File positioning management method and system based on artificial intelligence

Abstract

The invention provides a file positioning management method and system based on artificial intelligence, and relates to the technical field of artificial intelligence. According to the method, files are collected from multiple sources and subjected to standardization processing, an element set is generated in combination with multi-modal analysis of texts, images, audios, videos and tables, cross-modal alignment is achieved through a semantic representation model, hierarchical indexes of semantics, keywords and relations are constructed, and a unique traceability identifier is generated; in the query stage, intention recognition and joint retrieval are carried out, a result subjected to permission verification and traceability information labeling is output, online optimization and incremental reconstruction are executed based on user feedback, and comprehensiveness, accuracy, traceability and self-adaptive optimization of file positioning are achieved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a file location management method and system based on artificial intelligence. Background Technology

[0002] The organization's files are scattered across local disks, shared drives, collaboration platforms, and message attachments, with varying naming conventions, directory structures, and version management. Traditional searches rely heavily on filenames and a few text fields, offering limited understanding of images, scanned documents, audio / video files, tables, and other content. Coupled with the dispersed sources, hierarchical access permissions, and numerous versions, the common result is that while "relevant" results can be found, it is difficult to pinpoint the "correct, usable, and accessible" document.

[0003] Content intelligence, exemplified by deep learning, is gradually becoming a fundamental capability for file management. Multimodal parsing, such as text extraction, image recognition, speech transcription, and table parsing, allows for a unified representation of file content; semantic representation and vector retrieval elevate "word matching" to "intent matching"; structured modeling combined with version and citation relationships enables location tracking to consider not only content but also upstream and downstream connections and temporal evolution; and integration with permission systems and audit logs makes search results more interpretable, more traceable, and allows for continuous optimization based on user feedback.

[0004] The existing solution still has three shortcomings: First, the parsing and representation often tend to be single-modal, with insufficient consistency across text, images, audio and video, and tables, and the integration of source and relationship information is not deep enough; Second, query processing mostly stays at the keyword level, lacking the identification and joint retrieval of intents such as "which version to find, where it comes from, and what time and permission conditions are met"; Third, there is a lack of a mechanism to transform feedback such as clicks, corrections, and annotations into stable improvements, which makes it difficult to maintain the positioning accuracy and sorting stability in the long term when the scale expands and permissions change frequently. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a file location management method and system based on artificial intelligence. Through the collaborative design of multimodal parsing, semantic alignment, hierarchical indexing, intent recognition and online learning, the invention achieves comprehensiveness, accuracy, traceability and adaptive optimization of file location, significantly overcoming the limitations of existing technologies.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] An AI-based file location management method includes:

[0008] File resources and corresponding metadata are collected from at least one file source, and format normalization and field standardization are completed under a unified time base and identification rules to form standardized file objects and source identifiers;

[0009] The standardized document object is parsed to extract a set of elements that reflect the content, source, and relationships. The set of elements covers text fragments, image highlights, audio and video transcription fragments, table structure information, and relationship information related to version, citation, and project.

[0010] Based on a preset semantic representation model, the set of elements is mapped to a document semantic representation, and consistency constraints are applied between different modalities of the same document to obtain a unified semantic representation for retrieval.

[0011] Based on the unified semantic representation and the relationship between source, version and citation, a hierarchical index structure including semantic index, keyword index and relationship index is established, and a unique traceability identifier is generated for each file, and the unique traceability identifier is associated with and stored with the corresponding traceability information;

[0012] Upon receiving a query request, the system performs intent recognition and condition parsing, generates a semantic query representation, structured filtering conditions, and relational constraints, performs semantic retrieval, keyword retrieval, and relational retrieval on the hierarchical index structure, and fuses and sorts the candidate results according to the task intent to obtain the location result.

[0013] The location results are verified for permissions, and the results of the verification and the source information are output. Based on user clicks, error correction and annotation feedback, the intent recognition strategy, fusion strategy and ranking weight are updated online, and incremental reconstruction of error samples is triggered under the premise of meeting audit requirements.

[0014] Preferably, file resources and corresponding metadata are collected from at least one file source, and format normalization and field standardization are completed under a unified time base and identification rules to form standardized file objects and source identifiers, including:

[0015] The system performs format detection on the accessed files and metadata, and unifies files from different sources into a preset format paradigm.

[0016] Based on a unified time base and identification rules, the creation time, modification time, and source information of the files are normalized.

[0017] The source-side metadata is mapped to system-side standard fields according to the field mapping table to form the normalized file object;

[0018] For each of the standardized document objects, a source identifier is generated, and the source identifier is associated with the corresponding traceability information.

[0019] Preferably, the standardized document object is parsed to extract a set of elements reflecting its content, source, and relationships, including:

[0020] The text content in the normalized document object is subjected to word segmentation, paragraph segmentation, and named entity recognition to extract text elements;

[0021] Perform text recognition and layout analysis on the images or scanned documents in the standardized document object to extract image elements;

[0022] The audio and video content in the standardized file object is transcribed into speech and segmented by speaker to extract audio and video elements;

[0023] The table content in the standardized document object is subjected to header recognition and cell parsing to extract table elements;

[0024] The text elements, image elements, audio / video elements, and table elements are represented in a unified manner, and a complete set of elements is generated by combining the file version number, reference relationship, and project identifier.

[0025] Preferably, the element set is mapped to a file semantic representation based on a preset semantic representation model, and consistency constraints are applied across different modalities of the same file to obtain a unified semantic representation for retrieval, including:

[0026] Load a preset semantic representation model; the semantic representation model includes branch components for processing text elements, image elements, audio and video elements, and table elements;

[0027] Determine the public semantic space and the list of alignment rules; the list of alignment rules includes three categories: temporal proximity, page proximity, and reference correspondence.

[0028] The text elements, image elements, audio and video elements and table elements are encoded respectively to obtain corresponding intermediate semantic representations, and each intermediate semantic representation retains a source identifier and at least one of the following additional information: timestamp, page position or row and column position.

[0029] Based on the alignment rule list, intermediate semantic representations that originate from the same file and correspond to each other in time, space or reference relationship are associated to form alignment groups, and consistency constraints are applied in each alignment group to ensure that the representation of the same semantic unit remains consistent in different modalities.

[0030] Group-level representations are generated for each alignment group and merged with the intermediate semantic representations that are not grouped to obtain the document semantic representation. Redundancy removal and scale normalization are then performed on the document semantic representation to form a unified semantic representation for retrieval.

[0031] The unified semantic representation is subjected to interpretability and stability verification. Once the verification is passed, an association registration is established with the source information of the corresponding file.

[0032] Preferably, based on the unified semantic representation and the relationships between source, version, and citation, a hierarchical index structure including a semantic index, a keyword index, and a relationship index is established, and a unique source identification is generated for each file. The unique source identification is then associated with and stored with the corresponding source information, including:

[0033] The hierarchical index structure is composed of semantic layer, keyword layer and relation layer. The primary key field is a unique traceability identifier. A cross-layer mapping table is established to record the correspondence of the primary key in each layer.

[0034] Generate a unique traceability identifier for each file; the unique traceability identifier includes the file source, generation time, and serial number.

[0035] Using a unified semantic representation as input, a vector index is established to form a semantic layer record, and a unique traceability identifier is used as the primary key.

[0036] Keywords and high-frequency feature terms are extracted from the unified semantic representation, an inverted index is built, a keyword layer record is formed, and a unique traceability identifier is used as the primary key;

[0037] Based on the source, version, and citation relationships, a relationship graph structure is established, with nodes using a unique traceability identifier as the primary key and edges labeled with the relationship type;

[0038] Records corresponding to the same file in the semantic layer, keyword layer, and relation layer are linked together in the cross-layer mapping table using a unique traceability identifier. The unique traceability identifier is then associated with the traceability information of the file.

[0039] Preferably, the process involves receiving a query request, performing intent recognition and condition parsing, generating a semantic query representation, structured filtering conditions, and relational constraints, performing semantic retrieval, keyword retrieval, and relational retrieval on the hierarchical index structure, and fusing and ranking candidate results according to the task intent to obtain the location result, including:

[0040] Receive query requests, standardize character sets, languages ​​and time expressions, and extract source preferences, time ranges, version preferences and project clues related to the query;

[0041] The query task intent is determined based on a preset intent recognition strategy, and the corresponding intent tag is output; the task intent includes one or more of the following: locating a specific version, limiting the source, limiting the time range, and associating with projects;

[0042] The query request is broken down into three types of retrieval inputs: semantic query representation, structured filtering conditions, and relational constraints. The structured filtering conditions include one of the following: source, time, and file type. The relational constraints include one of the following: version relationship, reference relationship, or belonging to a project.

[0043] Semantic retrieval, keyword retrieval, and relation retrieval are performed on the hierarchical index structure to obtain the corresponding candidate set and hit clues.

[0044] Based on the intent tags and preset fusion rules, each candidate set is deduplicated and merged to form a fusion candidate set, while retaining the hit clues for interpretation;

[0045] The results are ranked based on a combination of content relevance, time freshness, source credibility, and intent matching. A summary of the criteria used for ranking is then output.

[0046] Preferably, the location results are subjected to permission verification, and the verification results and source information are output. Based on user clicks, error corrections, and annotation feedback, the intent recognition strategy, fusion strategy, and ranking weights are updated online. Incremental reconstruction of error samples is triggered, provided that audit requirements are met.

[0047] Based on the access control policy, each location result is compared with the access subject, resource tag and action type to form a set of location results that have passed the permission matching.

[0048] Source tracing information is generated from the set of location results;

[0049] Collect user clicks, error corrections, and annotation feedback, and bind the feedback with a unique traceability identifier, query intent tag, hit clue, and sorting position to generate feedback samples;

[0050] Based on the feedback samples, the intent recognition strategy, candidate fusion strategy, and ranking weight are updated online, and the updated strategy version identifier is recorded.

[0051] Provided that the audit requirements are met, when the feedback sample reaches the error threshold and the coverage threshold, incremental reconstruction is performed on the unified semantic representation and hierarchical index of the corresponding target file set.

[0052] Preferably, the tracing information includes the source path, a summary of the parsing steps, index hit elements, a summary of the version and reference relationship, permission verification results, and a unique tracing identifier.

[0053] An AI-based file location management system includes:

[0054] The resource access and standardization unit is used to collect file resources and corresponding metadata from at least one file source, and complete format normalization and field standardization under a unified time base and identification rules to form a standardized file object and source identifier;

[0055] The content parsing and element extraction unit is used to parse the content of the standardized document object and extract a set of elements that reflect the content, source and relationship; the set of elements covers text fragments, image highlights, audio and video transcription fragments, table structure information and relationship information related to version, citation and project;

[0056] The semantic representation and alignment unit is used to map the set of elements into a file semantic representation based on a preset semantic representation model, and to apply consistency constraints between different modalities of the same file to obtain a unified semantic representation for retrieval.

[0057] The hierarchical index construction unit is used to establish a hierarchical index structure containing semantic index, keyword index and relationship index based on the unified semantic representation and source, version and reference relationship, and to generate a unique traceability identifier for each file and associate the unique traceability identifier with the corresponding traceability information for storage.

[0058] The query understanding and joint retrieval unit is used to receive query requests, perform intent recognition and condition parsing, generate semantic query representations, structured filtering conditions and relational constraints, perform semantic retrieval, keyword retrieval and relational retrieval on the hierarchical index structure respectively, and fuse and sort candidate results according to task intent to obtain positioning results;

[0059] The permission verification and online learning unit is used to verify the permissions of the positioning results, output the verification results and traceability information, and update the intent recognition strategy, fusion strategy and ranking weight online based on user clicks, error correction and annotation feedback, and trigger incremental reconstruction of error samples under the premise of meeting audit requirements.

[0060] The present invention discloses the following technical effects:

[0061] This invention achieves comprehensive coverage of document information by uniformly parsing multimodal content such as text, images, audio and video, and tables, and combining version, citation, and project relationship information to generate standardized document objects and element sets, effectively improving the completeness of content understanding and subsequent retrieval.

[0062] This invention employs a pre-defined semantic representation model and applies consistency constraints between different modalities of the same file to obtain a unified semantic representation, thereby ensuring the alignment of multimodal data in the same semantic space and improving the accuracy of file matching and location.

[0063] This invention introduces intent recognition and condition parsing mechanisms into query processing, decomposing query requests into three types of retrieval inputs: semantic representation, structured filtering conditions, and relational constraints. It also combines semantic indexes, keyword indexes, and relational indexes to perform joint retrieval, achieving intent-driven result localization and significantly improving the accuracy and flexibility of retrieval.

[0064] This invention establishes a hierarchical index structure that includes semantic index, keyword index, and relationship index, and generates a unique traceability identifier for each file, which is associated with and stored with the corresponding traceability information. This enables the retrieval results to trace and explain the source, version, and citation relationship, thereby enhancing the credibility and auditability of the results.

[0065] This invention introduces permission verification and online learning after outputting the location results. It can dynamically optimize the intent recognition strategy, fusion strategy and ranking weight based on user clicks, error correction and annotation feedback, and trigger incremental reconstruction of error samples under the premise of meeting audit requirements, thereby forming a continuous optimization mechanism to ensure the stable performance of the system in dynamic environments. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 A flowchart of the method provided in an embodiment of the present invention;

[0068] Figure 2 This is a schematic diagram of the system structure provided in an embodiment of the present invention. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] The purpose of this invention is to provide a file location management method and system based on artificial intelligence, which realizes the integrated processing of multimodal parsing, semantic alignment, hierarchical indexing, intent recognition and online learning for file location, and significantly improves the comprehensiveness, accuracy, traceability and adaptability of the location.

[0071] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0072] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1As shown, the present invention provides a file location management method based on artificial intelligence, including:

[0073] Step 100: Collect file resources and corresponding metadata from at least one file source, and complete format normalization and field standardization under a unified time base and identification rules to form a standardized file object and source identifier;

[0074] Step 200: Perform content parsing on the standardized document object to extract a set of elements that reflect the content, source, and relationships; the set of elements covers text fragments, image highlights, audio and video transcription fragments, table structure information, and relationship information related to version, citation, and project;

[0075] Step 300: Based on the preset semantic representation model, the feature set is mapped to the document semantic representation, and consistency constraints are applied between different modalities of the same document to obtain a unified semantic representation for retrieval;

[0076] Step 400: Based on the unified semantic representation and source, version and reference relationship, establish a hierarchical index structure including semantic index, keyword index and relationship index, generate a unique traceability identifier for each file, and store the unique traceability identifier and the corresponding traceability information together.

[0077] Step 500: Receive the query request, perform intent recognition and condition parsing, generate semantic query representation, structured filtering conditions and relational constraints, perform semantic retrieval, keyword retrieval and relational retrieval on the hierarchical index structure respectively, and fuse and sort the candidate results according to the task intent to obtain the location result;

[0078] Step 600: Perform permission verification on the location results, output the verification results and traceability information, and update the intent recognition strategy, fusion strategy and ranking weight online based on user clicks, error correction and annotation feedback, and trigger incremental reconstruction of error samples under the premise of meeting audit requirements.

[0079] In this embodiment, step 100 is used to access and standardize file resources. Specifically, the system first collects file resources and corresponding metadata from at least one file source, which may include local storage, shared storage, or collaborative office platforms. The collected files and metadata are checked by a format detection module to identify the file type and encoding method, and files from different sources are unified into a preset format paradigm, such as unified character set encoding, unified timestamp format, and unified field separation method. To ensure the consistency of time information, the system further normalizes the file creation time and modification time according to a unified time benchmark to avoid retrieval errors caused by time differences from different source devices.

[0080] Building upon the above, this embodiment maps the source-side metadata to standard fields on the system side according to a pre-defined field mapping table. For example, fields such as "author," "department," and "version number" from the source side are mapped to unified names recognized by the system, and any missing fields are filled in. After completing the field standardization, a corresponding standardized file object is generated. A "standardized file object" refers to a file data unit formed after format normalization and field standardization. This data unit maintains consistency and operability in subsequent parsing, indexing, and retrieval processes. Furthermore, the system generates a source identifier for each standardized file object. This source identifier is used to uniquely identify the file source category, collection time, and access channel, and is associated with the corresponding traceability information to enable subsequent traceable queries and audits.

[0081] In this embodiment, step 200 is used to parse the content of the normalized document object to extract a set of elements that reflect the content, source, and relationships. This embodiment first performs word segmentation, paragraph segmentation, and named entity recognition operations sequentially on the text content of the normalized document object to extract text elements. "Text elements" refer to semantic units such as keywords, topic sentences, proper nouns, and their contextual positions obtained from the text content, used to accurately represent the semantic features of the document in subsequent searches.

[0082] In this embodiment, non-text content in standardized document objects is processed using a differentiated parsing method. Specifically, text recognition and layout analysis are performed on images or scanned documents to extract image elements, including recognized text information, layout structure, and spatial location; speech-to-text transcription and speaker segmentation are performed on audio and video content to extract audio and video elements, including transcribed text fragments, speaker identifiers, and corresponding timestamps; and header recognition and cell parsing are performed on table content to extract table elements, including field names, numerical cells, and logical relationships between rows and columns.

[0083] After extracting the aforementioned elements, this embodiment unifies the representation of text, image, audio / video, and table elements, and combines this with the file's version number, referencing relationships, and project identifier to generate a complete element set. The "element set" refers to a data structure organized in a unified format, consisting of semantic units and relational information from different modalities. It includes both the file's content features and source and relational attributes, supporting subsequent semantic representation modeling and index construction. Through this approach, this embodiment achieves the fusion and parsing of multimodal information and file relational information, ensuring the integrity, accuracy, and structured representation of the file content in subsequent processing.

[0084] In this embodiment, step 300 is used to map the set of elements extracted in step 200 into a unified document semantic representation. Specifically, this embodiment first loads a preset semantic representation model, which includes branch components set for text elements, image elements, audio / video elements, and table elements, respectively, for specialized encoding processing of data of different modalities. To ensure that different modalities can be unified to the same measurement system in the future, this embodiment also determines a common semantic space and establishes an alignment rule list during model initialization. The so-called "common semantic space" refers to a unified vector space to which the encoding results of different modalities are mapped, so as to perform cross-modal comparison and fusion; while the "alignment rule list" includes three types of rules: temporal proximity, page proximity, and reference correspondence, used to determine the correspondence between data of different modalities.

[0085] Based on this, this embodiment encodes text elements, image elements, audio / video elements, and table elements respectively to obtain corresponding intermediate semantic representations. Each intermediate semantic representation is accompanied by a source identifier and retains at least one location or time sequence information, such as a timestamp, page position, or row / column position, to support subsequent cross-modal alignment. This embodiment further establishes associations between intermediate semantic representations originating from the same file and corresponding in time, space, or citation relationship according to an alignment rule list, forming alignment groups. An "alignment group" refers to a set of intermediate semantic representations from multiple modalities aggregated according to alignment rules, used to represent the correspondence of the same semantic unit in different modalities. This embodiment applies consistency constraints within each alignment group to ensure that the representation of the same semantic unit remains consistent across different modalities.

[0086] After alignment, this embodiment generates group-level representations for each aligned group and merges them with the intermediate semantic representations that are not grouped, resulting in a complete file semantic representation. Subsequently, redundancy removal and scale normalization are performed on the file semantic representation to eliminate redundant features and unify the representation scale, forming a unified semantic representation usable for retrieval. Finally, this embodiment performs interpretability and stability checks on this unified semantic representation. If successful, it is associated with the source information of the corresponding file and registered. "Source information" refers to structured information recording the file's origin, processing steps, version relationships, and permission verification results, used to support subsequent result tracing and auditing. Through the above process, step 300 achieves unified alignment and searchable expression of multimodal semantics, providing a high-quality semantic foundation for subsequent hierarchical index construction.

[0087] In this embodiment, step 400 is used to establish a hierarchical index structure based on a unified semantic representation and source, version, and citation relationships, and to complete the association with traceability information. This embodiment first determines the composition of the hierarchical index structure, which consists of a semantic layer, a keyword layer, and a relationship layer, and unifies the primary key field as a unique traceability identifier. Simultaneously, this embodiment establishes a cross-layer mapping table to record the correspondence between unique traceability identifiers in each layer of the index. The so-called "cross-layer mapping table" refers to a data structure used to maintain consistency between multi-layer indexes, ensuring the correspondence of the same file recorded in the semantic layer, keyword layer, and relationship layer.

[0088] Within the framework of a hierarchical index structure, this embodiment generates a unique traceability identifier for each file. A "unique traceability identifier" is a unique marker composed of the file's source, generation time, and sequence number. It is used to ensure that there are no conflicts between different files and to provide a stable primary key for subsequent traceability information association. This embodiment uses this unique traceability identifier as the core index field across all layers to achieve unified management of multi-layered data.

[0089] This embodiment further utilizes unified semantic representation to establish a vector index in the semantic layer, forming semantic layer records, and storing the unique source identifier as the primary key. Simultaneously, keywords and high-frequency feature terms are extracted from the unified semantic representation to construct an inverted index, forming keyword layer records, also using the unique source identifier as the primary key. In the relationship layer, a relationship graph structure is constructed based on the file's source, version, and citation relationships. The "relationship graph structure" refers to a structure where unique source identifiers are used as nodes, and edges between nodes are labeled according to relationship types such as "version inheritance," "citation relationship," or "project affiliation," thereby reflecting the association and evolution path between files.

[0090] After constructing the semantic layer, keyword layer, and relationship layer, this embodiment establishes a one-to-one correspondence between records of the same file in the three layers through a cross-layer mapping table, ensuring the relevance and traceability of files in the multi-layer index. Finally, this embodiment establishes a relationship between the unique traceability identifier and the file's traceability information. "Traceability information" refers to structured information that records the file's source path, parsing step summary, hit elements, version and citation relationships, and permission verification results, used to provide interpretable and auditable support when outputting search results. Through this process, step 400 realizes the establishment of the hierarchical index and cross-layer association, enabling files to be quickly located using semantic features and accurately located using relationship information in subsequent searches.

[0091] In this embodiment, step 500 is used to receive the user's query request and convert it into multiple types of retrieval input that can be executed on a hierarchical index structure. This embodiment first receives the query request and performs normalization processing on the character set, language, and time representation to ensure that the input query content can be uniformly recognized and parsed. Simultaneously, during the normalization stage, the system extracts information related to the query, such as source preferences, time ranges, version preferences, and project clues, to assist in subsequent condition parsing.

[0092] After standardization, this embodiment identifies the query request based on a preset intent recognition strategy, determines the query's task intent, and outputs the corresponding intent tag. An "intent tag" refers to an identifier used to represent the user's true needs during query processing, such as "targeting a specific version," "limiting the source," "limiting the time range," or "associating with related items." By setting intent tags, the system can accurately match the query target during subsequent retrieval and sorting.

[0093] After intent recognition is completed, this embodiment decomposes the query request into three types of retrieval inputs: semantic query representation, structured filtering conditions, and relational constraints. The semantic query representation is used for comparison with the semantic index; the structured filtering conditions are used to filter by source, time, or file type; and the relational constraints are used to limit version relationships, reference relationships, or project affiliations between files. Through the generation of these three types of retrieval inputs, this embodiment can simultaneously perform semantic retrieval, keyword retrieval, and relational retrieval on a hierarchical index structure, obtaining the corresponding candidate set and hit clues.

[0094] After obtaining the candidate set, this embodiment processes the candidate results based on intent tags and preset fusion rules. Specifically, this embodiment first deduplicates and merges the candidate sets obtained from different search methods to form a fused candidate set. The so-called "fused candidate set" refers to the unified result set obtained after deduplication and merging of the results output by multiple search methods, while retaining the corresponding hit clues for each result for interpretation and auditing. Finally, this embodiment comprehensively sorts the results according to content relevance, time freshness, source credibility, and intent matching to obtain the final positioning result, and outputs a summary of the sorting criteria so that users can understand the rationality of the result sorting.

[0095] In this embodiment, step 600 is used to perform permission verification after outputting the location results, and to complete the generation of traceability information and online learning while ensuring compliance. This embodiment first compares each location result against the access subject, resource tag, and action type according to the access control policy, thereby filtering out the set of location results that pass the permission matching. The access subject refers to the user or system account that submitted the query, the resource tag is used to characterize the access level of the file, and the action type is used to specify the executable operation (such as reading, downloading, or modifying). Through the matching of these three, this embodiment can ensure that the output results are limited to files that conform to the permission rules.

[0096] After obtaining the set of verified location results, this embodiment generates traceability information for this set. "Traceability information" refers to a structured description recording the entire lifecycle information of a file, including the source path, a summary of parsing steps, index hit elements, a summary of version and reference relationships, permission verification results, and a unique traceability identifier. The unique traceability identifier is a unique number corresponding to each file, used to maintain consistency across different stages. By generating traceability information, this embodiment not only ensures the interpretability of the results but also provides evidentiary support for subsequent compliance audits.

[0097] After outputting the source information, this embodiment further collects user interaction feedback, including clicks, corrections, and annotations. This feedback is then bound to a corresponding unique source identifier, query intent tag, hit clue, and ranking position to form feedback samples. A "feedback sample" refers to a data unit composed of the user's actual operational behavior and the corresponding search results, used to guide system optimization. In this way, this embodiment can accurately capture user preferences and correction information during use, thereby providing a reliable basis for continuous model optimization.

[0098] Based on the feedback samples, this embodiment performs online updates to the intent recognition strategy, candidate fusion strategy, and ranking weights, and records the updated strategy version identifier. When the accumulated feedback samples reach the preset error threshold and coverage threshold, and meet the audit requirements, this embodiment triggers incremental reconstruction of the unified semantic representation and hierarchical index of the corresponding target file set. "Incremental reconstruction" refers to only partially updating the file set appearing in the error set, without rebuilding the entire index, thereby improving update efficiency and reducing system resource consumption. Through this mechanism, this embodiment can achieve adaptive optimization of retrieval performance while ensuring traceability and compliance.

[0099] Corresponding to the above methods, such as Figure 2 As shown, this embodiment also provides an artificial intelligence-based file location management system, including:

[0100] The resource access and standardization unit is used to collect file resources and corresponding metadata from at least one file source, and complete format normalization and field standardization under a unified time base and identification rules to form a standardized file object and source identifier;

[0101] The content parsing and element extraction unit is used to parse the content of the standardized document object and extract a set of elements that reflect the content, source and relationship; the set of elements covers text fragments, image highlights, audio and video transcription fragments, table structure information and relationship information related to version, citation and project;

[0102] The semantic representation and alignment unit is used to map the set of elements into a file semantic representation based on a preset semantic representation model, and to apply consistency constraints between different modalities of the same file to obtain a unified semantic representation for retrieval.

[0103] The hierarchical index construction unit is used to establish a hierarchical index structure containing semantic index, keyword index and relationship index based on the unified semantic representation and source, version and reference relationship, and to generate a unique traceability identifier for each file and associate the unique traceability identifier with the corresponding traceability information for storage.

[0104] The query understanding and joint retrieval unit is used to receive query requests, perform intent recognition and condition parsing, generate semantic query representations, structured filtering conditions and relational constraints, perform semantic retrieval, keyword retrieval and relational retrieval on the hierarchical index structure respectively, and fuse and sort candidate results according to task intent to obtain positioning results;

[0105] The permission verification and online learning unit is used to verify the permissions of the positioning results, output the verification results and traceability information, and update the intent recognition strategy, fusion strategy and ranking weight online based on user clicks, error correction and annotation feedback, and trigger incremental reconstruction of error samples under the premise of meeting audit requirements.

[0106] The beneficial effects of this invention are as follows:

[0107] (1) By unifying the time base and identification rules in the file access stage, and combining format normalization and field standardization, the present invention forms a standardized file object and source identifier, which solves the problems of inconsistent file formats, inconsistent timestamps and redundant fields in the prior art, so that subsequent processing can be carried out under a unified standard, ensuring the consistency and operability of file management.

[0108] (2) The present invention introduces a multimodal processing mechanism in the content parsing stage, which can simultaneously parse and extract elements from text, images, audio and video and table content, and generate a complete set of elements by combining relationship information such as version, reference and project, effectively overcoming the defect that the traditional method can only process a single modality and thus the search scope is limited, thereby achieving comprehensive coverage of the semantic and structural information of the document.

[0109] (3) This invention maps multimodal elements to a unified semantic representation through a pre-defined semantic representation model and cross-modal consistency constraints, ensuring that the feature representation of the same file remains consistent across different modalities. This mechanism not only improves the accuracy of semantic expression but also enhances the stability of cross-modal retrieval and comparison, solving the problem of difficulty in aligning multimodal information in existing technologies.

[0110] (4) The present invention adopts a hierarchical approach in the design of the index structure, establishing semantic index, keyword index and relation index respectively, and cross-layer association through unique traceability identifier, so that the retrieval can not only quickly achieve high recall rate by semantic matching, but also reveal the version inheritance and reference path between files by relying on the relation graph structure, so as to realize the traceability and interpretability of the location results, breaking through the limitations of the lack of relevance and verifiability of the results in the prior art.

[0111] (5) This invention introduces intent recognition and condition parsing in the query processing stage, decomposing the query request into three types of input: semantic query, structured conditions, and relational constraints. Combined with a hierarchical index, a joint retrieval is performed, and then the results are fused and sorted according to intent tags, achieving output based on the user's actual needs. This intent-driven retrieval method significantly improves the accuracy and flexibility of the location results, avoiding the problems of redundant results or deviation from user needs in traditional methods.

[0112] (6) The present invention introduces permission verification, traceability information generation and online learning mechanisms in the result output stage, which not only ensures the compliance and auditability of the location results, but also continuously optimizes the intent recognition strategy, fusion strategy and ranking weight based on user clicks, error correction and annotation feedback, and triggers incremental reconstruction when the threshold is reached, thereby forming a dynamically evolving and adaptively optimized file management system, ensuring the efficiency and stability of the system in long-term operation.

[0113] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0114] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. An artificial intelligence-based file positioning management method, characterized by, include: File resources and corresponding metadata are collected from at least one file source, and format normalization and field standardization are completed under a unified time base and identification rules to form standardized file objects and source identifiers; The standardized document object is parsed to extract a set of elements that reflect the content, source, and relationships. The set of elements covers text fragments, image highlights, audio and video transcription fragments, table structure information, and relationship information related to version, citation, and project. Based on a preset semantic representation model, the set of elements is mapped to a document semantic representation, and consistency constraints are applied between different modalities of the same document to obtain a unified semantic representation for retrieval. Based on the unified semantic representation and the relationship between source, version and citation, a hierarchical index structure including semantic index, keyword index and relationship index is established, and a unique traceability identifier is generated for each file, and the unique traceability identifier is associated with and stored with the corresponding traceability information; Upon receiving a query request, the system performs intent recognition and condition parsing, generates a semantic query representation, structured filtering conditions, and relational constraints, performs semantic retrieval, keyword retrieval, and relational retrieval on the hierarchical index structure, and fuses and sorts the candidate results according to the task intent to obtain the location result. The location results are verified for permissions, and the results of the verification and the source information are output. Based on user clicks, error correction and annotation feedback, the intent recognition strategy, fusion strategy and ranking weight are updated online, and incremental reconstruction of error samples is triggered under the premise of meeting audit requirements. 2.The artificial intelligence-based file positioning management method of claim 1, wherein, File resources and corresponding metadata are collected from at least one file source. Under a unified time base and identification rules, format normalization and field standardization are completed to form standardized file objects and source identifiers, including: The system performs format detection on the accessed files and metadata, and unifies files from different sources into a preset format paradigm. Based on a unified time base and identification rules, the creation time, modification time, and source information of the files are normalized. The source-side metadata is mapped to system-side standard fields according to the field mapping table to form the normalized file object; For each of the standardized document objects, a source identifier is generated, and the source identifier is associated with the corresponding traceability information. 3.The AI-based file positioning management method of claim 1, wherein, The standardized document object is parsed to extract a set of elements reflecting its content, source, and relationships, including: The text content in the normalized document object is subjected to word segmentation, paragraph segmentation, and named entity recognition to extract text elements; Perform text recognition and layout analysis on the images or scanned documents in the standardized document object to extract image elements; The audio and video content in the standardized file object is transcribed into speech and segmented by speaker to extract audio and video elements; The table content in the standardized document object is subjected to header recognition and cell parsing to extract table elements; The text elements, image elements, audio / video elements, and table elements are represented in a unified manner, and a complete set of elements is generated by combining the file version number, reference relationship, and project identifier. 4.The AI-based file positioning management method of claim 3, wherein, Based on a pre-defined semantic representation model, the element set is mapped to a file semantic representation, and consistency constraints are applied across different modalities of the same file to obtain a unified semantic representation for retrieval, including: Load a preset semantic representation model; the semantic representation model includes branch components for processing text elements, image elements, audio and video elements, and table elements; Determine the public semantic space and the list of alignment rules; the list of alignment rules includes three categories: temporal proximity, page proximity, and reference correspondence. The text elements, image elements, audio and video elements and table elements are encoded respectively to obtain corresponding intermediate semantic representations, and each intermediate semantic representation retains a source identifier and at least one of the following additional information: timestamp, page position or row and column position. Based on the alignment rule list, intermediate semantic representations that originate from the same file and correspond to each other in time, space or reference relationship are associated to form alignment groups, and consistency constraints are applied in each alignment group to ensure that the representation of the same semantic unit remains consistent in different modalities. Group-level representations are generated for each alignment group and merged with the intermediate semantic representations that are not grouped to obtain the document semantic representation. Redundancy removal and scale normalization are then performed on the document semantic representation to form a unified semantic representation for retrieval. The unified semantic representation is subjected to interpretability and stability verification. Once the verification is passed, an association registration is established with the source information of the corresponding file. 5.The artificial intelligence-based file positioning management method of claim 1, wherein, Based on the unified semantic representation and the relationships between source, version, and citation, a hierarchical index structure including semantic index, keyword index, and relationship index is established. A unique source identifier is generated for each file, and the unique source identifier is associated with and stored with the corresponding source information, including: The hierarchical index structure is composed of semantic layer, keyword layer and relation layer. The primary key field is a unique traceability identifier. A cross-layer mapping table is established to record the correspondence of the primary key in each layer. Generate a unique traceability identifier for each file; the unique traceability identifier includes the file source, generation time, and serial number. Using a unified semantic representation as input, a vector index is established to form a semantic layer record, and a unique traceability identifier is used as the primary key. Keywords and high-frequency feature terms are extracted from the unified semantic representation, an inverted index is built, a keyword layer record is formed, and a unique traceability identifier is used as the primary key; Based on the source, version, and citation relationships, a relationship graph structure is established, with nodes using a unique traceability identifier as the primary key and edges labeled with the relationship type; Records corresponding to the same file in the semantic layer, keyword layer, and relation layer are linked together in the cross-layer mapping table using a unique traceability identifier. The unique traceability identifier is then associated with the traceability information of the file. 6.The artificial intelligence-based file positioning management method of claim 1, wherein, Upon receiving a query request, the system performs intent recognition and condition parsing, generating a semantic query representation, structured filtering conditions, and relational constraints. Semantic retrieval, keyword retrieval, and relational retrieval are then performed on the hierarchical index structure. Finally, candidate results are fused and sorted according to the task intent to obtain the location results, including: Receive query requests, standardize character sets, languages ​​and time expressions, and extract source preferences, time ranges, version preferences and project clues related to the query; The query task intent is determined based on a preset intent recognition strategy, and the corresponding intent tag is output; the task intent includes one or more of the following: locating a specific version, limiting the source, limiting the time range, and associating with projects; The query request is broken down into three types of retrieval inputs: semantic query representation, structured filtering conditions, and relational constraints. The structured filtering conditions include one of the following: source, time, and file type. The relational constraints include one of the following: version relationship, reference relationship, or belonging to a project. Semantic retrieval, keyword retrieval, and relation retrieval are performed on the hierarchical index structure to obtain the corresponding candidate set and hit clues. Based on the intent tags and preset fusion rules, each candidate set is deduplicated and merged to form a fusion candidate set, while retaining the hit clues for interpretation; The results are ranked based on a combination of content relevance, time freshness, source credibility, and intent matching. A summary of the criteria used for ranking is then output. 7.The AI-based file positioning management method of claim 1, wherein, The location results are verified for permissions, and the results of successful verification and source tracing information are output. Based on user clicks, error corrections, and annotation feedback, the intent recognition strategy, fusion strategy, and ranking weights are updated online. Incremental reconstruction of error samples is triggered, provided that audit requirements are met, including: Based on the access control policy, each location result is compared with the access subject, resource tag and action type to form a set of location results that have passed the permission matching. Source tracing information is generated from the set of location results; Collect user clicks, error corrections, and annotation feedback, and bind the feedback with a unique traceability identifier, query intent tag, hit clue, and sorting position to generate feedback samples; Based on the feedback samples, the intent recognition strategy, candidate fusion strategy, and ranking weight are updated online, and the updated strategy version identifier is recorded. Provided that the audit requirements are met, when the feedback sample reaches the error threshold and the coverage threshold, incremental reconstruction is performed on the unified semantic representation and hierarchical index of the corresponding target file set. 8.The AI-based file positioning management method of claim 1, wherein, The tracing information includes the source path, a summary of the parsing steps, index hit elements, a summary of the version and reference relationship, permission verification results, and a unique tracing identifier.

9. An artificial intelligence-based file positioning management system, characterized by, include: The resource access and standardization unit is used to collect file resources and corresponding metadata from at least one file source, and complete format normalization and field standardization under a unified time base and identification rules to form a standardized file object and source identifier; The content parsing and element extraction unit is used to parse the content of the standardized document object and extract a set of elements that reflect the content, source and relationship; the set of elements covers text fragments, image highlights, audio and video transcription fragments, table structure information and relationship information related to version, citation and project; The semantic representation and alignment unit is used to map the set of elements into a file semantic representation based on a preset semantic representation model, and to apply consistency constraints between different modalities of the same file to obtain a unified semantic representation for retrieval. The hierarchical index construction unit is used to establish a hierarchical index structure containing semantic index, keyword index and relationship index based on the unified semantic representation and source, version and reference relationship, and to generate a unique traceability identifier for each file and associate the unique traceability identifier with the corresponding traceability information for storage. The query understanding and joint retrieval unit is used to receive query requests, perform intent recognition and condition parsing, generate semantic query representations, structured filtering conditions and relational constraints, perform semantic retrieval, keyword retrieval and relational retrieval on the hierarchical index structure respectively, and fuse and sort candidate results according to task intent to obtain positioning results; The permission verification and online learning unit is used to verify the permissions of the positioning results, output the verification results and traceability information, and update the intent recognition strategy, fusion strategy and ranking weight online based on user clicks, error correction and annotation feedback, and trigger incremental reconstruction of error samples under the premise of meeting audit requirements.

Images