Process data retrieval optimization method, system and equipment based on large model and medium

Abstract

The invention discloses a process data retrieval optimization method, system and device based on a large model and a medium, belongs to the technical field of artificial intelligence and enterprise data management crossing, and aims at solving the technical problem of how to overcome the defects of low semantic matching precision and rigid slicing strategy in enterprise process data retrieval in the prior art. According to the technical scheme, the method comprises the steps of data preprocessing and multi-granularity semantic feature construction, specifically, enterprise process data are obtained, multi-granularity slicing preprocessing operation is conducted on the process data, and multi-granularity semantic features are constructed; mixed retrieval: adopting a vector retrieval and keyword retrieval combined double-layer recall mechanism to obtain an initial retrieval result; performing large model enhanced reordering: performing context sensing reordering on the initial result through the large model, and combining position attention and diversity filtering to generate a final retrieval set; and multi-mode query interaction: constructing a query interaction system, and supporting convenient calling of process data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and enterprise data management, in particular to a process data retrieval optimization method, system, device and medium based on a large model. BACKGROUND

[0002] Process data retrieval is the core support of enterprise digital operation, and its ability to accurately extract target information from structured forms and unstructured documents directly affects business decision-making efficiency. The retrieval augmentation generation (RAG) technology driven by large models is the current mainstream solution, but for large and medium-sized enterprises with many subordinate units and complex processes, due to the "format is mixed, terms are many, and updates are fast", the existing technology has the following core deficiencies:

[0003] (1) Semantic and structural adaptation double absence: ① On the semantic level, traditional keyword retrieval cannot identify synonymous terms such as "work order suspension / pause", and general RAG models have an accuracy rate of less than 60% in matching professional terms such as "abnormal process jump rules", forming a "semantic gap"; ② On the structural level, fixed-size blocks such as 512 tokens easily split associated logic such as "supplier evaluation-contract signing", or introduce redundant content beyond the large model input window, disrupting the integrity of the process;

[0004] (2) Low efficiency of result sorting and efficiency: relying on fixed algorithms such as cosine similarity for sorting, which cannot identify business priorities, seriously affecting business response efficiency and failing to meet the needs of enterprise rapid decision-making.

[0005] Therefore, how to overcome the defects of low semantic matching precision and rigid slicing strategy in existing enterprise process data retrieval technology, and improve the precision, efficiency and business adaptability of process data retrieval is a technical problem that needs to be solved at present. SUMMARY

[0006] The technical task of the present application is to provide a process data retrieval optimization method, system, device and medium based on a large model to solve the problem of how to overcome the defects of low semantic matching precision and rigid slicing strategy in existing enterprise process data retrieval technology, and improve the precision, efficiency and business adaptability of process data retrieval.

[0007] The technical task of the present application is achieved in the following way: a process data retrieval optimization method based on a large model, which is specifically as follows:

[0008] Data preprocessing and multi-granularity semantic feature construction: obtaining enterprise process data, performing multi-granularity slicing preprocessing operations on the process data, and constructing multi-granularity semantic features;

[0009] Hybrid retrieval: using a two-layer recall mechanism combining vector retrieval and keyword retrieval to obtain initial retrieval results;

[0010] Large model enhanced re-ranking: Through a large model, the initial results are context-awarely re-ranked, combining position attention and diversity filtering to generate the final retrieval set;

[0011] Multi-mode query interaction: Build a flexible and efficient query interaction system to support convenient calling of process data and adapt to the query needs of employees in different positions of the enterprise.

[0012] As preferred, data preprocessing and multi-granularity semantic feature construction are as follows:

[0013] Dynamic term-aware Embedding: For the process approval terms of "process definition ID" and "purchase approval link", fine-tune the BERT-flow model with full approval corpus to build a dynamic term vector library. And for the new term of "cross-level approver", quickly update the vector through small sample learning to narrow the semantic gap between query and data, improve the accuracy of term matching, and introduce a term semantic similarity calculation model to balance the accuracy of semantic matching and literal matching. The formula of the term semantic similarity calculation model is as follows:

[0014]

[0015] Where t represents the term vector; q represents the query vector, and a represents the weight coefficient (0.7 is taken after validation set optimization);

[0016] Multi-granularity slicing: Adopt a "hierarchical blocking + dynamic adjustment" mode to balance logical integrity and redundancy control, and adapt to the hierarchical logic of process approval data. Specifically: coarse-grained segmentation: according to the document structure and "process definition→approval configuration→personnel assignment→submission specification" logic, form a complete macro unit; fine-grained segmentation: divide long paragraphs into 300-500 word semantic windows, combine syntax analysis to avoid breaking the "approval configuration-permission binding" key logic, focus on local rules, and determine the optimal slicing size based on slicing quality evaluation index Q to realize dynamic switching of "full-link query prefers coarse-grained, link rule query prefers fine-grained"; Where Q = β × R + (1-β) × C; Where R represents the recall rate of slice retrieval; C represents the context integrity of slice, which is obtained by cross-slice semantic coherence score; β is the balance coefficient, the value range is 0.4-0.6 (0.4 for process overview query, 0.6 for process step query);

[0017] Dynamic switching: Determine the optimal size through the validation dataset, and output coarse-grained for "full-link query" and fine-grained for "link rule query" preferentially.

[0018] As preferred, hybrid retrieval is as follows:

[0019] A two-layer collaborative recall mechanism: The vector retrieval layer uses an improved HNSW index and combines HyDE technology to generate hypothetical answer vectors (such as generating an extended vector containing dimensions such as "work order status" when querying "timeout work order processing") to improve the recall rate of short queries; the keyword retrieval layer supplements the precise matching capability of structured fields such as "work order number" and "approver ID" through regular expressions and fuzzy matching, making up for the shortcomings of vector retrieval, and merging the results of the vector retrieval layer and the keyword retrieval layer to form the initial search results;

[0020] Cross-modal data fusion: Extract the "process node-responsible person-time limit" entity relationship triples from unstructured documents and map them uniformly to the semantic space with structured form data, supporting cross-modal queries that associate structured fields with document descriptions using natural language.

[0021] As a preferred option, the large model enhancement reordering is as follows:

[0022] Position attention-guided sorting: Based on the large model of Llama-2, the search results are re-sorted and the position prompt "the Xth item contains key process information" is added. By leveraging the large model's attention preference for the beginning and end of the text, high-value documents (such as process operation details and approval node definitions) are prioritized to enhance the efficiency of key information delivery.

[0023] Diversity Intelligent Filtering: The Diversity Ranker tool is introduced, using a dual threshold system of cosine similarity (filtering duplicate content) and entity coverage (filtering low-relevance content) to eliminate redundant or irrelevant documents. This ensures that the context set of the input model not only matches the query requirements but also possesses information diversity, reducing the risk of "illusion" generation. The formula is as follows:

[0024]

[0025] Where n represents the initial number of search results; d i Let represent the i-th document vector; q represent the average similarity between the document and the query; w represent the domain experience threshold; γ is set to 0.6; when the document similarity is lower than T or the diversity score D is lower than 0.3, it is discarded to ensure that the context set input to the large model has both relevance and diversity.

[0026] Furthermore, this method also supports multi-mode query interaction. By constructing a flexible and efficient query interaction system, it supports convenient access to process data and adapts to the query habits of personnel in different positions within the enterprise, as detailed below:

[0027] Dual-entry query: Provides both natural language and structured query entry points. Business personnel can directly obtain results through a conversational query such as "Which procurement approval processes have timed out?", while operations or data personnel can accurately filter through structured conditions such as "Process definition ID number and approver name". The system automatically identifies the query type and matches the optimal retrieval strategy.

[0028] Intelligent assistance: Integrates query intent recognition and historical record management, that is, it automatically recommends related query terms ("overdue work order processing flow" "work order overdue responsibility definition") for fuzzy queries (such as "overdue work order"); and saves employees' high-frequency query records, supports one-click re-retrieval and result export (Excel / PDF format), and improves the efficiency of repeated queries.

[0029] A process data retrieval optimization system based on a large model, the system comprising:

[0030] The data preprocessing module is used to acquire and preprocess enterprise process data and build a dynamic terminology vector library.

[0031] The hybrid retrieval module is used to obtain initial retrieval results using a hybrid retrieval mechanism, realizing collaborative recall of vector retrieval and keyword retrieval, as well as cross-modal data fusion.

[0032] The intelligent re-ranking module is used to optimize the initial results by enhancing re-ranking through a large model, and obtain the final retrieval set that meets business needs. It achieves positional attention ranking and diversity filtering through a large model.

[0033] The interactive interface module provides dual query entry points and intelligent auxiliary functions.

[0034] Preferably, the data preprocessing module includes a multi-granularity slicing submodule and a dynamic term vector library construction submodule;

[0035] The multi-granularity slicing submodule is used to perform coarse-grained segmentation according to the business logic of "process definition → approval configuration → personnel assignment → submission specifications", and to perform fine-grained segmentation on long paragraphs with a semantic window of 300-500 words, combined with syntactic analysis to avoid the break of key logic; then, the optimal slice size is determined based on the slice quality evaluation index Q, realizing the dynamic switching of "full-link query priority coarse-grained and link detail query priority fine-grained", where Q=β×R+(1-β)×C, R is the slice retrieval recall rate, C is the context integrity, and β ranges from 0.4 to 0.6;

[0036] The dynamic terminology vector library construction submodule is used to pre-train models based on BERT-flow, and fine-tunes them using terminology corpora from the process domains of "process definition ID" and "procurement approval process". It also introduces a terminology semantic similarity calculation model to balance matching accuracy. The weight coefficient α in the terminology semantic similarity calculation is optimized to 0.7 based on the validation set. At the same time, new terms for "cross-level approvers" are added, and the vector representation is quickly updated through a few-shot learning mechanism.

[0037] More preferably, the hybrid retrieval mechanism is a two-layer collaborative recall mechanism, specifically: the vector retrieval layer uses an improved HNSW index and combines HyDE technology to generate hypothetical answer vectors to expand the semantics of short queries; the keyword retrieval layer supplements the precise matching capability of the structured fields "work order number" and "approver ID" through regular expression matching and fuzzy search, and the results of the two are merged to form the initial retrieval results;

[0038] The intelligent re-ranking module adds a "item X contains key process information" type location prompt when the initial result is input from Llama-2 to prioritize high-value documents; at the same time, the intelligent re-ranking module introduces the Diversity Ranker tool, which uses cosine similarity and entity coverage as dual thresholds to filter documents. Documents with relevance below the threshold T or diversity score D below 0.3 are removed, where γ is 0.6.

[0039] The interactive interface module provides dual query entry points for natural language and structured queries, automatically identifies query types and matches retrieval strategies; it also integrates query intent recognition functionality, automatically recommending related query terms for fuzzy queries; and saves high-frequency query records and supports exporting results in Excel or PDF format.

[0040] An electronic device includes: a memory and at least one processor;

[0041] The memory stores computer-executed instructions;

[0042] The at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to perform the process data retrieval optimization method based on the large model as described above.

[0043] A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the process data retrieval optimization method based on a large model as described above.

[0044] The process data retrieval optimization method, system, device, and medium based on a large model of the present invention have the following advantages:

[0045] (I) This invention implements multi-granularity slicing of process data, constructing a hierarchical structure of coarse-grained logical blocks and fine-grained semantic blocks; based on domain terminology, it fine-tunes the embedding model, generates a dynamic terminology vector library, and maps it to structured data fields into a unified semantic space; it adopts a two-layer recall mechanism combining vector retrieval and keyword retrieval to obtain initial retrieval results; through a large model, it performs context-aware reordering of the initial results, and combines positional attention and diversity filtering to generate the final retrieval set, solving the problems of low semantic matching accuracy and rigid slicing strategy in enterprise process data retrieval, improving the accuracy and efficiency of process data retrieval, and adapting to the multi-scenario process data management needs of enterprises;

[0046] (II) This invention achieves a comprehensive improvement in retrieval accuracy, efficiency and business adaptability through design such as term-aware Embedding fine-tuning, multi-granularity dynamic slicing and large-model enhanced reordering, providing support for efficient enterprise operation;

[0047] (III) This invention improves the retrieval accuracy of dynamic terms and cross-modal data by using term-aware Embedding fine-tuning and hybrid retrieval, thereby achieving high-precision semantic matching;

[0048] (iv) This invention uses an adaptive slicing strategy to dynamically adjust the granularity of text blocks based on business scenarios, balancing information integrity and noise control.

[0049] (V) This invention utilizes the contextual understanding capabilities of large models and combines a positional attention mechanism to optimize the ranking of search results, ensuring that key information is input first and achieving intelligent result re-ranking.

[0050] (vi) This invention constructs a unified framework for processing structured and unstructured process data, supports multimodal retrieval, and adapts to complex business queries;

[0051] (vii) Through the synergistic effect of terminology matching optimization, adaptive slicing strategy and intelligent result sorting, this invention improves the accuracy of professional terminology matching to over 92%, improves the overall retrieval effect by 32%, reduces the average query response time from 4.2 seconds to 2.5 seconds, increases the factual accuracy of large model answers by 28%, and reduces the occurrence rate of "illusion" problems such as missing process steps to 5%. It can be seamlessly adapted to multiple business scenarios such as procurement and approval, and takes into account both query convenience and data security, providing reliable support for efficient enterprise operation. Attached Figure Description

[0052] The invention will be further described below with reference to the accompanying drawings.

[0053] Appendix Figure 1 This is a schematic diagram of the structure of a process data retrieval and optimization system based on a large model. Detailed Implementation

[0054] The following detailed description of the process data retrieval optimization method, system, device, and medium based on a large model of the present invention is provided with reference to the accompanying drawings and specific embodiments.

[0055] Example 1:

[0056] This embodiment provides a process data retrieval optimization method based on a large model, as detailed below:

[0057] S1. Data Preprocessing and Multi-granularity Semantic Feature Construction: Obtain enterprise process data, perform multi-granularity slicing preprocessing on the process data, and construct multi-granularity semantic features;

[0058] S2. Hybrid retrieval: A two-layer recall mechanism combining vector retrieval and keyword retrieval is used to obtain initial search results;

[0059] S3, Large Model Enhanced Re-ranking: The initial results are re-ranked in a context-aware manner using a large model, and the final retrieval set is generated by combining positional attention and diversity filtering.

[0060] S4. Multi-mode query interaction: Construct a flexible and efficient query interaction system to support convenient access to process data and adapt to the query needs of employees in different positions within the enterprise.

[0061] S5 features multi-mode query interaction, which supports convenient access to process data and adapts to the query habits of employees in different positions within the enterprise by building a flexible and efficient query interaction system.

[0062] The data preprocessing and multi-granularity semantic feature construction in step S1 of this embodiment are as follows:

[0063] S101, Dynamic Terminology Awareness Embedding: For process approval terms such as "Process Definition ID" and "Procurement Approval Process," the BERT-flow model is fine-tuned using the full approval corpus to construct a dynamic terminology vector library. New terms are added for "cross-level approvers," and the vectors are quickly updated through small-sample learning to narrow the semantic gap between queries and data, improving terminology matching accuracy. A terminology semantic similarity calculation model is introduced to balance the accuracy of semantic matching and literal matching. The formula for the terminology semantic similarity calculation model is as follows:

[0064]

[0065] Where t represents the term vector; q represents the query vector; and α represents the weight coefficient (optimized to 0.7 based on the validation set).

[0066] S102, Multi-granularity slicing: Employing a "hierarchical segmentation + dynamic adjustment" model to balance logical integrity and redundancy control, adapting to the hierarchical logic of process approval data; specifically: Coarse-grained segmentation: Slicing according to document structure and the logic of "process definition → approval configuration → personnel assignment → submission specifications" to form complete macro-units; Fine-grained segmentation: Dividing long paragraphs into 300-500 character semantic windows, combining syntactic analysis to avoid breakage of the key logic of "approval configuration - permission binding," focusing on local details, determining the optimal slice size based on the slice quality evaluation index Q, and achieving dynamic switching between "full-link query prioritizing coarse-grained and process detail query prioritizing fine-grained"; where Q=β×R+(1-β)×C; where R represents the slice retrieval recall rate; C represents the slice context integrity, which is obtained through cross-slice semantic coherence scoring; β is the balance coefficient, ranging from 0.4 to 0.6 (0.4 for process overview query and 0.6 for process step query);

[0067] S103, Dynamic Switching: Determine the optimal size by verifying the dataset. "Full-link query" prioritizes outputting coarse-grained data, while "detailed process query" prioritizes outputting fine-grained data.

[0068] The hybrid retrieval in step S2 of this embodiment is as follows:

[0069] S201, Two-layer collaborative recall mechanism: The vector retrieval layer adopts an improved HNSW index and combines HyDE technology to generate hypothetical answer vectors (such as generating an extended vector containing dimensions such as "work order status" when querying "timeout work order processing"), thereby improving the recall rate of short queries; the keyword retrieval layer supplements the precise matching capability of structured fields such as "work order number" and "approver ID" through regular expressions and fuzzy matching, making up for the shortcomings of vector retrieval, and merging the results of the vector retrieval layer and the keyword retrieval layer to form the initial retrieval results;

[0070] S202, Cross-modal data fusion: Extract the "process node-responsible person-time limit" entity relationship triples from unstructured documents, and map them uniformly to the semantic space with structured form data, supporting cross-modal queries that associate natural language with structured fields and document descriptions.

[0071] The large model enhancement and reordering in step S3 of this embodiment is as follows:

[0072] S301, Position Attention-Guided Ranking: Based on the large model of Llama-2, the search results are re-ranked and the position prompt "the Xth item contains key process information" is added. By utilizing the large model's attention preference for the beginning and end of the text, high-value documents (such as process operation details and approval node definitions) are prioritized to enhance the efficiency of key information delivery.

[0073] S302, Diversity Intelligent Filtering: Introducing the Diversity Ranker tool, this tool uses a dual threshold system of cosine similarity (filtering duplicate content) and entity coverage (filtering low-relevance content) to eliminate redundant or irrelevant documents. This ensures that the context set of the input model not only matches the query requirements but also possesses information diversity, reducing the risk of "illusion" generation. The formula is as follows:

[0074]

[0075] Where n represents the initial number of search results; d i Let represent the i-th document vector; q represent the average similarity between the document and the query; w represent the domain experience threshold; γ is set to 0.6; when the document similarity is lower than T or the diversity score D is lower than 0.3, it is discarded to ensure that the context set input to the large model has both relevance and diversity.

[0076] The multi-mode query interaction in step S5 of this embodiment is as follows:

[0077] S501, Dual-entry query: Provides dual entry points for natural language and structured queries. Business personnel can directly obtain results through a conversational query such as "Which procurement approval processes have timed out?", while operations or data personnel can accurately filter through structured conditions such as "Process definition ID number and approver name". The system automatically identifies the query type and matches the optimal retrieval strategy.

[0078] S502 Intelligent Assistance: Integrates query intent recognition and historical record management, that is, it automatically recommends related query terms ("overdue work order processing flow" "work order overdue responsibility definition") for fuzzy queries (such as "overdue work order"); and saves employees' high-frequency query records, supports one-click re-retrieval and result export (Excel / PDF format), and improves the efficiency of repeated queries.

[0079] Example 2:

[0080] As attached Figure 1 As shown, this embodiment provides a process data retrieval optimization system based on a large model. The system includes:

[0081] The data preprocessing module is used to acquire and preprocess enterprise process data and build a dynamic terminology vector library.

[0082] The hybrid retrieval module is used to obtain initial retrieval results using a hybrid retrieval mechanism, realizing collaborative recall of vector retrieval and keyword retrieval, as well as cross-modal data fusion.

[0083] The intelligent re-ranking module is used to optimize the initial results by enhancing re-ranking through a large model, and obtain the final retrieval set that meets business needs. It achieves positional attention ranking and diversity filtering through a large model.

[0084] The interactive interface module provides dual query entry points and intelligent auxiliary functions.

[0085] The data preprocessing module in this embodiment includes a multi-granularity slicing processing submodule and a dynamic term vector library construction submodule;

[0086] The multi-granularity slicing submodule is used to perform coarse-grained segmentation according to the business logic of "process definition → approval configuration → personnel assignment → submission specifications", and to perform fine-grained segmentation on long paragraphs with a semantic window of 300-500 words, combined with syntactic analysis to avoid the break of key logic; then, the optimal slice size is determined based on the slice quality evaluation index Q, realizing the dynamic switching of "full-link query priority coarse-grained and link detail query priority fine-grained", where Q=β×R+(1-β)×C, R is the slice retrieval recall rate, C is the context integrity, and β ranges from 0.4 to 0.6;

[0087] The dynamic terminology vector library construction submodule is used to pre-train models based on BERT-flow, and fine-tunes them using terminology corpora from the process domains of "process definition ID" and "procurement approval process". It also introduces a terminology semantic similarity calculation model to balance matching accuracy. The weight coefficient α in the terminology semantic similarity calculation is optimized to 0.7 based on the validation set. At the same time, new terms for "cross-level approvers" are added, and the vector representation is quickly updated through a few-shot learning mechanism.

[0088] The hybrid retrieval mechanism in this embodiment is a two-layer collaborative recall mechanism. Specifically, the vector retrieval layer uses an improved HNSW index and combines HyDE technology to generate hypothetical answer vectors to expand the semantics of short queries; the keyword retrieval layer supplements the precise matching capability of the structured fields "work order number" and "approver ID" through regular expression matching and fuzzy search, and the results of the two are merged to form the initial retrieval results.

[0089] In this embodiment, the intelligent re-ranking module adds a "item X contains key process information" type location prompt when the initial result is input from Llama-2 to prioritize high-value documents. At the same time, the intelligent re-ranking module introduces the Diversity Ranker tool, which uses cosine similarity and entity coverage as dual thresholds for filtering. Documents with relevance below the threshold T or diversity score D below 0.3 are removed, where γ is 0.6.

[0090] The interactive interface module in this embodiment provides dual query entry points for natural language and structured queries, automatically identifies query types and matches retrieval strategies; it also integrates query intent recognition functions, automatically recommending related query terms for fuzzy queries; and it saves high-frequency query records and supports exporting results in Excel or PDF format.

[0091] This embodiment covers the entire optimization mechanism from data input to result output:

[0092] (1) Data preprocessing and multi-granularity representation construction. Data preprocessing is the foundation for achieving accurate retrieval. The core is to construct high-quality semantic representations through dynamic term optimization and multi-granularity slicing, specifically including the following technical means:

[0093] ① Dynamic Terminology Aware Embedding: For process approval terms such as "process definition ID" and "procurement approval process," the BERT-flow model is fine-tuned using the full approval corpus to construct a dynamic terminology vector library. For newly added terms such as "cross-level approvers," the vectors are quickly updated through small-sample learning to narrow the semantic gap between the query and the data, improve terminology matching accuracy, and introduce a terminology semantic similarity calculation model, as shown in the following formula:

[0094]

[0095] Where vector t is the term vector, vector q is the query vector, and α is the weight coefficient (optimized to 0.7 based on the validation set). This formula balances the accuracy of semantic matching and literal matching.

[0096] ② Multi-granularity slicing strategy: Adapting to the hierarchical logic of workflow approval data, a "hierarchical segmentation + dynamic adjustment" model is adopted to balance logical integrity and redundancy control: Coarse-grained slicing: Segmented according to document structure and the logic of "workflow definition → approval configuration → personnel assignment → submission specifications" to form complete macro units; Fine-grained slicing: Long paragraphs are segmented with semantic windows of 300-500 characters, and combined with syntactic analysis to avoid logical breaks such as "approval configuration - permission binding", focusing on local details. To scientifically determine the optimal slice size, a slice quality evaluation index Q is defined, with the following formula:

[0097] Q = β × R + (1 - β) × C;

[0098] Where R is the slice retrieval recall rate, C is the slice context integrity (obtained through cross-slice semantic coherence scoring), and β is the balance coefficient (0.4 for process overview queries and 0.6 for process step queries).

[0099] ③ Dynamic switching mechanism: The optimal size is determined by the validation set, and "full-link query" prioritizes outputting coarse-grained data, while "detailed process query" prioritizes outputting fine-grained data.

[0100] (2) Hybrid retrieval engine design. Hybrid retrieval works in conjunction with large-scale model re-ranking. The core achieves a balance between high accuracy and efficiency through "two-layer recall + cross-modal fusion," as detailed below:

[0101] ① Two-layer collaborative recall mechanism: The vector retrieval layer adopts an improved HNSW index and combines HyDE technology to generate hypothetical answer vectors (such as generating an extended vector containing dimensions such as "work order status" when querying "timeout work order processing") to improve the recall rate of short queries; the keyword retrieval layer supplements the precise matching capability of structured fields such as "work order number" through regular expressions and fuzzy matching to make up for the shortcomings of vector retrieval.

[0102] ② Cross-modal data fusion: Extract entity relationship triples such as "process node-responsible person-time limit" from unstructured documents, and map them to the semantic space in a unified manner with structured form data, supporting cross-modal queries that associate structured fields with document descriptions using natural language;

[0103] (3) Large-scale model enhancement and re-ranking. Large-scale model enhancement and re-ranking work in deep collaboration with hybrid search results. The core of this approach is to achieve a balance between the value and usability of the results through "context-awareness + diversity filtering," as detailed below:

[0104] ① Position Attention-Guided Ranking: Based on LLM such as Llama-2, a re-ranking module is built. When the search results are input, position prompts such as "the Xth item contains key process information" are added. By utilizing the model's attention preference for the beginning and end of the text, high-value documents (such as process operation details and approval node definitions) are prioritized to enhance the efficiency of key information delivery.

[0105] ② Diversity Intelligent Filtering: The Diversity Ranker tool is introduced, using a dual threshold of cosine similarity (filtering duplicate content) and entity coverage (filtering low-relevance content) to eliminate redundant or irrelevant documents. This ensures that the context set input to the large model not only matches the query requirements but also possesses information diversity, reducing the risk of "illusion" generation. The core formula is as follows:

[0106]

[0107] Where n is the initial number of search results, di is the i-th document vector, vector q is the average similarity between the document and the query, w is the domain experience threshold, and γ is 0.6. Documents with similarity below T or diversity score D below 0.3 are discarded to ensure that the context set input to the large model has both relevance and diversity.

[0108] (4) Multi-mode query interaction design. To adapt to the query habits of employees in different positions within the enterprise, a flexible and efficient query interaction system is constructed to support convenient access to process data, specifically including:

[0109] ① Dual-entry query support: Provides dual entry points for natural language and structured queries - business personnel can directly obtain results through conversational queries such as "which procurement approval processes have timed out"; operations or data personnel can accurately filter through structured conditions such as "process definition ID number, approver name", and the system automatically identifies the query type and matches the optimal retrieval strategy;

[0110] ② Intelligent auxiliary functions: Integrates query intent recognition and history management, automatically recommends related query terms ("overdue work order processing flow", "work order overdue responsibility definition") for fuzzy queries (such as "overdue work order"); saves employees' high-frequency query records, supports one-click re-retrieval and result export (Excel / PDF format), and improves the efficiency of repeated queries.

[0111] Example 3:

[0112] This embodiment also provides an electronic device, including: a memory and at least one processor;

[0113] The memory stores computer-executed instructions;

[0114] The at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to execute the process data retrieval optimization method based on a large model according to any one of the present invention.

[0115] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor can be a microprocessor or any conventional processor.

[0116] Memory is used to store computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, at least one application program required for a function, etc.; the data storage area can store data created based on the use of the terminal, etc. In addition, memory can also include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart memory cards (SMC), secure digital cards (SD cards), flash memory cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0117] Example 4:

[0118] This embodiment also provides a computer-readable storage medium storing multiple instructions, which are loaded by a processor to cause the processor to execute the large-model-based process data retrieval optimization method according to any embodiment of the present invention. Specifically, a system or apparatus equipped with a storage medium may be provided, on which software program code implementing the functions of any of the above embodiments is stored, and the computer (or CPU or MPU) of the system or apparatus may read and execute the program code stored in the storage medium.

[0119] In this case, the program code read from the storage medium can itself implement the function of any of the above embodiments, and therefore the program code and the storage medium storing the program code constitute part of the present invention.

[0120] Examples of storage media used to provide program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer via a communication network.

[0121] Furthermore, it should be clear that not only can the program code read by the computer be executed, but also the operating system or other components operating on the computer can be instructed based on the program code to perform some or all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0122] Furthermore, it is understood that the program code read from the storage medium is written to the memory set in the expansion board inserted into the computer or to the memory set in the expansion unit connected to the computer. Then, based on the instructions of the program code, the CPU or other components installed on the expansion board or expansion unit execute some and all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A process data retrieval optimization method based on a large model, characterized in that, The method is as follows: Data preprocessing and multi-granularity semantic feature construction: acquire enterprise process data, perform multi-granularity slicing preprocessing on the process data, and construct multi-granularity semantic features; Hybrid retrieval: Employs a two-layer recall mechanism combining vector retrieval and keyword retrieval to obtain initial search results; Large Model Enhanced Re-ranking: The initial results are re-ranked in a context-aware manner using a large model, and the final retrieval set is generated by combining positional attention and diversity filtering. Multi-mode query interaction: Construct a query interaction system to support convenient access to process data and adapt to the query needs of employees in different positions within the enterprise.

2. The process data retrieval optimization method based on a large model according to claim 1, characterized in that, The data preprocessing and multi-granularity semantic feature construction are detailed below: Dynamic Terminology Aware Embedding: For process approval terms such as "Process Definition ID" and "Procurement Approval Process," the BERT-flow model is fine-tuned using the full approval corpus to construct a dynamic terminology vector library. New terms are added for "cross-level approvers," and the vectors are quickly updated through small-sample learning to narrow the semantic gap between queries and data, improving terminology matching accuracy. A terminology semantic similarity calculation model is introduced to balance the accuracy of semantic matching and literal matching. The formula for the terminology semantic similarity calculation model is as follows: Where t represents the term vector; q represents the query vector; and α represents the weight coefficient. Multi-granular slicing: A "hierarchical segmentation + dynamic adjustment" model is adopted to balance logical integrity and redundancy control, adapting to the hierarchical logic of process approval data. Specifically: Coarse-grained segmentation: Segmentation is based on document structure and the logic of "process definition → approval configuration → personnel assignment → submission specifications" to form complete macro-units; Fine-grained segmentation: Long paragraphs are segmented using a 300-500 character semantic window, combined with syntactic analysis to avoid breakage in the key logic of "approval configuration - permission binding," focusing on local details. The optimal slice size is determined based on the slice quality evaluation index Q, achieving dynamic switching between "full-link query priority coarse-grained and link detail query priority fine-grained." Where Q = β × R + (1-β) × C; where R represents the slice retrieval recall rate; C represents the slice context integrity, which is obtained through cross-slice semantic coherence scoring; β is a balancing coefficient, ranging from 0.4 to 0.

6. Dynamic switching: The optimal size is determined by verifying the dataset. "Full-link query" prioritizes outputting coarse-grained data, while "detailed process query" prioritizes outputting fine-grained data.

3. The process data retrieval optimization method based on a large model according to claim 1, characterized in that, The hybrid search is as follows: Two-layer collaborative recall mechanism: The vector retrieval layer uses an improved HNSW index and combines HyDE technology to generate hypothetical answer vectors to improve the recall rate of short queries; the keyword retrieval layer supplements the precise matching capability of structured fields such as "work order number" and "approver ID" through regular expressions and fuzzy matching, making up for the shortcomings of vector retrieval, and merging the results of the vector retrieval layer and the keyword retrieval layer to form the initial retrieval results; Cross-modal data fusion: Extract the "process node-responsible person-time limit" entity relationship triples from unstructured documents and map them uniformly to the semantic space with structured form data, supporting cross-modal queries that associate structured fields with document descriptions using natural language.

4. The process data retrieval optimization method based on a large model according to claim 1, characterized in that, The large-scale model enhancement and reordering are detailed below: Position attention-guided ranking: Based on the large Llama-2 model, the ranking is re-ranked, and the search results are obtained and the position hint "the Xth item contains key process information" is added. By leveraging the large model's attention preference for the beginning and end of the text, high-value documents are prioritized and the efficiency of key information delivery is enhanced. Diversity Intelligent Filtering: The Diversity Ranker tool is introduced, using a dual threshold of cosine similarity and entity coverage to filter redundant or irrelevant documents. This ensures that the context set of the input large model not only matches the query requirements but also possesses information diversity, reducing the risk of "illusion" generation. The formula is as follows: Where n represents the initial number of search results; d i Let represent the i-th document vector; q represent the average similarity between the document and the query; w represent the domain experience threshold; γ is set to 0.6; when the document similarity is lower than T or the diversity score D is lower than 0.3, it is discarded to ensure that the context set input to the large model has both relevance and diversity.

5. The process data retrieval optimization method based on a large model according to any one of claims 1 to 4, characterized in that, This method also supports multi-mode query interaction. By building a flexible and efficient query interaction system, it supports convenient access to process data and adapts to the query habits of personnel in different positions within the enterprise, as detailed below: Dual-entry query: Provides both natural language and structured query entry points. Business personnel can directly obtain results through a conversational query such as "Which procurement approval processes have timed out?", while operations or data personnel can accurately filter through structured conditions such as "Process definition ID number and approver name". The system automatically identifies the query type and matches the optimal retrieval strategy. Intelligent assistance: It integrates query intent recognition and historical record management, that is, it automatically recommends related query terms for fuzzy queries; and saves employees' high-frequency query records, supporting one-click re-retrieval and result export.

6. A process data retrieval and optimization system based on a large model, characterized in that, The system includes: The data preprocessing module is used to acquire and preprocess enterprise process data and build a dynamic terminology vector library. The hybrid retrieval module is used to obtain initial retrieval results using a hybrid retrieval mechanism, realizing collaborative recall of vector retrieval and keyword retrieval, as well as cross-modal data fusion. The intelligent re-ranking module is used to optimize the initial results by enhancing re-ranking through a large model, and obtain the final retrieval set that meets business needs. It achieves positional attention ranking and diversity filtering through a large model. The interactive interface module provides dual query entry points and intelligent auxiliary functions.

7. The process data retrieval and optimization system based on a large model according to claim 6, characterized in that, The data preprocessing module includes a multi-granularity slicing submodule and a dynamic term vector library construction submodule; The multi-granularity slicing submodule is used to perform coarse-grained segmentation according to the business logic of "process definition → approval configuration → personnel assignment → submission specifications", and to perform fine-grained segmentation on long paragraphs with a semantic window of 300-500 words, combined with syntactic analysis to avoid the break of key logic; then, the optimal slice size is determined based on the slice quality evaluation index Q, realizing the dynamic switching of "full-link query priority coarse-grained and link detail query priority fine-grained", where Q=β×R+(1-β)×C, R is the slice retrieval recall rate, C is the context integrity, and β ranges from 0.4 to 0.6; The dynamic terminology vector library construction submodule is used to pre-train models based on BERT-flow, and fine-tunes them using terminology corpora from the process domains of "process definition ID" and "procurement approval process". It also introduces a terminology semantic similarity calculation model to balance matching accuracy. At the same time, it adds new terms for "cross-level approvers" and quickly updates the vector representation through a few-shot learning mechanism.

8. The process data retrieval and optimization system based on a large model according to claim 6 or 7, characterized in that, The hybrid retrieval mechanism is a two-layer collaborative recall mechanism, specifically: the vector retrieval layer uses an improved HNSW index and combines HyDE technology to generate hypothetical answer vectors to expand the semantics of short queries; the keyword retrieval layer supplements the precise matching capability of the structured fields "work order number" and "approver ID" through regular expression matching and fuzzy search, and the results of the two are merged to form the initial retrieval results; The intelligent re-ranking module adds a "item X contains key process information" type location prompt when the initial result is input from Llama-2 to prioritize high-value documents; at the same time, the intelligent re-ranking module introduces the Diversity Ranker tool, which uses cosine similarity and entity coverage as dual thresholds to filter documents. When the document relevance is lower than the threshold T or the diversity score D is lower than 0.3, it is removed, where γ is 0.

6. The interactive interface module provides dual query entry points for natural language and structured queries, automatically identifies query types and matches retrieval strategies; it also integrates query intent recognition functionality, automatically recommending related query terms for fuzzy queries; and saves high-frequency query records and supports exporting results in Excel or PDF format.

9. An electronic device, characterized in that, include: Memory and at least one processor; The memory stores computer-executed instructions; The at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to perform the process data retrieval optimization method based on a large model as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by the processor, implement the process data retrieval optimization method based on a large model as described in any one of claims 1 to 6.

Images