A method and system for constructing large language model text based on prompt enhancement

By constructing a three-level prompt meta-database and dynamically adjusting the weight adjustment coefficients, the static adaptability and feedback signal lag problems of existing prompt enhancement technologies are solved, realizing refined semantic guidance for text generation of large language models and improving the quality and efficiency of text generation in professional fields.

CN122964286BActive Publication Date: 2026-11-17CHANGSHA ZHIXIN WORKSHOP TECHNOLOGY CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611455938.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-18
Publication Date
2026-11-17
Estimated Expiration
2046-09-18

AI Technical Summary

Technical Problem

Existing prompting enhancement technologies suffer from weak static adaptability, delayed feedback signals, and insufficient hierarchical control capabilities. They are unable to provide refined semantic guidance for the text generation process of large language models while ensuring generation efficiency, and thus cannot meet the high-quality text construction needs of professional fields.

Method used

A three-level prompt meta-database was constructed, and semantic bias quantification was performed by combining the latent state representation vectors within the large language model. By dynamically adjusting the weight adjustment coefficients of the prompt data, closed-loop iterative optimization was achieved. Task anchoring, logical constraints, and expression style were decoupled in a hierarchical manner to generate text that meets the task requirements.

Benefits of technology

It significantly improves the domain adaptability, logical compliance, and style consistency of professional text generation, and achieves real-time semantic deviation assessment and refined semantic guidance with no lag and low noise, meeting the needs of high-quality automated text construction in vertical fields.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122964286B_ABST
    Figure CN122964286B_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for constructing text using a large language model based on cue enhancement. The method involves acquiring an initial task instruction, a three-level cue metadata database, and a prior semantic feature dataset for the target domain. Using the initial task instruction as the retrieval source, each layer of the three-level cue metadata database is searched to obtain initial enhanced cue data. This data is then input into the large language model for inference, and the hidden states of the intermediate layers are collected. These hidden states are matched with the prior semantic features to calculate the semantic deviation quantification result. If the deviation meets a threshold or reaches the maximum iteration count, the current enhanced cue data drives the model to generate the target constructed text. Otherwise, the weight adjustment coefficients of each layer are adjusted according to the deviation, the concatenation weights of the cue units are dynamically corrected, and the iteratively generated enhanced cue data is recombined and iterated again. This invention improves the domain adaptability and generation quality of the constructed text through hierarchical retrieval and dynamic weight correction, adaptively optimizing the cue data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a method and system for constructing large language model text based on prompt enhancement. Background Technology

[0002] With the rapid iteration and industrialization of large language model technology, automated text generation technology has become a core technological support for multiple vertical fields such as government document processing, legal compliance analysis, financial research report generation, and technical document production. Traditional template-based text generation methods rely on fixed frameworks and manual parameter filling. While they can achieve basic content output, the domain adaptability, logical compliance, and stylistic consistency of the generated text highly depend on the quality of input prompts. Against this backdrop, prompt enhancement technology, as a core means to optimize the generation effect of large models, has become a research and application hotspot in the field of natural language processing. Its technical performance directly determines the implementation effect and applicable boundaries of large language models in professional text generation scenarios.

[0003] Current mainstream prompt enhancement technologies are mainly divided into two categories: template-based prompt construction and retrieval-based prompt enhancement. Template-based solutions generate prompts by pre-setting a fixed prompt framework and filling it in with task parameters. While simple to implement, they lack flexibility and are difficult to adapt to diverse text construction needs. Retrieval-based solutions match candidate prompts from a prompt corpus based on semantic similarity and then concatenate them. They can adapt to different task scenarios, but generally use a static, single-pass concatenation processing logic, completing the prompt combination only based on the literal semantics of the task instructions, without dynamically optimizing based on the actual semantic response of the large model. Some solutions with feedback optimization capabilities often use the final generated text content as a feedback signal to adjust the prompts in reverse. This not only has feedback lag but is also susceptible to surface noise interference from the generated text and cannot accurately capture semantic shifts within the model. At the same time, existing prompts mostly adopt a flat structure, with prompt content from different dimensions such as task objectives, logical constraints, and expression styles being coupled together. It is difficult to accurately optimize for specific dimensions of semantic deviations, resulting in difficulty in guaranteeing the semantic accuracy of text construction in vertical domains.

[0004] In summary, existing cue enhancement techniques generally suffer from weak static adaptability, delayed feedback signals, and insufficient hierarchical control capabilities. They cannot simultaneously guarantee generation efficiency while providing refined semantic guidance for the text generation process of large models, thus failing to meet the high quality requirements of professional fields. Therefore, developing a cue enhancement text construction method capable of dynamic closed-loop optimization and precise hierarchical control based on the model's internal semantic representation is a key technical problem that needs to be solved in this field. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] (II) Technical Solution To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: S1. Obtain the initial task instructions, the pre-built three-level prompt meta-database, and the target domain prior semantic feature dataset corresponding to the text construction task; S2. Using the initial task instruction as the retrieval source, perform semantic retrieval with each layer of the three-level prompt meta-database to obtain initial enhanced prompt data; S3. Input the initial enhanced prompt data as the enhanced prompt data for the current round into the large language model to perform inference, and simultaneously collect the hidden state representation vector of the specified intermediate layer of the model; perform cosine similarity matching and difference operation between the hidden state representation vector and the prior semantic feature dataset of the target domain, and output the semantic deviation quantification result. If the semantic deviation quantization result meets the preset threshold condition or the current iteration round reaches the preset maximum number of iterations, then the enhanced prompt data of the current round is used as the final optimized enhanced prompt data, and step S5 is executed directly; S4. If the semantic deviation quantization result does not meet the preset threshold condition, the weight adjustment coefficients corresponding to each layer of the three-level prompt meta-database are adjusted according to the deviation quantization results of each layer. The splicing weights of each layer prompt unit are dynamically corrected based on the weight adjustment coefficients. The iteratively enhanced prompt data is recombined and generated. The iteratively enhanced prompt data is then returned to S3 to enter the next iteration. S5. The large language model is driven by the final optimized enhanced prompt data to complete inference, generate and output the target constructed text that meets the task requirements; This invention constructs a complete closed-loop system for constructing enhanced text for large language models, encompassing three-level semantic retrieval of prompt metadata, quantification of semantic bias in the hidden state of a large model, dynamic iterative correction of hierarchical weights, and enhanced prompt-driven text generation. It abandons the traditional static, single-time prompt concatenation and the prompt enhancement mode that relies on reverse optimization of the generated text's surface content. Addressing the shortcomings of existing technologies such as weak static adaptability, delayed feedback signals, susceptibility to surface noise interference, and insufficient precise hierarchical control, this invention directly captures the semantic offset state within the model by collecting the hidden state representation vector of the specified intermediate layer as a semantic alignment feedback signal. This achieves lag-free, low-noise real-time semantic bias assessment. This system utilizes a three-tiered prompt meta-database architecture that decouples semantics based on task anchoring, logical constraints, and expression style. It maps the quantified semantic deviations of each layer to corresponding weight adjustment coefficients, dynamically correcting the concatenation weights of prompt units at each layer and driving iterative optimization of the prompt data in a closed loop. This overcomes the limitations of flat prompt structures with semantic coupling and the inability to precisely optimize for deviation dimensions. While ensuring efficient text generation, it significantly improves the domain adaptability, logical compliance, and stylistic consistency of generated text in professional fields. It achieves quantifiable control and refined semantic guidance in the prompt enhancement process, fully meeting the application needs of high-quality automated text construction in vertical domains.

[0007] Preferably, step S1 includes the following steps: S11. Receive the input description of the text construction task, perform semantic parsing on the task objective, constraints, and output requirements, and extract the structured initial task instructions. S12. Obtain a pre-built three-level prompt meta-database. The three-level prompt meta-database is stored in layers according to the task anchoring layer, the logical constraint layer, and the expression style layer. Each layer contains several independent and callable prompt units, and there is no semantic overlap between the prompt units in each layer. S13. Obtain the target domain prior semantic feature dataset, wherein the target domain is the domain of legal documents, financial reports, or medical texts; the target domain prior semantic feature dataset is a set of benchmark semantic vectors extracted from the target domain standard corpus by a pre-trained language model and organized and stored according to the task intent dimension, logical compliance dimension, and expression style dimension, and is used as a reference benchmark for semantic alignment. By performing semantic parsing on the text construction task input to extract structured initial task instructions, a corresponding prompt meta-database with independent and non-overlapping task anchoring layer, logical constraint layer, and expression style layer was constructed, along with a target domain prior semantic feature benchmark dataset organized and stored in multiple dimensions. A complete basic data system for hierarchical prompt optimization was established, providing a standardized input source and reference benchmark for subsequent prompt retrieval, recall, and semantic alignment verification.

[0008] Preferably, step S12 includes the following steps: S121. Collect high-quality text samples in the target domain and extract three types of semantic fragments: task target class expression, logical rule class constraint, and style class feature. S122. Standardize and rewrite the three types of semantic fragments and deduplication them respectively to generate the initial prompt unit sets of the task anchoring layer, logical constraint layer and expression style layer. S123. Extract semantic vectors for all prompt units in each level and establish a retrieval index to form a three-level prompt meta-database that is stored independently in layers. By collecting high-quality text samples from the target domain and decomposing them into three types of semantic segments, and after standardization, rewriting and deduplication, three initial sets of prompt units are generated. Then, semantic vectors are extracted for each prompt unit and a retrieval index is established. This achieves hierarchical structured storage and semantic retrieval capabilities for prompt metadata, ensuring that the semantics of each level of prompts are independent and without redundancy, laying a data foundation for subsequent hierarchical and accurate recall of prompt units.

[0009] Preferably, step S2 includes the following steps: S21. Extract the semantic vector corresponding to the initial task instruction, and use the semantic vector as the retrieval vector to perform semantic cosine similarity retrieval on the task anchoring layer, logical constraint layer and expression style layer of the three-level prompt meta-database respectively. S22. For each level, sort the prompt units from high to low according to semantic similarity, and select a preset number of prompt units as the candidate prompt unit set for that level. S23. Following the fixed hierarchical order of task anchoring layer, logical constraint layer, and expression style layer, the candidate prompt units are sequentially concatenated to generate initial enhanced prompt data; By extracting the semantic vector of the initial task instruction, cosine similarity retrieval is performed on the three levels of the three-level prompt metadata database. A preset number of candidate prompt units are selected by similarity sorting, and then concatenated in a fixed hierarchical order to generate the initial enhanced prompt data. This achieves accurate prompt recall under the hierarchical dimension, ensuring that the initial enhanced prompts can simultaneously cover the three core requirements of task objectives, logical constraints, and expression style.

[0010] Preferably, step S3 includes the following steps: S31. Input the initial enhanced prompt data as the enhanced prompt data for the current round into the large language model, trigger the model to perform forward inference, and collect the hidden state representation vector output by the specified intermediate layer of the model during the inference process. S32. Perform batch cosine similarity matching between the hidden state representation vector and all benchmark semantic vectors in the target domain prior semantic feature dataset to calculate the global average semantic similarity value. S33. Using the difference between 1 and the global average semantic similarity value as the semantic deviation quantization result, output the semantic deviation quantization result of the current round; S34. If the semantic deviation quantization result meets the preset threshold condition or the current iteration round reaches the preset maximum number of iterations, then the enhanced prompt data of the current round is used as the final optimized enhanced prompt data, and step S5 is executed directly. By enhancing the input of the large language model to trigger forward inference, the hidden state representation vectors of the specified intermediate layers of the model are collected and batch-matched with the domain prior benchmark semantic vectors to calculate the global average semantic similarity. The difference between 1 and the similarity value is used as the semantic deviation quantification result and a threshold determination is made. Based on the hidden state inside the model, the degree of semantic alignment can be quantitatively evaluated, providing an objective and measurable judgment basis for the prompt optimization iteration.

[0011] Preferably, step S32 includes the following steps: S321. Perform dimensional splitting on the hidden state representation vector to extract sub-feature vectors for the task intent dimension, logical compliance dimension, and expression style dimension respectively. S322. Calculate the cosine similarity between the sub-feature vector and the corresponding dimension of the baseline semantic vector in the prior semantic feature dataset one by one to obtain the independent similarity value of each dimension. S323. The independent similarity values ​​of each dimension are weighted and summed according to the dimension weights to obtain the global average semantic similarity value; By splitting the latent state representation vector into sub-feature vectors according to three dimensions—task intent, logical compliance, and expression style—and calculating the cosine similarity with the baseline semantic vector of the corresponding dimension, the global average semantic similarity value is obtained by weighted summation according to dimension weights. This achieves multi-dimensional fine-grained semantic matching evaluation, which can differentiate and take into account the alignment requirements of different dimensions, effectively improving the accuracy and rationality of semantic deviation quantification results.

[0012] Preferably, step S4 includes the following steps: S41. If the semantic deviation quantization result does not meet the preset threshold condition, the semantic deviation quantization result is input into the preset hierarchical weight mapping function, and the weight adjustment coefficients corresponding to the task anchoring layer, the logical constraint layer, and the expression style layer are output respectively. S42. Based on the respective weight adjustment coefficients, the initial splicing weights of each candidate prompt unit in the level are dynamically corrected to obtain the corrected weight values ​​of each prompt unit. S43. Weighted fusion of the prompting units in each level according to the corrected unit weight value, then reorganized and spliced ​​in hierarchical order to generate iterated enhanced prompting data, and the reorganized enhanced prompting data generated by the iteration is returned to S3 to enter the next iteration. When the semantic deviation does not meet the threshold, the weight adjustment coefficients corresponding to each level are output through the hierarchical weight mapping function to dynamically correct the initial splicing weights of the candidate prompt units in each level. After weighted fusion, the prompt data is reorganized in hierarchical order to generate iteratively enhanced prompt data and returned for iterative verification. This realizes the adaptive dynamic adjustment of prompt weights, drives the prompt data to perform closed-loop iterative optimization along the semantic deviation direction, and gradually converges to the preset domain alignment requirements.

[0013] Preferably, step S41 includes the following steps: S411, Preset the deviation sensitivity coefficients for the task anchoring layer, logical constraint layer, and expression style layer respectively; S412. Multiply the independent deviation value of each dimension by the deviation sensitivity coefficient of each layer to obtain the weight adjustment increment of each layer; S413. Add the initial baseline weight of each layer to the corresponding weight adjustment increment, and output the corresponding weight adjustment coefficient of each layer after L1 normalization. By pre-setting differentiated deviation sensitivity coefficients for the three-layer prompt hierarchy, the weight adjustment increment is obtained by multiplying the independent deviation value of each dimension with the sensitivity coefficient of each layer. After normalization processing combined with the initial baseline weight, the weight adjustment coefficient of each layer is output. This realizes a differentiated deviation response mechanism for different prompt hierarchy, which can adapt to the different importance of each hierarchy in text construction and improve the targeting and optimization efficiency of weight adjustment.

[0014] Preferably, step S5 includes the following steps: S51. Obtain the final optimized enhanced prompt data, convert it into a word sequence, load it into the input of the large language model, and start the complete generation inference; S52. After the reasoning is completed, the enhanced cue word sequence of the input is stripped from the output sequence, and the text fragments corresponding to the pure generated content are extracted to obtain the initial generated text. S53. Perform format regularization and constraint verification on the initial generated text, remove redundant content that does not meet the task requirements, and obtain the standardized target construction text and output it. By converting the final optimized enhanced prompts into a lexical sequence and inputting it into a large language model to complete the generation inference, the initial generated text is extracted by stripping the input prompt lexical units. After format regularization and constraint verification to remove redundant content, the standardized target constructed text is output. This ensures the format compliance and task constraint matching of the final output text, effectively removes invalid and redundant content, and improves the usability and standardization of the output results.

[0015] Furthermore, a large language model text construction system based on prompt enhancement is provided, the system comprising a data acquisition and initialization module, a semantic retrieval and prompt generation module, a semantic bias evaluation module, a weight dynamic correction and iteration module, and a text generation and output module; Data acquisition and initialization module: Acquires initial task instructions, a three-level prompt meta-database, and a target domain prior semantic feature dataset, and parses, loads, and organizes them by dimension; Semantic retrieval and prompt generation module: Using the semantic vector of the initial task instruction as the retrieval source, it performs semantic retrieval on the prompt meta-database at each level, filters candidate prompt units, and concatenates them to generate initial enhanced prompt data; Semantic bias assessment module: Input the enhanced prompt data into the large language model, collect the specified intermediate layer hidden states, match and calculate with the prior semantic features, obtain the semantic bias quantification result, and determine whether the threshold condition is met. Weight dynamic correction and iteration module: If the deviation does not meet the threshold, the deviation quantification result is mapped to the weight adjustment coefficient of each level, the dynamic correction prompt unit splices the weights and reorganizes them to generate the enhanced prompt data after iteration, and returns to the semantic deviation evaluation module to continue the iteration. Text generation and output module: Obtains the final optimized enhanced prompt data, drives the large language model to generate the target text, and outputs it after formatting and verification.

[0016] (III) Beneficial Effects The present invention has the following beneficial effects: This invention constructs a closed-loop dynamic iterative mechanism. Initial enhanced prompts are input into a large language model, and the hidden states of intermediate layers are collected. These are then matched with prior semantic features of the target domain to generate a semantic deviation quantification result. This quantification drives the correction of weight coefficients at each level and the regeneration of prompt data until the deviation converges to a threshold range. Compared to existing retrieval schemes in the background, which rely solely on static, single-pass concatenation based on the literal semantics of task instructions, the dynamic closed-loop architecture of this invention enables the prompt combination strategy to be continuously optimized based on the real-time semantic response within the model. This elevates prompt enhancement from open-loop static retrieval to closed-loop dynamic guidance, effectively solving the problem of insufficient domain semantic accuracy caused by weak static adaptability, and significantly improving the effectiveness and robustness of prompt guidance.

[0017] This invention synchronously acquires the hidden state representation vectors of designated intermediate layers as feedback signals during the reasoning process of a large language model, rather than relying on the final generated text content for post-processing correction. Compared to existing feedback schemes in the background art, which inherently suffer from feedback lag and surface noise interference due to using generated text as the signal source, the intermediate layer hidden state feedback mechanism of this invention achieves pre-emptive perception and immediate response to semantic biases, fundamentally improving the timeliness and accuracy of the feedback signal. While ensuring generation efficiency, it effectively enhances the tracking ability and control sensitivity of the internal semantic state of the prompt enhancement process.

[0018] This invention stores the prompt metadata database in layers according to three dimensions: task anchoring, logical constraints, and expression style, with no semantic overlap between layers. During iterative optimization, the weight adjustment coefficients are adjusted based on the deviation quantification results and the differential sensitivity coefficients of each layer, dynamically correcting and reorganizing the splicing weights of each layer's prompt units. Compared to the shortcomings of the flat structure in the background technology, which leads to the coupling of multi-dimensional prompt content and difficulty in precise optimization, the layered decoupled storage and differential control mechanism of this invention allows semantic deviations in each dimension to be independently identified, responded to, and precisely corrected. From both the prompt structure design and control strategy levels, it achieves refined and independent guidance on task intent, logical compliance, and expression style during text construction, significantly improving the adaptation accuracy and control flexibility in professional domain text construction scenarios.

[0019] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating a method for constructing a large language model text based on prompt enhancement according to the present invention. Figure 2 This is a system architecture diagram of the present invention; Figure 3 This is a flowchart of the enhanced data generation and closed-loop iterative correction process for the present invention. Figure 4 This is a graph showing the ablation experiment analysis of the number of units screened according to the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the invention, and not all embodiments. Based on the embodiments of the invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the invention.

[0023] See Figure 1 This invention discloses a method for constructing large language model text based on prompt enhancement, comprising the following steps: S1. Obtain the initial task instructions, the pre-built three-level prompt meta-database, and the target domain prior semantic feature dataset corresponding to the text construction task; Step S1 includes the following steps: S11. Receive the input description of the text construction task, perform semantic parsing on the task objective, constraints, and output requirements, and extract the structured initial task instructions. S12. Obtain a pre-built three-level prompt meta-database. The three-level prompt meta-database is stored in layers according to the task anchoring layer, the logical constraint layer, and the expression style layer. Each layer contains several independent and callable prompt units, and there is no semantic overlap between the prompt units in each layer. Step S12 includes the following steps: S121. Collect high-quality text samples in the target domain and extract three types of semantic fragments: task target class expression, logical rule class constraint, and style class feature. S122. Standardize and rewrite the three types of semantic fragments and deduplication them respectively to generate the initial prompt unit sets of the task anchoring layer, logical constraint layer and expression style layer. Specifically, in embodiment S122: when the target domain is legal documents, the task anchoring layer focuses on the construction of prompting units for identifying litigation claims and extracting factual elements, the logical constraint layer focuses on the construction of prompting units for legal citation norms, completeness of evidence chains and procedural legality, and the expression style layer focuses on the construction of prompting units for the rigor of legal language and the standardization of judgment document format. When the target domain is financial reporting, the task anchoring layer focuses on the construction of prompting units for financial indicator analysis and risk assessment dimensions, the logical constraint layer focuses on the construction of prompting units for data source traceability, disclosure compliance and consistency of calculation methods, and the expression style layer focuses on the construction of prompting units for professional terminology standardization and report structure standardization. When the target domain is medical text, the task anchoring layer focuses on the construction of prompting units for symptom description and diagnosis and treatment path identification, the logical constraint layer focuses on the construction of prompting units for medical terminology usage standards, diagnosis and treatment process compliance and privacy protection, and the expression style layer focuses on the construction of prompting units for medical record writing standards and the rigor of medical expression. S123. Extract semantic vectors for all prompt units in each level and establish a retrieval index to form a three-level prompt meta-database that is stored independently in layers. Specifically, in embodiment S123: the extraction of semantic vectors uses the same pre-trained language model as in S13, and the pooled semantic vectors are extracted after each prompt unit text is input into the model; S13. Obtain the target domain prior semantic feature dataset, wherein the target domain is the domain of legal documents, financial reports, or medical texts; the target domain prior semantic feature dataset is a set of benchmark semantic vectors extracted from the target domain standard corpus by a pre-trained language model and organized and stored according to the task intent dimension, logical compliance dimension, and expression style dimension, and is used as a reference benchmark for semantic alignment. Specifically, in embodiment S13: the pre-trained language model architecture and parameters are as follows: a Chinese BERT-base model is used as the feature extraction base. The model contains a 12-layer Transformer encoder, a hidden layer dimension of 768, 12 attention heads, and a total of approximately 110 million parameters. The model is first pre-trained based on general Chinese corpus, and then fine-tuned using 120,000 corpora of authoritative normative texts, industry standard documents, and high-quality public samples from the target domain. The fine-tuning task uses contrastive learning loss to make the semantic vectors output by the model conform to the semantic distribution of the domain. 2,000 representative core texts are selected from the standard corpus of the target domain and input into the fine-tuned model one by one. The output vector at the token position is taken as a single baseline semantic vector. All baseline vectors are clustered and deduplicated to form a priori semantic feature dataset containing 1,800 baseline semantic vectors, which serves as a unified reference benchmark for semantic alignment. The clustering deduplication process employs the K-Means++ algorithm to cluster the extracted semantic vectors; the optimal number of clusters K is determined using the elbow rule; after clustering, the vector closest to the cluster center in each cluster is selected as the representative vector of that cluster, and other redundant vectors within the cluster are removed, ultimately forming the deduplicated prior semantic feature dataset. Furthermore, a source splitting and independent encoding strategy is adopted for each corpus. Specifically, experts accurately extract three types of text fragments from each standard corpus that do not overlap in terms of semantic function: task intent, logical rules, and expression style. These fragments are then independently input into the fine-tuned BERT-base model, generating independent 768-dimensional baseline vectors, which are then stored according to their dimensions. If a certain dimension is naturally missing, it is filled in using the average vector of that domain. S2. Using the initial task instruction as the retrieval source, perform semantic retrieval with each layer of the three-level prompt meta-database to obtain initial enhanced prompt data; Step S2 includes the following steps: S21. Extract the semantic vector corresponding to the initial task instruction, and use the semantic vector as the retrieval vector to perform semantic cosine similarity retrieval on the task anchoring layer, logical constraint layer and expression style layer of the three-level prompt meta-database respectively. Specifically, in embodiment S21: the extraction of semantic vectors uses the same pre-trained language model as in S13; S22. For each level, sort the prompt units from high to low according to semantic similarity, and select a preset number of prompt units as the candidate prompt unit set for that level. Specifically, in embodiment S22: Refer to Figure 4 The preset number of each layer of filtering is 3, which was determined through historical ablation experiments. The specific processing method is as follows: 100 sets of text construction tasks covering 10 different fields were selected to form a test set. Under the same experimental conditions, the final semantic deviation convergence speed and generated text quality were tested when 1, 2, 3, 4, 5, and 6 prompt units were filtered at each layer. According to statistics, when the number of filtering units at each layer is 3, the semantic deviation reduction per unit time reaches its peak. Further increasing the number of units improves the effect by less than 5%, but increases the inference time by more than 30%. To balance the optimization effect and inference efficiency, the default number of prompt units to be filtered at each layer is determined to be 3. The cumulative deviation reduction per unit time = (cumulative quality improvement value) / (cumulative time value) reaches a peak of 10.6 at n=3. S23. Following the fixed hierarchical order of task anchoring layer, logical constraint layer, and expression style layer, the candidate prompt units are sequentially concatenated to generate initial enhanced prompt data; S3. Input the initial enhanced prompt data as the enhanced prompt data for the current round into the large language model to perform inference, and simultaneously collect the hidden state representation vector of the specified intermediate layer of the model; perform cosine similarity matching and difference operation between the hidden state representation vector and the prior semantic feature dataset of the target domain, and output the semantic deviation quantification result. If the semantic deviation quantization result meets the preset threshold condition or the current iteration round reaches the preset maximum number of iterations, then the enhanced prompt data of the current round is used as the final optimized enhanced prompt data, and step S5 is executed directly; See Figure 3 Step S3 includes the following steps: S31. Input the initial enhanced prompt data as the enhanced prompt data for the current round into the large language model, trigger the model to perform forward inference, and collect the hidden state representation vector output by the specified intermediate layer of the model during the inference process. Specifically, in embodiment S31: the large language model adopts DeepSeek-Pro (65B); the designated intermediate layer is the 16th layer of the large language model decoder, and the collected hidden state representation vector is the output vector corresponding to the last token of this layer; the 16th layer was determined through historical comparison experiments, and the hidden states of the 8th, 16th and 24th layers of the model were collected for semantic bias calculation: the shallow hidden state is biased towards literal features and is not sensitive enough to semantic shift; the deep hidden state is biased towards the generation surface features and is easily affected by output noise; the 16th layer is the intermediate layer of the model, and has the best balance between semantic representation integrity and bias sensitivity. It can accurately capture the core semantic shift and is not affected by the generation end surface noise, so it is determined as the default collection layer; The acquisition method is to attach to the output of the 16th layer decoder through a hook mechanism during the forward inference process of the large language model. When the inference is executed to this layer, the output feature vector corresponding to the last token is directly captured. After L2 normalization, it is used as the hidden state representation vector of the current round. S32. Perform batch cosine similarity matching between the hidden state representation vector and all benchmark semantic vectors in the target domain prior semantic feature dataset to calculate the global average semantic similarity value. Step S321 includes the following steps: S321. Perform dimensional splitting on the hidden state representation vector to extract sub-feature vectors for the task intent dimension, logical compliance dimension, and expression style dimension respectively. Specifically, in embodiment S321: the method for dimensional splitting the hidden state representation vector is as follows: during the BERT-base domain fine-tuning stage, three independent linear projection matrices are trained in parallel, each corresponding to a semantic dimension; during inference, the collected 768-dimensional hidden state vector is input into these three projection layers respectively, and three independent 256-dimensional sub-feature vectors are obtained by forward calculation. Furthermore, the training employs a supervised contrastive learning strategy with dimension labels, performed concurrently with the domain fine-tuning phase in S13. Specifically, the three types of text fragments decomposed in S12 are used as independent training samples, each with its own dimension label. During training, a fragment is input into BERT-base to obtain a 768-dimensional vector, which is then fed into three projection layers. However, only the contrast loss between the output vector of the corresponding dimension projection layer and the vector center of the positive sample in the same batch is calculated. Simultaneously, the mutual exclusion loss between this output and the output vectors of the other two projection layers is calculated. The gradients of the three projection layers are independently backpropagated and synchronously updated, ensuring that each projection layer focuses only on learning the semantic compression mapping of its own dimension, ultimately forming three mutually independent dimensionality reduction projection matrices. Positive sample construction rule: Within the same training batch, the corresponding dimension projection vectors of all samples belonging to the same semantic dimension as the current sample constitute the positive sample set of the current sample; Negative sample construction rule: Divided into two categories, cross-dimensional negative samples: the output vectors of the current input sample in the other two non-corresponding dimension projection layers; cross-sample negative samples: the corresponding dimension projection vectors of samples from other semantic dimensions within the same batch; The mathematical expression of the mutual exclusion loss function is: ; in, The total number of samples in the batch; Let be the true semantic dimension of the i-th sample; Let be the output vector of the i-th sample projected onto the corresponding dimension of the layer. The projection output vector of the i-th sample onto the non-corresponding dimension d; () represents the cosine similarity calculation; The mutual exclusion margin threshold, with a default value of 0.1, is used to force semantic separation in different dimensional projection spaces; the total training loss is a weighted sum of the supervised contrastive loss and the mutual exclusion loss. ,in This is the mutual exclusion loss weight coefficient, with a default value of 0.3, used to balance the optimization intensity of dimension alignment and dimension mutual exclusion; Use the general formula for supervised comparison loss; S322. Calculate the cosine similarity between the sub-feature vector and the corresponding dimension of the baseline semantic vector in the prior semantic feature dataset one by one to obtain the independent similarity value of each dimension. Specifically, in embodiment S322: all the baseline semantic vectors in the prior dataset are input into the corresponding linear projection matrix and compressed to 256 dimensions; the 256-dimensional single-dimensional sub-feature vectors obtained by splitting are compared with the compressed baseline semantic vectors to obtain a set of similarity scores for that dimension; the arithmetic mean of the set of scores is then taken to obtain the single independent similarity value corresponding to that dimension. S323. The independent similarity values ​​of each dimension are weighted and summed according to the dimension weights to obtain the global average semantic similarity value; Specifically, in embodiment S323: global average semantic similarity value Calculate using the following formula: ; in, , , These are independent similarity values ​​across three dimensions: task intent, logical compliance, and expression style. , , The corresponding dimension weights are determined using the analytic hierarchy process (AHP) combined with industry expert scoring. Five experts in the field of text generation were invited to conduct pairwise comparisons and scores on the importance of the three dimensions. After constructing a judgment matrix and passing the consistency test, the aforementioned dimension weights were calculated. S33. Using the difference between 1 and the global average semantic similarity value as the semantic deviation quantization result, output the semantic deviation quantization result of the current round; S34. If the semantic deviation quantization result meets the preset threshold condition or the current iteration round reaches the preset maximum number of iterations, then the enhanced prompt data of the current round is used as the final optimized enhanced prompt data, and step S5 is executed directly. Specifically, in embodiment S34: the preset threshold condition is obtained by fitting historical annotation data, annotating 2000 groups of generated text with manual scores of domain conformity, and calculating the semantic deviation value of each group of text accordingly; after fitting, when the semantic deviation is ≤0.15, the domain conformity of the generated text can reach more than 92%, which meets the quality requirements of most text construction tasks, so the threshold is set to 0.15; Furthermore, the default value for the preset maximum number of iterations is 5. This value was determined through ablation experiments: 80 extreme text construction tasks with large domain biases were selected for testing, and the distribution of iteration convergence rounds was statistically analyzed. Among them, 98.7% of the tasks could converge to below the threshold of 0.15 within 5 rounds; after more than 5 rounds, the decrease in deviation per round was generally less than 0.01, the benefit approached zero, but the inference time increased linearly. Therefore, 5 rounds are set as the default maximum number of iterations, which can be customized within the range of 3-10 according to the task accuracy requirements. S4. If the semantic deviation quantization result does not meet the preset threshold condition, the weight adjustment coefficients corresponding to each layer of the three-level prompt meta-database are adjusted according to the deviation quantization results of each layer. The splicing weights of each layer prompt unit are dynamically corrected based on the weight adjustment coefficients. The iteratively enhanced prompt data is recombined and generated. The iteratively enhanced prompt data is then returned to S3 to enter the next iteration. Step S4 includes the following steps: S41. If the semantic deviation quantization result does not meet the preset threshold condition, the semantic deviation quantization result is input into the preset hierarchical weight mapping function, and the weight adjustment coefficients corresponding to the task anchoring layer, the logical constraint layer, and the expression style layer are output respectively. Step S41 includes the following steps: S411, Preset the deviation sensitivity coefficients for the task anchoring layer, logical constraint layer, and expression style layer respectively; Specifically, in embodiment S411: the deviation sensitivity coefficients corresponding to the three levels are: task anchoring layer 1.2, logic constraint layer 1.0, and expression style layer 0.8. These coefficients were determined through historical parameter tuning experiments and contribution statistics: the contribution of each level's weight adjustment to the reduction of semantic deviation was statistically analyzed during 1000 iterations. The task anchoring layer had the greatest impact on the overall semantic deviation, contributing approximately 50%, so it was set with the highest sensitivity coefficient; the expression style layer had the least impact on semantic deviation, contributing approximately 20%, so it was set with the lowest sensitivity coefficient; multiple coefficient combinations were tested under fixed initial conditions, and the above values ​​corresponded to the fewest average convergence rounds, so they were set as the default values. S412. Multiply the independent deviation value of each dimension by the deviation sensitivity coefficient of each layer to obtain the weight adjustment increment of each layer; Specifically, in embodiment S412: the independent deviation value of each layer dimension is 1 - the independent similarity value of each layer; S413. Add the initial baseline weight of each layer to the corresponding weight adjustment increment, and output the corresponding weight adjustment coefficient of each layer after L1 normalization. S42. Based on the respective weight adjustment coefficients, the initial splicing weights of each candidate prompt unit in the level are dynamically corrected to obtain the corrected weight values ​​of each prompt unit. Specifically, in embodiment S42: the dimensional deviation value corresponding to each level is calculated; wherein each dimensional deviation value is converted from the independent similarity value of each dimension, specifically, dimensional deviation value = 1 - independent similarity value of that dimension, the deviation values ​​of the three dimensions correspond to the adjustment benchmark of the three-level prompt meta-database respectively, and the value range is [0,1]. The larger the deviation value, the higher the degree of semantic deviation of that dimension from the domain benchmark. Within a single level, the cosine similarity between each candidate suggestion unit and the corresponding dimension's baseline semantic center vector is calculated one by one. The calculation result is used as the deviation correction fit of that suggestion unit, with the fit value ranging from [0,1]. The semantic vector of the candidate suggestion unit uses the same pre-trained language model as S13. The dimension's baseline semantic center vector is obtained from the arithmetic mean of all baseline semantic vectors in that dimension within the target domain's prior semantic feature dataset. Subsequently, the dimension deviation value of the current level is used as the baseline for the total adjustment magnitude. The independent adjustment increment is calculated by combining the difference between the fit of each unit and the average fit of the level. The calculation formula is as follows: ,in This is a preset incremental scaling factor, with a default value of 0.5, determined by historical parameter tuning experiments. It is used to control the fluctuation range of weights in a single round to avoid iterative divergence. This represents the dimensional deviation value corresponding to the current level. The deviation correction fit degree for the i-th prompt unit; The average fitness of all units within the current level is obtained by taking the arithmetic mean of the fitness of all units within the level; The original weight of each unit after correction is obtained by adding 1 / 3 of the initial equal weight to the corresponding adjustment increment and multiplying it by the corresponding weight adjustment coefficient. After the calculation is completed, a weight lower limit of 0.05 is set for truncation to avoid negative weights after correction. S43. Weighted fusion of the prompting units in each level according to the corrected unit weight value, then reorganized and spliced ​​in hierarchical order to generate iterated enhanced prompting data, and the reorganized enhanced prompting data generated by the iteration is returned to S3 to enter the next iteration. Specifically, in embodiment S43: the original weights of each prompt unit within the same level are first normalized by summing within the level; the prompt units are sorted from high to low according to the normalized weights, and the prompt unit ranked first is retained as the core prompt of the level, and the subsequent units are added to the end of the core prompt as supplementary prompts; prompt units with a weight lower than 0.1 are removed, and finally a single-level fused prompt text is formed; the fused prompt text of the three layers is spliced ​​together in a fixed hierarchical order of task anchoring layer first, logical constraint layer in the middle, and expression style layer last, and finally the iterated complete enhanced prompt data is generated; S5. The large language model is driven by the final optimized enhanced prompt data to complete inference, generate and output the target constructed text that meets the task requirements; Step S5 includes the following steps: S51. Obtain the final optimized enhanced prompt data, convert it into a word sequence, load it into the input of the large language model, and start the complete generation inference; Specifically, in embodiment S51: the complete generation reasoning adopts the standard autoregressive reasoning mechanism of the DeepSeek-Pro (65B) large language model, which is a mature existing technology: after the input sequence is converted into a feature vector by the embedding layer, it is sequentially processed by all Transformer decoder layers to complete the forward calculation. The semantic representation is transferred through the multi-head self-attention mechanism and the feedforward neural network. Finally, the probability distribution of the next word is obtained through the output layer vocabulary mapping. The target word is selected based on the sampling strategy and added to the output sequence. The generation process is executed cyclically until the termination marker is generated or the maximum generation length is reached, thus completing the complete reasoning. S52. After the reasoning is completed, the enhanced cue word sequence of the input is stripped from the output sequence, and the text fragments corresponding to the pure generated content are extracted to obtain the initial generated text. S53. Perform format regularization and constraint verification on the initial generated text, remove redundant content that does not meet the task requirements, and obtain the standardized target construction text and output it. Specifically, in embodiment S53: the verification rules come from two sources: first, task-level constraints parsed and extracted from the initial task instructions, including customized requirements such as output format, length range, and content boundaries; second, domain-wide compliance rules pre-set in the logical constraint layer of the three-level prompt meta-database, including general requirements such as domain-prohibited expressions, format specification standards, and logical consistency verification rules. See Figure 2 Furthermore, a large language model text construction system based on prompt enhancement is provided, the system comprising a data acquisition and initialization module, a semantic retrieval and prompt generation module, a semantic bias evaluation module, a weight dynamic correction and iteration module, and a text generation and output module; Data acquisition and initialization module: Acquires initial task instructions, a three-level prompt meta-database, and a target domain prior semantic feature dataset, and parses, loads, and organizes them by dimension; Semantic retrieval and prompt generation module: Using the semantic vector of the initial task instruction as the retrieval source, it performs semantic retrieval on the prompt meta-database at each level, filters candidate prompt units, and concatenates them to generate initial enhanced prompt data; Semantic bias assessment module: Input the enhanced prompt data into the large language model, collect the specified intermediate layer hidden states, match and calculate with the prior semantic features, obtain the semantic bias quantification result, and determine whether the threshold condition is met. Weight dynamic correction and iteration module: If the deviation does not meet the threshold, the deviation quantification result is mapped to the weight adjustment coefficient of each level, the dynamic correction prompt unit splices the weights and reorganizes them to generate the enhanced prompt data after iteration, and returns to the semantic deviation evaluation module to continue the iteration. Text generation and output module: Obtains the final optimized enhanced prompt data, drives the large language model to generate the target text, and outputs it after formatting and verification; In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0024] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for constructing large language model text based on cue enhancement, characterized in that, Includes the following steps: S1. Obtain the initial task instructions, the pre-built three-level prompt meta-database, and the target domain prior semantic feature dataset corresponding to the text construction task; S2. Using the initial task instruction as the retrieval source, perform semantic retrieval with each layer of the three-level prompt meta-database to obtain initial enhanced prompt data; S3. Input the initial enhanced prompt data as the enhanced prompt data for the current round into the large language model to perform inference, and simultaneously collect the hidden state representation vector of the specified intermediate layer of the model. The hidden state representation vector is matched with the prior semantic feature dataset of the target domain by cosine similarity and difference calculation, and the semantic deviation quantification result is output. If the semantic deviation quantization result meets the preset threshold condition or the current iteration round reaches the preset maximum number of iterations, then the enhanced prompt data of the current round is used as the final optimized enhanced prompt data, and step S5 is executed directly; S4. If the semantic deviation quantization result does not meet the preset threshold condition, the weight adjustment coefficients corresponding to each layer of the three-level prompt meta-database are adjusted according to the deviation quantization results of each layer. The splicing weights of each layer prompt unit are dynamically corrected based on the weight adjustment coefficients. The iteratively enhanced prompt data is recombined and generated. The iteratively enhanced prompt data is then returned to S3 to enter the next iteration. S5. Drive the large language model to complete inference through the final optimized enhanced prompt data, and generate and output the target constructed text that meets the task requirements.

2. The method for constructing a large language model text based on cue enhancement according to claim 1, characterized in that, S1 includes the following steps: S11. Receive the input description of the text construction task, perform semantic parsing on the task objective, constraints, and output requirements, and extract the structured initial task instructions. S12. Obtain a pre-built three-level prompt meta-database. The three-level prompt meta-database is stored in layers according to the task anchoring layer, the logical constraint layer, and the expression style layer. Each layer contains several independent and callable prompt units, and there is no semantic overlap between the prompt units in each layer. S13. Obtain the target domain prior semantic feature dataset, wherein the target domain is the domain of legal documents, financial reports, or medical texts; the target domain prior semantic feature dataset is a set of benchmark semantic vectors extracted from the target domain standard corpus by a pre-trained language model and organized and stored according to the task intent dimension, logical compliance dimension, and expression style dimension, and is used as a reference benchmark for semantic alignment.

3. The method for constructing a large language model text based on cue enhancement according to claim 2, characterized in that, In step S12, the three-level prompt metadata database is pre-constructed through the following steps: S121. Collect high-quality text samples from the target domain and extract three types of semantic fragments: task target class expression, logical rule class constraint, and style class feature. S122. Standardize and rewrite the three types of semantic fragments and deduplication them respectively to generate the initial prompt unit sets of the task anchoring layer, logical constraint layer and expression style layer. S123. Extract semantic vectors for all prompt units within each level and establish a retrieval index to form a three-level prompt meta-database with independent hierarchical storage.

4. The method for constructing a large language model text based on cue enhancement according to claim 2, characterized in that, S2 includes the following steps: S21. Extract the semantic vector corresponding to the initial task instruction, and use the semantic vector as the retrieval vector to perform semantic cosine similarity retrieval on the task anchoring layer, logical constraint layer and expression style layer of the three-level prompt meta-database respectively. S22. For each level, sort the prompt units from high to low according to semantic similarity, and select a preset number of prompt units as the candidate prompt unit set for that level. S23. Following the fixed hierarchical order of task anchoring layer, logical constraint layer, and expression style layer, the candidate prompt units are sequentially concatenated to generate initial enhanced prompt data.

5. The method for constructing a large language model text based on cue enhancement according to claim 4, characterized in that, S3 includes the following steps: S31. Input the initial enhanced prompt data as the enhanced prompt data for the current round into the large language model, trigger the model to perform forward inference, and collect the hidden state representation vector output by the specified intermediate layer of the model during the inference process. S32. Perform batch cosine similarity matching between the hidden state representation vector and all benchmark semantic vectors in the target domain prior semantic feature dataset to calculate the global average semantic similarity value. S33. Using the difference between 1 and the global average semantic similarity value as the semantic deviation quantization result, output the semantic deviation quantization result of the current round; S34. If the semantic deviation quantization result meets the preset threshold condition or the current iteration round reaches the preset maximum number of iterations, then the enhanced prompt data of the current round is used as the final optimized enhanced prompt data, and step S5 is executed directly.

6. The method for constructing a large language model text based on cue enhancement according to claim 5, characterized in that, S32 includes the following steps: S321. Perform dimensional splitting on the hidden state representation vector to extract sub-feature vectors for the task intent dimension, logical compliance dimension, and expression style dimension respectively. S322. Calculate the cosine similarity between the sub-feature vector and the corresponding dimension of the baseline semantic vector in the prior semantic feature dataset one by one to obtain the independent similarity value of each dimension. S323. The independent similarity values ​​of each dimension are weighted and summed according to the dimension weights to obtain the global average semantic similarity value.

7. The method for constructing large language model text based on cue enhancement according to claim 5, characterized in that, S4 includes the following steps: S41. If the semantic deviation quantization result does not meet the preset threshold condition, the semantic deviation quantization result is input into the preset hierarchical weight mapping function, and the weight adjustment coefficients corresponding to the task anchoring layer, logical constraint layer and expression style layer are output respectively. S42. Based on the respective weight adjustment coefficients, the initial splicing weights of each candidate prompt unit in the level are dynamically corrected to obtain the corrected weight values ​​of each prompt unit. S43. The prompt units in each level are weighted and merged according to the corrected unit weight values, and then reassembled and spliced ​​in hierarchical order to generate the iterative enhanced prompt data. The iterative enhanced prompt data generated by the reassembly is returned to S3 to enter the next round of iteration.

8. The method for constructing a large language model text based on cue enhancement according to claim 7, characterized in that, In step S41, the hierarchical weight mapping function adopts a mapping rule that differentiates based on bias sensitivity, specifically as follows: S411, Preset the deviation sensitivity coefficients for the task anchoring layer, logical constraint layer, and expression style layer respectively; S412. Multiply the independent deviation value of each dimension by the deviation sensitivity coefficient of each layer to obtain the weight adjustment increment of each layer; S413. Add the initial baseline weights of each layer to the corresponding weight adjustment increments, and output the corresponding weight adjustment coefficients of each layer after L1 normalization.

9. A method for constructing large language model text based on cue enhancement according to claim 7, characterized in that, S5 includes the following steps: S51. Obtain the final optimized enhanced prompt data, convert it into a word sequence, load it into the input of the large language model, and start the complete generation inference; S52. After the reasoning is completed, the enhanced cue word sequence of the input is stripped from the output sequence, and the text fragments corresponding to the pure generated content are extracted to obtain the initial generated text. S53. Perform format regularization and constraint verification on the initially generated text, remove redundant content that does not meet the task requirements, and obtain the standardized target construction text and output it.

10. A large language model text construction system based on prompting enhancement, characterized in that, The method for constructing large language model text based on prompt enhancement as described in any one of claims 1-9 includes a data acquisition and initialization module, a semantic retrieval and prompt generation module, a semantic bias evaluation module, a weight dynamic correction and iteration module, and a text generation and output module. Data acquisition and initialization module: acquires initial task instructions, a three-level prompt meta-database, and a target domain prior semantic feature dataset, and parses, loads, and organizes them by dimension; Semantic retrieval and prompt generation module: Using the semantic vector of the initial task instruction as the retrieval source, it performs semantic retrieval on the prompt meta-database at each level, filters candidate prompt units, and concatenates them to generate initial enhanced prompt data; Semantic bias assessment module: Input the enhanced prompt data into the large language model, collect the hidden states of the specified intermediate layer, match and calculate with the prior semantic features, obtain the semantic bias quantification result, and determine whether the threshold condition is met. Weight dynamic correction and iteration module: If the deviation does not meet the threshold, the deviation quantification result is mapped to the weight adjustment coefficient of each level, the dynamic correction prompt unit splices the weights and reorganizes them to generate the enhanced prompt data after iteration, and returns to the semantic deviation evaluation module to continue the iteration. Text generation and output module: Obtains the final optimized enhanced prompt data, drives the large language model to generate the target text, and outputs it after formatting and verification.

Citation Information

Patent Citations

  • Fusion caching method, system and equipment based on LLM cue word and medium

    CN120277069A

  • Large language model question and answer method and device for multi-task scene and medium

    CN121456088A