Large model driven enterprise collaborative office intelligent assistant method and system

Abstract

The present application relates to the field of enterprise informationization and artificial intelligence technology, in particular to a large model driven enterprise collaborative office intelligent assistant method and system, taking enterprise internal knowledge base, rules and regulations and historical project documents as domain corpus, continuously fine-tuning the embedding organization role vector on the basis of large language model to obtain enterprise vertical large model; through unified access layer, multi-source information flow of mail, instant messaging and document collaboration platform is aggregated and topic fingerprint is generated; multi-source information of the same topic fingerprint is aggregated in time sequence as topic recurrence track to replace semantic deduplication and keep cross-platform recurrence event sequence; recurrence attenuation-enhancement dynamics calculation is applied to calculate the closed-loop urgency score; combined with the reverse adaptation degree of the candidate undertaking object, action item dispatch is executed and the undertaking state is closed-loop written back to adjust the attenuation coefficient, which converts the cross-platform repeated appearance behavior signal into quantifiable urgency driving force from the principle, and realizes the active emergence of the closed-loop urgency topic and the automatic closed-loop achievement of the action item.

Description

Technical Field

[0001] This invention relates to the fields of enterprise informatization and artificial intelligence technology, specifically to a method and system for a large-scale model-driven intelligent assistant for collaborative office work. Background Technology

[0002] With the rapid penetration of large language model technology into enterprise informatization, enterprise collaborative office work is undergoing a paradigm shift from traditional process-driven to intelligent assistant-driven. Domain corpora comprised of internal enterprise knowledge bases, regulations, and historical project documents provide rich material for the vertical fine-tuning of large models, while the maturity of natural language command understanding has enabled intelligent assistants to replace menu-based human-computer interaction capabilities. However, the true characteristic of information flow in enterprise collaborative office scenarios is not isolated single queries or single-point commands, but rather the continuous flow of collaborative issues between email, instant messaging, and document collaboration platforms—the same issue is often re-initiated by the initiator on a different platform because it didn't receive a response the first time, is brought up again due to memory decay among relevant parties, and is re-questioned due to different departmental perspectives. How to accurately identify and amplify unresolved collaborative issues in multi-platform information flows, and prevent urgent matters from being buried, is a core challenge that has not yet been well addressed in the current intelligent upgrade of enterprise collaborative office work.

[0003] Chinese invention patent application CN117114521A discloses an enterprise decision support system based on a large model and knowledge graph. This system uses an enterprise data resource library to store structured, unstructured, and crowdsourced data. It constructs a vertical enterprise model through a three-stage training process: incremental pre-training, supervised fine-tuning, and reinforcement learning based on human feedback. Based on this vertical model, it performs entity recognition and relationship extraction on enterprise data to construct a knowledge graph covering five categories of concepts: personnel, machines, materials, methods, and environment. Finally, it empowers enterprise personnel through sub-modules for work handover, daily assistance, and exception handling. However, the business empowerment module of this solution relies on recommendation algorithms based on content, collaborative filtering, and popularity. Essentially, it still belongs to a recommendation paradigm oriented towards single static information or a single decision point. It fails to continuously track the same collaborative issues generated across multiple platforms and time periods, and it fails to identify the unclosed-loop meaning behind the repeated appearance of the same issue on different platforms. This results in urgent but unresolved collaborative matters not proactively appearing in the user's view.

[0004] Chinese invention patent application CN117743528A discloses an internal enterprise auxiliary model, its establishment method, and its application. This method involves collecting internal business scenario data and data from the entire R&D process of internet information services; establishing a GPU computing cluster; constructing a multi-task training dataset; selecting a base model and performing multi-task fine-tuning; establishing a human feedback dataset and an internal enterprise reward model; and finally completing training based on human feedback reinforcement learning. The resulting internal enterprise auxiliary model is then deployed within the internal GPU computing cluster, combined with a human-computer interaction module and a data storage module to construct an internal enterprise auxiliary system. However, this solution's auxiliary form remains at the level of single-round or multi-round question-and-answer human-computer interaction. Employees obtain answers generated by the large model based on internal data by asking questions. It does not involve proactive access and aggregation of information flows from multiple platforms, nor does it establish a complete automated chain from action item identification to assignment to closed-loop achievement. Therefore, it is difficult to address the typical pain point of repeated, unresolved cross-departmental collaborative issues in enterprise collaboration.

[0005] Chinese invention patent application CN117688946A discloses a method, apparatus, electronic device, and storage medium for intent recognition based on a large model. This method retrieves at least one candidate intent pair similar to the input statement, generates prompt information containing question-answer pairs and intent recognition range based on the candidate intent pairs, and inputs the prompt information and the input statement into a large language model to identify the target intent from the intent recognition range. It also supports storing intent pairs by domain on an Elasticsearch cluster and supports dynamic addition, updating, and deletion. However, this solution focuses on the localized problem of intent recognition for a single input statement and does not integrate the intent recognition results into a continuous tracking framework for cross-platform and cross-time-period collaborative issues. The identified intent itself is neither associated with historical recurring event sequences nor drives subsequent closed-loop dispatching and feedback writing of action items.

[0006] As can be seen from the existing technologies described above, semantic deduplication is currently the standard processing paradigm for multi-platform collaborative information aggregation. This involves comparing the content similarity of information on the same topic acquired across platforms and time periods, retaining only representative items to reduce repetitive interruptions for employees and improve the simplicity of the information list. However, this widely accepted processing paradigm suffers from a long-overlooked cognitive bias: the repeated appearance of the same collaborative topic across platforms and time periods is precisely a strong signal that the topic is not closed-loop or achieved, rather than simple redundant noise. Semantic deduplication treats this unclosed-loop signal as redundancy and removes it, causing urgent but unachieved collaborative tasks to be submerged in the simplified list after deduplication, while closed-loop, low-urgency topics are repeatedly presented to users because they are accompanied by a large amount of information when they first appear. This ultimately creates a paradoxical bottleneck where the more information is aggregated, the harder it is for urgent tasks to emerge. How to fundamentally break away from the inherent path of semantic deduplication and treat the recurrence itself as an observable physical signal driving the achievement of closed-loop action items is a key technical problem that existing technologies have not yet addressed. Summary of the Invention

[0007] Addressing the core bottleneck in existing technologies where semantic deduplication is commonly used in multi-platform information aggregation for enterprise collaborative work, leading to the redundant removal of strong signals of repeated occurrences of the same collaborative topic across platforms and the submergence of urgent but unfulfilled collaborative tasks, this invention provides a large-model-driven intelligent assistant method and system for enterprise collaborative work. This method replaces semantic deduplication with topic recurrence trajectories, quantifies the urgency of unclosed loops using recurrence attenuation-enhancement dynamics, matches candidate recipients using reverse adaptability, and dynamically adjusts the recurrence attenuation coefficient through closed-loop feedback. From a fundamental perspective, this transforms the previously redundant behavioral signal of repeated occurrences across platforms into a quantifiable urgency-driven quantity, enabling the proactive emergence of urgent unclosed loop topics and the automatic closure of action items.

[0008] The technical solution of this invention is as follows:

[0009] A large-scale model-driven intelligent assistant method for enterprise collaborative office work includes the following steps: Continuously fine-tuning a basic large-scale language model using internal enterprise knowledge bases, regulations, and historical project documents as domain corpora, and embedding organizational role vectors to obtain an enterprise vertical large-scale model; aggregating multi-source information flows from email, instant messaging, and document collaboration platforms through a unified access layer, and generating an issue fingerprint for each piece of information based on the enterprise vertical large-scale model, the issue fingerprint including the issue subject, action type, and responsible entity; aggregating multi-source information carrying the same issue fingerprint in chronological order to form an issue recurrence trajectory, the issue recurrence trajectory replacing semantic deduplication to retain the recurrence event sequence across platforms and time periods; calculating the unclosed-loop urgency score based on the time interval and platform diversity of the issue recurrence trajectory using recurrence decay-enhancement dynamics; extracting action items for the issue based on the unclosed-loop urgency score, assigning action items based on the reverse adaptability of candidate recipients, and writing the assigned acceptance status back to the issue recurrence trajectory as closed-loop feedback to adjust the subsequent recurrence decay coefficient.

[0010] This invention also provides a large-scale model-driven intelligent assistant system for enterprise collaborative office work, comprising: an enterprise vertical large-scale model construction module, used to continuously fine-tune a basic large-scale language model using the enterprise's internal knowledge base, rules and regulations, and historical project documents as domain corpora, and embed organizational role vectors to obtain an enterprise vertical large-scale model; a multi-platform information access module, used to aggregate multi-source information flows from email, instant messaging, and document collaboration platforms through a unified access layer, and generate an issue fingerprint for each piece of information based on the enterprise vertical large-scale model, wherein the issue fingerprint includes the issue subject, action type, and responsible party; and an issue recurrence trajectory tracking module, used to track the recurrence of issues. Multi-source information of the topic fingerprint is aggregated in time sequence to form the topic recurrence trajectory. The topic recurrence trajectory replaces semantic deduplication and retains the recurrence event sequence across platforms and time periods. The unclosed loop urgency assessment module is used to calculate the unclosed loop urgency score based on the time interval and platform diversity of the topic recurrence trajectory and apply recurrence decay-enhancement dynamics. The action item closed loop assignment module is used to extract action items for the topic based on the unclosed loop urgency score, perform action item assignment in combination with the reverse adaptability of the candidate receiving object, and write back the acceptance status after assignment as closed loop feedback to the topic recurrence trajectory tracking module to adjust the subsequent recurrence decay coefficient.

[0011] The beneficial effects of this invention include:

[0012] First, this invention uses topic recurrence trajectories instead of semantic deduplication to aggregate multi-source information streams. The mechanism lies in the fact that the multiple occurrences of the same topic fingerprint across platforms are not compressible redundant signals, but rather observable physical evidence of unclosed collaborative loops. Compared to the limitations of content-based, collaborative filtering, and popularity-based recommendation algorithms, which can only provide recommendations for single static pieces of information and whose single-round or multi-round question-and-answer human-computer interaction cannot support continuous cross-platform tracking, this invention retains each cross-platform, cross-time-period recurrence event and arranges them chronologically as a recurrence event sequence. This makes each recurrence a usable original observation for subsequent urgency assessment, fundamentally avoiding the inherent flaw in existing technologies where urgent, unclosed issues are semantically deduplicated due to initial response failure and ultimately buried in the aggregation list.

[0013] Second, this invention calculates the urgency score of unclosed-loop issues based on the issue recurrence trajectory using recurrence decay-enhancement dynamics. The mechanism involves decomposing the temporal event sequence of recurrence behavior into a decay term reflecting time expiration and an enhancement term reflecting cross-platform diversity amplification, dynamically switching the weights according to cross-platform diversity conditions. Compared to the limitations of existing technologies that rank urgency based on single-point timestamps or single-time content vectors, this invention enables unclosed-loop issues that repeatedly occur across platforms and time periods to receive a higher urgency ranking than expired issues on a single platform, thus proactively bringing urgent but repeatedly unresponsive issues to the forefront.

[0014] Third, this invention introduces the reverse fit of candidate recipients as a selection criterion in the action item assignment process. The mechanism involves transforming the traditional one-way top-down skill matching into a four-dimensional weighted score based on the candidate recipient's historical response timeliness, issue type acceptance rate, current workload saturation, and the strength of their departmental association with the responsible entity. This reverse fit significantly improves the initial assignment hit rate and prevents assignment failures from leading to repeated accumulation of contamination attenuation coefficients.

[0015] Fourth, this invention uses the acceptance status after dispatch as a closed-loop feedback to write back to the issue reproduction trajectory and dynamically adjust the reproduction attenuation coefficient. Its mechanism is to break the one-way push paradigm of action item dispatch in the existing technology and establish a learning closed loop: when the acceptance is confirmed or the achievement is completed, the reproduction attenuation is accelerated and the urgency is reduced. When there is no response or rejection, the enhancement coefficient is further increased, which reverses the urgency and triggers re-dispatch, so that the system has the ability to adapt to collaborative closed loop.

[0016] Fifth, the aforementioned four beneficial effects work synergistically to form a nonlinear gain of 1+1>2: the issue reproduction trajectory retains the unclosed-loop signal, the reproduction attenuation-enhancement dynamics are transformed into an urgency signal, the reverse adaptation improves the dispatch hit rate, and the closed-loop feedback write-back gives the system continuous learning ability. The four constitute a deeply coupled synergistic closed loop, enabling the present invention to achieve a significant overall improvement over the existing distributed optimization technology in both the dimensions of collaborative efficiency and employee information burden. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the large-model-driven intelligent assistant method for enterprise collaborative office work provided in an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of the architecture of the large-model-driven intelligent assistant system for enterprise collaborative office work provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The described embodiments are only some embodiments of the present invention and not all embodiments of the present invention. Based on the embodiments described in the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection claimed by the present invention.

[0020] like Figure 1 As shown, the large-model-driven intelligent assistant method for enterprise collaborative office provided in this embodiment of the invention includes steps S1 to S5. The method is deployed in the private computing environment of the enterprise intranet and provides all employees of the enterprise with the ability to intelligently track and achieve closed-loop collaboration on cross-platform issues.

[0021] Step S1: Using the enterprise's internal knowledge base, rules and regulations, and historical project documents as domain corpora, continuously fine-tune the basic large language model and embed organizational role vectors to obtain the enterprise vertical large model.

[0022] The goal of this step is to enable the general-purpose basic large language model to understand the specific business context of an enterprise and to enable its downstream inference process to be aware of the roles and permission boundaries within the organization. The basic large language model is deployed privately within the enterprise's intranet, and can be selected from open-source base models with suitable parameter sizes, such as Qwen-14B, ChatGLM3-6B, and Baichuan2-13B, deployed within the enterprise's internal GPU computing cluster to ensure that enterprise data does not leave the intranet.

[0023] The preparation process for the domain corpus includes: extracting published product documents, technical specifications, and best practice manuals from the enterprise knowledge base system; extracting current policy documents from business lines such as human resources, finance, procurement, R&D, and IT security from the enterprise rules and regulations system; and extracting full lifecycle documents such as requirement documents, design schemes, meeting minutes, and debriefing reports from completed projects over the years from the historical project archive. After preprocessing such as removing watermarks, uniform encoding, removing personal privacy fields, and segmenting into fragments of 512 to 2048 token lengths according to knowledge density, the corpus forms an enterprise-specific pre-training corpus with a total size typically ranging from 5GB to 50GB.

[0024] The continuous fine-tuning process includes: first, performing unsupervised incremental pre-training on the basic large language model to inject enterprise domain language distribution features; then, performing supervised fine-tuning based on instruction-response pairs labeled by domain experts. The instruction-response pair dataset typically ranges in size from 30,000 to 150,000 entries, covering typical enterprise collaboration scenarios such as cross-departmental coordination, process initiation, information retrieval, and action item assignment. The supervised fine-tuning employs the LoRA low-rank adaptation method to reduce memory usage and training costs. Optional hyperparameters include rank (8), scaling factor (16), learning rate (3e-4), batch size (16 to 64), training epochs (2 to 5), and cosine annealing for learning rate scheduling.

[0025] The embedding process of organizational role vectors includes: extracting structured attributes such as employee department affiliation, job responsibilities, reporting relationships, and skill tags from the enterprise's human resources system; vectorizing and encoding the structured attributes into dense vectors of 128 or 256 dimensions, which are the organizational role vectors; and inserting the organizational role vectors as learnable parameters before the input embedding layer of the enterprise's vertical large model, so that the large model can adjust its semantic parsing and responsibility identification preferences based on the organizational role vectors of the initiator of the input information during the inference stage.

[0026] Continuous fine-tuning is repeated according to a set cycle: every 7 days or every 14 days, incremental fine-tuning is performed on newly added documents and historical interaction data within the enterprise. Each incremental fine-tuning only trains about 80% of the incremental data in each round to avoid overfitting; organizational role vectors are updated synchronously when personnel changes occur in the enterprise's organizational structure. The corresponding role vectors are regenerated and replaced by subscribing to personnel change events in the human resources system.

[0027] The enterprise vertical big model obtained through the above complete process serves as the underlying intelligent foundation for all subsequent steps of this invention, and its output consistently carries organizational context and role perception capabilities during the inference stage.

[0028] Step S2: Aggregate multi-source information flows from email, instant messaging, and document collaboration platforms through a unified access layer, and generate topic fingerprints for each message based on the enterprise vertical big model.

[0029] The goal of this step is to unify and aggregate information flows scattered across multiple office platforms within an enterprise into a semantically aligned data channel, and assign a fingerprint identifier to each message that can be used for cross-platform topic-based judgment. The unified access layer is built on the open APIs of the enterprise's existing office suite, using IMAP / ExchangeWebServices protocols for email platforms, open interfaces from vendors such as WeChat Work / DingTalk / Lark for instant messaging platforms, and open interfaces from vendors such as Confluence / Notion / Yuque / Office365 for document collaboration platforms. The multi-source information flows carry fields such as platform identifier, sending timestamp, sender identifier, receiver identifier, content body, and attachment references in a unified JSON Schema format.

[0030] The generation of issue fingerprints is based on the semantic parsing capabilities of the enterprise vertical big model. The issue fingerprint comprises three fields: issue subject, action type, and responsible entity, which together constitute a unique anchor identifier for cross-platform and cross-time period determination of the same issue. The issue subject is obtained by performing named entity recognition on each piece of information. This named entity recognition employs a few-shot prompting method based on the enterprise vertical big model. The prompt template is designed to guide the model to extract the specific project name, product code, customer name, or matter title from the information content as the issue subject. The action type is obtained by performing a four-category classification on each piece of information. The candidate sets for the four categories include action obligation, information sharing, status reporting, and decision request. Action obligation corresponds to information that explicitly requires the recipient to complete a specific action within a specified time limit; information sharing corresponds to information that only shares facts or data without explicit execution requirements; status reporting corresponds to information describing the current progress or result of a matter; and decision request corresponds to information requesting the recipient to make a choice among multiple candidate solutions. The responsible party is identified by combining organizational role vectors with the identity of the information sender. For information involving action obligations, the responsible party is the primary employee of the recipient to whom the information is directed. For information involving status reports, the responsible party is the business line manager of the information initiator. For information involving decision requests, the responsible party is the superior position with decision-making authority to which the information is directed.

[0031] During the issue fingerprint generation process, the enterprise vertical large model typically controls the processing latency of each message to within 200ms. For a typical email or instant messaging message, the three fields of issue subject, action type, and responsible party can be output simultaneously in a single forward inference of the large model. The issue fingerprint generation result, along with the platform identifier, timestamp, content summary, and other fields of the original information, is written into the issue fingerprint index library. The issue fingerprint index library uses triplet keys as bucket keys to support fast retrieval in subsequent steps.

[0032] Meeting recordings, as a special form of multi-source information flow, are accessed by a corresponding preprocessing sub-process. The preprocessing sub-process performs automatic speech recognition on the meeting recordings to obtain the meeting text. Based on the enterprise vertical big model, a meeting summary is generated from the meeting text and meeting action items are extracted. The meeting action items are injected into the multi-source information flow in a triplet key-value structure consistent with the topic fingerprint to participate in subsequent topic reproduction trajectory aggregation.

[0033] Step S3: Aggregate multi-source information carrying the same topic fingerprint into topic recurrence trajectory in chronological order. The topic recurrence trajectory replaces semantic deduplication and retains the recurrence event sequence across platforms and time periods.

[0034] This step is the core element that distinguishes this invention from existing semantic deduplication paradigms. Its design intent is to transform the behavioral signal of repeated occurrences of the same topic across platforms and time periods—originally considered redundant noise—into observable and quantifiable physical evidence. The aggregation rule for the topic recurrence trajectory no longer relies on similarity determination of information content vectors or retention of representative entries. Instead, it uses the triplet key value of the topic fingerprint as the unique aggregation key. All cross-platform information sharing the same triplet key value is included in the same recurrence trajectory without performing content deduplication.

[0035] Formal definition, for fingerprints with the same issue The collection of all cross-platform information, and the trajectory of its recurring issues. Represented as a sequence of recurring events arranged in ascending order of their timestamps:

[0036] ,in: The issue recurrence trajectory is an ordered set of recurring event tuples sorted by time, directly calculated by this aggregation formula, representing the fingerprint of the same issue. Complete reproduction of event sequences across platforms and time periods; The issue fingerprint is a triple key value consisting of the issue subject, action type, and responsible party. It is a string type and is generated by the aforementioned step S2. It is used as a unique anchor identifier for determining the same issue. For the first Each recurring event is a quintuple structure, directly constructed by this aggregation formula, and is used to record complete observation information for a single recurrence. For the first The platform identifier for each reproducible event is a discrete enumeration type that takes values ​​from four candidate values: email, instant messaging, document collaboration, or meeting. It is extracted from the original information by the unified access layer and used to support the subsequent calculation of the cross-platform diversity index. For the first The timestamp of each reproducible event is a floating-point number, with a value range of typical UNIX timestamp intervals and a unit of seconds. It is extracted from the original information by the unified access layer and used to support the calculation of the time interval for subsequent decay terms. For the first The information content summary of each reproduced event is a string of no more than 200 characters, which is generated by the enterprise vertical big model by performing a summary on the original information, and is used for subsequent message dispatch and manual verification; For the first The current status field of each reproducible event is a discrete enumeration type that takes five candidate values: not closed loop, confirmed acceptance, completed achievement, no response, and rejection. It is initialized to not closed loop by default and is updated by subsequent closed loop feedback. It is used to drive the dynamic adjustment of the attenuation coefficient. For the first The basic weight of each reproducible event is a floating-point scalar, dimensionless, with a value range of [value missing]. The interval, initialized to 1.0 by default and dynamically adjusted by subsequent closed-loop feedback, is used to adjust the contribution strength of the event in the decay term; To carry issue fingerprints The total number of recurring events is a positive integer, dimensionless, and its value ranges from 1 to 2. The recurrence factor is obtained from real-time statistics of the issue fingerprint index library and is used to support the recurrence frequency of subsequent enhancements.

[0037] The essential difference between the topic recurrence trajectory described in this invention and traditional semantic deduplication is that traditional semantic deduplication only retains representative entries and discards the rest when the same topic appears multiple times, causing the statistical signal of the recurrence behavior itself to be permanently lost; the recurrence trajectory of this invention retains a complete observation record of each recurrence event, thus ensuring that the number of recurrences is accurately recorded. Statistics such as cross-platform distribution and time interval distribution can be directly used in subsequent steps, thereby transforming the reproduction behavior into observable and quantifiable physical evidence.

[0038] To support the urgency calculation in subsequent step S4, a cross-platform diversity index is maintained simultaneously during the issue reproduction trajectory tracking process. Consistency coefficient with the responsible party Two summary statistics. The cross-platform diversity index. Defined as the ratio of the number of different platform identifiers appearing in the reproduced event sequence to the total number of all platforms connected to the system, with a value range of [value missing]. The interval; the consistency coefficient of the responsible parties. Defined as the proportion of the number of events in a reproducible event sequence that share the most frequent responsible party value to the total number of reproducible events, with the value range also being [missing information]. Interval.

[0039] Step S4: Based on the time interval of the recurrence trajectory of the issue and the diversity of platforms, apply the recurrence decay-enhancement dynamics to calculate the unclosed loop urgency score.

[0040] The goal of this step is to transform the temporal event sequence of the issue recurrence trajectory into a single scalar unclosed-loop urgency score through dynamic modeling. This allows for a comparable global ranking of different issues based on this score, and enables unclosed-loop issues that recur multiple times across platforms to proactively emerge at the top of the ranking. The recurrence decay-enhancement dynamics are formed by a weighted fusion of decay and enhancement terms. The decay term performs a monotonically decreasing exponential decay based on the time interval between the recurring event and the current moment to reflect the time aging effect of information. The enhancement term constructs a monotonically increasing nonlinear gain based on the product of three factors: the number of recurrences, the cross-platform diversity index, and the consistency coefficient of the responsible party, to reflect the strength of this unclosed-loop signal that repeatedly occurs across platforms.

[0041] The formula for calculating the unclosed-loop urgency score is as follows:

[0042] ,

[0043] Among them, the attenuation term Defined as:

[0044] ,

[0045] Enhancement Defined as:

[0046] ,

[0047] The weight switching rule is defined as follows:

[0048] ,

[0049] in: Fingerprint for the issue At the present moment The unclosed-loop urgency score is a floating-point scalar, dimensionless, and its range falls within a certain range after normalization. The interval, directly calculated by this formula, is used to drive the triggering determination and priority sorting of subsequent action item closed-loop distribution; The definition is the same as the issue fingerprint of the issue reproduction trajectory in step S3 above; The timestamp is the current time, a floating-point scalar in seconds, with a value range typical of UNIX timestamps. It is provided by the system's real-time clock and is used to calculate the time interval for the decay term. The attenuation term is a floating-point scalar, dimensionless, with a range of values ​​of 1. The interval, calculated by the decay term formula, is used to reflect the weighting effect of recurring events as they expire over time. This is an enhancement term, a floating-point scalar, dimensionless, with a range of values ​​of [value range missing]. The interval, calculated by the enhancement term formula, is used to reflect the strength of the unclosed-loop signal that is reproduced multiple times across platforms; and These are the dynamic weights for the attenuation and enhancement terms, both of which are floating-point scalars, dimensionless, and have a range of values ​​of [missing value]. interval and satisfy The weight switching rule is based on the current The value is dynamically determined to enable dynamic switching between overdue and recurrence-dominated urgency assessment behavior; The cross-platform diversity index is defined in the same way as step S3 above, and is used to trigger weight switching. The cross-platform diversity index switching threshold is a floating-point scalar, dimensionless, with a value range of [value range missing]. The range, typically 0.5, is initialized by the system configuration and can be adjusted by the operations and maintenance side. If it is too large, it will be difficult to trigger the enhanced dominance even when reproducing cross-platform issues. If it is too small, it will cause expired issues on a single platform to be mistakenly triggered by the enhanced dominance. To enhance the weight of the enhancement item in the dominant mode, it is a floating-point scalar, dimensionless, with a typical value of 0.7. If it is too large, it will weaken the time expiration effect and make it difficult for old issues to decay; if it is too small, it will make the weight switching ineffective. The weight of the enhancement term in the decay-dominated mode is a floating-point scalar with no dimension. A typical value is 0.2. If it is too large, it will weaken the decay-dominated effect. If it is too small, it will completely eliminate the enhancement of issues that still exist on a single platform. The definition is the same as the basic weight of the reproduced event in step S3 above, and is dynamically adjusted by the closed-loop feedback in step S5. is the exponential decay coefficient, is a floating-point scalar, and its unit is . The range of values ​​is The interval typically takes the value of The corresponding half-life is about 10 days, which is determined by the typical urgency window of the enterprise's collaborative matters. If it is too large, the value of recent events will be rapidly depreciated, resulting in a distortion of the urgency judgment. If it is too small, old events will occupy the top position for a long time. The definition is the same as the timestamp of the event in step S3 above; The definition is the same as that of carrying issue fingerprints in step S3 above. The total number of recurring events; The recurrence frequency exponential factor is a floating-point scalar, dimensionless, with a range of values ​​of [value range missing]. The interval, typically taking a value of 0.5 corresponding to square root growth, is used to avoid extreme recurrence of issues suppressing other issues; and The definition is the same as in step S3 above; This is a cross-platform diversity index factor, a floating-point scalar, dimensionless, with a value range of [value range missing]. The interval, typically taking a value of 1.5, produces a significant nonlinear gain for cross-platform diversity; The consistency index factor of the responsible parties is a floating-point scalar, dimensionless, with a value range of [value range missing]. The interval typically takes a value of 1.0, corresponding to a linear gain.

[0050] The design motivation for the weight switching rule stems from the core insight of this invention: when the cross-platform diversity index... At this point, the issue has significantly spread across platforms, and the strength of the unclosed-loop signal carried by the reproducible behavior is significantly higher than the time expiration effect. Therefore, a higher weight should be assigned to the enhancement term so that the urgency increases with the increase of reproducibility. At this time, the issue is still limited to a single platform or a few platforms. The reproduction behavior may only be a normal follow-up within that platform rather than an unclosed-loop request for help. Therefore, a higher weight should be given to the decay term so that the urgency level decreases reasonably over time. This switching mechanism is the core algorithmic support for the technical effect of this invention, which is to obtain a higher urgency ranking for unclosed-loop issues reproduced multiple times across platforms than for expired issues on a single platform.

[0051] Normalization of unclosed-loop urgency scores: For all candidate issues in the global issue pool during system operation... The values ​​are normalized using min-max normalization, and the normalized values ​​fall within... The normalized unclosed-loop urgency score within the specified interval serves as the input for triggering action item assignment in subsequent step S5. When the normalized unclosed-loop urgency score exceeds the trigger threshold... At that time, the corresponding issue is incorporated into the closed-loop distribution process of action items; the trigger threshold The typical value is 0.6, which is dynamically adjusted by the enterprise's operations and maintenance side based on the processing capacity of collaborative tasks.

[0052] Step S5: Extract action items for the issue based on the unclosed-loop urgency score, distribute action items in combination with the reverse adaptability of the candidate accepting object, and write the acceptance status after distribution back to the issue reproduction trajectory as closed-loop feedback to adjust the subsequent reproduction attenuation coefficient.

[0053] The goal of this step is to transform the unclosed-loop urgency signal generated in the preceding steps into action items assigned to specific responsible parties, and then use a closed-loop feedback mechanism to write the assignment results back to the attenuation coefficient of the issue reproduction trajectory, thereby completing the closure of the entire collaborative closed loop. This step includes three tightly coupled sub-steps: action item extraction, reverse fit calculation and assignment selection, and closed-loop feedback writing.

[0054] The action item extraction sub-step extracts an executable action item description from the recurring event sequence of the corresponding issue based on the enterprise vertical big model. The input to the extraction process is a summary of the information content of all recurring events in the recurrence trajectory of the issue. The extraction process also includes the triplet key value of the issue fingerprint. The output of the extraction process is an action item description of no more than 150 characters, including the specific action to be completed, the specific objects involved, and the suggested completion deadline. The action item description is stored in the action item dispatch queue in the form of structured fields.

[0055] The reverse adaptation calculation sub-step determines the set of candidate receiving objects for each action item and calculates the reverse adaptation degree for each candidate object in the set. The formula for calculating the reverse adaptation degree is expressed as follows:

[0056] ,in: Candidate recipients Regarding the issue The inverse adaptivity is a floating-point scalar, dimensionless, with a range of values ​​of 1. The interval, directly calculated by this formula, is used to support the priority ranking of distribution selection; The index of the candidate receiving object is a positive integer, with a value range of 1. interval, For the topic The size of the corresponding set of candidate recipients is determined by the organizational role vector and the subject of the issue. This is the issue index, a positive integer used to identify the issue currently being processed; The sub-fraction index is a positive integer with a value range of 1. These correspond to four sub-fractions; For the first The weights of the sub-fractions are floating-point scalars, dimensionless, and their range is [value range missing]. interval and satisfy Typical value To reflect the paramount importance of historical timeliness of response; Candidate recipients Regarding the issue The historical response timeliness rate is a floating-point scalar, dimensionless, with a value range of [value missing]. The interval is defined as the proportion of the number of times the candidate has responded to the same type of issue within a specified time limit in its history to the total number of times it has received the same type of issue. It is obtained in real time by the enterprise action item closed-loop database and is used to reflect the reliability of the candidate in handling the same type of issue. Candidate recipients Regarding the issue The topic type acceptance rate is a floating-point scalar, dimensionless, with a value range of [value range missing]. The interval is defined as the proportion of the number of times the candidate has undertaken the same type of issue in history to the total number of times the candidate has received the same type of issue, which is used to reflect the candidate's willingness and ability to work on this type of issue. Candidate recipients The inverse value of the current workload saturation, is a floating-point scalar, dimensionless, and its value range is [value range missing]. An interval is defined as Subtract the ratio of the number of currently incomplete action items to the historical average number of action items processed for the candidate object. The higher the saturation, the lower the negative value, which is used to avoid assigning new action items to objects with excessive workload. Candidate recipients The affinity strength between the entity and the department of the responsible entity is a floating-point scalar, dimensionless, with a value range of [value missing]. The interval is defined as the normalized value of the reciprocal of the departmental distance between the candidate object and the responsible entity in the organizational structure. The value is 1.0 when the two belong to the same second-level department, 0.6 when they belong to the same first-level department but different second-level departments, and 0.2 when they belong to different first-level departments. It is used to reflect the ease of cross-departmental coordination between the two.

[0057] The dispatch selection rule is: select the candidate receiving object with the highest reverse adaptability. As the formal recipient, the message delivery includes a summary of the recurring event sequence and the urgency score for the unclosed loop, allowing the formal recipient to understand the basis for the complete cross-platform recurring event sequence and the urgency score upon receiving the delivery. At that time The typical value for the reverse fit threshold is 0.4. The action item assignment triggers a cross-departmental coordination process rather than direct assignment. The coordination path is determined by the superior manager of the responsible entity to avoid the high risk of non-response caused by low fit assignment.

[0058] The closed-loop feedback write-back sub-stage is the final closing stage of the collaborative closed loop of this invention. After receiving the dispatch message, the formal accepting entity writes back its acceptance status through the intelligent assistant interface. This acceptance status includes four enumerated values: confirmed acceptance, completed achievement, no response, and rejected. The write-back of the acceptance status triggers a basic weighting of the corresponding recurring event in the issue's recurrence trajectory. The dynamic adjustment, the adjustment rule is expressed as:

[0059] ,

[0060] in: The updated base weight for the corresponding reproducible event after the closed-loop feedback triggers a write-back is a floating-point scalar, dimensionless, with a value range of [value range missing]. The interval is directly calculated by this piecewise function and participates in the recalculation of the attenuation term in the subsequent step S4; The basic weights before the closed-loop feedback is triggered are defined in the same way as in step S3 above. ; The attenuation acceleration factor under the confirmed receiving state is a floating-point scalar, dimensionless, with a value range of [value missing]. The interval, typically 0.5, causes the unclosed loop urgency score to quickly drop back to half of its original value after acceptance and confirmation; The urgency amplification coefficient in the unresponsive state is a floating-point scalar, dimensionless, with a value range of [value missing]. The interval, typically 1.5, causes the unclosed loop urgency score to rise in the opposite direction to 1.5 times its original value after no response. The urgency enhancement coefficient under rejection status is a floating-point scalar, dimensionless, with a value range of [value range missing]. The interval, typically taking the value 2.0 and satisfying... This causes the urgency triggered by rejection to increase in intensity more strongly than a simple lack of response; The definition is the same as the current state field of the reproduced event in step S3 above.

[0061] The redistribution trigger rule is as follows: when the acceptance status of a reproducible event is written back as unresponsive or rejected, and the urgency score of the unclosed loop of the issue, after being recalculated after the write-back, exceeds the redistribution threshold. When the typical value is 0.75, the redistribution process for the issue is triggered. The redistribution process will exclude the previously redistributed formal recipient from the candidate recipient set, re-execute the reverse adaptability calculation, and select a new formal recipient.

[0062] Through the three tightly coupled sub-steps of action item extraction, reverse adaptation calculation and dispatch selection, and closed-loop feedback write-back, this invention achieves a complete collaborative closed loop that dynamically adjusts the reproduction attenuation coefficient from the initial closed-loop urgency signal to the dispatch of specific action items and then to the write-back of the acceptance status. This complete collaborative closed loop enables the entire system to possess a learning-adaptive capability where the dispatch results influence the urgency calculation in reverse, thereby continuously optimizing the dispatch hit rate.

[0063] Furthermore, the method embodiments of the present invention also include a sub-process for processing natural language commands: receiving natural language commands input by the user, performing intent recognition on the natural language commands based on the enterprise vertical big model, and scheduling corresponding capabilities in multi-platform information access, issue recurrence trajectory query, or action item assignment according to the identified target intent, replacing the menu-style operations in traditional office systems with natural language commands. The candidate intent set for intent recognition includes intent types corresponding to typical enterprise collaboration scenarios such as tracking the progress of an issue, querying one's own pending action items, querying the current urgent issues of a responsible entity, and initiating cross-departmental coordination requests; the intent recognition results are directly connected to the issue fingerprint index library and the action item assignment queue to drive the rapid invocation of corresponding capabilities.

[0064] like Figure 2 As shown, this embodiment of the invention also provides a large model-driven intelligent assistant system for enterprise collaborative office work. The system includes an enterprise vertical large model construction module 1, a multi-platform information access module 2, an issue reproduction trajectory tracking module 3, an unclosed loop urgency assessment module 4, and an action item closed loop distribution module 5. Each module forms a deeply coupled collaborative closed loop according to the data flow direction and is deployed in the enterprise intranet private computing environment.

[0065] The Enterprise Vertical Large Model Construction Module 1 is deployed within the enterprise's internal GPU computing cluster, providing an enterprise vertical large model inference API for other modules to call. The Enterprise Vertical Large Model Construction Module 1 includes a corpus acquisition submodule, an incremental pre-training submodule, a supervised fine-tuning submodule, an organizational role embedding submodule, and a continuous fine-tuning scheduling submodule. The corpus acquisition submodule extracts domain corpora in batches by connecting to the enterprise's knowledge base system, regulations system, and historical project archive, and performs preprocessing such as watermark removal, unified encoding, removal of personal privacy fields, and segmentation according to knowledge density. The incremental pre-training submodule performs unsupervised domain incremental pre-training on the basic large language model based on the preprocessed corpora. The supervised fine-tuning submodule performs LoRA low-rank adaptation supervised fine-tuning on the dataset based on instruction-response pairs annotated by domain experts. The organizational role embedding submodule connects to the enterprise's human resources system to extract structured attributes such as employee department affiliation, job responsibilities, reporting relationships, and skill tags, and encodes them into 128-dimensional or 256-dimensional dense vectors, inserting them before the input embedding layer of the enterprise vertical large model. The continuous fine-tuning scheduling submodule triggers incremental fine-tuning and synchronizes the role vector updates according to a set period. The output of the enterprise vertical large model construction module 1 serves as the unified intelligent foundation for all subsequent modules.

[0066] The multi-platform information access module 2 constructs a unified access layer based on the open APIs of the enterprise's existing office suite. It connects to email platforms, instant messaging platforms, document collaboration platforms, and meeting recording streams, converting the raw information streams from each platform into a unified JSON Schema format carrying fields such as platform identifier, timestamp, sender, receiver, content body, and attachment references. The multi-platform information access module 2 further calls the inference API of the enterprise's vertical large-scale model construction module 1 to generate an issue fingerprint for each uniformly formatted message, including three fields: issue subject, action type, and responsible party. The issue fingerprint, along with the raw information, is written into the issue fingerprint index library.

[0067] The issue recurrence trajectory tracking module 3 subscribes to new events in the issue fingerprint index library, and aggregates multi-source information sharing the same issue fingerprint triple key value into an issue recurrence trajectory in ascending order of occurrence timestamp. This issue recurrence trajectory retains a complete observation record of each recurrence event without performing content deduplication. The issue recurrence trajectory tracking module 3 simultaneously maintains two summary statistics: a cross-platform diversity index and a responsibility subject consistency coefficient, and provides them to the downstream non-closed-loop urgency assessment module 4. The issue recurrence trajectory tracking module 3 also subscribes to the acceptance status write-back events of the action item closed-loop dispatch module 5, and dynamically adjusts the basic weights of the corresponding recurrence events in the issue recurrence trajectory based on the write-back status.

[0068] The unclosed-loop urgency assessment module 4 retrieves the issue reproduction trajectory and its summary statistics from the issue reproduction trajectory tracking module 3. It applies reproduction decay-enhancement dynamics to all unclosed-loop issues at a set calculation period, typically every 5 minutes or every 15 minutes, to calculate the unclosed-loop urgency score. It then performs min-max normalization on the unclosed-loop urgency scores of all candidate issues in the global issue pool to obtain a normalized unclosed-loop urgency score. When the normalized unclosed-loop urgency score of an issue exceeds a trigger threshold, the unclosed-loop urgency assessment module 4 pushes the issue, its reproduction trajectory, and its unclosed-loop urgency score to the downstream action item closed-loop dispatch module 5.

[0069] The action item closed-loop dispatch module 5 receives high-urgency issues and their recurrence trajectories pushed by the unclosed-loop urgency assessment module 4. It calls the inference API of the enterprise vertical large model construction module 1 to extract an executable action item description from the recurring event sequence. Based on the organizational role vector and the issue subject, it determines a set of candidate recipients and calculates the reverse fit for each candidate. The candidate with the highest reverse fit is selected as the formal recipient and dispatched. The dispatch message carries a summary of the recurring event sequence and the unclosed-loop urgency score for this issue. The action item closed-loop dispatch module 5 receives the acceptance status written back by the formal recipient through the intelligent assistant interface. It writes the acceptance status back to the issue recurrence trajectory tracking module 3 as closed-loop feedback to adjust the subsequent recurrence attenuation coefficient. When the acceptance status of a recurring event is written back as unresponsive or rejected and the unclosed-loop urgency score of the issue is higher than the redispatch threshold after recalculation, the redispatch process is triggered to exclude the previous formal recipient from the candidate recipient set, re-execute the reverse fit calculation, and select a new formal recipient.

[0070] The data flow direction between modules is as follows: the enterprise vertical large model construction module 1 uses the enterprise vertical large model as the inference base for other modules to call on demand; the topic fingerprint and raw information of the multi-platform information access module 2 are fed into the topic reproduction trajectory tracking module 3; the topic reproduction trajectory and summary statistics of the topic reproduction trajectory tracking module 3 are fed into the non-closed-loop urgency assessment module 4; the high-urgency topics and non-closed-loop urgency scores of the non-closed-loop urgency assessment module 4 are fed into the action item closed-loop distribution module 5; the acceptance status of the action item closed-loop distribution module 5 is used as closed-loop feedback and written back to the topic reproduction trajectory tracking module 3 to complete the closure of the entire collaborative closed loop. In addition, the multi-platform information access module 2 also performs automatic speech recognition on the meeting recordings through the meeting recording preprocessing sub-process to obtain the meeting text, generates a meeting summary based on the enterprise vertical large model, and extracts meeting action items. The meeting action items are injected into the multi-source information flow with a triple key-value structure consistent with the topic fingerprint to participate in the topic reproduction trajectory aggregation.

[0071] The hardware deployment of the large-model-driven enterprise collaborative office intelligent assistant system is as follows: An internal GPU computing cluster provides training and inference computing power for the enterprise vertical large-model construction module 1, typically configured as 4 to 8 physical machines equipped with Nvidia A100 or H100 GPUs; an internal application server cluster deploys a multi-platform information access module 2, an issue reproduction trajectory tracking module 3, an unclosed-loop urgency assessment module 4, and an action item closed-loop dispatch module 5, typically configured as CPU computing nodes combined with distributed message queues and distributed key-value storage; the issue fingerprint index library and issue reproduction trajectory storage library are implemented using an Elasticsearch cluster and a time-series database, respectively; employees access the intelligent assistant interface through a browser, WeChat / DingTalk / Lark sidebar plugin, or an internal enterprise instant messaging client. The intelligent assistant interface supports natural language command input, issue reproduction trajectory query, action item to-do list viewing, and status write-back, among other interactive capabilities.

[0072] To more clearly demonstrate the collaborative closed-loop effect of this invention, the following example illustrates an end-to-end implementation scenario: A project manager requests confirmation of a product prototype review time from the R&D head via internal instant messaging, but the R&D head does not reply at the time. Three days later, the project manager sends the same request to the R&D department again via email, explicitly referring to the same product prototype. Two days later, a marketing colleague inquires about the same review progress in the project collaboration document's comment section. These three cross-platform events on the same topic are uniformly accessed by the multi-platform information access module 2 and generated by the enterprise's vertical big data model, resulting in identical topic fingerprint triples (topic subject: product prototype review; action type: decision request; responsible entity: R&D head). The topic recurrence trajectory tracking module 3 aggregates the three cross-platform, cross-time period events into a single topic recurrence trajectory instead of performing deduplication and erasing two of them. The unclosed-loop urgency assessment module 4, based on the cross-platform diversity index exceeding the switching threshold across three different platforms, triggers an enhanced dominant mode, significantly increasing the unclosed-loop urgency score above the trigger threshold. The action item closed-loop dispatch module 5 extracts the action item description as "Please confirm the product prototype review time," calculates the reverse adaptability of the R&D manager and their subordinate technical directors, and finds that the R&D manager's current uncompleted action items are saturated, while the subordinate technical directors have high departmental relevance and a high historical response time rate. Finally, the action item is dispatched to the subordinate technical director with a summary of the reproduced event sequence to let them know that this issue has been unclosed three times, spanning 3 platforms and 5 working days. The subordinate technical director confirms acceptance and completes the review time confirmation within 24 hours. The acceptance status is written back as "completed," and the issue reproduction trajectory tracking module 3 updates the status field of the three reproduced events to "closed." The unclosed urgency score is removed from the global issue pool sorting, and the entire collaborative issue completes automatic closed-loop. This typical scenario demonstrates the collaborative efficiency improvement of this invention compared to traditional semantic deduplication paradigms: In traditional deduplication schemes, three cross-platform messages are aggregated into a single representative entry and repeatedly pushed to the R&D manager in one direction. Due to the inherent inertia of being ignored after the first non-response, the R&D manager continues not to respond, and multiple cross-platform inquiries from project managers and marketing colleagues go unanswered. In contrast, this invention retains the sequence of reproducible events and automatically reassigns them to the subordinate technical director through reverse adaptation, thus closing the loop in one fell swoop.

[0073] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A large-model-driven intelligent assistant method for enterprise collaborative office work, characterized in that: include: By using the enterprise's internal knowledge base, rules and regulations, and historical project documents as domain corpora, the basic large language model is continuously fine-tuned and organizational role vectors are embedded to obtain the enterprise vertical large model. By aggregating multi-source information flows from email, instant messaging, and document collaboration platforms through a unified access layer, and generating an issue fingerprint for each message based on the enterprise vertical big model, the issue fingerprint includes the issue subject, action type, and responsible party. Multi-source information carrying the same issue fingerprint is aggregated in chronological order to form an issue recurrence trajectory. The issue recurrence trajectory replaces semantic deduplication and retains the recurrence event sequence across platforms and time periods. Based on the time intervals and platform diversity of the recurrence trajectories of the issues, a recurrence decay-enhancement dynamics is applied to calculate the urgency score of unclosed loop issues. The recurrence decay-enhancement dynamics includes a decay term and an enhancement term. The decay term performs monotonically decreasing exponential decay based on the time interval between each event in the recurrence event sequence and the current time. The enhancement term constructs a monotonically increasing nonlinear gain based on the product of three factors: the number of recurrences in the recurrence event sequence, the cross-platform diversity index, and the consistency coefficient of the responsible parties. The urgency score of unclosed loop issues is a weighted fusion of the decay term and the enhancement term. When the cross-platform diversity index is higher than a set threshold, the weight of the enhancement term is higher than the weight of the decay term, so that unclosed loop issues that recur multiple times across platforms receive a higher urgency ranking than expired issues on a single platform. Based on the unclosed-loop urgency score, action items are extracted for the issue, and action items are assigned in combination with the reverse adaptability of the candidate recipient. The assigned acceptance status is then written back to the issue reproduction trajectory as closed-loop feedback to adjust the subsequent reproduction attenuation coefficient. The calculation of the reverse fit includes: for the set of candidate recipients extracted from the action item, calculating four sub-scores for each candidate recipient: historical response timeliness rate, issue type acceptance rate, current workload saturation, and departmental correlation strength with the responsible entity. The four sub-scores are normalized and weighted to obtain the reverse fit. The action item dispatch selects the candidate recipient with the highest reverse fit as the formal recipient and carries a summary of the recurring event sequence and the urgency score of the unclosed loop in the dispatch message. When the reverse fit of the formal recipient is lower than a set threshold, the action item dispatch triggers a cross-departmental coordination process instead of direct dispatch.

2. The method according to claim 1, characterized in that, The generation of the issue fingerprint includes: The enterprise vertical big model performs named entity recognition on each piece of information to obtain the subject of the issue, performs action type classification to obtain one of the following as the action type: action obligation, information sharing, status reporting or decision request, and identifies the responsible subject to which the information initiator points based on the organizational role vector. The topic subject, the action type, and the responsible party together constitute the triplet key value of the topic fingerprint, serving as the unique anchor identifier for determining the same topic across platforms and time periods.

3. The method according to claim 2, characterized in that, The aggregation rule for the issue reproduction trajectory is as follows: For multiple pieces of information obtained across platforms and time periods, the triplet key value of the issue fingerprint is used as the aggregation key. Information that shares the same triplet key value is not deduplicated, but is arranged in ascending order by the timestamp of occurrence to form a sequence of reproducible events. Each event in the reproduced event sequence retains the platform identifier, timestamp, information content summary, and current status field. The current status field is marked as not closed by default and is updated by subsequent closed-loop feedback.

4. The method according to claim 1, characterized in that, The closed-loop feedback adjusts the reproduction attenuation coefficient as follows: When the acceptance status is written back as confirmed acceptance, the attenuation term coefficient of the corresponding issue fingerprint is accelerated to cause the unclosed loop urgency score to drop rapidly. When the acceptance status is written back as completed, the recurrence event sequence of the corresponding issue fingerprint is marked as closed and the calculation of the unclosed urgency score is exited. When the acceptance status is written back as unresponsive or rejected, the enhancement coefficient of the corresponding issue fingerprint is further increased to reverse the unclosed loop urgency score and trigger the redistribution process.

5. The method according to claim 1, characterized in that, The multi-source information stream also includes a conference recording stream, and the processing of the conference recording stream includes: Automatic speech recognition is performed on the meeting recording to obtain the meeting text. Based on the enterprise vertical big model, a meeting summary is generated from the meeting text and meeting action items are extracted. The meeting action items are injected into the multi-source information stream in a triple key-value structure consistent with the topic fingerprint to participate in the topic reproduction trajectory aggregation.

6. The method according to claim 1, characterized in that, The method further includes: The system receives natural language commands input by the user, performs intent recognition on the natural language commands based on the enterprise vertical big model, and schedules corresponding capabilities in multi-platform information access, issue reproduction trajectory query, or action item assignment according to the identified target intent, replacing menu-style operations with natural language commands.

7. The method according to claim 1, characterized in that, The basic large language model is deployed in a private environment on the enterprise intranet. The continuous fine-tuning is performed incrementally based on newly added internal enterprise documents and historical interaction data according to a set period. The organizational role vector is updated synchronously with changes in the enterprise's organizational structure.

8. A large-scale model-driven intelligent assistant system for enterprise collaborative office work, used to implement the method described in any one of claims 1-7, characterized in that, include: The Enterprise Vertical Model Construction Module is used to continuously fine-tune the basic large language model using the enterprise's internal knowledge base, rules and regulations, and historical project documents as domain corpora, and to embed organizational role vectors to obtain the enterprise vertical model. The multi-platform information access module is used to aggregate multi-source information flows from email, instant messaging and document collaboration platforms through a unified access layer, and generate issue fingerprints for each piece of information based on the enterprise vertical big model. The issue fingerprint includes the issue subject, action type and responsible party. The topic recurrence trajectory tracking module is used to aggregate multi-source information carrying the same topic fingerprint into a topic recurrence trajectory in chronological order. The topic recurrence trajectory replaces semantic deduplication and retains the recurrence event sequence across platforms and time periods. The unclosed-loop urgency assessment module is used to calculate the unclosed-loop urgency score based on the time interval of the issue's recurrence trajectory and platform diversity, applying recurrence decay-enhancement dynamics. The recurrence decay-enhancement dynamics includes a decay term and an enhancement term. The decay term performs monotonically decreasing exponential decay based on the time interval between each event in the recurrence event sequence and the current time. The enhancement term constructs a monotonically increasing nonlinear gain based on the product of three factors: the number of recurrences in the recurrence event sequence, the cross-platform diversity index, and the consistency coefficient of the responsible parties. The unclosed-loop urgency score is a weighted fusion of the decay term and the enhancement term. When the cross-platform diversity index is higher than a set threshold, the weight of the enhancement term is higher than the weight of the decay term, so that unclosed-loop issues that recur multiple times across platforms receive a higher urgency ranking than expired issues on a single platform. The action item closed-loop dispatch module is used to extract action items for issues based on the unclosed-loop urgency score, dispatch action items in combination with the reverse fit of candidate recipients, and write back the acceptance status after dispatch as closed-loop feedback to the issue reproduction trajectory tracking module to adjust the subsequent reproduction attenuation coefficient. The calculation of the reverse fit includes: for each candidate recipient set obtained from the action item extraction, calculating four sub-scores for each candidate recipient: historical response timeliness rate, issue type acceptance rate, current workload saturation, and the strength of association with the responsible entity within the same department. These four sub-scores are normalized and weighted to obtain the reverse fit. The action item dispatch selects the candidate recipient with the highest reverse fit as the formal recipient, and carries a summary of the reproduction event sequence and the unclosed-loop urgency score in the dispatch message. When the reverse fit of the formal recipient is lower than a set threshold, the action item dispatch triggers a cross-departmental coordination process instead of direct dispatch.

Images