A large language model workflow visual arrangement system and method

Abstract

The present application relates to node linkage constraint and text length adaptation technical field, specifically, it relates to a kind of big language model workflow visual arrangement system and method, in the present application, abstract constraint modeling module identifies the serial relationship of technical abstract generation node and multilingual translation node, and extracts the total capacity of multilingual translation node context window from serial relationship, and deducts control instruction occupied capacity, calculates abstract output length upper threshold, to establish the negative mapping rule of temperature parameter and length, temperature threshold binding module linkage parameter adjustment and threshold calculation, feedback to configuration interface by visual identification and early warning prompt;Abstract intercept execution module intercepts over-limit text when workflow executes, and complete sentence is positioned by triple semantic boundary verification, and redundant content is cut off in conjunction with technical parameter coherence check, ensure abstract adaptation translation node capacity, guarantee workflow smoothly executes, improve arrangement accuracy and reliability.

Description

Technical Field

[0001] This invention relates to the field of node linkage constraints and text length adaptation technology, specifically, to a visualization and orchestration system and method for a large language model workflow. Background Technology

[0002] Node linkage constraints and text length adaptation are important technologies, specifically applied to workflow scenarios that link technical summary generation and multilingual translation. The core is to dynamically constrain the output length of the technical summary to adapt to the processing capabilities of the multilingual translation nodes, ensuring smooth workflow execution. In large language model workflows, technical summary generation nodes and multilingual translation nodes often form a direct chain relationship. The context window capacity of the multilingual translation node has a fixed limit, and executing the translation task requires a portion of the capacity to load control instructions. Furthermore, an increase in the temperature parameter of the technical summary generation node can lead to a non-linear increase in the output text length. Because the correlation between the temperature parameter and the summary length is not precisely adapted to the effective capacity of the translation node, the generated technical summary may exceed the processing limit of the translation node, resulting in truncated translation content, semantic distortion, or even workflow task failure. To address this issue, we provide a large language model workflow visualization orchestration system and method. Summary of the Invention

[0003] The purpose of this invention is to provide a large language model workflow visualization and orchestration system and method to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, a large language model workflow visualization orchestration system is provided, including a workflow orchestration unit, a node configuration unit, and a workflow execution engine unit. Its distinguishing feature is the addition of a calibration unit, which includes:

[0005] The abstract constraint modeling module is used to identify the serial relationship between the technical abstract generation node and the multilingual translation node during the process orchestration stage. It extracts the total capacity of the context window of the multilingual translation node from the serial relationship, deducts the capacity occupied by the preset translation control instructions, calculates the upper limit threshold of the output length of the technical abstract generation node, and simultaneously establishes a negative mapping rule between the temperature parameter adjustment amount and the upper limit threshold of the output length. It also defines a calibration logic in which the upper limit threshold of the output length decreases inversely when the temperature parameter increases.

[0006] The temperature threshold binding module is used to associate temperature parameters with the upper limit threshold of output length in the node configuration unit. When the temperature parameters of the technical summary generation node are adjusted, the upper limit threshold of output length corresponding to the current temperature value is calculated in real time according to the negative mapping rule, and the upper limit threshold of output length is fed back to the configuration interface of the node configuration unit in real time with a visual upper limit identifier.

[0007] The summary interception and execution module is used to monitor the output content of the technical summary generation node during the workflow execution phase. When the actual output length exceeds the upper limit threshold of the output length corresponding to the current temperature value, the module uses a sentence boundary detection algorithm to locate the end position of the first complete sentence from the end of the text and cuts off the continuous sentences that exceed the upper limit threshold of the output length, ensuring that the length of the summary text input to the multilingual translation node meets the upper limit threshold constraint of the output length.

[0008] The second objective of this invention is to provide a method for implementing a large language model workflow visualization and orchestration system, including any one of the above-described features, comprising the following steps:

[0009] S1. Identify the connection between the technical summary generation node and the multilingual translation node, extract the total capacity of the context window of the multilingual translation node and deduct the capacity occupied by the translation control instructions, calculate the upper limit threshold of the output length of the technical summary generation node, and simultaneously establish a negative dynamic mapping rule between the temperature parameter adjustment amount and the upper limit threshold of the output length.

[0010] S2. When the user adjusts the temperature parameters of the technical summary generation node, the current output length upper limit threshold is calculated in real time according to the negative dynamic mapping rule, and a visual upper limit indicator and warning prompt are rendered in the configuration interface.

[0011] S3. The temperature parameter is mapped to the upper limit threshold of the output length through the prediction model, and the final upper limit threshold of the output length is generated by combining the safety redundancy coefficient. The linkage between the temperature control and the upper limit of the length is bound in real time.

[0012] S4. During workflow execution, intercept the original summary text stream, count the number of tokens and compare them with the current output length upper limit threshold. If the limit is exceeded, locate the end position of the sentence based on triple semantic boundary verification, cut off the excess text, and then perform tail technical parameter coherence verification to ensure that the text input to the translation node meets the length constraint.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0014] This invention achieves precise adaptation between technical summary generation and multilingual translation nodes by adding a calibration unit, effectively solving the translation distortion problem caused by exceeding length limits. The summary constraint modeling module identifies the serial relationship between two nodes, deducts the token occupation of translation control instructions, and establishes a negative mapping between temperature parameters and the upper limit of output length by combining a prediction model trained on historical samples, ensuring that constraints fit the effective capacity of translation nodes. The temperature threshold binding module links parameter adjustment and threshold calculation in real time, allowing users to intuitively control capacity boundaries and avoid misoperation through visual indicators and early warning prompts. The summary interception and control execution module intercepts text exceeding limits during workflow execution, locates complete sentences through triple semantic boundary verification, and removes redundant content by combining end-of-line technical parameter coherence verification, maximizing the retention of core technical information. This achieves dynamic adaptation of temperature parameters, summary length, and translation node capacity, ensuring smooth workflow execution and improving the accuracy and reliability of large language model workflow orchestration. Attached Figure Description

[0015] Figure 1 This is an overall block diagram of the present invention;

[0016] Figure 2 This is the overall flowchart of the present invention.

[0017] The meanings of the labels in the diagram are as follows:

[0018] 1. Process orchestration unit; 2. Node configuration unit; 3. Workflow execution engine unit; 4. Calibration unit; 41. Summary constraint modeling module; 42. Temperature threshold binding module; 43. Summary interception execution module. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention provides a large language model workflow visualization and orchestration system. Please refer to [link / reference]. Figure 1 As shown, it includes a process orchestration unit 1, a node configuration unit 2, and a workflow execution engine unit 3. Its distinguishing feature is the addition of a calibration unit 4, which includes:

[0021] The abstract constraint modeling module 41 is used to identify the connection relationship between the technical abstract generation node and the multilingual translation node during the process orchestration stage. Specifically, it includes traversing the node connection graph generated by the process orchestration unit 1 to locate the direct data channel between the output port of the technical abstract generation node and the input port of the multilingual translation node. When it is detected from the direct data channel that the output data of the technical abstract generation node is configured as the only input source of the multilingual translation node, it is determined that the two form a connection relationship, and the connection relationship is recorded as a directed edge in the node relationship graph.

[0022] It needs further explanation that after the workflow orchestration unit completes the drag-and-drop and initial connection of nodes in the large language model workflow, the primary task of the abstract constraint modeling module 41 is to accurately identify the connection between the technical abstract generation node and the multilingual translation node. This relationship is the core prerequisite for the subsequent calculation of the upper limit threshold of the output length. Only by clarifying the data flow relationship between the two can we ensure that the subsequent constraint rules accurately adapt to the context window capacity of the multilingual translation node and avoid constraint failure due to ambiguity in the data link. The specific implementation method is as follows:

[0023] The technical summary generation node is a functional node in the workflow specifically designed to condense and summarize technical documents, patent texts, and engineering solutions. Its core output is structured technical summary text. The node has a built-in summary generation algorithm based on a large language model, and the level of detail and professionalism of the output text can be adjusted through parameter configuration. The multilingual translation node is a functional node that takes the output text from the upstream node and accurately translates it into the target language. Its operation depends on the context window capacity of the target large language model it is bound to. This capacity determines the total number of text tokens that the node can process in a single run. Exceeding the capacity will lead to truncation of the translation content, semantic distortion, or even task failure. Therefore, the length of its input text must be strictly limited. The connection relationship here specifically refers to the direct unidirectional data flow between the two nodes. The relationship is such that the output of the technical summary generation node serves as the sole input source for the multilingual translation node. Data flowing out of the technical summary generation node is directly injected into the multilingual translation node without any other nodes intervening for relaying, filtering, or supplementation, forming a closed-loop data link from the technical summary generation node to the multilingual translation node. This relationship is not expressed through mathematical expressions but is determined primarily by the port connection status and data flow direction between nodes. It represents a clear functional dependency in the workflow. The identification process is based on the node connection diagram generated by the workflow orchestration unit. The node connection diagram is a structured data model describing all nodes, ports, and connections in the workflow, including each node's unique ID, function type identifier, port list, port connection pairs, and data... Based on key information such as flow direction, the abstract constraint modeling module 41 uses a depth-first traversal algorithm to perform a full traversal of the node connection graph. During the traversal, the module focuses on retrieving the output port connections of all technical abstract generation nodes. The module first filters out all technical abstract generation nodes by their functional type identifiers, and then extracts the connection records of each node's output port one by one to locate the node type of the target port corresponding to that port. When a direct connection record is found between the output port of a technical abstract generation node and the input port of a multilingual translation node, and further verification of the multilingual translation node's input port connection list reveals that it is only bound to this one connection record from the technical abstract generation node, with no other node's output port connected to it. If the input port is identified, it is determined that the two nodes form a series relationship. To facilitate the rapid invocation of this relationship by subsequent modules, the module will record it as a directed edge in the node relationship graph. The node relationship graph is a graph structure database that stores the relationships between all nodes in the workflow. The starting point of the directed edge is the unique ID of the technical summary generation node, and the ending point is the unique ID of the corresponding multilingual translation node. The attribute fields of the edge are clearly marked as series relationships, and also include the unique identifier of the data channel and the connection establishment timestamp. This ensures that when extracting the context window capacity of the multilingual translation node and calculating the upper limit threshold of the output length, the associated technical summary generation node can be quickly located through the directed edge, providing clear node relationship support for the accurate establishment of constraint rules.

[0024] The abstract constraint modeling module 41 extracts the total capacity of the context window of the multilingual translation node by accessing the metadata interface of the target large language model bound to the multilingual translation node, and deducts the capacity occupied by the preset translation control instructions. It then calculates the upper limit threshold of the output length of the technical abstract generation node. In this process, when deducting the capacity occupied by the preset translation control instructions, a preset translation instruction template library is loaded and the total token occupancy required by the multilingual translation node to perform the translation task is calculated. The upper limit threshold of the output length of the technical abstract generation node is generated by subtracting the total token occupancy from the total capacity of the context window.

[0025] It needs further explanation that after successfully extracting the total capacity of the context window for the multilingual translation node, not all of the capacity can be used to hold the output text of the technical summary generation node. When the multilingual translation node performs translation tasks, it also needs to load the system's built-in translation control instructions. These instructions will occupy a certain amount of context window resources. This occupied capacity is the preset capacity occupied by the translation control instructions. Its core is the total number of tokens required by system instructions, format control characters, and meta-instructions. It needs to be accurately calculated and deducted from the total capacity to ensure that the total number of tokens for the summary text and control instructions does not exceed the window limit. The specific implementation method is as follows:

[0026] First, it's necessary to clarify the core function and structure of the pre-built translation instruction template library. It's a pre-stored set of structured instructions, covering standard instruction templates for different translation scenarios. Each template is optimized to ensure it contains three core elements essential for performing translation tasks: system instructions, used to clarify translation objectives (e.g., translating input text into English while maintaining the accuracy and consistency of technical terminology and ignoring meaningless punctuation redundancy); format control characters, used to standardize the output format of translation results, ensuring clear output text structure and adaptability to subsequent use cases; and meta-instructions, used to assist in translation task execution and result feedback, while also allowing users to customize and save existing templates, ensuring rapid matching of the specific needs of the current translation task and computational complexity. When determining the total token usage required for a language translation node to execute a translation task, the system retrieves matching instruction templates from a pre-built translation instruction template library based on the current multilingual translation node's task configuration. If a template perfectly matches the configuration, it is directly invoked; otherwise, templates are combined to generate a suitable instruction set. The system's built-in tokenization engine is then invoked, and tokenization statistics are performed segment by segment according to the token counting rules of the target large language model. The token partitioning logic is consistent across different large language models: English words are divided into individual tokens by spaces, Chinese characters are counted as a single token, and punctuation marks, instruction keywords, and meta-instruction identifiers are each counted as a single token. For example, the system instruction translates the input technical summary into English and retains... The paragraph structure is broken down into 11 tokens for translating the input technical summary into English while retaining the paragraph structure. Formatting controls include 3 newline characters and 2 paragraph breaks, totaling 5 tokens. The meta-instruction's translation confidence score is broken down into 4 tokens for returning the translation confidence score. The token statistics for system instructions, formatting controls, and meta-instructions are summed to obtain the total token usage required for this translation task. The total token usage essentially represents the total number of tokens after all translation control instructions are converted into tokens recognizable by the target large language model. It is a core indicator for measuring the context window resources occupied by control instructions, ensuring that the statistical results accurately reflect the actual usage. After obtaining the total token usage, the total capacity of the previously extracted multilingual translation node context window is subtracted... The total token usage amount is used to generate the upper limit threshold for the output length of the technical summary generation node. The core of this calculation logic is total capacity allocation. The total capacity of the context window is the upper limit of the total number of tokens that the multilingual translation node can process in a single operation. After subtracting the token usage required for translation control instructions, the remaining capacity is the effective resource available to carry the technical summary text. This avoids the total number of tokens for the summary text and control instructions exceeding the window capacity due to an excessive number of tokens. If the total capacity of the context window of the multilingual translation node is 4096 tokens, and the total token usage for this translation task is calculated to be 210 tokens, then the upper limit threshold for the output length of the technical summary generation node is 3886 tokens, clearly defining the maximum number of tokens for the summary text.This provides a precise basis for subsequent binding of temperature parameters and length thresholds, as well as text interception control during the execution phase.

[0027] A negative mapping rule between the temperature parameter adjustment and the upper limit threshold of the output length is established simultaneously. This is achieved by training a prediction model for temperature and length using historical task samples. Specifically, this includes:

[0028] Collect abstract samples and their corresponding lengths generated by the technical abstract generation node under different temperature parameters, construct a regression dataset of temperature parameters and output length, train a neural network model based on the regression dataset, fit the negative inverse proportional function relationship that the output length increases non-linearly with increasing temperature, and encode the neural network model as a calibration function that maps the temperature parameter adjustment amount to the upper limit threshold of the output length, and define the calibration logic that the upper limit threshold of the output length decreases in an inverse proportional relationship when the temperature parameter increases.

[0029] Further explanation is needed: After calculating the upper limit threshold of the output length of the technical summary generation node, in order to address the issue that the output length of the summary increases non-linearly due to the rise in temperature parameters, thus exceeding the threshold, it is necessary to simultaneously establish a negative mapping rule between the temperature parameter adjustment amount and the upper limit threshold of the output length. This rule is implemented by training a neural network model using historical task samples. The core is to let the model learn the non-linear relationship between temperature and output length, and then transform it into a calibration function that can be called in real time. The specific implementation method is as follows:

[0030] Constructing a regression dataset for temperature parameters and output length is fundamental to model training. The core is collecting enough samples to cover the common range of temperature parameters and the common length of summary outputs. First, the gradient of temperature parameter values ​​is determined, selecting a range commonly used in large language model technical summary generation scenarios (0.1 to 1.0). Ten gradients are created with a step size of 0.1, and two intermediate values ​​are added to each gradient, resulting in 28 different temperature parameter values. For each temperature parameter value, other parameters of the technical summary generation node are kept consistent. 100 sets of technical documents on different topics are input, and the node generates 100 corresponding technical summaries. Each summary records two core data points: the corresponding temperature parameter value and the output length. After collection, the samples were preprocessed to remove invalid samples and filter outliers, ultimately retaining 80-100 valid samples for each temperature parameter, forming an original dataset containing approximately 2500 valid records. This dataset was then divided into a training set (for model training) and a validation set (for monitoring training effectiveness) in a 7:3 ratio. The dataset was organized according to input features and labels. The input features were single-dimensional temperature parameter values, and the labels were the corresponding output length tokens, ensuring the model could directly learn the mapping relationship between the two. When training the neural network model based on this regression dataset, a model structure suitable for single-input, single-output nonlinear regression tasks was first selected. The input layer consisted of one neuron, corresponding to the temperature... For parameter input, the hidden layer has three layers with 32, 64, and 32 neurons per layer, respectively. The ReLU activation function is used to fit the non-linear relationship between temperature and output length. The output layer has one neuron, corresponding to the predicted output length tokens, and a linear activation function is used to ensure the output value covers a reasonable range of summary length. The loss function is mean squared error, used to measure the deviation between the predicted output length and the actual sample length. The Adam optimizer is used, with an initial learning rate of 0.001. A learning rate decay strategy is employed to avoid training oscillations. Training is batch-based, with the training set divided into batches of 32 samples. Each iteration inputs the batch data into the model. The input layer is passed to the hidden layer, where it undergoes a non-linear transformation using the ReLU activation function. The predicted output length is then output by the output layer. The mean squared error between the predicted value and the true label is calculated, and this error is propagated back from the output layer to each hidden layer via backpropagation. The weights and bias parameters of each layer are updated, and this process is repeated for 50 iterations. The validation loss is then calculated using the validation set. If the validation loss does not decrease after 100 consecutive iterations, an early stopping mechanism is triggered to terminate training and prevent overfitting. The entire training process continues until both the training and validation losses stabilize. At this point, the model has fully learned the non-linear relationship between temperature parameters and output length. The core objective of model training is to fit the pattern of non-linear growth in output length caused by increasing temperature.This supports a negative mapping rule where the upper limit threshold of the output length decreases inversely with increasing temperature. This inverse proportional relationship is not a strict mathematical formula, but rather refers to the inverse trend of the two. As the temperature parameter gradually increases from low to high, the output length of the technical summary generation node exhibits non-linear growth, while the upper limit threshold of the output length must decrease accordingly, with the rate of decrease matching the rate of increase in output length. This ensures that regardless of temperature adjustments, the summary output length will not exceed the threshold. The model automatically fits this non-linear relationship by learning from the patterns in the training set. This data-driven, automatically learned relationship perfectly matches the non-linear growth of temperature and output length. The established pattern provides model support for the subsequent establishment of negative mapping rules. When the temperature parameter is adjusted, the model can accurately predict the corresponding increase in output length, and then inversely deduce the decrease in the upper limit threshold of output length. After training, the neural network model is encoded as a calibration function that maps the temperature parameter adjustment to the upper limit threshold of output length. The core is to solidify the training parameters of the model, such as weights and biases, and transform them into computational logic that the system can call in real time. First, the trained model parameters are extracted, including the weight matrix between the input layer and the first hidden layer, the weight matrix between each hidden layer, the weight matrix between the hidden layer and the output layer, and the bias vector of each hidden layer. These parameters are stored in binary format. The model parameter library is retrieved to ensure fast access during calls. The core logic of the calibration function is then written. The function's input is the temperature parameter adjustment amount, i.e., the difference between the user's current adjusted temperature parameter value and the initial temperature parameter value, or the current temperature parameter value can be directly input. Internally, the function first standardizes the input temperature parameter value, converting it to the input format used during training. Then, it performs forward propagation calculations according to the model's network structure, inputting the standardized temperature parameter to the input layer. This parameter is then multiplied sequentially by the weight matrices of each layer and superimposed with the bias vector. After processing by the ReLU activation function, the final predicted output length value is obtained through the linear activation function of the output layer. Based on this predicted output length value, combined with the previously calculated multilingual... The available token capacity of the translation node is used to calculate the upper limit threshold of the output length according to a negative mapping rule. For example, if the available capacity is 3886 tokens and the model predicts an output length of 1500 tokens at a temperature of 0.5, then the upper limit threshold of the output length is set inversely proportionally as 3886 × (initial temperature output length / current temperature predicted output length). This ensures that the threshold decreases inversely as the temperature increases. Finally, this entire calculation logic is encapsulated into an independent calibration function and integrated into the summary constraint modeling module 41. This ensures that when the user adjusts the temperature parameters, the function can receive the parameter values ​​in real time, quickly complete the calculation, and output the corresponding upper limit threshold of the output length, providing an accurate calculation basis for the temperature threshold binding module 42.

[0031] The temperature threshold binding module 42 is used to associate temperature parameters with the upper limit threshold of output length in the node configuration unit 2. When the temperature parameter of the technical summary generation node is adjusted, the upper limit threshold of output length corresponding to the current temperature value is calculated in real time according to the negative mapping rule. When the temperature threshold binding module 42 calculates the upper limit threshold of output length corresponding to the current temperature value according to the negative mapping rule, it calls the prediction model and uses the temperature parameter input value adjusted by the user in real time as the independent variable of the prediction model. The predicted output length value is obtained through forward propagation calculation, and the predicted output length value is multiplied by the preset safety redundancy coefficient to generate the final upper limit threshold of output length.

[0032] It needs further explanation that when the temperature threshold binding module 42 calculates the upper limit threshold of the output length corresponding to the current temperature value according to the negative mapping rule, the prediction model called is a neural network model that has been trained with historical samples. Its internal weights, biases and other parameters are all fixed. The core of the forward propagation calculation is to convert the user's real-time adjusted temperature parameter input value into a format that the model can process, and output the predicted output length value after layer-by-layer calculation. The whole process does not require backpropagation to update parameters, and only needs to complete the data flow according to fixed logic. The specific implementation method is as follows:

[0033] The first step in forward propagation computation is to standardize the input temperature parameter value. After the user adjusts the temperature parameter through the parameter slider in the node configuration unit, the system obtains the original input value of the parameter (e.g., 0.6). Since the temperature parameters used during model training have all been normalized, mapping the temperature range of 0.1-1.0 to the interval of 0-1, the current input value needs to be transformed according to the same normalization rule to avoid prediction bias caused by inconsistent input scales. The transformation logic is to subtract the minimum temperature value (0.1) of the training dataset from the current temperature parameter value, and then divide by the difference between the maximum temperature value (1.0) and the minimum temperature value (0.1) of the training dataset to obtain the standardized input value, such as 0. The standardized value corresponding to .6 is (0.6-0.1) / (1.0-0.1)=0.556. This standardized value serves as the initial input data for forward propagation and is input to the model's input layer. The model's input layer contains only one neuron, whose function is to receive the standardized temperature parameter value and pass it to the first hidden layer. The first hidden layer contains 32 neurons. Each neuron receives the output value of the input layer and performs weighted summation and bias superposition operations. Specifically, each neuron calls its corresponding weight parameter (a value fixed during training) to multiply with the input layer's output value, and then superimposes the neuron's bias parameter (also a value fixed after training) to obtain 32 preliminary operation results. For example, if the input layer output value is 0.556, the weight of a certain neuron is 1.2, and the bias is -0.3, then the computation result of this neuron is 0.556 × 1.2 + (-0.3) = 0.367. Subsequently, these 32 computation results are input one by one into the ReLU activation function for nonlinear transformation. The ReLU function sets all negative results to 0, while keeping positive results unchanged, thereby introducing nonlinear features so that the model can fit the nonlinear relationship between temperature and output length. For example, the result of 0.367 above is still 0.367 after ReLU processing, but if the computation result of a certain neuron is -0...If the result is 2, it becomes 0 after processing. The first hidden layer outputs 32 non-linearly transformed feature values. These 32 feature values ​​serve as input to the second hidden layer, which contains 64 neurons. The computational logic is the same as the first hidden layer: each neuron multiplies its 32 input feature values ​​by its corresponding 32 weight parameters (the weight matrix dimension is 64×32), and then adds its own bias parameters, resulting in 64 linear combinations. These are then subjected to a ReLU activation function for non-linear transformation, outputting 64 higher-level feature values. This step, by increasing the number of neurons and feature dimensions, further enhances the model's ability to fit complex non-linear relationships. The 64 feature values ​​from the second hidden layer are then passed to the third hidden layer, and so on. The third hidden layer contains 32 neurons, and its operational logic is the same as the first two hidden layers. Each neuron receives 64 input feature values, which are then weighted, summed, and activated by ReLU, outputting 32 refined feature values. The core function of this layer is to reduce and filter the high-dimensional features of the second hidden layer, retaining the key features most relevant to the output length prediction and avoiding redundant features from affecting prediction accuracy. The 32 feature values ​​of the third hidden layer are finally passed to the output layer, which contains only one neuron and uses a linear activation function, meaning the output value equals a linear combination of the inputs without any additional nonlinear transformations. This neuron multiplies each of the 32 input feature values ​​by its corresponding 32 weight parameters, and then adds the bias parameters to obtain the final linear value. The combined result is the standardized predicted output length value. Subsequently, the system will restore this standardized predicted value to the actual output length tokens according to the inverse normalization rule used during training. For example, if the standardized predicted value is 0.6, and the maximum output length of the training dataset is 3000 and the minimum is 500, then the actual predicted output length value is 0.6 × (3000 - 500) + 500 = 2000. This completes the entire forward propagation calculation process, obtaining the predicted output length value corresponding to the user's current temperature parameter. After obtaining the predicted output length value, it needs to be multiplied by a preset safety redundancy coefficient to generate the final upper limit threshold of the output length. This safety redundancy coefficient is a fixed coefficient between 0.85 and 0.95, with a default value of [missing value]. The coefficient is 0.9, which can be adjusted according to the actual application scenario. Its core definition is the capacity redundancy ratio reserved to compensate for the prediction error of the prediction model. In essence, it avoids the translation task failure caused by the number of summary text tokens exceeding the context window capacity of the multilingual translation node due to the model's predicted value being lower than the actual output length (i.e., prediction error). The value of this coefficient is calibrated through a large number of experiments: 100 test samples under different temperature parameters are selected, and the error rate between the model's predicted output length and the actual output length is calculated. The maximum error rate is statistically analyzed (usually not exceeding 15%), and the safety redundancy coefficient is set to 1 minus 1.2 times the maximum error rate (e.g., if the maximum error rate is 10%, then the coefficient is 1 - 10% × 1).The redundancy coefficient (2=0.88) ensures that even in extreme error scenarios, the length of the summary text is strictly controlled within the window capacity. For example, if the predicted output length is 2000 tokens and the safety redundancy coefficient is 0.9, the final output length upper limit threshold is 2000 × 0.9 = 1800 tokens. This provides sufficient redundancy for prediction errors without excessively reducing the effective capacity and causing loss of summary information. The safety redundancy coefficient is used to compensate for the boundary overflow risk caused by prediction model errors. The output length upper limit threshold is fed back to the configuration interface of node configuration unit 2 in real time as a visual upper limit indicator. This includes rendering the visual upper limit indicator in the parameter panel of the technical summary generation node of node configuration unit 2, covering the over-limit area of ​​the temperature parameter slider track, and simultaneously displaying the output length upper limit threshold value corresponding to the current temperature value and the percentage of remaining available capacity. When the user adjusts the temperature parameter to make the predicted output length reach the window capacity limit, a threshold warning animation is activated to enhance the visual prompt.

[0034] The summary interception and execution module 43 is used to intercept the original summary text stream during the callback phase after the technical summary generation node is completed when monitoring the output content of the technical summary generation node during the workflow execution phase. It calls the tokenization engine to count the number of tokens in the original summary text stream in real time and compares the number of tokens with the current output length upper limit threshold generated by the temperature threshold binding module 42. If the output length upper limit threshold is exceeded, the semantic boundary truncation process is triggered; otherwise, the complete summary text is passed to the multilingual translation node.

[0035] It needs further explanation that, during the process of the workflow execution engine unit driving the large language model workflow according to the orchestration logic, the summary interception execution module 43 needs to accurately control the output stage of the technical summary generation node. Through a series of operations such as interception, statistics, and comparison, it ensures that the length of the summary text meets the constraints. This process relies on precise control of specific stages, an efficient data interception mechanism, and a clear logical judgment process. The specific implementation method is as follows:

[0036] First, it's important to clarify that the workflow execution phase refers to the entire process from workflow startup to all nodes sequentially completing their tasks and outputting the final result. This encompasses all key stages, including node initialization, parameter loading, task execution, result output, and data transfer between nodes. It is the core stage of the workflow lifecycle. The monitoring and interception operations of the summary interception execution module 43 are triggered at specific node output stages within this phase. The technical summary generation node's execution completion callback stage is a specific sub-stage within the workflow execution phase belonging to the technical summary generation node. Specifically, it refers to the moment after the node completes the summary text generation task, it reports its execution status (success / failure) to the workflow execution engine unit and prepares to output the result. This stage is the optimal time to intercept the original summary text stream. At this point, the summary text has been fully generated and has not yet been transmitted to the next node (multilingual translation node). This allows for interference-free interception without affecting the operation of other nodes, while avoiding data inconsistencies caused by late interception leading to partial text transmission. The specific implementation of the original summary text stream relies on the linkage between the callback mechanism of the module and the workflow execution engine unit. During system initialization, the summary interception and control execution module 43 registers the execution completion callback function of the technical summary generation node with the workflow execution engine unit. This callback function is granted data interception permissions, and its trigger condition is set to the technical summary generation node returning the 'execution successful' status code. When the technical summary generation node completes the summary generation task, it will automatically send feedback information containing the execution status code and the original summary text stream to the workflow execution engine unit. After receiving this information, the workflow execution engine unit immediately calls the registered callback function. The callback function then starts the data interception process. By reading the text stream data pointer in the feedback information, it directly obtains the memory address of the original summary text stream and creates a read-only copy of the text stream based on this address. This avoids node output abnormalities caused by directly manipulating the original data, while ensuring that the intercepted text stream is completely consistent with the original text generated by the node, without any data loss or tampering. The raw summary text stream here refers to the unprocessed output text data of the technical summary generation node, stored as a continuous byte stream. It contains all generated technical summary content, original punctuation, paragraph structure, and other information, serving as the foundational data for subsequent processing. After text stream interception, the tokenization engine is used to count the number of tokens in the summary text. The tokenization engine is a built-in text transcoding and counting tool. Its core function is to convert natural language text into the smallest semantic unit (i.e., token) recognizable by the target large language model according to the token partitioning rules, and then count the total number of tokens. Its working principle is as follows:

[0037] First, the token dictionary of the target large language model bound to the multilingual translation node is loaded. This dictionary contains all the pre-set tokens and corresponding text fragments of the model. Then, the intercepted raw summary text stream is scanned and parsed character by character and word by word. For Chinese text, each individual Chinese character and punctuation mark is usually counted as one token. For English text, each word, hyphenated word, and punctuation mark is counted as one token. Special format characters are counted as tokens individually or in combination according to the model rules. For example, the Chinese summary "the core parameter of this technology is 50MPa" will be parsed as "the core parameter of this technology is 50MPa", which is 11 tokens; the English phrase "technical parameter 100kW" will be parsed as "technical parameter 100kW". The meter100kW has a total of 4 tokens. After parsing, the tokenization engine directly outputs the total number of tokens obtained from the statistics. This value directly reflects the size of the target large language model context window occupied by the summary text. Then, the token count comparison and logical judgment stage begins. The core is to accurately compare the obtained token count with the current output length upper limit threshold generated by the temperature threshold binding module 42. Based on the comparison result, different processing logic is executed. First, the summary interception execution module 43 retrieves the current output length upper limit threshold calculated and stored by the temperature threshold binding module 42 in real time through the system's internal data interface. This threshold is determined by combining the multilingual translation node context window capacity, the amount of translation control instructions occupied, the negative mapping rule of temperature parameters, and... The final constraint value generated by the security redundancy coefficient is the sole criterion for determining whether the summary text length is compliant. During comparison, the module compares the total number of tokens output by the tokenization engine with this threshold. If the total number of tokens is less than or equal to the current output length upper limit threshold, it indicates that the summary text length is within the compliant range and will not exceed the processing capacity of the multilingual translation node. At this time, the module releases the intercepted original summary text stream and transmits the complete, unmodified summary text to the input port of the multilingual translation node through the workflow's preset data transmission channel—the direct data channel between the previously identified technical summary generation node and the multilingual translation node—ensuring that the translation task can start normally based on the complete technical summary. If the total number of tokens is greater than the current output length upper limit threshold, it indicates that the summary text length is within the compliant range and will not exceed the processing capacity of the multilingual translation node. If the length of the abstract text exceeds the limit, continued transmission will cause the context window of the multilingual translation node to overflow, leading to translation failure or semantic distortion. At this time, the module immediately triggers the semantic boundary truncation process. The startup logic of this process is that the module sends a length limit trigger signal to the system, locks the intercepted original abstract text stream, and prevents it from being transmitted to the multilingual translation node. Then, it calls the sentence boundary detection algorithm and truncation execution logic to prepare for semantic lossless truncation of the excess text, ensuring that the length of the abstract text input to the multilingual translation node meets the constraints, while preserving core technical information to the maximum extent. The entire comparison and judgment process is executed automatically, with the response time controlled in milliseconds, without affecting the overall execution efficiency of the workflow, achieving a balance between length constraints and smooth process.

[0038] When the actual output length exceeds the upper limit threshold corresponding to the current temperature value, the sentence boundary detection algorithm locates the end position of the first complete sentence from the end of the text backwards and performs triple semantic boundary verification. After triggering the semantic boundary truncation process, the core is to accurately locate the end position of the complete sentence to avoid semantic fragmentation or loss of technical information due to truncation. This operation relies on the sentence boundary detection algorithm, a technology specifically designed to identify sentence boundaries in text. Its core logic combines text format features, punctuation rules, and semantic coherence to determine the start and end points of sentences. It is particularly suitable for technical abstract texts with dense technical terminology and relatively standardized sentence structures, but which may contain missing punctuation, ensuring that the located boundaries conform to language expression habits without compromising technical information. To ensure the integrity of the technical information, the localization process proceeds from the end of the text forward, following a triple semantic boundary verification logic with decreasing priority. This ensures that reliable sentence ending positions can be found under different text formats. The triple semantic boundary verification includes scanning from the end of the text forward, prioritizing the identification of paragraph separators and line breaks as primary boundaries. If there are no paragraph separators, combinations of periods, question marks, exclamation marks, and subsequent spaces are detected as secondary boundaries. If the secondary boundaries fail, a pre-trained punctuation completion model is used to insert virtual sentence separators as tertiary boundaries. This ensures the semantic integrity before and after the localization point by removing continuous sentences exceeding the output length upper limit threshold, ensuring that the length of the summary text input to the multilingual translation node meets the output length upper limit threshold constraint. Specifically:

[0039] First, a primary boundary verification is initiated, prioritizing the identification of paragraph separators and line breaks as the core criteria for sentence ending positions. Paragraph separators include pre-defined paragraph marks and special segmentation symbols in the text, while line breaks are control characters used for line breaks. These symbols naturally mark the end of a semantic unit and are the most reliable sentence boundaries. During localization, the summary interception execution module 43 starts from the last character of the original summary text stream and scans forward character by character, matching the feature codes of paragraph separators and line breaks in real time. Once a corresponding symbol is detected, the scanning is immediately paused, and the position of that symbol is determined as the end position of the first complete sentence, without proceeding to subsequent verification. The verification process ensures both efficiency and accuracy in positioning. If the first-level boundary verification fails to detect a valid separator, the second-level boundary verification is automatically triggered, detecting combinations of periods, question marks, exclamation marks, and subsequent spaces. These three types of punctuation are standard symbols indicating the end of a sentence in natural language. The constraint of following a space is to avoid misjudging special punctuation marks in technical scenarios. During scanning, the process also searches character by character from the end of the text backward. When a period, question mark, or exclamation mark is detected, it is immediately checked whether the next character after the punctuation mark is a space, tab, or other whitespace character. If the combination of punctuation and whitespace character is satisfied, the punctuation mark position is determined to be the end of the sentence. If the punctuation mark is followed by a space, the next character is checked to see if it is the end of the sentence. Directly connecting letters, numbers, or technical symbols is considered a non-sentence-ending punctuation mark, and the scan continues forward until a punctuation position that meets the combination conditions is found. This ensures that punctuation marks in technical parameters are not misjudged as sentence boundaries. If secondary boundary verification still fails to find a valid boundary, tertiary boundary verification is initiated. A virtual sentence separator is inserted using a pre-trained punctuation missing completion model. This model is a natural language processing model trained on a large amount of technical summary text. The training data covers various technical document types such as patents and engineering solutions, including samples with complete sentence-ending punctuation and missing punctuation. The model can learn the semantic pause patterns of technical texts. In specific operation, the model processes the text... The text truncated from the end (within the range of tokens corresponding to the upper limit threshold of the output length) is semantically analyzed to identify semantically complete phrases or clauses. Virtual sentence separators are inserted at positions that conform to the sentence end logic, and the position of the separator is determined as the sentence end position. This ensures that even without explicit punctuation, the boundary can be located based on semantic integrity. Through this triple-priority decreasing verification logic, the end position of the first complete sentence can be accurately located from the end of the text. This not only ensures the reliability of boundary location but also adapts to scenarios such as non-standard formatting and missing punctuation in technical abstract texts, laying the foundation for subsequent accurate truncation and preservation of core semantics.

[0040] After triggering the semantic boundary truncation process, the summary truncation control execution module 43 performs sentence cutting based on the sentence end position, including truncating all continuous text from the sentence end position forward until the total number of tokens drops to the upper limit threshold of the output length, and performing tail semantic coherence verification on the truncated summary text. If technical parameter keywords are detected to be cut off, it will fall back to the previous sentence boundary to re-truncate and discard the text segment that failed the verification.

[0041] It needs further explanation that after accurately locating the end of the first complete sentence using the sentence boundary detection algorithm, the summary truncation execution module 43 immediately initiates the sentence truncation process. The core objective is to reduce the total number of summary text tokens to within the upper limit threshold of the output length while preserving the core semantics and complete technical information of the technical summary to the maximum extent possible, avoiding semantic fragmentation or loss of key technical parameters due to truncation. The specific implementation method is as follows:

[0042] Sentence truncation begins at the located sentence end position and proceeds backward through all continuous text. This sequential design prioritizes preserving the essential technical information at the beginning of the text, removing only redundant portions exceeding the capacity at the end. During truncation, the module continuously calls the tokenization engine to count the number of tokens in the remaining text. Each time a character or semantic unit is truncated, the token count is updated synchronously until the total number of tokens falls within the output length limit threshold. At this point, truncation stops. For example, if the output length limit threshold is 1800 tokens and the text at the located sentence end position has 2200 tokens, truncation proceeds backward from that position, counting tokens after each segment. When the remaining text has 1780 tokens, the threshold requirement is met, and truncation stops. At this point, the text content corresponding to the 420 tokens exceeding the capacity has been preliminarily removed. The truncation operation is complete. Instead of directly outputting the result, it performs a tail semantic coherence verification on the truncated abstract text. Tail semantic coherence verification is a semantic validation process specifically designed for technical abstract texts. Its core is to check whether the end of the truncated text is semantically complete, without half-sentences, split technical terms, or incomplete technical parameters, ensuring that the output abstract text accurately conveys technical information without ambiguity. During verification, the module loads a pre-set technical parameter keyword library and simultaneously calls a semantic analysis model to perform semantic parsing on the tail of the truncated text. First, it identifies whether the tail has a complete grammatical structure, then matches it against the technical parameter keyword library to check if any keywords have been truncated. If the semantic analysis model detects technical... If a parameter keyword is truncated or the semantics at the end are incomplete, a rollback mechanism is immediately triggered, reverting to the previous sentence boundary for re-truncation. The previous sentence boundary refers to the end position of the previous complete sentence located through the previous triple semantic boundary verification, preceding the end position of the current sentence. For example, if the currently located sentence end position is a period at the second-level boundary, the previous sentence boundary might be the previous paragraph separator or a period. During the rollback process, the module automatically discards the text segment between the current sentence end position and the previous sentence boundary, i.e., the text segment that failed verification, because this part of the text either has semantic incompleteness or contains truncated technical parameters. Retaining it would affect the accuracy of the summary. After re-truncation, the module is re-adjusted. The tokenization engine is used to count the number of tokens to ensure that it meets the upper limit threshold constraint of the output length. At the same time, the semantic coherence of the tail is verified again until a semantically complete summary text with a compliant number of tokens is obtained. For example, after the initial truncation, it is found that the core weight of the model at the tail is 0.8, which is cut off to the core weight of the model. The technical parameter 0.8 is detected to be missing, so it goes back to the end of the previous sentence and truncates to that position. The part of the text in the middle where the core weight of the model is 0.8 fails the verification. If the number of tokens of the remaining text meets the threshold requirement and the semantics of the tail are complete, the final truncation is completed, ensuring that the summary text of the input multilingual translation node does not exceed the limit and can completely retain the core technical information.

[0043] In this invention, the abstract constraint modeling module 41 identifies the concatenation relationship between the technical abstract generation node and the multilingual translation node, extracts the total capacity of the multilingual translation node context window from the concatenation relationship, deducts the capacity occupied by control instructions, and calculates the upper limit threshold of the abstract output length to establish a negative mapping rule between temperature parameters and length. The temperature threshold binding module 42 links parameter adjustment and threshold calculation, and provides feedback to the configuration interface through visual indicators and early warning prompts. The abstract interception and control execution module 43 intercepts text exceeding the limit during workflow execution, locates complete sentences through triple semantic boundary verification, and removes redundant content in combination with technical parameter coherence verification to ensure that the abstract adapts to the capacity of the translation node, ensures smooth workflow execution, and improves the accuracy and reliability of arrangement.

[0044] The second objective of this invention is to provide a method for implementing a large language model workflow visualization and orchestration system, including any of the above-mentioned features, comprising the following steps:

[0045] S1. Identify the connection between the technical summary generation node and the multilingual translation node, extract the total capacity of the context window of the multilingual translation node and deduct the capacity occupied by the translation control instructions, calculate the upper limit threshold of the output length of the technical summary generation node, and simultaneously establish a negative dynamic mapping rule between the temperature parameter adjustment amount and the upper limit threshold of the output length.

[0046] S2. When the user adjusts the temperature parameters of the technical summary generation node, the current output length upper limit threshold is calculated in real time according to the negative dynamic mapping rule, and a visual upper limit indicator and warning prompt are rendered in the configuration interface.

[0047] S3. The temperature parameter is mapped to the upper limit threshold of the output length through the prediction model, and the final upper limit threshold of the output length is generated by combining the safety redundancy coefficient. The linkage between the temperature control and the upper limit of the length is bound in real time.

[0048] S4. During workflow execution, intercept the original summary text stream, count the number of tokens and compare them with the current output length upper limit threshold. If the limit is exceeded, locate the end position of the sentence based on triple semantic boundary verification, cut off the excess text, and then perform tail technical parameter coherence verification to ensure that the text input to the translation node meets the length constraint.

[0049] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A large language model workflow visualization orchestration system, comprising a workflow orchestration unit (1), a node configuration unit (2), and a workflow execution engine unit (3), characterized in that, A calibration unit (4) is added, which includes: The abstract constraint modeling module (41) is used to identify the serial relationship between the technical abstract generation node and the multilingual translation node in the process orchestration stage, extract the total capacity of the context window of the multilingual translation node from the serial relationship, deduct the capacity occupied by the preset translation control instructions, calculate the upper limit threshold of the output length of the technical abstract generation node, establish a negative mapping rule between the temperature parameter adjustment amount and the upper limit threshold of the output length, and define the calibration logic that the upper limit threshold of the output length decreases in an inverse proportion when the temperature parameter increases. The temperature threshold binding module (42) is used to associate temperature parameters with the upper limit threshold of output length in the node configuration unit (2). When the temperature parameters of the technical summary generation node are adjusted, the upper limit threshold of output length corresponding to the current temperature value is calculated in real time according to the negative mapping rule, and the upper limit threshold of output length is fed back to the configuration interface of the node configuration unit (2) in real time with a visual upper limit identifier. The summary interception execution module (43) is used to monitor the output content of the technical summary generation node during the workflow execution phase. When the actual output length exceeds the upper limit threshold of the output length corresponding to the current temperature value, the end position of the first complete sentence is located from the end of the text forward based on the sentence boundary detection algorithm, and the continuous sentences exceeding the upper limit threshold of the output length are cut off to ensure that the length of the summary text input to the multilingual translation node meets the upper limit threshold constraint of the output length.

2. The large language model workflow visualization and orchestration system according to claim 1, characterized in that: The abstract constraint modeling module (41) is used to identify the connection between the technical abstract generation node and the multilingual translation node during the workflow orchestration stage, specifically including: The process orchestration unit (1) is used to generate a node connection graph. By traversing the node connection graph, it locates the direct data channel between the output port of the technical summary generation node and the input port of the multilingual translation node. When the output data of the technical summary generation node is detected from the direct data channel as the only input source of the multilingual translation node, it is determined that the two form a series relationship and the series relationship is recorded as a directed edge in the node relationship graph.

3. The large language model workflow visualization and orchestration system according to claim 2, characterized in that, The abstract constraint modeling module (41) extracts the total capacity of the context window of the multilingual translation node by accessing the metadata interface of the target large language model bound to the multilingual translation node, and deducts the capacity occupied by the preset translation control instructions, and calculates the upper limit threshold of the output length of the technical abstract generation node. When deducting the capacity occupied by the preset translation control instructions, a preset translation instruction template library is loaded and the total token occupancy of system instructions, format control characters and meta instructions required by the multilingual translation node to perform translation tasks is calculated. The upper limit threshold of the output length of the technical abstract generation node is generated by subtracting the total token occupancy from the total capacity of the context window.

4. The large language model workflow visualization and orchestration system according to claim 3, characterized in that: The abstract constraint modeling module (41) synchronously establishes a negative mapping rule between the temperature parameter adjustment amount and the upper limit threshold of the output length, which is achieved by training the temperature and length prediction model through historical task samples. Specifically, it includes: We collect abstract samples and their corresponding lengths generated by the technical abstract generation node under different temperature parameters, construct a regression dataset of temperature parameter values ​​and output length, train a neural network model based on the regression dataset, fit the negative inverse proportional function relationship that the output length increases nonlinearly with increasing temperature, and encode the neural network model as a calibration function that maps the temperature parameter adjustment amount to the upper limit threshold of the output length.

5. The large language model workflow visualization and orchestration system according to claim 4, characterized in that: When the temperature threshold binding module (42) calculates the upper limit threshold of the output length corresponding to the current temperature value according to the negative mapping rule, it calls the prediction model to use the temperature parameter input value adjusted by the user in real time as the independent variable of the prediction model, calculates the predicted output length value through forward propagation, and multiplies the predicted output length value by the preset safety redundancy coefficient to generate the final upper limit threshold of the output length. The safety redundancy coefficient is used to compensate for the boundary overflow risk caused by the prediction error of the prediction model.

6. The large language model workflow visualization and orchestration system according to claim 5, characterized in that: When the temperature threshold binding module (42) feeds back the upper limit of the output length as a visual upper limit identifier to the configuration interface of the node configuration unit (2) in real time, the visual upper limit identifier is rendered in the parameter panel of the technical summary generation node of the node configuration unit (2), covering the over-limit area of ​​the temperature parameter slider track. At the same time, the upper limit of the output length corresponding to the current temperature value and the percentage of remaining available capacity are displayed in a floating manner. When the user adjusts the temperature parameter to make the predicted output length reach the window capacity limit, the threshold warning animation is activated to enhance the visual prompt.

7. The large language model workflow visualization and orchestration system according to claim 1, characterized in that: The summary interception and execution module (43) is used to intercept the original summary text stream during the node execution completion callback stage when monitoring the output content of the technical summary generation node during the workflow execution stage. It calls the tokenization engine to count the number of tokens in the summary text in real time and compares the number of tokens with the current output length upper limit threshold generated by the temperature threshold binding module (42). If the output length upper limit threshold is exceeded, the semantic boundary truncation process is triggered. Otherwise, the complete summary text is passed to the multilingual translation node.

8. The large language model workflow visualization and orchestration system according to claim 7, characterized in that: When the sentence boundary detection algorithm locates the end of the first complete sentence from the end of the text backwards, it performs triple semantic boundary verification: Scanning from the end of the text forward, paragraph separators and line breaks are identified first-level boundaries. If there are no paragraph separators, combinations of periods, question marks, exclamation marks, and spaces are detected as second-level boundaries. If the second-level boundaries fail, virtual sentence separators are inserted as third-level boundaries using a pre-trained punctuation completion model to ensure semantic integrity before and after the location points.

9. A large language model workflow visualization and orchestration system according to claim 8, characterized in that: After triggering the semantic boundary truncation process, the summary truncation execution module (43) performs statement truncation based on the sentence end position: Truncate all continuous text forward from the end of the sentence until the total number of tokens drops to the upper limit of the output length threshold. Perform tail semantic coherence verification on the truncated summary text. If technical parameter keywords are detected to be cut off, backtrack to the previous sentence boundary, re-truncate, and discard the text segment that failed the verification.

10. A method for implementing a large language model workflow visualization and orchestration system as described in any one of claims 1-9, characterized in that: Includes the following steps: S1. Identify the connection between the technical summary generation node and the multilingual translation node, extract the total capacity of the context window of the multilingual translation node and deduct the capacity occupied by the translation control instructions, calculate the upper limit threshold of the output length of the technical summary generation node, and simultaneously establish a negative dynamic mapping rule between the temperature parameter adjustment amount and the upper limit threshold of the output length. S2. When the user adjusts the temperature parameters of the technical summary generation node, the current output length upper limit threshold is calculated in real time according to the negative dynamic mapping rule, and a visual upper limit indicator and warning prompt are rendered in the configuration interface. S3. The temperature parameter is mapped to the upper limit threshold of the output length through the prediction model, and the final upper limit threshold of the output length is generated by combining the safety redundancy coefficient. The linkage between the temperature control and the upper limit of the length is bound in real time. S4. During workflow execution, intercept the original summary text stream, count the number of tokens and compare them with the current output length upper limit threshold. If the limit is exceeded, locate the end position of the sentence based on triple semantic boundary verification, cut off the excess text, and then perform tail technical parameter coherence verification to ensure that the text input to the translation node meets the length constraint.

Images