Search

CN-121615664-B - Visual arrangement system and method for workflow of large language model

CN121615664BCN 121615664 BCN121615664 BCN 121615664BCN-121615664-B

Abstract

The invention relates to the technical field of node linkage constraint and text length adaptation, in particular to a large language model workflow visualization arrangement system and method, wherein a abstract constraint modeling module identifies a serial connection relation between a technical abstract generation node and a multi-language translation node, extracts the total capacity of a multi-language translation node context window from the serial connection relation, deducts the occupied capacity of a control instruction, calculates an abstract output length upper limit threshold value to establish a negative mapping rule of temperature parameters and length, and adjusts and calculates the linkage parameters and the threshold value of a temperature threshold value binding module, and feeds back the negative mapping rule to a configuration interface through a visualization mark and an early warning prompt; the abstract interception and control execution module intercepts overrun text when the workflow is executed, positions complete sentences through triple semantic boundary verification, and combines technical parameter consistency verification to cut redundant contents, so that the capacity of abstract adaptation translation nodes is ensured, smooth execution of the workflow is ensured, and the arrangement accuracy and reliability are improved.

Inventors

  • WANG XINZHENG
  • CHEN XINLEI
  • JIA XIAOJIE
  • CAI YUAN
  • ZHAO YUBING

Assignees

  • 企商在线(北京)数据技术股份有限公司
  • 光子云(北京)网络技术有限公司

Dates

Publication Date
20260512
Application Date
20260130

Claims (10)

  1. 1. The large language model workflow visualization arrangement system comprises a flow arrangement unit (1), a node configuration unit (2) and a workflow execution engine unit (3), and is characterized in that a calibration unit (4) is additionally arranged, and the calibration unit (4) comprises: The abstract constraint modeling module (41) is used for identifying a series connection relation between a technical abstract generating node and a multi-language translating node in a process arranging stage, extracting the total capacity of a context window of the multi-language translating node from the series connection relation, deducting the occupied capacity of a preset translating control instruction, calculating to obtain an output length upper limit threshold value of the technical abstract generating node, synchronously establishing a negative mapping rule of a temperature parameter regulating quantity and the output length upper limit threshold value, and defining a calibration logic that the output length upper limit threshold value is reduced according to an inverse proportion relation when the temperature parameter is increased; The temperature threshold binding module (42) is used for associating temperature parameters with output length upper limit thresholds in the node configuration unit (2), calculating the output length upper limit thresholds corresponding to the current temperature values according to the negative mapping rules in real time when adjusting the temperature parameters of the technical abstract generation nodes, and feeding back the output length upper limit thresholds to a configuration interface of the node configuration unit (2) in real time by using a visual upper limit mark; The abstract interception and control execution module (43) is used for monitoring the output content of the technical abstract generation node in the workflow execution stage, and when the actual output length exceeds the output length upper limit threshold corresponding to the current temperature value, locating the end position of the first complete sentence from the tail of the text forward based on the sentence boundary detection algorithm, cutting off the continuous sentences exceeding the output length upper limit threshold, and ensuring that the abstract text length of the input multi-language translation node accords with the output length upper limit threshold constraint.
  2. 2. The large language model workflow visualization orchestration system of claim 1, wherein the abstract constraint modeling module (41) is configured to identify a tandem relationship between a technical abstract generating node and a multi-language translation node during a process orchestration stage, specifically comprising: The flow arranging unit (1) is used for generating a node connection graph, locating a direct data channel between an output port of a technical abstract generating node and an input port of a multi-language translation node by traversing the node connection graph, determining that the technical abstract generating node and the multi-language translation node form a series relation when detecting that output data of the technical abstract generating node is configured as a unique input source of the multi-language translation node from the direct data channel, and recording the series relation as a directed edge in a node relation graph.
  3. 3. The workflow visualization orchestration system according to claim 2, wherein the abstract constraint modeling module (41) extracts a total capacity of a context window of the multi-language translation node by accessing a metadata interface of a target large language model bound by the multi-language translation node, deducts a preset translation control instruction occupation capacity, calculates an output length upper limit threshold of the technical abstract generating node, and loads a preset translation instruction template library and calculates a total token occupation amount of system instructions, format controllers and meta instructions required by the multi-language translation node to execute a translation task when deducting the preset translation control instruction occupation capacity, and subtracts the total token occupation amount from the total capacity of the context window to generate the output length upper limit threshold of the technical abstract generating node.
  4. 4. The large language model workflow visualization orchestration system according to claim 3, wherein the abstract constraint modeling module (41) establishes negative mapping rules of temperature parameter adjustment and output length upper limit threshold synchronously, and training temperature and length prediction models through historical task samples, specifically comprising: And collecting summary samples and corresponding lengths of the summary samples generated by the technical summary generating node under different temperature parameters, constructing a regression data set of temperature parameter values and output lengths, training a neural network model based on the regression data set, fitting a negative inverse proportion function relation of nonlinear increase of the output lengths caused by temperature rise, and encoding the neural network model into a calibration function of mapping temperature parameter adjustment amounts to an upper limit threshold of the output lengths.
  5. 5. The workflow visualization orchestration system for a large language model according to claim 4, wherein when the temperature threshold binding module (42) calculates an output length upper limit threshold corresponding to a current temperature value according to the negative mapping rule, the prediction model is called to take an input value of the temperature parameter adjusted in real time by a user as a prediction model argument, a prediction output length value is obtained through forward propagation calculation, the prediction output length value is multiplied by a preset safety redundancy coefficient to generate a final output length upper limit threshold, and the safety redundancy coefficient is used for compensating boundary overflow risks caused by prediction errors of the prediction model.
  6. 6. The large language model workflow visualization orchestration system according to claim 5, wherein when the temperature threshold binding module (42) feeds back the output length upper limit threshold value to the configuration interface of the node configuration unit (2) in real time with the visualization upper limit identification, the visualization upper limit identification is rendered in the parameter panel of the technical abstract generation node of the node configuration unit (2), the overrun area of the temperature parameter slider track is covered, the output length upper limit threshold value corresponding to the current temperature value and the remaining available capacity percentage are displayed in a floating mode, and when the user adjusts the temperature parameter to enable the predicted output length to reach the window capacity limit, the threshold warning animation strengthening visual prompt is activated.
  7. 7. The large language model workflow visualization orchestration system according to claim 1, wherein the abstract interception execution module (43) is configured to intercept an original abstract text stream at a node execution completion callback stage when monitoring output content of a technical abstract generation node at a workflow execution stage, invoke a tokenization engine to count token number of abstract text in real time, compare the token number with a current output length upper limit threshold generated by a temperature threshold binding module (42), and trigger a semantic boundary interception flow if the output length upper limit threshold is exceeded, otherwise, transfer the complete abstract text to a multi-language translation node.
  8. 8. The large language model workflow visualization orchestration system of claim 7, wherein the triple semantic boundary verification is performed when the sentence-based boundary detection algorithm locates the end position of the first complete sentence from the end of the text forward: scanning forwards from the tail of the text, preferentially identifying paragraph separators and line-wrapping symbols as primary boundaries, detecting combinations of periods, question marks, exclamation marks and follow-up space characters as secondary boundaries if no paragraph separators exist, and inserting virtual sentence separators as tertiary boundaries through a pre-trained punctuation deletion complement model if the secondary boundaries fail, so that the front-rear semantic integrity of positioning points is ensured.
  9. 9. The large language model workflow visualization orchestration system of claim 8, wherein after triggering a semantic boundary truncation process, the abstract truncation execution module (43) performs sentence excision based on sentence ending location: And cutting all continuous texts forwards from the sentence end position until the total number of tokens is reduced to be within the range of the upper limit threshold of the output length, executing tail semantic consistency verification on the cut abstract text, and if the technical parameter keywords are detected to be cut off, returning to the boundary of the previous sentence to cut off again and discarding the text segment with verification failure.
  10. 10. A method for implementing a large language model workflow visualization orchestration system according to any one of claims 1-9, comprising the steps of: S1, identifying a serial connection relation between a technical abstract generating node and a multi-language translating node, extracting the total capacity of a context window of the multi-language translating node, deducting the occupied capacity of a translating control instruction, calculating an output length upper limit threshold value of the technical abstract generating node, and synchronously establishing a negative dynamic mapping rule of a temperature parameter regulating quantity and the output length upper limit threshold value; S2, when a user adjusts the temperature parameter of the technical abstract generating node, calculating the current output length upper limit threshold value in real time according to a negative dynamic mapping rule, and rendering a visual upper limit mark and an early warning prompt on a configuration interface; s3, mapping the temperature parameter into an output length upper limit threshold value through a prediction model, generating a final output length upper limit threshold value by combining a safety redundancy coefficient, and binding the linkage relation between the temperature control and the length upper limit in real time; s4, intercepting the original abstract text flow when the workflow is executed, counting the number of tokens and comparing the number with the current output length upper limit threshold value, if the number exceeds the threshold value, positioning the end position of the sentence based on triple semantic boundary verification, and executing tail technical parameter consistency verification after cutting off the text of the exceeding part so as to ensure that the text of the input translation node accords with the length constraint.

Description

Visual arrangement system and method for workflow of large language model Technical Field The invention relates to the technical field of node linkage constraint and text length adaptation, in particular to a large language model workflow visualization arrangement system and method. Background Node linkage constraint and text length adaptation are an important technology, and are particularly applied to workflow scenes formed by serially connecting technical abstract generation and multi-language translation, and the core is that the processing capacity of the multi-language translation node is adapted through dynamic constraint of the technical abstract output length, so that smooth execution of the workflow is ensured, in a large language model workflow, the technical abstract generation node and the multi-language translation node always form a direct serial connection relationship, the capacity of a context window of the multi-language translation node is fixedly limited, partial capacity is occupied for loading control instructions when a translation task is executed, the temperature parameter of the technical abstract generation node is increased to cause nonlinear increase of the output text length, and the generated technical abstract is possibly beyond the upper processing limit of the translation node due to the fact that the correlation of the temperature parameter and the abstract length is not accurately adapted to the effective capacity of the translation node, so that the translation content is intercepted, semantic distortion and even workflow task failure are caused. Disclosure of Invention The invention aims to provide a large language model workflow visualization arrangement system and method for solving the problems in the background technology. In order to achieve the above object, a large language model workflow visualization layout system is provided, which includes a flow layout unit, a node configuration unit and a workflow execution engine unit, and is characterized in that a calibration unit is added, and the calibration unit includes: The abstract constraint modeling module is used for identifying a series connection relation between a technical abstract generating node and a multi-language translating node in a process arrangement stage, extracting the total capacity of a context window of the multi-language translating node from the series connection relation, deducting the occupied capacity of a preset translating control instruction, calculating to obtain an output length upper limit threshold value of the technical abstract generating node, synchronously establishing a negative mapping rule of a temperature parameter adjustment quantity and the output length upper limit threshold value, and defining a calibration logic that the output length upper limit threshold value is reduced according to an inverse proportion relation when the temperature parameter is increased; The temperature threshold binding module is used for associating temperature parameters with output length upper limit thresholds in the node configuration unit, calculating the output length upper limit thresholds corresponding to the current temperature values according to the negative mapping rules in real time when adjusting the temperature parameters of the technical abstract generation nodes, and feeding back the output length upper limit thresholds to a configuration interface of the node configuration unit in real time by using a visual upper limit mark; The abstract interception and control execution module is used for monitoring the output content of the technical abstract generation node in the workflow execution stage, when the actual output length exceeds the output length upper limit threshold corresponding to the current temperature value, locating the end position of the first complete sentence from the tail of the text forward based on the sentence boundary detection algorithm, cutting off the continuous sentences exceeding the output length upper limit threshold, and ensuring that the abstract text length of the input multi-language translation node accords with the output length upper limit threshold constraint. It is a second object of the present invention to provide a method for implementing a large language model workflow visualization orchestration system comprising any one of the above, comprising the steps of: S1, identifying a serial connection relation between a technical abstract generating node and a multi-language translating node, extracting the total capacity of a context window of the multi-language translating node, deducting the occupied capacity of a translating control instruction, calculating an output length upper limit threshold value of the technical abstract generating node, and synchronously establishing a negative dynamic mapping rule of a temperature parameter regulating quantity and the output length upper limit threshold value; S2, when a user adjusts the temperatu