An archive data intelligent retrieval strategy optimization method based on a large model

Abstract

The application discloses a kind of based on big model's file data intelligent retrieval strategy optimization method, comprising: S1, initial retrieval strategy tree is generated with identity token by collecting query text;S2, is distributed to text semantic and structured metadata double-channel parallel probe recall, obtains initial candidate set and feature matrix;S3, compliance conflict detection is carried out, and batch critical feature vector is generated by identifying defect;S4, input improved DreamerV3 model is mapped into directed causal graph, and positioning causal source node executes counterfactual mask operation, and the target optimization strategy is output after multi-step deduction;S5, is parsed into structured query language and executed secondary fusion retrieval, and obtains target archive set;S6, parameter fine-tuning is carried out by extracting log and constructing sample, and strategy library closed loop self-evolution is realized.The application realizes the cause and effect level of retrieval defect and self-repair, effectively improves retrieval accuracy and authority control ability.

Description

Technical Field

[0001] This invention relates to the intersection of artificial intelligence and digital archival management, and in particular to an optimization method for intelligent archival data retrieval strategy based on a large model. Background Technology

[0002] Intelligent archival retrieval systems, with their deep analytical capabilities for massive amounts of unstructured text and structured metadata, have been widely applied in recent years in fields such as digital government, enterprise knowledge bases, and confidentiality reviews, becoming an important development direction for realizing the assetization of archival data. However, in practical applications, archival retrieval scenarios face many challenges, such as strict access boundaries, dynamic changes in retention periods, and complex ownership of archival collections. The optimization effect of existing intelligent retrieval strategies is still constrained by many factors.

[0003] Most current archival retrieval methods rely on single-channel feature matching or fixed-weight dual-track retrieval, which struggles to simultaneously leverage the generalization ability of text semantic vectors and the precise location capabilities of structured metadata. Furthermore, they cannot dynamically adjust the matching thresholds of the dual-track channels based on query intent. Some systems only apply hard-score filtering after the retrieval results are output, ignoring the underlying causes of multi-dimensional compliance defects such as exceeding security level limits and violating time limits. They lack the ability to locate and intervene in the causal sources of defects in the initial retrieval strategy, thus limiting the adaptive correction level of the retrieval strategy. Simultaneously, existing large-model-based retrieval logic processes lack transparent explanation paths, making it difficult to provide archivists with clear evidence for tracing the origins of defects, thus affecting the security and reliability of the retrieval results.

[0004] Furthermore, most existing retrieval strategy optimization mechanisms are static rule rewriting or directly rely on large-scale black-box model generation, failing to construct closed-loop training samples containing defect features and optimization trajectories for fine-tuning parameters in local systems. Due to the lack of counterfactual deduction and multi-step forward imaginative evaluation of the internal world model, the system cannot predict the actual compliance effects of different strategies after intervening in potential states. It is difficult to transform corrective experience into weighted memories at the system's underlying level, making it difficult for the retrieval strategy library to adapt to the continuous evolution of document classification rules and security boundaries. This severely impacts the practical value and stability of intelligent retrieval models in real-world, highly confidential scenarios.

[0005] Therefore, how to provide an optimization method for intelligent retrieval strategy of archival data based on large models is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an intelligent retrieval strategy optimization method for archival data based on a large model. This invention fully integrates key steps such as dual-track probe recall, compliance conflict detection, improved DreamerV3 model deduction, counterfactual mask intervention, and local parameter fine-tuning. It constructs a closed-loop self-evolving process for retrieval strategies, featuring defect feature quantification, directed causal graph structure tracing, multi-step forward imagination evaluation, and dynamic Boolean logic rewriting. This achieves accurate diagnosis and adaptive correction of archival retrieval defects in complex permission boundary environments. By mapping the policy defect critique feature vector to the directed causal graph structure of the improved DreamerV3 model's internal world model, this invention utilizes backward gradient tracing to accurately locate the causal source nodes causing defects such as boundary violations and performs blocking and deduction in the latent space. It possesses advantages such as transparent causal tracing paths, precise dual-track weight allocation, strong adaptive dynamic threshold adjustment, and closed-loop evolution of the strategy library. This significantly improves the compliance of retrieval results, the timeliness of defect correction, and the credibility of model output in high-security scenarios, effectively solving problems such as the solidification of dual-track channel weights, passive compliance interception, and black-box optimization processes in existing methods.

[0007] An optimization method for intelligent retrieval strategy of archival data based on a large model according to an embodiment of the present invention includes the following steps: S1. Synchronously collect natural language query text and identity authentication token, perform semantic parsing on natural language query text to extract initial search keyword set, and generate initial search strategy tree; S2. Distribute the initial retrieval strategy tree to both the text semantic retrieval channel and the structured metadata retrieval channel, and perform probe-style recall in parallel to obtain the corresponding multi-dimensional initial file candidate set and associated metadata feature matrix. S3. Perform compliance conflict detection on the multi-dimensional initial archive candidate set and metadata feature matrix, identify at least one structural retrieval defect among the following: overstepping of confidentiality level, violation of retention period, and incorrect ownership of archive number, and generate a strategy defect critical feature vector. S4. Input the strategy defect critique feature vector into the improved DreamerV3 model and map it into the directed causal graph structure of the internal world model of the improved DreamerV3 model. Track and locate the causal source node that causes structural retrieval defects in the initial retrieval strategy, perform counterfactual masking operation, block the original error state propagation path, and re-perform multi-step forward imagination deduction to output the target optimized retrieval strategy. S5. Parse the target optimization retrieval strategy into an executable structured query language and calculate the text channel similarity threshold. Perform a secondary fusion retrieval in the text semantic retrieval channel and the structured metadata retrieval channel to obtain the target archive dataset. S6. Extract the generation process log of the target archive dataset, construct training samples from the natural language query text, strategy defect critique feature vector, and target optimization retrieval strategy, fine-tune and update the parameters of the local retrieval strategy system, and realize the closed-loop self-evolution of the retrieval strategy library.

[0008] Optionally, S1 specifically includes: S11. Synchronously capture the natural language query text and identity authentication token input by the user through the system front-end interaction interface, parse the access permission attributes containing the security level and department code from the identity authentication token, call the preset stop word list, remove the characters existing in the stop word list from the natural language query text word by word, divide the remaining text string after removal into multiple words by spaces, and summarize to generate the initial search keyword set. S12. Read the locally stored archive classification terminology, calculate the character overlap between each word in the initial search keyword set and each category name in the archive classification terminology, and assign each word to the category name with the highest character overlap. S13. Concatenate all words belonging to the same category name using a preset connector to generate a candidate search path in the form of a string. Iterate through all category names to generate multiple candidate search paths. S14. Establish a tree-like connection relationship containing the root node and multi-level branch nodes, store multiple candidate search paths in the multi-level branch nodes respectively, and output the initial search strategy tree.

[0009] Optionally, S2 specifically includes: S21. Copy the multiple candidate retrieval paths in the initial retrieval strategy tree into two copies, distribute the first copy of the path to the text semantic retrieval channel, and distribute the second copy of the path to the structured metadata retrieval channel. S22. In the text semantic retrieval channel, each word in the path copy is read sequentially. The 256-dimensional numerical array corresponding to each Chinese character is found in the pre-stored character vector mapping table. The numerical arrays of all Chinese characters in the same word are concatenated first and last in order. If the word length is less than the preset second threshold, zeros are added to the end. If it exceeds the preset second threshold, it is truncated to generate the word vector of each word. S23. Concatenate all word vectors in the path copy in word order to generate a long vector. Calculate the dot product of the long vector and each existing file vector in the file vector library and divide it by the product of their lengths to obtain the cosine similarity. Extract the file record identifiers with a cosine similarity greater than the preset first threshold and generate text recall. S24. In the structured metadata retrieval channel, extract the words connected by AND in the path copy, and search for strings that are exactly the same as the words in the title column and the responsible person column of the database index table in turn. Extract the archive record identifier of the row containing the same string and generate a metadata recall list. S25. Merge the file record identifiers in the text recall list and the metadata recall list into the same set. Iterate through each file record identifier in the set. If it exists in both the text recall list and the metadata recall list, keep only one. Combine the file data corresponding to all the deduplicated file record identifiers to generate a multi-dimensional initial file candidate set. S26. Traverse each file in the multi-dimensional initial candidate file set, read the security classification value, retention period value, and archival group number value of each file in the database, multiply the security classification value by 100, the retention period value by 10, and the archival group number value by 1, sum the three multiplications to obtain a total value, write the total value of each file into a one-dimensional array according to the order of the file data in the candidate set, and generate the associated metadata feature matrix.

[0010] Optionally, S3 specifically includes: S31. Read the identity authentication token of the currently logged-in user, extract the security level permission value and the total number permission value owned by the user from the identity authentication token, and read the current system timestamp at the same time. S32. For each piece of file data in the multi-dimensional initial file candidate set, compare the security level value in the file data with the security level permission value that the user possesses. If the security level value in the file data is greater than the security level permission value, mark the corresponding file data as having a security level permission overstepping. S33. For each piece of archive data in the multi-dimensional initial archive candidate set, calculate the difference between the current system timestamp and the retention period value in the archive data. If the difference is greater than the preset third threshold, mark the corresponding archive data as having a retention period violation. S34. For each piece of archive data in the multi-dimensional initial archive candidate set, compare the group number value in the archive data with the group number permission value owned by the user. If the two values ​​are not equal, mark the corresponding archive data as having an incorrect group number ownership. S35. Initialize a one-dimensional array with the same number of elements as the number of archive data entries in the multi-dimensional initial archive candidate set, and with all elements being zero, as a defect marker vector. Iterate through all archive data. If the archive data is marked as exceeding the confidentiality level, modify the value at the corresponding position in the defect marker vector to 1. If it is marked as violating the retention period, modify it to 2. If it is marked as having an incorrect record number, modify it to 3. S36. Multiply the sum of the values ​​at each position in the metadata feature matrix with the values ​​at the same position in the defect marker vector, arrange the products in the original order to generate a new one-dimensional array, and use the new one-dimensional array as the policy defect critique feature vector.

[0011] Optionally, S4 specifically includes: S41. Divide the strategy defect criticism feature vector into a first sub-vector and a second sub-vector according to the data column. Input each value in the first sub-vector into the first fully connected layer for linear transformation and input it into the ReLU function. Arrange the results of the ReLU function on all values ​​in the first sub-vector in order to generate a continuous state matrix. S42. Read each integer value in the second sub-vector, divide each integer value by the constant 3 and round down, and use the rounded result as the category number. Find the mean and standard deviation of the corresponding category number in the Gaussian probability distribution formula with the same number of integer values. Draw a random number from the standard normal distribution random number table, multiply the random number by the standard deviation value and add the mean value to calculate the floating-point number with the same number of integer values, and arrange them in order to generate a random state vector. S43. Add each value in the continuous state matrix to the corresponding value in the random state vector digit by digit to generate the initial world state vector. Calculate the covariance between the i-th value and the j-th value in the initial world state vector. Divide the covariance by the product of the standard deviation of the i-th value and the standard deviation of the j-th value to obtain the Pearson correlation coefficient. S44. If the Pearson correlation coefficient is greater than the preset fourth threshold, then create a pointing line from the i-th value to the j-th value, traverse all the values ​​in the initial world state vector and combine them in pairs, treat all the values ​​as dots and draw all the pointing lines as directed line segments in the two-dimensional coordinate system to generate a directed causal graph structure. S45. In the two-dimensional coordinate system, find the point at the end of the directed line segment. Determine whether the value of the point at the corresponding position in the policy defect criticism feature vector is greater than 0. If it is greater than 0, mark the point as a defect result node. Starting from the defect result node, move along the directed line segment against the direction of the arrow to the starting point. Calculate the product of the Pearson correlation coefficients corresponding to all directed line segments on the moving path. If the product is less than the preset fifth threshold, delete the corresponding moving path and mark the starting point of the remaining valid moving path as the causal source node. S46. Create a one-dimensional mask array, read the causal source nodes and calculate the intervened state vector, and use the dual-track weight allocation coefficient adjustment and Boolean logic expression to rewrite the output target to optimize the retrieval strategy.

[0012] Optionally, S46 specifically includes: S461. Create a one-dimensional mask array in which all elements are 1. Read the position index of the causal source node in the initial world state vector. Modify the value of the corresponding position in the one-dimensional mask array to 0. Multiply each value in the initial world state vector with the value of the same position in the one-dimensional mask array. Use the product as the new value and arrange it in the original order to generate the intervened state vector. S462. The state vector to be intervened is taken as the input of a step size. The first half of the value of the state vector to be intervened in the current step size is multiplied by the weight coefficients in the first row of the state transition matrix and summed. The second half of the value of the state vector to be intervened in the current step size is multiplied by the weight coefficients in the second row of the state transition matrix and summed. The two sums are added to the transition bias constant to generate the predicted state vector of the next step size. This multiplication and summation calculation step is repeated a preset number of times to generate the predicted state vectors from step 1 to the preset step size in sequence. S463. Input the predicted state vectors from step 1 to the preset step into the second fully connected layer in sequence. Perform a linear transformation on each predicted state vector to generate scalar values ​​from step 1 to the preset step. Sequentially determine whether the scalar values ​​from step 1 to the preset step are greater than 0. If they are greater than 0, record the value 1. If they are less than or equal to 0, record the value -1. Add the preset number of recorded values ​​to generate the compliance reward evaluation result. S464. In the compliance reward assessment results, find the record value with the largest sum and the corresponding step size. Extract the first two values ​​in the predicted state vector corresponding to the step size. Divide the first and second values ​​by a constant 5 to generate the first and second proportional coefficients. Write them into the same array in order to generate the dual-track weight allocation coefficient adjustment. S465. Extract the 3rd to the preset value from the predicted state vector corresponding to the record value with the largest sum. Sequentially determine whether each value from the 3rd to the preset value is greater than the preset sixth threshold. If a value is greater than the preset sixth threshold, search for the preset connector AND in the string of the initial retrieval strategy tree stored in memory and replace the preset connector AND with the character OR. S466. If the value is less than or equal to the preset sixth threshold, keep the preset AND connector unchanged. After traversal, extract the complete string that has been replaced in memory and generate a Boolean logical expression to rewrite it. S467. Simultaneously write the Boolean logic expression rewrite and the dual-track weight allocation coefficient adjustment into the same key-value pair data structure, and output the target optimized retrieval strategy.

[0013] Optionally, S5 specifically includes: S51. Read the Boolean logic expression rewrite in the target optimization retrieval strategy, identify the preset connector AND and character OR in the Boolean logic expression rewrite, map the preset connector AND to the database intersection query instruction, map the character OR to the database union query instruction, and use the words in the Boolean logic expression rewrite as search terms to fill the corresponding positions in the database query template containing the title column and the responsible person column, and generate an executable structured query language. S52. Read the dual-track weight allocation coefficient adjustment in the target optimization retrieval strategy, extract the first proportional coefficient and multiply it by the preset first threshold to obtain the text channel similarity threshold. S53. Read the executable Structured Query Language string, send the string character by character to the hard disk where the database is located, compare the characters corresponding to the query conditions row by row in the header column of the database index table stored on the hard disk, and compare the characters corresponding to the query conditions row by row in the responsible column of the database index table, extract the file record identifier of the row where the comparison is successful in the header column or the responsible column, write all the extracted file record identifiers into a one-dimensional array in order from top to bottom, and generate a secondary metadata recall list. S54. Read the first path copy in the initial retrieval strategy tree, concatenate them in the text semantic retrieval channel according to word order to generate a long vector, calculate the dot product of the long vector and each existing file vector in the file vector library and divide it by the product of their lengths to obtain the cosine similarity, extract the file record identifiers with cosine similarity greater than the text channel similarity threshold, and generate a secondary text recall list by arranging them in rows. S55. Merge the file record identifiers in the secondary metadata recall list and the secondary text recall list into the same deduplication set. Iterate through each file record identifier in the deduplication set. If the file record identifier exists in both the secondary metadata recall list and the secondary text recall list, keep only one. Read the complete file data corresponding to all deduplicated file record identifiers from the database hard disk and combine them to generate the target file dataset.

[0014] Optionally, S6 specifically includes: S61. Read each operation instruction and return result recorded in the generation process log of the target file dataset in chronological order, concatenate the operation instructions and return results in chronological order to generate a plain text string, and use the plain text string as background environment data. S62. Extract the natural language query text input by the user, read all the values ​​in the strategy defect criticism feature vector and convert them into comma-separated numeric strings in the original order, read the Boolean logic expression rewriting and dual-track weight allocation coefficient adjustment in the target optimization retrieval strategy, and store the natural language query text, numeric strings, Boolean logic expression rewriting and dual-track weight allocation coefficient adjustment in a fixed order into the same one-dimensional array to generate a training sample. S63. Divide the one-dimensional array in the training samples into input segment and label segment according to position. Replace each character in the input segment with the corresponding index value in the preset character dictionary in turn. Arrange the index values ​​in order to generate the sample input matrix. Extract the text in the label segment directly to generate the sample label matrix. S64. Input the sample input matrix into the neural network inside the local retrieval strategy system, calculate the difference between the predicted value output by the last layer of the neural network and the value in the sample label matrix, and divide the square of the difference by 2 to calculate the loss value. S65. Starting from the last layer of the neural network, calculate the partial derivative of the loss value with respect to each weight coefficient layer by layer. Subtract the corresponding partial derivative from the weight coefficient of each layer and multiply it by the preset learning rate to generate the updated weight coefficient. Write the updated weight coefficient back into the storage area of ​​the local retrieval strategy system and execute the closed-loop self-evolution of the retrieval strategy library.

[0015] The beneficial effects of this invention are: This invention addresses the challenges of coordinating semantic generalization and precise localization, as well as the reactive and lagging compliance interception, in archival retrieval scenarios by constructing a dual-track parallel probe recall and compliance conflict detection mechanism. It employs parallel probe recall via text semantic channels and structured metadata channels, combining rules based on security level, time limit, and archival group number to perform defect marking and feature matrix multiplication, generating a policy defect critique feature vector. This vector is then mapped to a directed causal graph structure of the improved DreamerV3 model's internal world model. Pearson correlation coefficients are used to construct node connections, and backgrading is combined to precisely locate the causal sources causing structural retrieval defects. The system performs counterfactual masking on causal source nodes to block error propagation and re-performs multi-step forward imaginative deductions on the intervened potential space branches. Based on the compliance reward assessment results, it outputs a target optimization retrieval strategy that includes rewriting Boolean logic expressions and adjusting dual-track weight allocation coefficients. In the secondary fusion retrieval stage, it parses the rewritten logic expressions and performs dual-channel constraint retrieval with dynamic thresholds to obtain target archive datasets that conform to physical permission boundaries. Furthermore, through a closed-loop self-evolution mechanism, it extracts generation process logs and optimization strategies to construct training samples, and performs partial derivative calculations and parameter fine-tuning updates on the local retrieval strategy system. Ultimately, it achieves precise causal tracing of archive data retrieval defects, strategy counterfactual deduction correction, and closed-loop self-evolution of the retrieval strategy library, effectively improving the accuracy of permission control, self-repair capability of strategy defects, and adaptive evolution level of the system model under complex compliance boundaries. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an intelligent retrieval strategy optimization method for archival data based on a large model proposed in this invention; Figure 2 This is a flowchart of the dual-track probe-based recall and multi-dimensional initial file candidate set generation process proposed in this invention. Figure 3 This is a flowchart of the counterfactual deduction optimization process based on the directed causal graph structure and the improved DreamerV3 model proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-3 An optimization method for intelligent retrieval strategy of archival data based on large model includes the following steps: S1. Synchronously collect natural language query text and identity authentication token, perform semantic parsing on the natural language query text to extract the initial search keyword set, and combine it with the archive classification thesaurus to generate an initial search strategy tree containing multiple candidate search paths; S2. Distribute each candidate retrieval path in the initial retrieval strategy tree to the text semantic retrieval channel and the structured metadata retrieval channel simultaneously, and perform probe-style recall in parallel to obtain the corresponding multi-dimensional initial file candidate set and associated metadata feature matrix. S3. Perform compliance conflict detection on the multi-dimensional initial archive candidate set and metadata feature matrix, identify at least one structural retrieval defect in the candidate set, such as exceeding the scope of confidentiality permissions, violating the retention period, and incorrectly attributing the record number, and generate a strategy defect critical feature vector. S4. Input the strategy defect critique feature vector into the improved DreamerV3 model, and simultaneously map it to the directed causal graph structure of the internal world model of the improved DreamerV3 model. Locate the causal source node that causes structural retrieval defects in the initial retrieval strategy through reverse gradient tracing. Perform counterfactual masking operation on the causal source node to block the original error state propagation path. Then, re-perform multi-step forward imaginary deduction on the intervened potential space branch. Based on the compliance reward evaluation results of the deduction trajectory, output the target optimized retrieval strategy, which includes Boolean logic expression rewriting and dual-track weight allocation coefficient adjustment. S5. Parse the target optimization retrieval strategy into an executable structured query language and calculate the text channel similarity threshold. Perform a secondary fusion retrieval in the text semantic retrieval channel and the structured metadata retrieval channel to obtain the target archive dataset that meets the dual constraints of physical permission boundaries and semantic matching. S6. Extract the generation process log of the target archive dataset, construct training samples from the natural language query text, strategy defect critique feature vector, and target optimization retrieval strategy, fine-tune and update the parameters of the local retrieval strategy system, and realize the closed-loop self-evolution of the retrieval strategy library.

[0019] This embodiment can significantly improve the accuracy of file retrieval and the ability of compliance control in scenarios with high security constraints. Through the dual-track probe-based recall mechanism, parallel matching of semantic generalization and structured metadata constraints is realized, which effectively overcomes the problems of information omission or weight imbalance caused by a single retrieval channel. After obtaining the initial candidate set, it can automatically perform in-depth compliance conflict detection, accurately identify complex structural defects such as out-of-bounds security classification and violation of storage period, and avoid the exposure of violation data. After introducing the improved DreamerV3 model, the defect features are mapped into a directed causal graph structure, which realizes causal-level backward tracking and accurate positioning of logical errors in the retrieval strategy. Through counterfactual masking operation and multi-step forward imagination deduction, the target optimization strategy including Boolean logic rewriting and dual-track weight adjustment is dynamically generated in the latent space. Combined with compliance reward evaluation, it ensures that the results of the secondary fusion retrieval strictly meet the dual constraints of physical permission and semantic matching. By extracting logs for parameter fine-tuning, the closed-loop self-evolution of the retrieval strategy library is realized. This method fundamentally solves the pain points of rigid traditional static retrieval strategies and difficult tracing of defects, shows strong robustness in the multi-source heterogeneous data environment, greatly reduces the cost of manual review, and comprehensively improves the security and adaptive optimization level of intelligent retrieval under complex permission boundaries.

[0020] In this embodiment, S1 specifically includes: S11: Synchronously capture the natural language query text and identity authentication token input by the user through the front-end interactive interface of the system, parse out the access permission attributes including security classification and department code from the identity authentication token, call a preset stop word list, remove characters existing in the stop word list from the natural language query text character by character, split the remaining text string after removal into multiple words by spaces, and aggregate to generate an initial retrieval keyword set, wherein the preset stop word list consists of 5 single Chinese characters including "的 (of)", "了 (modal particle)", "在 (at / in)", "是 (is)", "和 (and)" and punctuation marks; S12: Read the locally stored file classification thesaurus, sequentially calculate the character coincidence degree between each word in the initial retrieval keyword set and each category name in the file classification thesaurus, and assign each word to the category name with the highest character coincidence degree, wherein the calculation step of the character coincidence degree is to count the number of same Chinese characters between the word and the category name, divide the number of same Chinese characters by the total number of Chinese characters of the word, and obtain a decimal between 0 and 1; S13: Splice all words belonging to the same category name through a preset connector to generate a candidate retrieval path in the form of a character string, traverse all category names to generate multiple candidate retrieval paths, wherein the preset connector is set to AND; S14: Establish a tree-like connection relationship including a root node and multi-layer branch nodes, store multiple candidate retrieval paths in the multi-layer branch nodes respectively, and output an initial retrieval strategy tree including multiple candidate retrieval paths.

[0021] In this embodiment, S2 specifically includes: S21. Copy the multiple candidate retrieval paths in the initial retrieval strategy tree into two copies, distribute the first copy of the path to the text semantic retrieval channel, and distribute the second copy of the path to the structured metadata retrieval channel. S22. In the text semantic retrieval channel, each word in the path copy is read sequentially. The 256-dimensional numerical array corresponding to each Chinese character is found in the pre-stored character vector mapping table. The numerical arrays of all Chinese characters in the same word are concatenated first and last in order. If the word length is less than the preset second threshold, zeros are added to the end. If it exceeds the preset second threshold, it is truncated. The word vector of each word is generated. The preset second threshold is 64. S23. Concatenate all word vectors in the path copy in word order to generate a long vector. Calculate the dot product of the long vector and each existing file vector in the file vector library and divide it by the product of their lengths to obtain the cosine similarity. Extract the file record identifiers with a cosine similarity greater than a preset first threshold and generate a text recall list. The preset first threshold is 0.75. S24. In the structured metadata retrieval channel, extract the words connected by AND in the path copy, and search for strings that are exactly the same as the words in the title column and the responsible person column of the database index table in turn. Extract the archive record identifier of the row containing the same string and generate a metadata recall list. S25. Merge the file record identifiers in the text recall list and the metadata recall list into the same set. Iterate through each file record identifier in the set. If it exists in both the text recall list and the metadata recall list, keep only one. Combine the file data corresponding to all the deduplicated file record identifiers to generate a multi-dimensional initial file candidate set. S26. Traverse each file in the multi-dimensional initial candidate file set, read the security classification value, retention period value, and archival group number value of each file in the database, multiply the security classification value by 100, the retention period value by 10, and the archival group number value by 1, sum the three multiplications to obtain a total value, write the total value of each file into a one-dimensional array according to the order of the file data in the candidate set, and generate the associated metadata feature matrix.

[0022] In this embodiment, S3 specifically includes: S31. Read the identity authentication token of the currently logged-in user, extract the security level permission value and the total number permission value owned by the user from the identity authentication token, and read the current system timestamp at the same time. S32. For each piece of file data in the multi-dimensional initial file candidate set, compare the security level value in the file data with the security level permission value that the user possesses. If the security level value in the file data is greater than the security level permission value, mark the corresponding file data as having a security level permission overstepping. S33. For each piece of archive data in the multi-dimensional initial archive candidate set, calculate the difference between the current system timestamp and the retention period value in the archive data. If the difference is greater than the preset third threshold, mark the corresponding archive data as having a retention period violation. The preset third threshold is the number of seconds corresponding to 30 years. S34. For each piece of archive data in the multi-dimensional initial archive candidate set, compare the group number value in the archive data with the group number permission value owned by the user. If the two values ​​are not equal, mark the corresponding archive data as having an incorrect group number ownership. S35. Initialize a one-dimensional array with the same number of elements as the number of archive data entries in the multi-dimensional initial archive candidate set, and with all elements being zero, as a defect marker vector. Iterate through all archive data. If the archive data is marked as exceeding the confidentiality level, modify the value at the corresponding position in the defect marker vector to 1. If it is marked as violating the retention period, modify it to 2. If it is marked as having an incorrect record number, modify it to 3. S36. Multiply the sum of the values ​​at each position in the metadata feature matrix with the values ​​at the same position in the defect marker vector, arrange the products in the original order to generate a new one-dimensional array, and use the new one-dimensional array as the policy defect critique feature vector.

[0023] In this embodiment, S4 specifically includes: S41. Divide the strategy defect criticism feature vector into a first sub-vector and a second sub-vector according to the data column. Input each value in the first sub-vector into the first fully connected layer for linear transformation and input it into the ReLU function. Arrange the results of the ReLU function on all values ​​in the first sub-vector in order to generate a 256-dimensional continuous state matrix. S42. Read each integer value in the second sub-vector, divide each integer value by the constant 3 and round down, and use the rounded result as the category number. Find the mean and standard deviation of the corresponding category number in the Gaussian probability distribution formula with the same number of integer values. Draw a random number from the standard normal distribution random number table, multiply the random number by the standard deviation value and add the mean value to calculate the floating-point number with the same number of integer values, and arrange them in order to generate a random state vector. S43. Add each value in the continuous state matrix to the corresponding value in the random state vector digit by digit to generate the initial world state vector. Calculate the covariance between the i-th value and the j-th value in the initial world state vector. Divide the covariance by the product of the standard deviation of the i-th value and the standard deviation of the j-th value to obtain the Pearson correlation coefficient. S44. If the Pearson correlation coefficient is greater than the preset fourth threshold, then create a pointing line from the i-th value to the j-th value, traverse all the values ​​in the initial world state vector and combine them in pairs, treat all the values ​​as dots and draw all the pointing lines as directed line segments in the two-dimensional coordinate system to generate a directed causal graph structure. The preset fourth threshold is 0.6. S45. In the two-dimensional coordinate system, find the endpoint of the directed line segment and determine whether the value of the corresponding position of the endpoint in the strategy defect criticism feature vector is greater than 0. If it is greater than 0, mark the endpoint as a defect result node. Starting from the defect result node, move along the directed line segment against the direction of the arrow to the starting endpoint. Calculate the product of the Pearson correlation coefficients corresponding to all directed line segments on the moving path. If the product is less than the preset fifth threshold, delete the corresponding moving path and mark the starting endpoint of the remaining valid moving path as a causal source node. The preset fifth threshold is 0.1. S46. Create a one-dimensional mask array, read the causal source nodes and calculate the intervened state vector, and use the dual-track weight allocation coefficient adjustment and Boolean logic expression to rewrite the output target to optimize the retrieval strategy.

[0024] In this embodiment, S46 specifically includes: S461. Create a one-dimensional mask array containing 256 elements, all of which have a value of 1. Read the position index of the causal source node in the initial world state vector. Modify the value of the corresponding position in the one-dimensional mask array to 0. Multiply each value in the initial world state vector with the value of the same position in the one-dimensional mask array. Use the product as a new value and arrange it in the original order to generate the intervened state vector. S462. The state vector to be intervened is used as the input of a step size. The first 128 values ​​of the state vector to be intervened in the current step size are multiplied by the 128 weight coefficients in the first row of the state transition matrix and summed. The last 128 values ​​of the state vector to be intervened in the current step size are multiplied by the 128 weight coefficients in the second row of the state transition matrix and summed. The two summation results are added to the transition bias constant to generate the predicted state vector for the next step size. This multiplication and summation calculation step is repeated a preset number of times to generate the predicted state vectors from step 1 to the preset number of steps. The preset number of times is 10. S463. Input the predicted state vectors from step 1 to the preset step into the second fully connected layer in sequence. Perform a linear transformation on each predicted state vector to generate scalar values ​​from step 1 to the preset step. Sequentially determine whether the scalar values ​​from step 1 to the preset step are greater than 0. If they are greater than 0, record the value 1. If they are less than or equal to 0, record the value -1. Add the preset number of recorded values ​​to generate the compliance reward evaluation result. S464. In the compliance reward assessment results, find the record value with the largest sum and the corresponding step size. Extract the first two values ​​in the predicted state vector corresponding to the step size. Divide the first and second values ​​by a constant 5 to generate the first and second proportional coefficients. Write them into the same array in order to generate the dual-track weight allocation coefficient adjustment. S465. Extract the 3rd to the preset value from the predicted state vector corresponding to the record value with the largest sum. Sequentially determine whether each value from the 3rd to the preset value is greater than the preset sixth threshold. If a value is greater than the preset sixth threshold, search for the preset connector AND in the string of the initial retrieval strategy tree stored in memory and replace the preset connector AND with the character OR. The preset sixth threshold is 0.5. S466. If the value is less than or equal to the preset sixth threshold, keep the preset AND connector unchanged. After traversal, extract the complete string that has been replaced in memory and generate a Boolean logical expression to rewrite it. S467. Simultaneously write the Boolean logic expression rewriting and the dual-track weight allocation coefficient adjustment into the same key-value pair data structure, and output the target optimization retrieval strategy that includes the Boolean logic expression rewriting and the dual-track weight allocation coefficient adjustment.

[0025] This invention introduces an improved DreamerV3 model's internal world model and a directed causal graph structure to achieve precise causal-level tracing and dynamic counterfactual deduction of retrieval strategy defects. The defect feature vector is segmented and dimensionality reduced, then transformed by a fully connected layer and fused with Gaussian distribution sampling to generate an initial world state vector. A directed causal graph is constructed by calculating the Pearson correlation coefficient, and invalid paths are dynamically pruned using a coefficient product threshold to accurately locate the causal source node causing the violation. Counterfactual intervention is performed on the source node using a one-dimensional mask array to block error propagation, and multi-step forward imaginative deduction is performed in the latent space using a state transition matrix. Based on the compliance reward evaluation of the deduction trajectory, adaptive Boolean logic rewriting and dual-track weight adjustment are performed. This invention effectively solves the pain points of traditional rule matching being black-boxed and defects being difficult to eradicate under complex permission constraints, realizing a transformation of retrieval strategies from passive interception to proactive self-evolution, significantly improving the security and fault tolerance of highly confidential file retrieval.

[0026] The improved DreamerV3 model of this invention is similar to the original DreamerV3 model in that both retain the core architecture of the world model, namely, dynamic environmental modeling through state representation of the latent space, using state transition matrices for multi-step forward imaginative deduction to predict future state trajectories, and relying on neural networks to extract features and output evaluation results of policy optimization.

[0027] The difference lies in that this invention breaks through the limitation of the original DreamerV3 model, which relies solely on black-box recurrent neural networks for general state prediction and representation updates. It introduces an explicit directed causal graph structure and a counterfactual intervention mechanism, constructing a white-box causal attribution decision-making system. Building upon the original model's direct input of state vectors into the transition matrix, this invention pre-calculates the Pearson correlation coefficient and constructs the directed graph in steps S43 to S45, explicitly mapping continuous latent states as nodes and directed line segments in a two-dimensional coordinate system. Then, in step S45, reverse tracing based on path coefficient products and threshold pruning are performed, forcing the model to accurately locate the causal source nodes causing defects in the complex state space. Subsequently, in steps S461 to S463, instead of directly inferring from the original state, a one-dimensional mask array is used to perform counterfactual masking operations on the causal source nodes, blocking error propagation paths. Ten rigorous forward imaginary deductions are performed in the intervened branch space, and the optimal step size is extracted based on compliance reward evaluation. Finally, in steps S464 to S467, the predicted trajectory inversion is resolved by rewriting specific Boolean logic expressions and adjusting the weight coefficients of the dual tracks, rather than abstract action instructions.

[0028] Based on the aforementioned improvements, the beneficial effects of this invention are that by introducing a directed causal graph and a counterfactual mask, the improved DreamerV3 model can adaptively trace the root cause of underlying logical errors in complex compliance conflict scenarios, such as exceeding security level limits or violating deadlines. This breaks the limitation of the original model, which can only passively adapt to state changes and cannot explain the causes of strategy defects, and achieves interpretability and causal-level targeted repair of retrieval strategy optimization. This design significantly enhances the model's depth of exploration of structural defects and can accurately cut off the propagation links of erroneous states in the potential space. The causal tracing combined with the counterfactual inference mechanism improves the accuracy of strategy repair while effectively avoiding blind disturbances of global parameters, enhancing the security and robustness of the system's retrieval strategy evolution under strict permission constraints.

[0029] In this embodiment, S5 specifically includes: S51. Read the Boolean logic expression rewrite in the target optimization retrieval strategy, identify the preset connector AND and character OR in the Boolean logic expression rewrite, map the preset connector AND to the database intersection query instruction, map the character OR to the database union query instruction, and use the words in the Boolean logic expression rewrite as search terms to fill the corresponding positions in the database query template containing the title column and the responsible person column, and generate an executable structured query language. S52. Read the dual-track weight allocation coefficient adjustment in the target optimization retrieval strategy, extract the first proportional coefficient and multiply it by the preset first threshold to obtain the text channel similarity threshold. S53. Read the executable Structured Query Language string, send the string character by character to the hard disk where the database is located, compare the characters corresponding to the query conditions row by row in the header column of the database index table stored on the hard disk, and compare the characters corresponding to the query conditions row by row in the responsible column of the database index table, extract the file record identifier of the row where the comparison is successful in the header column or the responsible column, write all the extracted file record identifiers into a one-dimensional array in order from top to bottom, and generate a secondary metadata recall list. S54. Read the first path copy in the initial retrieval strategy tree, concatenate them in the text semantic retrieval channel according to word order to generate a long vector, calculate the dot product of the long vector and each existing file vector in the file vector library and divide it by the product of their lengths to obtain the cosine similarity, extract the file record identifiers with cosine similarity greater than the text channel similarity threshold, and generate a secondary text recall list by arranging them in rows. S55. Merge the file record identifiers in the secondary metadata recall list and the secondary text recall list into the same deduplication set. Iterate through each file record identifier in the deduplication set. If the file record identifier exists in both the secondary metadata recall list and the secondary text recall list, keep only one. Read the complete file data corresponding to all deduplicated file record identifiers from the database hard disk and combine them to generate the target file dataset.

[0030] In this embodiment, S6 specifically includes: S61. Read each operation instruction and return result recorded in the generation process log of the target file dataset in chronological order, concatenate the operation instructions and return results in chronological order to generate a plain text string, and use the plain text string as background environment data. S62. Extract the natural language query text input by the user, read all the values ​​in the strategy defect criticism feature vector and convert them into comma-separated numeric strings in the original order, read the Boolean logic expression rewriting and dual-track weight allocation coefficient adjustment in the target optimization retrieval strategy, and store the natural language query text, numeric strings, Boolean logic expression rewriting and dual-track weight allocation coefficient adjustment in a fixed order into the same one-dimensional array to generate a training sample. S63. Divide the one-dimensional array in the training samples into input segment and label segment according to position. Replace each character in the input segment with the corresponding index value in the preset character dictionary in turn. Arrange the index values ​​in order to generate the sample input matrix. Extract the text in the label segment directly to generate the sample label matrix. S64. Input the sample input matrix into the neural network inside the local retrieval strategy system, calculate the difference between the predicted value output by the last layer of the neural network and the value in the sample label matrix, and divide the square of the difference by 2 to calculate the loss value. S65. Starting from the last layer of the neural network, calculate the partial derivative of the loss value with respect to each weight coefficient layer by layer. Subtract the corresponding partial derivative from the weight coefficient of each layer and multiply it by the preset learning rate to generate updated weight coefficients. Rewrite the updated weight coefficients into the storage area of ​​the local retrieval strategy system and execute the closed-loop self-evolution of the retrieval strategy library. The preset learning rate is 0.001.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to the core intelligent archive retrieval platform of a large-scale provincial comprehensive government data management center. This center is responsible for the secure aggregation, classification, archiving, and cross-departmental sharing and access of government data from more than ten cities across the province. The platform's underlying storage contains over 4.5 million unstructured and structured mixed archive data, including policy documents, meeting minutes, approval work orders, and planning drawings. The archive access scenarios involve hundreds of staff members from dozens of commissions and bureaus of provincial government departments, with an average daily search request volume exceeding 8,000. Government archive management has extremely stringent access boundaries and compliance requirements. Typical restrictions include top-secret documents being accessible only to designated personnel via specific IP terminals, cross-departmental collaborative data being limited to specific archival group numbers, and a large number of historical documents with ten- or thirty-year confidentiality periods being prohibited from external exposure. Traditional archive retrieval systems have long relied on manually configured static Boolean logic and basic database field matching. When faced with ambiguous natural language queries from staff, they often fail to accurately balance the relationship between semantic generalization and precise metadata constraints. This rigid retrieval mode makes it easy for the system to trigger an imbalance in channel weight allocation when processing complex requests such as "viewing top-secret meeting minutes involving cross-city transportation planning in the past three years". This can lead to a large number of unauthorized access prompts or the return of a large amount of irrelevant waste data. It may even incorrectly expose sensitive file fragments that are within the confidentiality period as summaries to ordinary staff.

[0032] In practical deployment, the method of this invention is fully integrated into the retrieval chain of the government data center. When business personnel input a query request, the system simultaneously captures the identity authentication token and natural language statement, removes stop words, extracts keywords, and matches them with the government classification thesaurus to generate an initial retrieval strategy tree. Subsequently, the system distributes probes in parallel through the text semantic retrieval channel and the structured metadata retrieval channel to obtain a multi-dimensional initial candidate set of archives containing features such as security level, retention period, and archival number. If the candidate set contains non-compliant archives within the confidentiality period, the system immediately extracts the strategy defect critique feature vector and maps it to the improved DreamerV3 model's internal world model to construct a directed causal graph structure containing retrieval conditions, permission fields, and the correlation of hit results. Through Pearson correlation coefficient calculation and reverse gradient tracing, the system can accurately locate which connector or threshold setting in the initial logic caused the out-of-bounds error, and then perform counterfactual masking operations in the latent space. After multiple steps of forward imaginative deduction, it dynamically outputs the rewritten Boolean logic expression and dual-track weight allocation coefficients, realizing the closed-loop self-evolution of the entire retrieval strategy library. Table 1 below shows the comparative data of the method of this invention and the original static rule retrieval method in typical government affairs scenarios during a three-month trial period: Table 1. Comparative Data on the Multidimensional Evaluation of the Performance of Intelligent Retrieval Strategies for Government Archives under Complex Compliance Constraints

[0033] As can be seen from the comparative data in Table 1 above, the policy counterfactual inference optimization method based on the directed causal graph structure and the improved DreamerV3 model proposed in this invention demonstrates an overwhelming technical advantage in high-security government retrieval scenarios. In the most critical scenario of cross-departmental access to top-secret documents, traditional methods, unable to penetrate the complex coupling relationships of permissions, triggered 389 compliance interceptions in the face of over 1,200 monthly requests. However, only 11 of these interceptions actually completed logical tracing and policy repair, resulting in a repair rate as low as a paltry 2.9%. This means that the vast majority of errors would recur the next day. In contrast, this invention, through precise tracing and masking inference using the directed causal graph structure, successfully transformed 407 out of 414 interceptions into fine-tuning of underlying parameters and rewriting of Boolean logic, boosting the repair rate to 96.8%, fundamentally blocking the recurrence path of similar violations.

[0034] In terms of response timeliness, the advantages of this invention are equally evident. Government affairs inquiries often occur at high-pressure points involving multi-departmental collaboration. Traditional methods, when faced with complex queries, are hampered by multiple manual interventions to adjust rules and inefficient hard-scan filtering, resulting in an average response time of 24 to 31 seconds, severely slowing down government workflow efficiency. In contrast, this invention, relying on the improved DreamerV3 model's rapid deduction capabilities in the latent space, directly outputs the optimal solution, comprehensively compressing the average response time of the four core scenarios to within 12 seconds. In cross-departmental top-secret access, it achieves an ultra-fast response of 7.2 seconds, greatly improving the operational experience for business personnel.

[0035] The rationality of dual-track weight allocation is a key indicator for measuring whether a system can balance semantic understanding and field control. Traditional methods, due to fixed weights, have long seen this indicator hover around 50%, leading to either inaccurate or incomplete searches. The dynamic feedback mechanism of this invention stabilizes this indicator at over 90%, completely resolving the trade-off between text semantic generalization and precise positioning of structured metadata. Even more remarkable is the dramatic reduction in the false alarm rate for violation exposure. Traditional systems have a false alarm rate as high as 16.5% in retrieving historical archives with long-term confidentiality, resulting in the incorrect blocking of numerous compliant documents. This invention, however, firmly controls the false alarm rate to around 2%, directly reducing the workload of manual review by nearly 88%. Overall, this invention effectively solves the pain points in government archive management, such as complex and intricate permission boundaries, rigid and black-box retrieval strategies, and lagging compliance correction, providing a solid and reliable technical foundation for the utilization of high-security data assets.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing intelligent retrieval strategies for archival data based on a large model, characterized in that, Includes the following steps: S1. Synchronously collect natural language query text and identity authentication token, perform semantic parsing on natural language query text to extract initial search keyword set, and generate initial search strategy tree; S2. Distribute the initial retrieval strategy tree to both the text semantic retrieval channel and the structured metadata retrieval channel, and perform probe-style recall in parallel to obtain the corresponding multi-dimensional initial file candidate set and associated metadata feature matrix. S3. Perform compliance conflict detection on the multi-dimensional initial archive candidate set and metadata feature matrix, identify at least one structural retrieval defect among the following: overstepping of confidentiality level, violation of retention period, and incorrect ownership of archive number, and generate a strategy defect critical feature vector. S4. Input the strategy defect critique feature vector into the improved DreamerV3 model and map it into the directed causal graph structure of the internal world model of the improved DreamerV3 model. Track and locate the causal source node that causes structural retrieval defects in the initial retrieval strategy, perform counterfactual masking operation, block the original error state propagation path, and re-perform multi-step forward imagination deduction to output the target optimized retrieval strategy. S5. Parse the target optimization retrieval strategy into an executable structured query language and calculate the text channel similarity threshold. Perform a secondary fusion retrieval in the text semantic retrieval channel and the structured metadata retrieval channel to obtain the target archive dataset. S6. Extract the generation process log of the target archive dataset, construct training samples from the natural language query text, strategy defect critique feature vector, and target optimization retrieval strategy, fine-tune and update the parameters of the local retrieval strategy system, and realize the closed-loop self-evolution of the retrieval strategy library.

2. The method for optimizing intelligent retrieval strategies for archival data based on a large model according to claim 1, characterized in that, S1 specifically includes: S11. Synchronously capture the natural language query text and identity authentication token input by the user through the system front-end interaction interface, parse the access permission attributes containing the security level and department code from the identity authentication token, call the preset stop word list, remove the characters existing in the stop word list from the natural language query text word by word, divide the remaining text string after removal into multiple words by spaces, and summarize to generate the initial search keyword set. S12. Read the locally stored archive classification terminology, calculate the character overlap between each word in the initial search keyword set and each category name in the archive classification terminology, and assign each word to the category name with the highest character overlap. S13. Concatenate all words belonging to the same category name using a preset connector to generate a candidate search path in the form of a string. Iterate through all category names to generate multiple candidate search paths. S14. Establish a tree-like connection relationship containing the root node and multi-level branch nodes, store multiple candidate search paths in the multi-level branch nodes respectively, and output the initial search strategy tree.

3. The method for optimizing intelligent retrieval strategies for archival data based on a large model according to claim 1, characterized in that, S2 specifically includes: S21. Copy the multiple candidate retrieval paths in the initial retrieval strategy tree into two copies, distribute the first copy of the path to the text semantic retrieval channel, and distribute the second copy of the path to the structured metadata retrieval channel. S22. In the text semantic retrieval channel, each word in the path copy is read sequentially. The 256-dimensional numerical array corresponding to each Chinese character is found in the pre-stored character vector mapping table. The numerical arrays of all Chinese characters in the same word are concatenated first and last in order. If the word length is less than the preset second threshold, zeros are added to the end. If it exceeds the preset second threshold, it is truncated to generate the word vector of each word. S23. Concatenate all word vectors in the path copy in word order to generate a long vector. Calculate the dot product of the long vector and each existing file vector in the file vector library and divide it by the product of their lengths to obtain the cosine similarity. Extract the file record identifiers with a cosine similarity greater than the preset first threshold and generate text recall. S24. In the structured metadata retrieval channel, extract the words connected by AND in the path copy, and search for strings that are exactly the same as the words in the title column and the responsible person column of the database index table in turn. Extract the archive record identifier of the row containing the same string and generate a metadata recall list. S25. Merge the file record identifiers in the text recall list and the metadata recall list into the same set. Iterate through each file record identifier in the set. If it exists in both the text recall list and the metadata recall list, keep only one. Combine the file data corresponding to all the deduplicated file record identifiers to generate a multi-dimensional initial file candidate set. S26. Traverse each file in the multi-dimensional initial candidate file set, read the security classification value, retention period value, and archival group number value of each file in the database, multiply the security classification value by 100, the retention period value by 10, and the archival group number value by 1, sum the three multiplications to obtain a total value, write the total value of each file into a one-dimensional array according to the order of the file data in the candidate set, and generate the associated metadata feature matrix.

4. The method for optimizing intelligent retrieval strategies for archival data based on a large model according to claim 1, characterized in that, S3 specifically includes: S31. Read the identity authentication token of the currently logged-in user, extract the security level permission value and the total number permission value owned by the user from the identity authentication token, and read the current system timestamp at the same time. S32. For each piece of file data in the multi-dimensional initial file candidate set, compare the security level value in the file data with the security level permission value that the user possesses. If the security level value in the file data is greater than the security level permission value, mark the corresponding file data as having a security level permission overstepping. S33. For each piece of archive data in the multi-dimensional initial archive candidate set, calculate the difference between the current system timestamp and the retention period value in the archive data. If the difference is greater than the preset third threshold, mark the corresponding archive data as having a retention period violation. S34. For each piece of archive data in the multi-dimensional initial archive candidate set, compare the group number value in the archive data with the group number permission value owned by the user. If the two values ​​are not equal, mark the corresponding archive data as having an incorrect group number ownership. S35. Initialize a one-dimensional array with the same number of elements as the number of archive data entries in the multi-dimensional initial archive candidate set, and with all elements being zero, as a defect marker vector. Iterate through all archive data. If the archive data is marked as exceeding the confidentiality level, modify the value at the corresponding position in the defect marker vector to 1. If it is marked as violating the retention period, modify it to 2. If it is marked as having an incorrect record number, modify it to 3. S36. Multiply the sum of the values ​​at each position in the metadata feature matrix with the values ​​at the same position in the defect marker vector, arrange the products in the original order to generate a new one-dimensional array, and use the new one-dimensional array as the policy defect critique feature vector.

5. The method for optimizing intelligent retrieval strategies for archival data based on a large model according to claim 1, characterized in that, S4 specifically includes: S41. Divide the strategy defect criticism feature vector into a first sub-vector and a second sub-vector according to the data column. Input each value in the first sub-vector into the first fully connected layer for linear transformation and input it into the ReLU function. Arrange the results of the ReLU function on all values ​​in the first sub-vector in order to generate a continuous state matrix. S42. Read each integer value in the second sub-vector, divide each integer value by the constant 3 and round down, and use the rounded result as the category number. Find the mean and standard deviation of the corresponding category number in the Gaussian probability distribution formula with the same number of integer values. Draw a random number from the standard normal distribution random number table, multiply the random number by the standard deviation value and add the mean value to calculate the floating-point number with the same number of integer values, and arrange them in order to generate a random state vector. S43. Add each value in the continuous state matrix to the corresponding value in the random state vector digit by digit to generate the initial world state vector. Calculate the covariance between the i-th value and the j-th value in the initial world state vector. Divide the covariance by the product of the standard deviation of the i-th value and the standard deviation of the j-th value to obtain the Pearson correlation coefficient. S44. If the Pearson correlation coefficient is greater than the preset fourth threshold, then create a pointing line from the i-th value to the j-th value, traverse all the values ​​in the initial world state vector and combine them in pairs, treat all the values ​​as dots and draw all the pointing lines as directed line segments in the two-dimensional coordinate system to generate a directed causal graph structure. S45. In the two-dimensional coordinate system, find the point at the end of the directed line segment. Determine whether the value of the point at the corresponding position in the policy defect criticism feature vector is greater than 0. If it is greater than 0, mark the point as a defect result node. Starting from the defect result node, move along the directed line segment against the direction of the arrow to the starting point. Calculate the product of the Pearson correlation coefficients corresponding to all directed line segments on the moving path. If the product is less than the preset fifth threshold, delete the corresponding moving path and mark the starting point of the remaining valid moving path as the causal source node. S46. Create a one-dimensional mask array, read the causal source nodes and calculate the intervened state vector, and use the dual-track weight allocation coefficient adjustment and Boolean logic expression to rewrite the output target to optimize the retrieval strategy.

6. The method for optimizing intelligent retrieval strategies for archival data based on a large model according to claim 5, characterized in that, Specifically, S46 includes: S461. Create a one-dimensional mask array in which all elements are 1. Read the position index of the causal source node in the initial world state vector. Modify the value of the corresponding position in the one-dimensional mask array to 0. Multiply each value in the initial world state vector with the value of the same position in the one-dimensional mask array. Use the product as the new value and arrange it in the original order to generate the intervened state vector. S462. The state vector to be intervened is taken as the input of a step size. The first half of the value of the state vector to be intervened in the current step size is multiplied by the weight coefficients in the first row of the state transition matrix and summed. The second half of the value of the state vector to be intervened in the current step size is multiplied by the weight coefficients in the second row of the state transition matrix and summed. The two sums are added to the transition bias constant to generate the predicted state vector of the next step size. This multiplication and summation calculation step is repeated a preset number of times to generate the predicted state vectors from step 1 to the preset step size in sequence. S463. Input the predicted state vectors from step 1 to the preset step into the second fully connected layer in sequence. Perform a linear transformation on each predicted state vector to generate scalar values ​​from step 1 to the preset step. Sequentially determine whether the scalar values ​​from step 1 to the preset step are greater than 0. If they are greater than 0, record the value 1. If they are less than or equal to 0, record the value -1. Add the preset number of recorded values ​​to generate the compliance reward evaluation result. S464. In the compliance reward assessment results, find the record value with the largest sum and the corresponding step size. Extract the first two values ​​in the predicted state vector corresponding to the step size. Divide the first and second values ​​by a constant 5 to generate the first and second proportional coefficients. Write them into the same array in order to generate the dual-track weight allocation coefficient adjustment. S465. Extract the 3rd to the preset value from the predicted state vector corresponding to the record value with the largest sum. Sequentially determine whether each value from the 3rd to the preset value is greater than the preset sixth threshold. If a value is greater than the preset sixth threshold, search for the preset connector AND in the string of the initial retrieval strategy tree stored in memory and replace the preset connector AND with the character OR. S466. If the value is less than or equal to the preset sixth threshold, keep the preset AND connector unchanged. After traversal, extract the complete string that has been replaced in memory and generate a Boolean logical expression to rewrite it. S467. Simultaneously write the Boolean logic expression rewrite and the dual-track weight allocation coefficient adjustment into the same key-value pair data structure, and output the target optimized retrieval strategy.

7. The method for optimizing intelligent retrieval strategies for archival data based on a large model according to claim 1, characterized in that, S5 specifically includes: S51. Read the Boolean logic expression rewrite in the target optimization retrieval strategy, identify the preset connector AND and character OR in the Boolean logic expression rewrite, map the preset connector AND to the database intersection query instruction, map the character OR to the database union query instruction, and use the words in the Boolean logic expression rewrite as search terms to fill the corresponding positions in the database query template containing the title column and the responsible person column, and generate an executable structured query language. S52. Read the dual-track weight allocation coefficient adjustment in the target optimization retrieval strategy, extract the first proportional coefficient and multiply it by the preset first threshold to obtain the text channel similarity threshold. S53. Read the executable Structured Query Language string, send the string character by character to the hard disk where the database is located, compare the characters corresponding to the query conditions row by row in the header column of the database index table stored on the hard disk, and compare the characters corresponding to the query conditions row by row in the responsible column of the database index table, extract the file record identifier of the row where the comparison is successful in the header column or the responsible column, write all the extracted file record identifiers into a one-dimensional array in order from top to bottom, and generate a secondary metadata recall list. S54. Read the first path copy in the initial retrieval strategy tree, concatenate them in the text semantic retrieval channel according to word order to generate a long vector, calculate the dot product of the long vector and each existing file vector in the file vector library and divide it by the product of their lengths to obtain the cosine similarity, extract the file record identifiers with cosine similarity greater than the text channel similarity threshold, and generate a secondary text recall list by arranging them in rows. S55. Merge the file record identifiers in the secondary metadata recall list and the secondary text recall list into the same deduplication set. Iterate through each file record identifier in the deduplication set. If the file record identifier exists in both the secondary metadata recall list and the secondary text recall list, keep only one. Read the complete file data corresponding to all deduplicated file record identifiers from the database hard disk and combine them to generate the target file dataset.

8. The method for optimizing intelligent retrieval strategies for archival data based on a large model according to claim 1, characterized in that, S6 specifically includes: S61. Read each operation instruction and return result recorded in the generation process log of the target file dataset in chronological order, concatenate the operation instructions and return results in chronological order to generate a plain text string, and use the plain text string as background environment data. S62. Extract the natural language query text input by the user, read all the values ​​in the strategy defect criticism feature vector and convert them into comma-separated numeric strings in the original order, read the Boolean logic expression rewriting and dual-track weight allocation coefficient adjustment in the target optimization retrieval strategy, and store the natural language query text, numeric strings, Boolean logic expression rewriting and dual-track weight allocation coefficient adjustment in a fixed order into the same one-dimensional array to generate a training sample. S63. Divide the one-dimensional array in the training samples into input segment and label segment according to position. Replace each character in the input segment with the corresponding index value in the preset character dictionary in turn. Arrange the index values ​​in order to generate the sample input matrix. Extract the text in the label segment directly to generate the sample label matrix. S64. Input the sample input matrix into the neural network inside the local retrieval strategy system, calculate the difference between the predicted value output by the last layer of the neural network and the value in the sample label matrix, and divide the square of the difference by 2 to calculate the loss value. S65. Starting from the last layer of the neural network, calculate the partial derivative of the loss value with respect to each weight coefficient layer by layer. Subtract the corresponding partial derivative from the weight coefficient of each layer and multiply it by the preset learning rate to generate the updated weight coefficient. Write the updated weight coefficient back into the storage area of ​​the local retrieval strategy system and execute the closed-loop self-evolution of the retrieval strategy library.

Images