedicated output device; transmitting the user input and the read text instruction to a Large Language Model using a dedicated send device; receiving a response from the Large Language Model using a dedicated receive device; and outputting the response to the text area and / or control area using a dedicated output device.

[0017] The input device, the output device, the transmitting device, the receiving device and the output device are therefore also part of and included in the device, and the event handler triggers the interaction of the devices.

[0018] Accordingly, a device for generating text output with at least one Large Language Model is proposed, comprising: - a text area and a control area, wherein the text area and the control area can be displayed on a user interface of an end device, - at least one actuable button arranged in the control area, - an instruction database in which at least one text instruction is stored, wherein the at least one text instruction is assigned to the at least one button, - an event handler configured to detect an actuation of the at least one button and to trigger the following steps: i. Reading user input from the text area and / or from the control area with a reading device configured for this purpose, ii.

[0019] iii. Reading the text instruction assigned to the pressed button using a dedicated reading device, iii. Transmitting the user input and the read text instruction to a Large Language Model using a dedicated sending device, iv. Receiving a response from the Large Language Model using a dedicated receiving device, and v. Outputting the response in the text area and / or control area using a dedicated output device.

[0020] Preferably, the user input is text selected by the user, and the device also includes an output setting means with which it is optionally possible to set that the response of the Large Language Model is output before the selected text, output after the selected text, output on a new line, and / or that the response of the Large Language Model replaces the selected text.

[0021] Preferably, the device also has a context window setting means with which a reading context window of the reading device can be set.

[0022] User input is particularly preferred if the text is selected by the user, and the context window setting allows the user to optionally configure the input context window to correspond to the selected text, the input context window to correspond to a text section in which the selected text is located, the input context window to correspond to a predetermined number of preceding or following text sections, and / or the input context window to correspond to the entire text of the text area.

[0023] Accordingly, it is preferably proposed that the device also has a context window setting means with which a reading context window of the reading device can be set, wherein preferably the user input is text marked by the user and with the context window setting means it is optionally set that the reading context window corresponds to the marked text, the reading context window corresponds to a text section in which the marked text is arranged, the reading context window corresponds to a predetermined number of preceding or following text sections, and / or the reading context window corresponds to the entire text of the text area.

[0024] Preferably, the device includes a buffer for storing a read-in context window, and before reading user input from the text area and / or the control area with a dedicated read-in device, the following additional step is performed: First, reading a first user input from the text area and / or the control area and storing the first user input as a read-in context window in the buffer, and second, reading a second user input from the text area and / or the control area with a dedicated read-in device, and transmitting the first user input, the read-in context window, and the extracted text instruction to the Large Language Model with the dedicated send-in device.

[0025] Preferably, the device also has a display means to visually indicate a reading context window of the reading device to the user, in particular the visually perceptible highlighting is achieved by means of displayed special characters, by means of color highlighting or by means of font type change or the like.

[0026] Preferably, the device includes an output language button with which an output language for the response of the Large Language Model can be selected. It is therefore proposed to provide an additional button with which the output language of the LLM can be switched.

[0027] Preferably, the device includes a comparison device for comparing the input user input with the received response of the Large Language Model and a difference highlighting device for highlighting differences between the user input and the received response in a way that is perceptible to a user.

[0028] The difference highlighting device is particularly preferred if it is designed with a deterministic algorithm.

[0029] Preferably, the device includes a receiving device for receiving user feedback on a received response from the Large Language Model and an update device for automatically adjusting the stored text instruction based on the received feedback.

[0030] Particularly preferably, the automated adaptation includes transmitting the stored text instruction and user feedback to a Large Language Model with a sending device, wherein the Large Language Model is configured to generate a modified text instruction based on the stored text instruction and user feedback, and an adaptation device for adapting or overwriting the stored text instruction with the modified text instruction.

[0031] Preferably, the device also includes a conversion device for adapting the received response of the Large Language Model, wherein the conversion device is at least configured to convert quotations and / or paragraph numbers and / or page references into operable buttons and / or hyperlinks.

[0032] Preferably, the device has a comment input button, and the user input is a comment text assignable to the text area, and a text instruction is read out using the readout device assigned to the comment input button, and the comment text and the read text instruction are transmitted to a Large Language Model using the sending device, and the output of the response in the text area depends on the comment text and the text instruction.

[0033] Preferably, the device also includes a register data retrieval unit, and an input field for a patent number and at least one register data selection button are arranged in the control area. The register data retrieval unit is configured to perform a register data query based on the patent number and to retrieve register data from a public patent register, wherein the register retrieval unit is also designed to store the retrieved register data as individual options in the register data selection button in order to make the register data individually selectable and perceptible to the user in the register data selection button.

[0034] Particularly preferred is an operable button arranged in the control area which, when pressed, outputs the content of the selected register data selection button option into the text editing area and / or copies the content of the selected register data selection button option to an electronic clipboard.

[0035] According to another aspect of the disclosure, a computer-implemented method for generating text output using a Large Language Model is proposed, namely a computer-implemented method for generating text output using a Large Language Model, comprising the steps of: - providing several operable buttons arranged in a control area of ​​a user interface, - providing an instruction database in which several different text instructions are stored for retrieval, wherein one of the text instructions is assigned to each button, and after pressing one of the buttons, at least the following steps are performed: i. reading text selected in a text editing area of ​​the user interface, ii. reading the stored text instruction assigned to the pressed button, iii.1. Transmitting the selected text and the extracted text instruction to a Large Language Model, iv. Receiving a response from the Large Language Model, and v. Outputting the response in the text processing area.

[0036] The process comprises several steps, which are described below. The first step involves providing several clickable buttons arranged in a control area of ​​a user interface.

[0037] Furthermore, the step involves providing an instruction database in which several different text instructions are stored and can be retrieved, with each button being assigned one of the text instructions in terms of data.

[0038] It is therefore proposed to set up a database containing prompts intended for use by a Learning Management Device (LLM). Each button is then assigned one of the stored text instructions. This assignment can be implemented in various ways, for example, by having each button have an event listener that, upon triggering an event such as a click, calls a function to retrieve a predefined text instruction. In short, the proposal suggests setting up a prompt database with multiple prompts and permanently and uniquely assigning each prompt to a button. This instruction database can be stored locally on the user's device or in the cloud on a server.

[0039] Furthermore, after pressing one of the buttons, at least the following steps are performed: reading text selected in a text editing area of ​​the user interface, reading the stored text instruction assigned to the pressed button, transmitting the selected text and the read text instruction to a Large Language Model, receiving a response from the Large Language Model and outputting the response in the text editing area.

[0040] Accordingly, the step involves reading in text that has been selected in a text editing area of ​​the user interface.

[0041] The text editing area is a visually perceptible area for the user. It is explicitly suggested that the text be selected through user input.

[0042] The text editing area is therefore a different area than the control area described above. Both areas can be arranged side by side or overlapping. The area in which the user performs text editing is thus the text editing area, and the control area is a distinct area on a user interface containing the buttons that allow the user to perform specific actions or adjust settings.

[0043] The preferred approach is to display the text editing area and the control area side-by-side simultaneously, making them visually perceptible to the user, or alternatively, to display them overlapping. Simultaneous display saves click resources by eliminating the need to manually access the control area, for example, by right-clicking. The overlapping display can be accessed, for instance, by right-clicking, so that the control area appears as an overlapping area in the foreground. In other words, it is proposed that a user manually selects text in the text editing area by entering text, and that the selected text is then automatically read in for further processing.

[0044] Furthermore, the stored text instruction assigned to the pressed button is read out.

[0045] This step can also be understood as reading the text instruction assigned to the pressed button from the instruction database. It is therefore suggested that the system detects which button is pressed and then reads the text instruction associated with that button.

[0046] The next step is then: transmitting the marked text and the extracted text instruction to a Large Language Model.

[0047] As part of the revelation, the text selected by the user, along with the extracted text instruction, is transmitted to an LLM. This transmission can be implemented, for example, via an API call if a cloud-based LLM is used, or via a local call to a local LLM. Therefore, in accordance with the revelation, a local or cloud-based LLM is preferably used. It is also possible to use multiple different Large Language Models, each assigned to different buttons.

[0048] In a further step, a response from the Large Language Model is received.

[0049] The output of the LLM is therefore based on the input, which is formed from the text instruction assigned to the pressed button and from the text marked by the user via user input.

[0050] In a further step, the answer is then displayed in the text editing area.

[0051] It is therefore proposed that the response be displayed in a way that is perceptible to the user, namely in the text editing area. The output will not appear in a chat window or similar, but directly in the text editing area.

[0052] Preferably, the method includes the additional step of comparing the read-in marked text with the received response of the Large Language Model using a comparison algorithm and highlighting differences between the marked text and the received response in a way that is perceptible to a user.

[0053] A comparison algorithm is specifically an algorithm for matching strings, such as the Levenshtein algorithm, the Jaro-Winkler algorithm, the Hamming distance algorithm, or similar algorithms. This advantageously ensures that discrepancies between the original text and the generated response are quickly and easily recognizable to the user, thereby increasing the efficiency and accuracy of response verification and validation.

[0054] Preferably, the method includes the additional step of receiving user feedback on the response generated by the Large Language Model and automatically adjusting the stored text instruction based on this feedback. This advantageously ensures that the quality and relevance of the generated responses are continuously improved without requiring manual intervention. This leads to more efficient model optimization and reduces the need for time-consuming manual post-processing.

[0055] The automated adaptation process is particularly advantageous when it involves the following steps: transmitting the stored text instruction and user feedback to a Large Language Model (LLM), which is configured to generate a modified text instruction based on the stored text instruction and user feedback, and adapting or overwriting the stored text instruction with the modified text instruction. This effectively allows the text instruction to be dynamically and automatically adapted to the user's needs, and the instructions to adapt to user preferences. Therefore, the user does not need to be a prompt expert.

[0056] Preferably, the comparison algorithm comprises at least one algorithm from the list of algorithms: an LCS algorithm, wherein the LCS algorithm is an algorithm that determines a longest common subsequence between the tagged text and the LLM response; a sentence embedding algorithm, wherein the sentence embedding algorithm comprises at least the steps of: calculating a sentence embedding of the tagged text, calculating a sentence embedding of the Large Language Model response, and comparing the calculated sentence embeddings; a token alignment algorithm, wherein the token alignment algorithm decomposes the tagged text and the received response into individual tokens, compares the tokenized sequences, and identifies discrepancies in the form of missing or altered tokens;a hash-based verification algorithm, wherein the hash-based verification algorithm calculates a checksum for both the marked text and the received response and compares the checksums; a Levenshtein distance algorithm, wherein the Levenshtein distance algorithm calculates the minimum number of edit operations between the marked text and the received response, thereby identifying potential character loss.

[0057] Additionally or alternatively, it is preferably proposed that the highlighting, perceptible to the user, is achieved by outputting matching parts between the selected text and the received response in a deactivated change mode, and by outputting non-matching parts between the selected text and the received response in an activated change mode.

[0058] Additionally or alternatively, it is particularly preferred that non-matching parts between the highlighted text and the received response be highlighted by means of color highlighting, by changing a font, by displaying some other indicator or the like.

[0059] Preferably, the procedure prior to the output step includes the additional step of checking the received response for hallucination using a hallucination detection algorithm, and adapting the received response of the Large Language Model or retransmitting the marked text and the extracted text instruction to a Large Language Model if the hallucination detection algorithm detects a hallucination.

[0060] The hallucination testing algorithm is particularly preferred if it includes at least one algorithm from the following list: Self-evaluation algorithm, where the self-evaluation algorithm is an algorithm that retransmits the received response to a large language model that verifies the content of the response; cross-referencing algorithm, where the cross-referencing algorithm is an algorithm that verifies the received response against external sources to detect implausible or inconsistent statements; multi-model verification algorithm, where the multi-model verification algorithm is an algorithm in which several independent large language models (LLMs) or different instances of the same LLM are queried with identical or slightly varied prompts, and the results are evaluated for contradictions or discrepancies;Chain-of-Thought Validation Algorithm, wherein the Chain-of-Thought Validation Algorithm is an algorithm that requests the addressed LLM to provide a chain of thought or additional explanations, which are then systematically examined for internal inconsistencies, missing intermediate steps, or fabricated facts (hallucinations); and Confidence-Scoring Algorithm, wherein the Confidence-Scoring Algorithm is an algorithm that applies a thresholding or ranking procedure that assigns a confidence or trust rating to each statement in the response, and optionally questions statements below a defined trust rating through additional inquiries or external verification.

[0061] Preferably, the procedure includes the step of displaying a confidence score for the LLM response in a way that is perceptible to the user. A confidence score for the LLM response is a numerical, percentage, or other value that quantifies the reliability or accuracy of a response generated by the Large Language Model (LLM). This score is typically based on statistical, probabilistic, or heuristic methods that assess the consistency, plausibility, and agreement of the response.

[0062] Preferably, as an alternative to the step of reading the stored text instruction, the method includes the step of reading an individual instruction that is stored by a user through user input in an instruction input field, wherein the individual instruction is assigned to the pressed button.

[0063] It is therefore proposed that buttons be assigned not a text instruction from the instruction database, but rather a text instruction that a user enters into the instruction input field. This allows users to assign and use individual prompts for their buttons. A particularly desirable feature is the additional step of saving the individual instruction in the instruction database. Saving the individual instruction in the instruction database enables efficient management and permanent retrieval of the prompts. This eliminates the need for the user to re-enter the individual prompt each time, and the prompt is permanently assigned and saved to the button. Therefore, it is proposed that a user can enter an individual prompt in an input field, which is then assigned to a button.When a user then presses the button, the individual prompt is read and sent to the LLM.

[0064] Preferably, the method includes the further step of setting, using an instruction setting device, that the stored text instruction is read from the instruction database and transmitted to the Large Language Model, and / or that the individual instruction is read from the instruction input field and transmitted to the Large Language Model.

[0065] It is therefore proposed that a user be able to select, using a setting, whether a prompt from the instruction database should be used or a custom prompt from the instruction input field. The instruction setting could be, for example, a checkbox, a switch box, a choice field, or something similar.

[0066] Preferably, the method includes the further step of setting the selection of the Large Language Model using a model selection tool, wherein a different API endpoint is changed by the model selection tool depending on the user selection, in particular to be able to use Large Language Models from different manufacturers.

[0067] It is therefore proposed that a user be able to select which Large Language Model they wish to use with the model selection tool. This would allow different specialized LLMs to be used for LLM-supported text generation.

[0068] Preferably, the method comprises the further step of setting the temperature of the Large Language Model to a lower half of an adjustable temperature interval. Preferably, the temperature is set to a value below 0.5 for a temperature interval of 0 to 1, or to a value below 1 for a temperature interval of 0 to 2, particularly to obtain the most factual response possible from the Large Language Model. The temperature of a Large Language Model relates to a hyperparameter that controls the probability distribution of generated texts. Lower temperatures produce more conservative or less diverse texts, while higher temperatures can lead to more creative or diverse results.

[0069] The text processing device preferably comprises an output setting means that can be actuated by user input, with which the user can optionally set that the response of the Large Language Model is output before the marked text, the response of the Large Language Model is output after the marked text, and / or the response of the Large Language Model replaces the marked text, in particular as described above.

[0070] According to a further aspect of the invention, a computer program product is proposed for outputting text outputs generated with at least one Large Language Model, wherein the computer program product comprises commands which, when executed by an end device, cause the end device to perform the following steps: displaying a control area on a user interface of the end device, displaying at least one actuable button located in the control area, wherein at least one text instruction is assigned to the at least one button and stored in an instruction database, detecting an actuation of the at least one button, and after detection of actuation, i. reading user input from a text area and / or from the control area, ii. retrieving the text instruction assigned to the actuated button, iii.1. Transmitting the input user command and the retrieved text instruction to a Large Language Model, iv. Receiving a response from the Large Language Model, and v. Outputting the response in the text area and / or control area.

[0071] Preferably, several buttons are displayed in the control area.

[0072] Preferably, a computer program product for outputting text output to a text editing area of ​​a word processing program is proposed.

[0073] Dynamically adjustable text output position: Preferably, the user input is text selected by the user, and the computer program is configured to perform the following step: Displaying an output setting that allows the user to optionally configure the response of the Large Language Model to be output before the selected text, after the selected text, on a new line, and / or to replace the selected text. The output setting can also be understood as an output setting button. A user can therefore easily change the location or position of the text output using the output setting. Outputting before and after the selected text is to be understood generically and does not only mean immediately before or after it. For example, the text can be output at the end of the paragraph, even though the selected text is only part of the paragraph.

[0074] Statically adjustable input context window: Preferably, the computer program is configured to perform the following step: Displaying a context window setting tool with which an input context window for the text area can be set. The context window setting tool can also be understood as a context window setting button.

[0075] User input is particularly preferred if the text is selected by the user, and the context window setting allows the user to optionally configure the input context window to correspond to the selected text, the input context window to correspond to a text section in which the selected text is located, the input context window to correspond to a predetermined number of preceding or following text sections, and / or the input context window to correspond to the entire text of the text area.

[0076] Dynamically selectable input context window: Preferably, the computer program product is configured to perform the following steps: reading a first selected text from the text area and storing the first selected text as an input context window in a buffer, reading a second selected text from the text area, and transmitting the input context window, the second selected text, and the retrieved text instruction to a Large Language Model.

[0077] Cursor-dependent context window display: Preferably, the computer program product is configured to perform the following step: Visually highlighting a read-in context window, wherein the visual highlighting depends on a cursor position and a set read-in context window.

[0078] Highlighting is preferably achieved by means of displayed special characters, color highlighting, or font type change, or the like, and the highlighting includes the following additional steps: Automated addition and removal of special characters, color highlighting and / or font type change depending on the cursor position and the set input context window.

[0079] Split indicators and subsequence queries: Preferably, the computer program product is configured to perform the following steps: detecting split indicators in the input user input, wherein the split indicators can be added by the user in the text area; splitting the input user input into subsequences depending on the split indicators; transmitting the input subsequences and the retrieved text instruction to a Large Language Model in parallel or sequentially; receiving multiple responses for each subsequence from a Large Language Model; and outputting the received responses in the text area and / or control area.

[0080] The split indicator is particularly preferred to be a user-added underline or one or more special characters in a selected text within the text area.

[0081] Kl-Markups: Preferably, the computer program product is configured to perform the following step: comparing the input user input with the received response of the Large Language Model, and visually highlighting differences between the user input and the received response for a user.

[0082] Particularly preferred is the comparison performed using a deterministic comparison algorithm, in particular the comparison algorithm comprising at least one algorithm from the list of algorithms: i. an LCS algorithm, wherein the LCS algorithm is an algorithm that determines a longest common subsequence between the marked text and the response of the LLM; ii. a sentence embedding algorithm, wherein the sentence embedding algorithm comprises at least the steps of: calculating a sentence embedding of the marked text, calculating a sentence embedding of the response of the Large Language Model, and comparing the calculated sentence embeddings; iii. a token alignment algorithm, wherein the token alignment algorithm decomposes the marked text and the received response into individual tokens, compares the tokenized sequences, and identifies deviations in the form of missing or altered tokens; iv.a hash-based verification algorithm, wherein the hash-based verification algorithm calculates a checksum for the marked text and for the received response, and compares the checksums; v. a Levenshtein distance algorithm, wherein the Levenshtein distance algorithm calculates the minimum number of editing operations between the marked text and the received response, thereby identifying a potential loss of characters.

[0083] The user-visual highlighting is particularly preferred by displaying matching parts between the user input and the received response in a deactivated editing mode, and by displaying non-matching parts between the highlighted text and the received response in an activated editing mode.

[0084] Preferably, the computer program product is configured to perform the step of displaying a markup button that allows the visually perceptible highlighting to be turned on and off.

[0085] Automated prompt adaptation: Preferably, the computer program product is configured to perform the following step: Submitting user feedback to a received response from the Large Language Model, and automatically adapting the stored text instruction based on the received feedback.

[0086] Particularly preferred, the automated adaptation includes the following steps: transmitting the stored text instruction and user feedback to a Large Language Model, where the Large Language Model is configured to generate a modified text instruction based on the stored text instruction and user feedback, and adapting or overwriting the stored text instruction with the modified text instruction.

[0087] Smart Jump Links: Preferably, the computer program product is configured to perform the following steps: detecting citations, paragraph numbers, and / or page references in the Large Language Model response; converting or transforming these citations, paragraph numbers, and / or page references into clickable buttons or hyperlinks; and issuing a control command to a text preview program product that causes the text preview program product to jump to the citation, paragraph, and / or page, and optionally causes the text preview program product to visually highlight the citation and / or paragraph. These steps are performed, for example, using a transformation device or algorithm.

[0088] Register data quick access: Preferably, the computer program product is configured to perform the following steps: Displaying an input field for a patent number in the control area, displaying at least one register data selection button in the control area, triggering a register data query to retrieve register data from a public patent register based on the entered patent number, storing the retrieved register data as individual options in the register data selection button to make the register data individually selectable and perceptible to the user in the register data selection button, and optionally displaying an operable button in the control area which, when pressed, outputs the content of the selected register data selection button option to the text area and / or, when pressed, copies the content of the selected register data selection button option to an electronic clipboard.

[0089] Automated comment processing: Preferably, the computer program product is configured to perform the following step: Displaying a comment processing button in the control area, wherein, after activation of the comment processing button, the following sub-steps are performed: Reading a comment text that is assigned to a text section in the text area, reading at least the text section to which the comment is assigned, retrieving a text instruction that is assigned to the comment processing button, transmitting the read comment text, the read text section and the retrieved text instruction to a Large Language Model, and outputting the response in the text area depending on the comment text and the text instruction.

[0090] Dynamically changeable language setting: Preferably, the computer program product is configured to perform the step of: Displaying an output language button with which an output language of the output response of the Large Language Model can be selected.

[0091] Dynamic module areas: Preferably, the control area has several module areas, in each of which a portion of the actuable buttons are arranged, wherein each module area is designed to be variable in size and has a folded-in state in which no buttons are visually recognizable to the user in the module area, and wherein the module area has an unfolded state in which buttons are visually recognizable to the user, and wherein the module area is designed such that a user input causes a change between the folded-in and unfolded state.

[0092] Individual user prompts: Preferably, the device has an instruction input field in which the user can enter an individual text instruction, and the reading device is configured to read the individual text instruction from the instruction input field. Preferably, the computer program is configured to display the aforementioned instruction input field.

[0093] Hallucination detection and display: Preferably, the device also includes a testing device for checking the received response for hallucinations and a display device for showing a hallucination score, wherein the hallucination score is a numerical value or a scale that represents a quantitative measure of a hallucination that occurred. Preferably, the computer program product is configured to perform the aforementioned testing and display.

[0094] Preferably, a first button is assigned a first text instruction that causes the LLM to output the selected text in simplified language.

[0095] Preferably, a second button is assigned a second text instruction that causes the LLM to output a spell-corrected version of the selected text.

[0096] Preferably, a third button is assigned a third text instruction that causes the LLM to output a translation of the selected text.

[0097] Preferably, a fourth button is assigned a fourth text instruction that causes the LLM to output synonymous terms for the selected text.

[0098] Preferably, a fifth button is assigned a fifth text instruction that causes the LLM to output generic terms related to the selected text.

[0099] Preferably, a sixth button is assigned a sixth text instruction that causes the LLM to output the selected text in a reworded form.

[0100] Preferably, a seventh button is associated with a seventh text instruction that causes the LLM to generate examples for the selected text.

[0101] Preferably, an eighth button is assigned an eighth text instruction that causes the LLM to generate a definition for the selected item.

[0102] Preferably, a ninth button is assigned a ninth text instruction that prompts the LLM to state an advantage of the selected text.

[0103] Preferably, a tenth button is assigned a tenth text instruction that causes the LLM to generate a claim support for the selected text.

[0104] Preferably, an eleventh button is associated with an eleventh text instruction that causes the LLM to remove or add a reference mark from the selected text.

[0105] Preferably, a twelfth button is associated with a twelfth text instruction that causes the LLM to output a summary of a referenced document.

[0106] The computer program product can also be described synonymously as a computer program.

[0107] The computer program product is preferably designed to execute the process steps of the computer-implemented procedure described above.

[0108] The computer program product can be implemented locally on the end device or on a server and run in the CPU or GPU.

[0109] The computer program product may additionally or alternatively comprise several program parts, wherein a first part of the computer program product is executed locally on an end device and a second part of the computer program product is executed on a server. Accordingly, a distributed computer program product is proposed for outputting text output generated using at least one Large Language Model, wherein the computer program product includes commands that, when the computer program product is executed on an end device and / or on a server, cause the end device and / or the server to perform the described steps.

[0110] Particularly preferred is a first part of the computer program product, an add-in for a word processing program that can be executed locally on an end device and includes commands that, when the add-in is executed by an end device, cause it to establish a data exchange with a server, and a second part of the computer program product is a server computer program hosted on a server, which includes commands that, when the server computer program is executed by the server, cause it to establish a data exchange with the add-in and to establish a data exchange with data interfaces of third-party services.

[0111] According to a further aspect of the invention, a computer-readable data carrier is claimed, wherein the computer program product according to the preceding embodiment is stored on the computer-readable data carrier. Reference is made to all explanations, definitions, and advantages relating to the device, method, and computer program product described above.

[0112] According to a further aspect of the invention, a communication system is proposed for outputting text outputs generated with at least one Large Language Model, comprising: an end device, wherein the end device comprises a computer program product configured according to one of the preceding or following embodiments, and a cloud server, wherein a Large Language Model is hosted on the cloud server, and wherein the end device and the cloud server are configured to exchange data via a communication network. Reference is made to all explanations, definitions, and advantages relating to the device, method, and computer program product described above, etc.

[0113] It is explicitly pointed out that the explanations, definitions, and embodiments of the aspects of the invention apply analogously. This means, for example, that all explanations, definitions, and embodiments described for the device and the method also apply analogously to the computer program product, and conversely, all explanations, definitions, and embodiments relating to the computer program product also apply to the device, the method, the communication network, and the computer-readable data carrier.

[0114] The aspects of the invention produce several advantageous technical effects. As described above, user input is read into the text area and / or control area. Subsequently, the LLM's response is also output in the text area and / or control area. Input and output thus coincide. This eliminates the need for an additional input field for the LLM. Normally, a separate prompt input field is provided for the LLM, into which the user enters the prompt. Therefore, an advantageous dual use of the text area and / or control area is proposed here. This saves resources.

[0115] Furthermore, assigning instructions to buttons is advantageous because it allows individual and different instructions to be permanently stored for each button and retrieved automatically without additional effort. This eliminates the need to re-enter instructions each time, and the assignment of instructions to buttons also makes it very easy to address different Large Language Models or trigger specific LLM behaviors.

[0116] Another advantage is improved and simpler access to the LLM, because displaying the response directly in the text area and / or control area eliminates the need to open a new tab or window, thus saving resources in the form of memory and improving performance. Furthermore, it saves time and effort, as the LLM response is displayed directly in the text area and / or control area, eliminating the need to copy and paste it. This also saves resources.

[0117] The output setting tool advantageously allows the position of the output of the response from the Large Language Model to be flexibly adapted to the user's needs.

[0118] The context window setting, or configurable input context windows, offers the advantage of allowing the input context window of the input device to be flexibly adapted to different requirements or conditions, thereby increasing the efficiency and accuracy of the input processes. The user can thus optimize the context window via the context window setting and, for example, adjust it if it is too small or too large, provided the LLM response is unsatisfactory. For instance, if the user selects the text "Cache" in the text area, this may be too imprecise a context for generating a technical definition, as "caches" can mean anything. If the context window setting in this example is set to "Text section in which the selected text is located," the LLM is also provided with the context of the text section, and the response becomes more accurate.A statically adjustable context window is therefore proposed. This allows users to optimize the context window for the LLM and, for example, dynamically adjust it if it is too small or too large, provided the LLM response is unsatisfactory.

[0119] The display method, or the visually perceptible highlighting of the input context window, advantageously enables the user to precisely and intuitively identify the input context window of the input device, thus facilitating the checking of its size and the selection of a suitable context window. For example, visual highlighting can be achieved using displayed special characters, such as... or by temporarily coloring the text green.

[0120] The output language button allows you to easily switch between different languages, for example between German and English or Chinese, without having to use different LLMs, thus saving resources.

[0121] Using a difference highlighting feature trained with a deterministic algorithm advantageously ensures that the results are free from statistical errors. Users can therefore be confident that the unhighlighted parts actually originate from the user and are not the result of a hallucination.

[0122] The comparison device and the difference highlighting device advantageously achieve that differences between the user input and the received response of the Large Language Model are made immediately and clearly recognizable to the user, making the hallucinations in the response of the Large Language Model easier to recognize. Glossary

[0123] This section includes definitions of terms that are explanatory in nature and apply to all embodiments and aspects of the invention.

[0124] "Store in a retrievable format": Stored in a retrievable format, in the sense of disclosure, means that the stored text instructions can be read by machine.

[0125] "Response": Response can also be understood as the output or response of the LLM. The output of the LLM is therefore based on the input, which consists of the text instruction assigned to the pressed button and the text selected by the user.

[0126] “Display device”: A display device is a device used to visually present information, typically through the use of screens, light-emitting diodes (LEDs) or other optical display means.

[0127] "Reading device for reading the text instruction": A reading device is a device designed to capture selected text passages or user input. The input device is, for example, a software module, a program routine, a script, or a function that reads the selected text passage or input field.

[0128] "Output device": An output device for displaying the response of the Large Language Model in the text area and / or control area according to the disclosure is a device or system configured to provide the generated response of a Large Language Model either in a textual display area or in a control-related area. The output device is, for example, a software module, a program routine, a script, or a function that outputs an API response of an LLM API endpoint in the text area or control area.

[0129] “Output setting tool”: An adjustable output tool could be, for example, a software button, such as a drop-down menu or selection menu, which the user can modify to specify where the output of the Large Language Model is displayed.

[0130] “Commands”: Commands refers to data-related commands, such as control commands, retrieval commands, write commands, delete commands or the like, and can also be understood as function calls.

[0131] "Active button": An actuated button refers to graphical elements in user interfaces that trigger certain actions or functions when clicked or selected by the user.

[0132] "Providing clickable buttons": Providing clickable buttons can also be understood as displaying or showing the buttons on the user interface.

[0133] "Providing an instruction database": Providing an instruction database can also be understood as setting up or building an instruction database.

[0134] “Computer-implemented method”: A computer-implemented method for generating text outputs within the meaning of the disclosure refers to a method executed by a computer or a computer-controlled device, the result of which is text output. A computer according to the disclosure is to be understood generally as an electronic device that processes and executes data according to a set of instructions. The term includes both specialized and general-purpose computing units, including, but not limited to, desktop computers, servers, laptops, tablets, smartphones, and embedded systems.

[0135] "Computer program product": A computer program product is a computer-readable medium or one or more files comprising instructions which, when executed by a computing unit, are designed to perform a predetermined function or procedure. A computer program product is, in particular, an add-in, a plugin, a software application, an operating system, a driver, a firmware update, or the like.

[0136] "Data-based mapping of a text instruction to a button": Data-based mapping of a text instruction to a button refers to the data-based assignment of a predefined button to a predefined instruction. For example, this assignment can be achieved by assigning a unique identifier or class to the button and retrieving the text instruction based on this identifier or class. Alternatively, the button could execute a function call with an identifier for a predefined text instruction, or something similar.

[0137] "End device": An end device is, in particular, a desktop computer, a laptop, a smartphone, a tablet or the like.

[0138] "Reading in": Reading in a selected text, in the sense of disclosure, means that a user manually selects a digital text or individual words, for example with a cursor or the like, and this is read in by machine.

[0139] "Reading device for capturing user input": A reading device is a device designed to capture selected text passages or user input. The reading device can be, for example, a software module, a program routine, a script, or a function that reads the selected text passage or an input field.

[0140] "Read context window": A read context window is a defined data window or memory area used to read in a specific context or set of data for processing, analysis, or interpretation. It is typically used to selectively choose and provide only relevant information from a data source.

[0141] "Receiving": Receiving means receiving data, for example in the form of an API response or the like.

[0142] "Receiving device": A receiving device for receiving a response from a Large Language Model is a technical device designed to receive and process data or signals generated by a Large Language Model. The device may include communication interfaces, storage, and processing units to analyze, store, or forward the received responses to downstream systems.

[0143] "Event handler": An event handler is a device or software component designed to detect the activation of at least one button and subsequently trigger a corresponding action or processing. An event handler is, for example, a software module, a program routine, a script, or a function designed to recognize user input such as mouse clicks, key presses, touchscreen gestures, or voice commands. The event handler can also be understood as a trigger device. Triggering can also be understood as activating.

[0144] "Instruction database": An instruction database is an electronic database that stores and provides instructions or prompts for performing specific tasks of the LLM (Learning Management Module). The instruction database can also be called a prompt database. It can be local or cloud-based, with cloud-based being particularly preferred. A cloud-based instruction database is a database-driven system that runs in a cloud computing environment and serves to centrally store, manage, and provide prompts. It allows access to this data via networked devices and typically supports features such as versioning, search, and user management.

[0145] "Instruction input field": An input field on a user interface into which a user can enter a custom text instruction.

[0146] "Context window setting device": A context window setting device is a device or mechanism designed to allow the adjustment of a reading context window of a reading device. It enables the adjustment of the area from which the reading device captures or processes data.

[0147] "Large Language Model (LLM)": A Large Language Model is an artificial neural network trained on large amounts of text to perform natural language processing tasks such as text generation, translation, and language understanding. The Large Language Model is a text-processing Large Language Model and can also be referred to as a generative language model or text generation model. Technically, in the context of this revelation, a Large Language Model is a transformer-based model designed and trained for Natural Language Processing (NLP). Therefore, the Large Language Model is configured to process text as input and output text as output (text-to-text model).Well-known models include, for example, Text-to-Text Transfer Transformer (T5), Bidirectional and Auto-Regressive Transformers (BART), Generative Pre-trained Transformer (GPT), Unifying Language Learning Paradigm, and Conditional Transformer Language (CTRL). LLM can be implemented locally on a user's device or hosted on a cloud-based server.

[0148] “Highlighted text”: Highlighted text refers to individual words, sentences, clauses, paragraphs or sections, or the like.

[0149] “User input”: User input includes, for example, text selected by the user, in particular text selected in the text editing area or text selected in a PDF; user input entered by the user in an input field of the control area, in particular a patent number, a chat input, a search input or a user prompt; a comment, in particular a saved comment; parts of or all of the text present in the text area; and / or files uploaded by the user, in particular PDF or Word files.

[0150] "User interface": User interface refers to the surface of a user interface that is interactively used by users to interact with a computer program, application, or electronic device.

[0151] "Button": A button, for example, is a graphical control element on the user interface that is configured to trigger a predefined function or action when activated by user input, such as a mouse click or a touch. A button is, for example, a control button.

[0152] "Control Panel": The control panel of a user interface refers to a part of the user interface that contains controls or buttons that allow the user to perform specific actions or adjust settings. For example, the control panel might be a side panel or drop-down menu displayed next to or within the text editing area.

[0153] "Storage": Storing refers to the technical storage of data.

[0154] "Sending Device": A sending device for transmitting user input and extracted text instructions is a device or system configured to transmit data entered by a user and text instructions extracted from a source in a format processable by a Large Language Model. The sending device may include communication interfaces and protocols that ensure efficient and error-free data transmission. For example, the sending device could be a software module, a program routine, a script, or a function that makes an API call (POST request) to a configured API endpoint.

[0155] "Splitting indicator": A splitting indicator is an indicator that shows a subsequence within the user input. Splitting indicators show the beginning and end of a subsequence. For example, a splitting indicator could be an underline or one or more special characters that indicate a subsequence.

[0156] "Text output": Text output refers to the output of text perceptible to humans on a screen or other output device.

[0157] "Text area": ​​A text area refers to an area of ​​a user interface where a user can view text (e.g., a PDF preview) or to the text editing area (e.g., the text editing area in a word processing program or an editable rich text or plain text input field).

[0158] "Text editing area": ​​An area or input field where a user can modify, edit, and optionally format text. This area can also be understood as a text editing area.

[0159] “Text instruction”: Text instruction can be understood synonymously as prompt or instruction.

[0160] “Text preview program product”: A software program for text preview, for example a PDF viewer or the like.

[0161] “Transmitting”: Transmitting generally refers to the transfer of information, data or signals from a sender to a receiver, via a wireless or wired connection.

[0162] "Comparison algorithm": A comparison algorithm is, for example, an algorithm that analyzes text characters or strings to identify similarities, differences, or matches between two or more texts. It applies specific rules and procedures to compare the structure, order, frequency, or semantic meaning of the characters or strings. Such algorithms can be based on various methods, including, but not limited to, lexical analysis, syntactic comparison, semantic analysis, or statistical procedures.

[0163] “Perceptible highlighting”: Perceptible highlighting is, in particular, visual highlighting through color change, acoustic highlighting through a signal tone, haptic highlighting through vibration or the like.

[0164] "Assigning a text instruction to a button": Assigning a text instruction to a button refers to the act of linking a specific text instruction to a button, so that the text instruction is used when the button is pressed. Fiquren description

[0165] The present invention will now be explained in more detail below by way of example embodiments with reference to the accompanying figures, whereby the same reference numerals are used for identical or similar assemblies. Reference sign 10 End device 100 User interface 110 Control area 120 text area 130 Instruction database 140 output setting devices 150 context window settings 160 Import context windows 165 Cursor position 170 special characters 180 Output language button 190 Al Markup Button 200 comparison algorithm LLM Large Language Model S1, S2, Sn button Inst1, Inst2, Instn text instruction Ant1, Ant2, Antn Response Fig. Figure 1 shows a schematic representation of a computer program product for outputting text output according to one embodiment, executed on a terminal device. Fig. Figure 2 shows a further development of the computer program product and the principle of a dynamically adjustable text output position and the output language button. Fig. Figure 3 shows a further development of the computer program product and the principles of a statically adjustable input context window as well as cursor-dependent context window display. Fig. Figure 4 shows a further development of the computer program product and the principle of AI markups. Fig. Figure 5 shows a further embodiment of the computer program product and the principle of smart jump links.

[0166] Fig. Figure 1 shows a computer program product running on an end device 10. As can be seen, a control area 110 is displayed on a user interface 100 of an end device 10, and at least one operable button S1, S2, Sn is displayed, which is arranged in the control area 110. At least one text instruction Inst1 is assigned to the at least one button S1, which is stored in an instruction database 130. Furthermore, the activation of the at least one button Inst1 is detected, which is illustrated by the mouse pointer and lightning bolt. After activation is detected, user input is read from a text area 120, where, in the illustrated embodiment, the user input is a highlighted text Text2. The highlighting of the text Text2 is illustrated by a gray area.Furthermore, the text instruction Inst1, which is assigned to the pressed button S1, is retrieved, and the input text2 and the called text instruction Inst1 are transmitted to a Large Language Model (LLM). Additionally, a response Ant1 from the Large Language Model (LLM) is received and displayed in text area 120.

[0167] Fig. Figure 2 shows a computer program product executed on a terminal device 10 according to a further embodiment and the principle of a dynamically adjustable text output position. The user input is a text Text2 selected by the user, and an output setting device 140 is displayed with which it can be selected that the response Ant1 of the Large Language Model LLM is output before the selected text, output after the selected text, output on a new line, and / or that the response of the Large Language Model replaces the selected text. In the illustrated embodiment, the output setting device 140 is set to output the response Ant1 after the selected text Text1.

[0168] Furthermore, the Fig. 2. An output language button 180, with which an output language DE can be selected for the output response of the Large Language Model LLM. In the example shown, DE, i.e., German, is set as the output language.

[0169] Fig. Figure 3 shows a computer program product executed on an end device 10 according to a further embodiment and the principles of a statically adjustable input context window and the cursor-dependent context window display. In the embodiment shown, the user input is a selected text Text2, and a context window setting device 150 is also displayed with which an input context window 160 of the text area 160 can be set. In the example shown, the user presses a setting button, and a popup appears in which the context window setting device 150 is displayed.The user can use the context window settings to selectively configure the input context window 160 to correspond to the selected text "Text2", to a text section containing the selected text "Text2", to a predetermined number of preceding or following text sections, and / or to the entire text within the text area. In the illustrated example, the user has configured three paragraphs as a static input context window. This means that the paragraph in which the user selects "Text2" is read, as well as the preceding and following paragraphs. Changing the context window settings adjusts the size of the context window accordingly.

[0170] Furthermore, the input context window 160 is visually highlighted, with the highlighting depending on the cursor position 165 and the selected input context window 160, specifically by means of displayed special characters 170. If the user were to click elsewhere in the text area, the software would cause the device 10 to automatically add and remove the special characters 170 depending on the cursor position 165 and the selected input context window 160. The special characters 170 are deleted when the user changes the cursor position 165 and then reinserts them at the new position, with the spacing determined by the input context window setting.

[0171] Fig. Figure 4 shows a computer program product executed on an end device 10 according to a further embodiment and the principle of AI markups. As can be seen, the input user text 2 is compared with the received response Ant 1 of the Large Language Model (LLM) using a comparison algorithm 200. Differences between the user input text 2 and the received response Ant 1 are visually highlighted for the user. This highlighting is achieved by displaying matching parts of the user input and the received response in a deactivated change mode and by displaying non-matching parts of the highlighted text and the received response in an activated change mode. The AI ​​markups can be activated or deactivated using a button 190.

[0172] Fig. Figure 5 shows a computer program product executed on a terminal device 10 according to a further embodiment and the principle of smart jump links.The computer program product comprises commands which, when executed by the terminal device 10, cause the terminal device 10 to perform the following steps: displaying a control area 110 on a user interface 100 of the terminal device 10; displaying at least one operable button S1 located in the control area 110, wherein at least one text instruction Inst1 is assigned to the at least one button S1 and is stored in an instruction database 130; detecting an actuation of the at least one button S1; and, after detection of actuation, reading user input (text) from the control area 110; retrieving the text instruction Inst1 assigned to the actuated button S1; transmitting the read user input (Text2) and the called text instruction Inst1 to a Large Language Model (LLM); and receiving a response from the Large Language Model (Ant). 1,ohne hyperlinksand outputting the answer Ant 1,mit hyperlinks into tax area 110, whereby recognition of quotations, paragraph numbers and / or page references in the response of the Large Language Model Ant 1,ohne hyperlinks and converting or transforming the quotations, paragraph numbers, and / or page numbers into clickable buttons or hyperlinks, and issuing a control command to a text preview program that causes the text preview program to jump to the quotation, paragraph, and / or page and optionally highlight it visually. Control area 110 is, in this embodiment, the Fig.5. A chat dialog with a text input field Text2 and a button S1 for sending a message. After clicking the S1 button, Text2 and the instruction Inst1 are provided to the LLM. The LLM outputs a response Ant1 without hyperlinks. Subsequently, quotations, paragraph numbers, and page references are recognized in the response Ant1 without hyperlinks, and the recognized quotations, paragraph numbers, and / or page references are displayed in the response Ant1 with hyperlinks in the control area of ​​a chat window. If the user clicks on the hyperlink, the text preview, for example, of a PDF viewer, jumps to the clicked text passage.

Claims

[1] Computer program product for outputting text outputs generated using at least one Large Language Model (LLM), wherein the computer program product includes commands which, when executed by an end device (10), cause the end device to perform the following steps: - Displaying a control area (110) on a user interface (100) of the terminal device (10), - Representing at least one actuable button (S1, S2, Sn) arranged in the control area (110), wherein at least one text instruction (Inst1, Inst2, Instn) is assigned to the at least one button (S1, S2, Sn) in a data-technical manner and is stored in an instruction database (130), - Detect activation of at least one button (S1, S2, Sn), and after activation has been detected: i. Reading user input (Text2) from a text area (120) and / or from the control area (110), ii. Retrieving the text instruction (Inst1, Inst2, Instn) assigned to the pressed button (S1, S2, Sn), iii. Transmitting the input user command and the retrieved text instruction (Inst1, Inst2, Instn) to a Large Language Model (LLM), iv. Receiving a response from the Large Language Model (LLM), and v. Outputting the answer in the text area (120) and / or control area (110). [2] Computer program product according to claim 1, wherein the user input is text (120) selected by the user, and the computer program product is configured to perform the step: - Displaying an output setting means (140) in the control area (110) with which it is optionally possible to set that the response of the Large Language Model (LLM) is output before the marked text (120), is output after the marked text (120), is output on a new line, and / or that the response of the Large Language Model (LLM) replaces the marked text (120). [3] Computer program product according to claim 1 or 2, wherein the computer program product is configured to perform the step: - Representing a context window setting means (150) with which a reading context window (160) of the text area (120) can be set, wherein preferably the user input is text selected by the user and with the context window setting means (150) it can be optionally set that the reading context window (160) corresponds to the selected text, the reading context window (160) corresponds to a text section in which the selected text is arranged, the reading context window (160) corresponds to a predetermined number of preceding or following text sections, and / or the reading context window (160) corresponds to the entire text of the text area (120). [4] Computer program product according to any of the preceding claims, wherein the computer program product is configured to perform the step: - Reading in the first selected text from the text area (120) and saving the first selected text as a reading context window (160) in a temporary storage area, - Reading a second selected text from the text area (120), and - Transmitting the input context window (160), the second selected text and the retrieved text instruction (Inst1, Inst2, Instn) to a Large Language Model (LLM). [5] Computer program product according to any of the preceding claims, wherein the computer program product is configured to perform the step: - Visually perceptible highlighting of a read-in context window (160), wherein the visual highlighting depends on a cursor position (165) and a set read-in context window (160), Preferably, the highlighting is achieved by means of displayed special characters (170), by means of color highlighting, or by means of font type change or the like, and the highlighting includes the following additional steps: - Automated addition and removal of special characters (170), color highlighting and / or font change depending on the cursor position (165) and the set input context window (160). [6] Computer program product according to any of the preceding claims, wherein the computer program product is configured to perform the step: - Detect division indicators in the input user data, whereby the division indicators can be added by the user in the text area (120), - Dividing the input from the user into sub-sequences depending on the division indicators, - Parallel or sequential transmission of the read-in subsequences and the retrieved text instruction (Inst1, Inst2, Instn) to a Large Language Model (LLM), - Receiving multiple responses to each subsequence from a Large Language Model (LLM), and - Outputting the received responses to the text area (120) and / or control area (110), Preferably, the division indicator is a user-added underline in selected text or one or more special characters (170) in selected text. [7] Computer program product according to any of the preceding claims, wherein the computer program product is configured to perform the step: - Comparing the input user data with the received response of the Large Language Model (LLM), preferably by means of a deterministic comparison algorithm (200), in particular the comparison algorithm (200) is at least one algorithm from the list of algorithms comprising: i. an algorithm that determines a longest common subsequence between the marked text and the LLM response, for example an LCS algorithm; ii. a sentence embedding algorithm, wherein the sentence embedding algorithm comprises at least the following steps: computing a sentence embedding of the marked text, computing a sentence embedding of the response of the Large Language Model, and comparing the computed sentence embeddings; iii. a token alignment algorithm, wherein the token alignment algorithm decomposes the marked text and the received response into individual tokens, compares the tokenized sequences, and identifies deviations in the form of missing or altered tokens; iv. a hash-based verification algorithm, wherein the hash-based verification algorithm calculates a checksum for the marked text and for the received response, and compares the checksums; v. a Levenshtein distance algorithm, wherein the Levenshtein distance algorithm computes the minimum number of editing operations between the marked text and the received response, thereby identifying potential character loss; and - Visually highlighting differences between user input and received response, preferably by outputting matching parts between user input and received response in a deactivated change mode and outputting non-matching parts between the highlighted text and received response in an activated change mode. [8] Computer program product according to any of the preceding claims, wherein the computer program product is configured to perform the step: - Identifying quotations, paragraph numbers and / or page references in the Large Language Model (LLM) response and - Converting or transforming quotations, paragraph numbers and / or page numbers into clickable buttons or hyperlinks, and - Issuing a control command to a text preview program product that causes the text preview program product to jump to the quoted location, paragraph and / or page, and optionally causes the text preview program product to visually highlight the quoted location and / or paragraph. [9] Computer program product according to any of the preceding claims, wherein the computer program product is configured to perform the step: - Displaying an input field for a patent number in the tax area (110), - Display at least one register data selection button in the control area (110), - Triggering a registry data query to retrieve registry data from a public intellectual property registry based on the entered intellectual property number, - Storing the retrieved register data as individual options in the register data selection button, in order to make the register data individually selectable and perceptible to the user in the register data selection button, and optionally displaying an actuable button (S1, S2, Sn) in the control area (110), which, when actuated, outputs the content of the selected register data selection button option to the text area (120) and / or, when actuated, copies the content of the selected register data selection button option to an electronic clipboard. [10] Computer program product according to any of the preceding claims, wherein the computer program product is configured to perform the step: - Displaying a comment input button (S1) in the control area (110), whereby after pressing the comment input button (S1) the further sub-steps are executed: - Reading a comment text that is assigned to a text section in the text area (120), - Reading at least the text section to which the comment is assigned, - Retrieving a text instruction (Inst1) that is assigned to the comment incorporation button (S1), - Transmitting the extracted comment text, the extracted text segment, and the retrieved text instruction (Inst1) to a Large Language Model (LLM), and - Outputting the answer in the text area (120) depending on the comment text and the text instruction (Inst1).

Images