Enterprise consultation intelligent question and answer system based on large language model

Abstract

The present application relates to the technical field of computer data processing, in particular to an enterprise consultation intelligent question and answer system based on a large language model. It comprises: extracting structured static operating characteristic parameters based on instruction messages; a threshold analysis unit recalls policy texts and constructs a qualification examination threshold matrix; a bottom layer hardware comparison is performed, and a time lag rejection identifier is configured when the numerical comparison fails; based on the time lag rejection identifier, the edge computing power consumption drift rate is calculated, the total R&D expenses are compensated for feature weight to obtain dynamic R&D parameters; the dynamic R&D parameters and the threshold matrix are used to perform secondary comparison, the comparison results and the drift rate are synchronously input into a generative language model, and dynamic compliance conclusion text data streams are output. The present application uses physical computing power to dynamically compensate for static account time lag attenuation, and intervenes in the attention mechanism of the large model through tensor bias constraint, completely establishes physical state binding, and guarantees the reliability and engineering landing of the compliance conclusion.

Description

Technical Field

[0001] This invention relates to the field of computer data processing technology, and more specifically, to an intelligent question-and-answer system for enterprise consultation based on a large language model. Background Technology

[0002] As enterprises delve deeper into digital transformation, retrieval enhancement generation systems based on large language models are being widely deployed on public service platforms in comprehensive consulting services, including high-tech enterprise certification, CMMI system assessment, and system cloud architecture migration based on specific ecosystems. These systems are used to parse natural language queries and output customized consulting solutions.

[0003] In the context of composite assessment, existing systems forcibly map the enterprise's underlying structured operational data (such as the proportion of R&D expenses) and unstructured policy guidelines into the same high-dimensional vector space. This mechanism fails to establish rigid constraints between "discrete entity data" and "continuous policy thresholds," making it easy for the weights of structured engineering parameters to be diluted by unstructured semantics, leading to recall distortion and logical breaks in compliance assessment.

[0004] Secondly, the rigid interception mechanism of static accounting sets introduced to compensate for the aforementioned vector dilution defects can lead to secondary defects at the execution level: there is an inherent time lag between the enterprise's actual physical R&D activities (such as cloud integration testing of edge nodes) and static financial accounting settlements. Relying solely on historical static accounting data as a comparison benchmark can result in misjudgments and rejections of genuinely occurring high-frequency R&D activities during compliance assessments. Existing systems lack dynamic compensation methods based on the underlying real computing power load status, causing the compliance conclusions output by the model to lack causal traceability and dynamic entity binding. Therefore, there is an urgent need for an intelligent question-and-answer system for enterprise consultation based on a large language model. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent question-and-answer system for enterprise consultation based on a large language model, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention aims to provide an intelligent question-and-answer system for enterprise consultation based on a large language model, including a status acquisition unit. The status acquisition unit is deployed in the physical memory of the application server. It receives natural language consultation instruction messages based on a time-constrained gateway and extracts structured static operational feature parameters by triggering the physical network card to address and call the underlying hardware database of the ERP system associated with the target enterprise. Among them, the natural language consultation instruction message contains the target company identification information and the intent of compliance inquiry; The threshold parsing unit is deployed in an AI inference server equipped with a graphics processor. It is used to receive the natural language consultation instruction message, determine the target qualification category label, recall unstructured policy text that matches the target qualification category label from a distributed vector database pre-installed on a solid-state drive, perform semantic segmentation extraction on the unstructured policy text, construct a qualification review threshold matrix, and write it into the cache system. The static evaluation unit is used to receive the structured static operational feature parameters and the qualification review threshold matrix, and generate a first static compliance evaluation vector data stream containing a Boolean state identifier sequence based on the comparison result by triggering the underlying hardware comparison instruction. When any numerical comparison fails, the corresponding Boolean state identifier is configured as a time-delay rejection identifier, and the first static compliance evaluation vector data stream is input to the system memory bus. The compliance generation unit, based on the time-delay rejection identifier, listens to the physical computing power load status of the underlying edge computing nodes and calculates the edge computing power consumption drift rate. It then uses the edge computing power consumption drift rate as a dynamic weight factor to compensate the total R&D expenses for feature weights, thereby obtaining dynamic R&D parameters. The unit performs a secondary comparison between the dynamic R&D parameters and the qualification review threshold matrix, and simultaneously inputs the comparison result and the edge computing power consumption drift rate into the generative language model for decoding, outputting a dynamic compliance conclusion text data stream.

[0007] As a further improvement to this technical solution, the status acquisition unit includes a consultation instruction receiving module and a data capture module; The consultation instruction receiving module is used to receive the natural language consultation instruction message and parse the time-series review sliding window parameters by calling the time-named entity recognition algorithm in the microprocessor. The data capture module is used to generate SQL query messages with hard time constraints based on the time-series review sliding window parameters, send query requests to the underlying hardware database through the API gateway, capture the corresponding financial general ledger account table and R&D project initiation ledger data, and extract structured static operational characteristic parameters; wherein, the structured static operational characteristic parameters include the total annual R&D expenses stored on the disk medium. Number of intellectual property rights and the most recent main business revenue .

[0008] As a further improvement to this technical solution, the threshold parsing unit includes an intent classification routing module and a policy text regular expression extraction module; The intent classification routing module is used to send the natural language consultation instruction message into a preset large language model, extract the query entity features through an attention mechanism, output the corresponding intent category, and match and map it to the target qualification category label. The policy text regular expression extraction module is used to perform regular expression matching and text block parsing on unstructured policy texts retrieved from a distributed vector database that match the target qualification category tags, remove modifying semantic features, and filter out the lower limit threshold of R&D expense ratio with physical review constraints. With the lower limit threshold of the number of intellectual property rights Construct a qualification review threshold matrix consisting of continuous numerical boundaries.

[0009] As a further improvement to this technical solution, the neural network architecture of the pre-set large language model includes a word embedding layer, a multi-head self-attention encoder, and a fully connected classification layer; then the intent classification routing module inputs the natural language consultation instruction message into the word embedding layer, outputs the corresponding intent category, and matches and maps it to the target qualification category label. The specific data flow steps involved are as follows: The word embedding layer receives the natural language consultation instruction message, calls a preset word segmentation dictionary to parse it into a discrete token sequence, and maps the discrete token sequence into a continuous high-dimensional input vector sequence; The multi-head self-attention encoder receives the high-dimensional input vector sequence, constructs a query matrix, a key matrix, and a value matrix based on the high-dimensional input vector sequence, obtains the attention weight distribution by calculating the inner product of the query matrix and the key matrix, and uses the attention weight distribution to perform a weighted summation on the value matrix to extract the contextual semantic feature matrix representing the compliant query entity. The fully connected classification layer performs a linear transformation on the context semantic feature matrix, inputs a normalized exponential function (Softmax) for probabilistic activation, and outputs a probability distribution vector corresponding to the predetermined compliant business space. The microprocessor is invoked to extract the classification node corresponding to the maximum likelihood probability in the probability distribution vector as the corresponding intent category, and an addressing instruction is triggered to perform key-value matching in the intent mapping hash table of static random access memory, rigidly replacing the intent category with a unique machine-readable hexadecimal code to obtain the target qualification category label.

[0010] As a further improvement to this technical solution, the policy text regular expression extraction module extracts the lower limit threshold and constructs the qualification review threshold matrix. The specific steps involved are as follows: Based on the unstructured policy text retrieved from the distributed vector database that matches the target qualification category label, it is loaded into the microprocessor's memory buffer and segmented into multiple discrete syntactic logic blocks based on preset punctuation identifiers. Feature traversal is performed on each of the syntactic logic blocks based on a pre-set multimodal regular expression constraint template; wherein, the multimodal regular expression constraint template includes a first regular expression template and a second regular expression template, and each of the first regular expression template and the second regular expression template contains text prefix anchors for filtering modifying semantics and regular expression capture groups for locking numerical boundaries; The first regular expression template performs character traversal on each of the syntactic logic blocks. When it matches the text prefix anchor representing the R&D funding constraint, it triggers the corresponding regular expression capture group to extract the floating-point string containing the percent sign, and calls the type casting function to translate it into a double-precision floating-point number, which is used as the lower limit threshold of the R&D expense ratio. ; The second regular expression template performs character traversal on each of the syntactic logic blocks. When it matches the text prefix anchor representing intellectual property constraints, it triggers the corresponding regular expression capture group to extract continuous Arabic numeral strings and calls a type casting function to translate them into unsigned integer values, which are used as the lower limit threshold for the number of intellectual property rights. ; Extract a contiguous physical address space from the system memory bus and set the lower limit threshold for the R&D expense ratio. With respect to the lower limit threshold of the number of intellectual property rights According to the preset memory offset, the values ​​are sequentially pushed into the continuous physical address space to construct the qualification review threshold matrix for direct use by subsequent logic gates.

[0011] As a further improvement to this technical solution, the static evaluation unit includes a heterogeneous feature mapping module and a rigid numerical comparison module; The heterogeneous feature mapping module is used to establish a data type alignment and physical addressing mapping channel between the structured static operational feature parameters and the qualification review threshold matrix between system memory and registers, namely the heterogeneous mapping channel; The rigid numerical comparison module is used to extract the corresponding feature parameters and threshold factors along the heterogeneous mapping channel to perform machine-level Boolean comparison. When the numerical comparison of the R&D investment dimension fails, the corresponding Boolean status identifier is forcibly configured as the time-delay rejection identifier to generate the first static compliance evaluation vector data stream and send the data stream into the system memory bus.

[0012] As a further improvement to this technical solution, the rigid numerical comparison module performs machine-level Boolean comparison along the heterogeneous mapping channel, involving the following specific steps: The total annual R&D expenses extracted from the structured static operational characteristic parameters along the heterogeneous mapping channel. Compared with the most recent main business revenue And the lower limit threshold of R&D expense ratio extracted from the qualification review threshold matrix. Determine the total amount of annual R&D expenses collected. Compared with the most recent main business revenue Is the ratio greater than or equal to the lower limit threshold for the proportion of R&D expenses? ; Simultaneously, the rigid numerical comparison module extracts the number of intellectual property rights from the structured static operational characteristic parameters along the heterogeneous mapping channel. And extract the lower limit threshold for the number of intellectual property rights from the qualification review threshold matrix. Determine the number of intellectual property rights. Is it greater than or equal to the lower limit threshold for the number of intellectual property rights? ; If the total annual R&D expenditure is determined to be discrete Compared with the most recent main business revenue The ratio is lower than the lower limit threshold for the proportion of R&D expenses. If the historical static financial accounting data has a physical lag defect in representing the R&D status, then the corresponding Boolean status identifier is used as the time-lag rejection identifier to generate the first static compliance assessment vector data stream.

[0013] As a further improvement to this technical solution, the compliance generation unit includes an edge computing power compensation module and an offset intervention decoding module; The edge computing power compensation module establishes a network hardware addressing path with the target node based on the time delay rejection identifier. Within the time-series review sliding window parameter, it listens in real time to the physical computing power load status of the target node to solve and construct the edge computing power consumption drift rate. The edge computing power consumption drift rate is used as a dynamic weight factor to perform nonlinear feature weight compensation on the total R&D expenses in the structured static operation feature parameters, and outputs the dynamic R&D parameters bound to the physical state. The bias intervention decoding module is used to receive the dynamic R&D parameters, call the microprocessor to re-perform machine-level secondary comparison with the qualification review threshold matrix, and instantiate a physical bias weight matrix based on the edge computing power consumption drift rate mapping. This matrix is ​​used to apply hardware-level bias intervention constraints to the autoregressive decoding calculation of the preset generative language model. Finally, the generative language model autoregressively generates and outputs the dynamic compliance conclusion text data stream supported by the physical computing power evidence chain.

[0014] As a further improvement to this technical solution, the edge computing power compensation module outputs dynamic R&D parameters bound to the physical state, and the specific steps involved are as follows: Based on the hardware interrupt signal triggered by the time-delay rejection flag, the network topology routing calibration code corresponding to the target enterprise identification information is extracted, and the disk read mechanism is bypassed to directly activate the hardware performance counter on the corresponding underlying edge computing node motherboard. Within the time interval defined by the timing review sliding window parameter, high-frequency sampling is used to listen for and capture the number of concurrent CPU threads in active running state of the underlying edge computing node. Variance of concurrent requests per second with cloud API hardware gateway ; The number of concurrent CPU threads is controlled via the data bus. The variance of concurrent requests per second compared to the cloud API hardware gateway Write the data into the corresponding temporary data register, call the arithmetic logic unit to perform optimization and solve the problem, and construct the edge computing power consumption drift rate. ; The edge computing power consumption drift rate As the dynamic weighting factor, and by invoking a pre-set nonlinear exponential smoothing compensation equation, the extracted total annual R&D expenditure is aggregated. Perform numerical compensation to compensate for the time-delay decay of static accounting data through exponential changes in physical computing load, and output the dynamic R&D parameters. .

[0015] As a further improvement to this technical solution, the specific steps involved in the bias intervention decoding module performing a secondary comparison and outputting the dynamic compliance conclusion text data stream are as follows: The arithmetic logic unit of the microprocessor is invoked to process the dynamic research and development parameters. Compared with the most recent main business revenue Perform a floating-point division operation to generate a dynamic R&D ratio, and then compare the dynamic R&D ratio with the lower limit threshold of R&D expense ratio in the qualification review threshold matrix. Execute machine-level Boolean comparison instructions; Based on the logical judgment result of the machine-level Boolean comparison instruction, the status register is triggered to perform the corresponding physical level flip to output the secondary comparison pass status result represented by high and low levels, and the original time-delay rejection flag is overwritten as the secondary comparison success status code. The pre-built generative language model deployed in the graphics processor memory array is invoked, with the edge computing power consuming drift rate. As a decay control variable, the mapping instantiation generates a dimension-matched physical bias weight matrix. Wherein, when the edge computing power consumption drift rate The closer the physical bias weight matrix is ​​to the upper limit of its effective range, the more significant its bias becomes. Internally, it corresponds to the key matrix associated with the failed accounting dimension in the structured static operational characteristic parameters. The greater the weight of the negative penalty applied; The successful secondary comparison status code is compared with the dynamic R&D parameters. The input is fed into the generative language model, and the underlying tensor operation unit is invoked to process the physical bias weight matrix. Tensor multiplication is performed, and the zero-value suppression characteristic of the negative penalty weights is used to autoregressively decode the generative language model to output the dynamic compliance conclusion text data stream supported by the underlying physical computing power evidence chain.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. In this enterprise consultation intelligent question-and-answer system based on a large language model, a single-cycle machine-level Boolean comparison is performed along the heterogeneous mapping channel to generate a first static compliance assessment vector data stream containing a Boolean state identifier sequence; an absolute rigid constraint is established between the discrete values ​​of business entities and the continuous threshold boundary of policies, thereby blocking the logical discontinuity in compliance assessment caused by the dilution of structured engineering parameters by unstructured semantics during the retrieval stage.

[0017] 2. In this enterprise consultation intelligent question-and-answer system based on a large language model, the time-delay rejection flag is used as a hardware-level interrupt signal to forcibly interrupt the edge computing power compensation calculation branch. The edge computing power consumption drift rate is obtained through optimization. The edge computing power consumption drift rate is used as a dynamic weighting factor and substituted into the nonlinear exponential smoothing compensation equation to perform numerical compensation on the total annual R&D expense collection. Through the exponential change of real-time physical computing power load, the time-delay decay of static accounting data is accurately compensated, and the compliance comparison benchmark is rigidly bound to the actual physical entity state.

[0018] 3. In this enterprise consulting intelligent question-and-answer system based on a large language model, the underlying tensor operation unit performs operations on the physical bias weight matrix with physical bias constraints. By utilizing the zero-value suppression characteristic generated by applying negative penalty weights to the key matrix associated with the failed accounting dimension, the system forcibly deprives the model of computing resources for old and lagging accounting data. This allows the generative language model to focus its attention entirely on dynamic R&D parameters, ultimately outputting a dynamic compliance conclusion text data stream supported by the underlying physical computing power evidence chain. This thoroughly establishes a causal tracing and state binding mechanism between the model's output character sequence and the real physical cloud resource configuration status. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the overall process of the present invention.

[0020] The meanings of the labels in the diagram are as follows: 1. Status acquisition unit; 11. Consultation instruction receiving module; 12. Data capture module; 2. Threshold parsing unit; 21. Intent classification and routing module; 22. Policy text regular expression extraction module; 3. Static evaluation unit; 31. Heterogeneous feature mapping module; 32. Rigid numerical comparison module; 4. Compliance generation unit; 41. Edge computing power compensation module; 42. Bias intervention decoding module. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 As shown, an intelligent question-and-answer system for enterprise consultation based on a large language model is provided, including a status acquisition unit 1. The status acquisition unit 1 is deployed in the physical memory of the application server. It receives natural language consultation instruction messages based on a time-constrained gateway and calls the underlying hardware database of the ERP system associated with the target enterprise through the physical network card to extract structured static operational feature parameters. Among them, the natural language consultation instruction message contains the target company identification information and the intent of compliance inquiry; In this embodiment, the state acquisition unit 1 is a hardware execution entity invoked by the microprocessor to convert front-end fuzzy natural language instructions into an engineering addressing control flow that can be directly executed by the back-end underlying hardware. The state acquisition unit 1 resides in the physical memory of the application server. Based on the timing constraint gateway, it receives natural language consultation instruction messages containing target enterprise identification information and compliance inquiry intent, and calls the underlying hardware database of the ERP system associated with the target enterprise through the physical network card to capture and extract the total annual R&D expenses. Number of intellectual property rights and the most recent main business revenue The structured static operational characteristic parameters constituted.

[0023] Specifically, the status acquisition unit 1 includes an inquiry instruction receiving module 11 and a data acquisition module 12; The consultation instruction receiving module 11 receives natural language consultation instruction messages based on the timing constraint gateway. After the natural language consultation instruction message enters the timing constraint gateway, the consultation instruction receiving module 11 calls a lightweight parsing script, preferably a combination of a lightweight sequence labeling algorithm (such as the BiLSTM-CRF named entity recognition algorithm) and a physical address mapping dictionary stored in the microprocessor's L1 / L2 cache, to forcibly deconstruct it into a set of feature parameters with hardware execution orientation.

[0024] The physical address mapping dictionary is a dynamic key-value hash table in the microprocessor's static random access memory (SRAM) or a high-speed memory database (such as Redis). In the data structure of this dictionary, the 'Key' field consists of standardized entity names (such as the full name of the enterprise's social credit code and standardized qualification name) automatically retrieved by the system from the target enterprise's ERP general ledger configuration table and the national policy bulletin interface during the initialization phase; the 'Value' field corresponds one-to-one with the physical shard index of the storage sector, the MAC address of the edge node, and the vector library addressing label automatically allocated by the underlying hardware of the system.

[0025] Specifically, the execution logic of the script is as follows: First, the named entity recognition algorithm is called to perform lexical stripping on the natural language consultation instruction message to extract key text entities such as enterprise name, qualification category and time adverbial; then, a memory lookup instruction is triggered to perform key-value matching in the preset physical address mapping dictionary, and the text entities are rigidly converted into corresponding physical shard indexes, authentication tokens and hash tags.

[0026] Furthermore, the natural language consultation instruction message is deconstructed into low-level control parameters by the consultation instruction receiving module; The target enterprise identification information is converted into: enterprise account physical shard index, API gateway authentication token, and network topology routing identification code. Among them, the enterprise account physical shard index calls the hash algorithm in the microprocessor to perform arithmetic mapping on the enterprise main identifier text, which directly points to the sector of a specific physical disk medium, driving the disk controller to perform precise addressing of the underlying data; API Gateway Authentication Token: An asymmetric encrypted string generated based on login verification status, used as a one-way read-only communication front-end packet to drive the physical network card to configure protocol access rules at the link layer; Network topology routing identification code: It is generated by extracting the network hardware coordinates of the edge computing nodes and forcibly latching them into the physical register mapping space of the application server as a unique physical addressing location.

[0027] The intent of the compliance inquiry is transformed into: discrete hash labels for target qualification categories, time-series review sliding window parameters, and Boolean trigger bits for core verification dimensions; Among them, the target qualification category discrete hash label: calls the named entity recognition algorithm to translate the qualification name into hexadecimal code, writes it into the address register of the threshold parsing unit, and drives the distributed vector database to perform rigid semantic recall; Time-series inspection sliding window parameter: compares the time adverb with the hardware real-time clock, deconstructs it into an absolute Unix timestamp interval, packages it into an SQL query message and uses it as a probe to listen for the time trigger limit of computing load; Core verification dimension Boolean trigger bit: Based on the instruction core pointer, the corresponding evaluation dimension register inside the central processing unit is set to a logic high level, driving the heterogeneous feature mapping module to perform low-level comparison computing power allocation.

[0028] It should be noted that the consultation instruction receiving module 11 calls the Time-Named Entity Recognition (Time-NER) algorithm in the microprocessor to perform lexical stripping of the time adverbial in the electrical signal and perform benchmark alignment calculation with the real-time clock (RTC) on the system physical motherboard to parse out the timing review sliding window parameter (i.e., timing review sliding window parameter) composed of absolute Unix timestamp intervals.

[0029] By performing dimensionality reduction analysis on the aforementioned feature parameters, the fuzzy natural language queries from the user's front end are transformed into rigid engineering control commands that can be directly invoked by the underlying network card, bus, and logic operation unit, thus ensuring the objectivity and physical binding characteristics of the subsequent measurement process from the data source.

[0030] Furthermore, to clarify how the natural language consultation command message is converted into the underlying engineering control sequence in this embodiment, the following explanation is provided in conjunction with a specific application scenario: Scenario setting: The user (corporate CFO) enters a natural language consultation command on the system's interactive terminal: "I am from *** company (credit code: 91110108MAxxxx). Our company has been working on low-code cloud integration and testing of our R&D system for the past three months. Please help me calculate the current qualification rate for applying for national high-tech enterprise status." When the system receives the instruction signal, the status acquisition unit 1 calls the microprocessor to execute the parsing script, converting "91110108MAxxxx" into a physical fragment index 0x8F9A4C... to lock the fourth disk array in the cloud. At the same time, it generates an API gateway authentication token and captures the physical network card IP (10.24.1.100) of the joint adjustment point as the network topology routing identification code. Simultaneously, it maps "National High-tech Enterprise" to a discrete hash label Tag_HNTE_2026, synchronizes "Last Three Months" to an absolute timestamp sliding window [1712534400,1720396800], and sets Flag_RD and Flag_IP in the static evaluation register to high level.

[0031] Specifically, the data capture module 12 receives the set of feature parameters parsed and sent by the consultation instruction receiving module 11; generates an SQL query message with hard time constraints based on the time-series review sliding window parameters; and sends a query request to the underlying hardware database through the API gateway and physical network card by combining the enterprise account physical shard index and the API gateway authentication token.

[0032] It is worth noting that, driven by the aforementioned characteristic parameters, the status acquisition unit 1 sends an SQL query message with time-series hard constraints to the underlying hardware database through the physical network card; the underlying hardware database responds to the query message, performs data extraction and field cleaning from the locked physical sectors, and completes the physical mapping of each core parameter in the structured static operational characteristic parameters. The specific extraction mechanism is as follows: Data capture module 12, based on the time-series review sliding window parameters [1712534400, 1720396800], traverses and searches the R&D auxiliary ledger storage partition to extract all cumulative debit amounts under the general ledger codes "R&D Expenditure - Expensed Expenditure" and "R&D Expenditure - Capitalized Expenditure". The extracted raw values ​​are then converted to discrete continuous values ​​representing the company's static R&D investment scale by the central processing unit, and these values ​​are bound and defined as parameters. It is then entered into the first cache queue of the system memory.

[0033] The data capture module 12 calls the underlying interface of the external asset database and adds Boolean logic filtering conditions to filter and extract valid asset features; Specifically, the Boolean logic filtering condition is as follows: A logical operator is used to perform a logical condition judgment on the ownership status field identifier retrieved from the target asset database, and the condition is met only if the value of the ownership status field identifier satisfies the identity operator "". If the condition is a Boolean truth value indicating "authorized" or "certified", the current asset data record is determined to be true.

[0034] The data capture module 12 performs an accumulation count operation on the filtered valid matching data records, outputs an unsigned integer value representing the actual intellectual property reserve, and defines it as a parameter. Through the filtering constraints of the Boolean logic mentioned above, non-compliant time-delay interference data in the "applying for" or "expired" state can be automatically excluded at the hardware level during the data reading stage, ensuring the absolute data accuracy when the subsequent arithmetic logic unit performs the weight confirmation quantity comparison.

[0035] Data capture module 12 performs a reverse retrieval based on the hardware timestamp of the current physical clock, locates the profit and loss statement storage module of the most recent closed financial and tax cycle, accurately extracts the double-precision numerical stream corresponding to the "Main Business Revenue" field, and defines it as a parameter. This parameter, which serves as the core denominator for subsequent evaluation of the R&D expense ratio, is synchronously locked in the reference register of the memory bus.

[0036] It is worth noting that the underlying field mapping and data type conversion implemented by the aforementioned state acquisition unit 1 rigidly transforms the discrete accounting data that originally existed in the heterogeneous financial system into standard engineering parameters that can be directly called by the subsequent central processing unit's logic operation unit. , and .

[0037] Furthermore, the threshold parsing unit 2, based on the parsed tag Tag_HNTE_2026, drives the graphics processor to lock the review boundary in the vector database and extract the lower limit threshold of R&D expense ratio. (i.e., 0.05, dimensionless); Status acquisition unit 1, in collaboration with the underlying ERP database, retrieves the total static R&D expenses within the corresponding time-series sliding window for the enterprise. 10,000 yuan (revenue) (10,000 yuan, static percentage 2%) Subsequently, the static evaluation unit 3 opens the hardware comparison channel according to the trigger bit Flag_RD=1 (before performing numerical comparison, the preprocessing normalization logic is called in advance to perform unit alignment and normalization processing on the extracted structured static operational feature parameters), determines that the static comparison fails (2% < 5%), and writes the time-delay rejection flag in the memory bus; after receiving the interrupt signal, the compliance generation unit 4 wakes up the target edge node probe according to the network topology routing calibration code (10.24.1.100), and listens to the underlying computing load in the time-series sliding window [1712534400, 1720396800] interval; Finally, compliance generation unit 4 uses a nonlinear compensation equation to perform feature weight compensation on static book R&D expenses, calculates dynamic R&D parameters that meet compliance boundaries, and applies them to the physical bias matrix. Under the constraints, the large language model is guided to decode and output a dynamic compliance conclusion text data stream containing a computational power evidence chain.

[0038] The system also includes a threshold parsing unit 2, which is deployed in an AI inference server equipped with a graphics processor. The threshold parsing unit 2 is used to receive natural language consultation instruction messages, determine the target qualification category label, and recall unstructured policy text that matches the target qualification category label from a distributed vector database pre-installed on a solid-state drive. The unstructured policy text is then used to perform semantic segmentation extraction, construct a qualification review threshold matrix, and write it to a cache system.

[0039] In this embodiment, the intent classification and routing module 21 calls a pre-set large language model to perform hardware-level dimensionality reduction extraction of the instruction intent. Specifically, the large language model is stored in the graphics processing unit's video memory array (VRAM), and its neural network architecture is preferably a bidirectional encoder representation model based on the Transformer architecture (such as BERT or RoBERTa network architecture), mainly including a word embedding layer, a multi-head self-attention encoder, and a fully connected classification layer. The specific data flow steps involved are as follows: The word embedding layer receives the natural language consultation instruction message, invokes a preset word segmentation dictionary (preferably a static hash dictionary constructed based on the WordPiece algorithm, in which a preset number (specifically preferably 21128) of root morpheme mapping tables are solidified in the read-only memory (ROM) of the microprocessor, and the underlying data structure of the root morpheme mapping table is represented as a rigid binding array of literal characters and integer index IDs, for example, mapping the Chinese character 'yan' (research) to index ID 4261 and 'fa' (development) to index ID 1355; by invoking this mapping table, the ambiguous natural language character string is deterministically translated into a digital sequence executable by underlying hardware) to segment and parse the text string in the electrical signal into a discrete token sequence (i.e., a one-dimensional array composed of integer index IDs, with [CLS] and [SEP] tokens forcibly appended to the head and tail of the sequence), then, the word embedding layer maps each integer ID in the discrete token sequence to a word vector, and superimposes a segment embedding vector and an absolute position encoding vector, so as to map to a continuous high-dimensional input vector sequence (specifically, a dimension of high-dimensional floating-point tensor, where is the length of the instruction sequence, and 768 is the preset feature dimension of the model hidden layer ).

[0040] Subsequently, the multi-head self-attention encoder (the actual network parameters of which are preferably configured as: the number of attention heads , the number of coding network layers )receives the high-dimensional input vector sequence, multiplies the tensor by three pre-trained weight matrices respectively to construct a query matrix , a key matrix and a value matrix ; by calculating the inner product of the transpose matrix of the query matrix and the key matrix , dividing by the scaling factor (wherein, is the dimension of a single attention head, such as 64), and then inputting the result into a Softmax function for normalization processing to obtain an attention weight distribution matrix; the system uses the distribution matrix to perform weighted summation of matrix multiplication on the value matrix (that is, the underlying microprocessor executes the scaled dot product formula) to extract a context semantic feature matrix representing the compliant query entity: ) The context semantic feature matrix is specifically a dense feature matrix that contains bidirectional global semantic dependencies ).

[0041] Furthermore, the fully connected classification layer (which is actually configured as a linear feedforward network containing multiple hidden units, and whose final output dimension is equal to the total number of compliant business intent categories preset by the system, such as 64 dimensions) specifically extracts the first position (i.e., the [CLS] marker) in the aforementioned context semantic feature matrix. The dimensional vector serves as a global feature representation of the entire instruction; the fully connected classification layer uses this vector... The dimensional vector undergoes a linear transformation (i.e., it calls the classification weights and bias parameters inherent within the classification layer and fine-tuned using compliant business data to perform affine calculations; specifically, the initial range of the classification weights is rigidly constrained within a certain range). Within a uniformly distributed interval, and after supervised optimization, its value range is preferably limited to [value range missing]. The standardized floating-point range; the bias parameter is initialized to a zero vector, and its value range is preferably limited to... (The rational floating-point interval); then, the result of the linear transformation is input into the normalized exponential function (Softmax) for probabilistic activation, and outputs a probability distribution vector corresponding to the predetermined compliant business space with a sum of 1.

[0042] After obtaining the probability distribution, the microprocessor extracts the classification node (such as the 42nd classification node) corresponding to the maximum likelihood probability in the probability distribution vector as the corresponding intent category, and triggers an addressing instruction to perform key-value matching in the intent mapping hash table in static random access memory (SRAM), rigidly replacing the intent category with a unique machine-readable hexadecimal code (such as Tag_HNTE_2026) as the target qualification category label.

[0043] In the microprocessor's memory system, the intention mapping hash table consists of fixed key-value pairs. The data type of the mapping key is forcibly declared as an unsigned short integer, which is used specifically to receive the classification node index output by the underlying neural network. The data type of the mapping value is declared as a fixed-length hexadecimal string, which is used specifically to initiate precise physical address retrieval to the distributed vector database later.

[0044] As a preferred embodiment of the present invention, the local mapping data structure of the hash table in the static random access memory is as follows: When the input mapping key is a category node index (such as Node_42) representing the intent of 'high-tech enterprise application', its hard mapping output value is a unique hexadecimal code 0x484E5445; When the input mapping key is a category node index (such as Node_15) representing the intent of 'Specialized, Refined and Innovative Enterprise Recognition', the mapping output value is the hexadecimal code 0x53524449.

[0045] Through the rigid one-to-one mapping at the underlying physical storage layer, the risk of addressing ambiguity caused by the probabilistic uncertainty and uninterpretability of the large language model output to subsequent system database retrieval is completely blocked.

[0046] It is worth noting that the intent mapping hash table is completely independent of the physical address mapping dictionary called by the aforementioned state acquisition unit 1 in terms of physical storage area and addressing logic. The former is dedicated to the dimensionality reduction transformation of semantic probability of large language model to vector retrieval coordinates (mapping "intent" to "tag"); the latter is dedicated to the addressing mapping of external entities to hardware addresses (mapping "enterprise code" to "disk sector"). The two achieve computing power isolation through different memory addressing buses.

[0047] Furthermore, the lower limit threshold is extracted by the policy text regular expression extraction module 22, and a qualification review threshold matrix is ​​constructed. The specific steps involved are as follows: The system receives an unstructured policy text data stream retrieved from a distributed vector database that matches the target qualification category label, loads it into the microprocessor's memory buffer, and segments it into multiple discrete syntactic logic blocks based on preset punctuation identifiers.

[0048] Furthermore, feature traversal is performed on each of the syntactic logic blocks based on a preset multimodal regular expression constraint template; wherein, the multimodal regular expression constraint template includes a first regular expression template and a second regular expression template, and each of the regular expression templates includes text prefix anchors for filtering modifier semantics and regular expression capture groups for locking numerical boundaries.

[0049] Specifically, the first regular expression template is used to perform character traversal on each of the syntactic logic blocks. When a text prefix anchor representing the R&D funding constraint is matched, the corresponding regular expression capture group is triggered to extract a floating-point string containing a percentage sign. In a specific embodiment, the preferred code structure of the first regular expression template is: (?<=R&D expenses.*?proportion not less than)\s*(\d+(\.\d+)?)%.

[0050] When the microprocessor recognizes a text prefix anchor such as "the proportion of total R&D expenses to total sales revenue in the same period is not less than", it immediately triggers the capture group to extract the following string such as "5%".

[0051] Instead of directly calculating the string, the system removes the trailing "%" character and calls the first type casting function (preferably the `std::stod()` function in the C++ standard library) pre-installed in the system standard library. This function forces the string to be compiled into a double-precision floating-point number occupying 8 bytes of memory in the underlying registers, and then translates it into a discrete value (e.g., 0.05) by scaling it by a percentage. This value is then defined as the lower limit threshold for R&D expenses. .

[0052] Similarly, the second regular expression template (preferred structure: (?<= own.*?(?: software copyright).*? not less than)\s*(\d+)\s*(?: item|item)) is used to perform character traversal on each of the syntactic logic blocks; when a corresponding anchor point is matched, the capture group is triggered to extract a continuous Arabic numeral string (such as "15"); for this numeral string, the microprocessor calls the preset second type casting function (preferably the std::stoul() string to unsigned long integer function in the C++ standard library) to rigidly translate it into a positive integer without a sign bit at the hardware level, which serves as the lower limit threshold for the number of intellectual property rights. .

[0053] Specifically, after completing the above-mentioned forced type conversion, the policy text regular expression extraction module 22 requests the system motherboard to extract a continuous physical address space in the system memory bus, and sets the R&D expense ratio lower limit threshold. (e.g., 0.05) and the lower limit threshold for the number of intellectual property rights. (e.g., 15) Based on a predetermined memory offset (rigidly calibrated based on the system's underlying bus width and data type byte length; in this embodiment, considering the lower limit threshold of the R&D cost ratio after the preceding translation) It is a 64-bit double-precision floating-point number (occupying 8 bytes of physical memory), and the lower limit threshold for the number of intellectual property rights. During storage, the data is also forcibly converted and aligned to a 64-bit unsigned long integer. To prevent memory addressing misalignment and to match the single-cycle read step size of the microprocessor's arithmetic logic unit, the predetermined memory offset is preferably calibrated and hard-coded as 8 bytes (i.e., 64 bits) and sequentially pushed into the contiguous physical address space, thereby constructing a qualification review threshold matrix (a mathematical tensor array containing a set of one-dimensional continuous addressing features) for direct access by the subsequent microprocessor arithmetic logic unit. ; In the formula, This indicates that the starting pointer address (index 0) in physical memory stores a floating-point threshold factor with continuous proportional attributes; This indicates that an integer threshold factor with discrete absolute quantity attributes is stored at the subsequent physical address (index 1) after the predetermined memory offset of the 8 bytes.

[0054] The system also includes a static evaluation unit 3, which receives structured static operational characteristic parameters and qualification review threshold matrix. By triggering the underlying hardware comparison instruction, it generates a first static compliance evaluation vector data stream containing a Boolean state identifier sequence based on the comparison result. When any numerical comparison fails, the corresponding Boolean state identifier is configured as a time-delay rejection identifier, and the first static compliance evaluation vector data stream is input to the system memory bus.

[0055] In this embodiment, because the target company has not yet completed the financial settlement of the cloud server rental fees and R&D personnel salaries in the ERP system when performing digital migration and compliance consulting at the end of the month, the historical static financial accounting data captured by the status acquisition unit 1 will have a serious lag (for example, the accounting R&D expenses are much lower than the actual amount incurred); the static evaluation unit 3 is used to intercept and mark such defects: The static evaluation unit 3 includes a heterogeneous feature mapping module 31 and a rigid numerical comparison module 32; wherein, the heterogeneous feature mapping module 31 is used to establish a data type alignment and physical addressing mapping channel between the structured static operational feature parameters and the qualification review threshold matrix between system memory and registers, i.e., a heterogeneous mapping channel.

[0056] Specifically, due to the inherent heterogeneity between the structured static operational feature parameters (discrete fixed-point integer or double-precision data from the ERP underlying hardware database) and the qualification review threshold matrix (floating-point contiguous memory array generated from a large model) in terms of underlying storage structure, byte length, and physical storage medium (one is the system main memory, and the other is the graphics processor's video memory), the central processing unit cannot directly address them and issue single-cycle comparison instructions. To address this addressing isolation caused by heterogeneity, the heterogeneous feature mapping module 31 first executes a channel construction instruction, retrieving the structured static data from the system data bus. The starting physical memory pointer (Pointer_A) of the state operation characteristic parameters and the starting physical memory pointer (Pointer_B) of the qualification review threshold matrix are then used. Subsequently, the heterogeneous feature mapping module 31 requests a completely independent, address-contiguous physical storage sector from the microprocessor's cache as a data buffer physical stack, and hard-binds the addressing logic of Pointer_A and Pointer_B to the input bus of the buffer physical stack, thereby instantiating and constructing a physical addressing path across the main memory and video memory space at the hardware level, namely the heterogeneous mapping channel.

[0057] Furthermore, after the heterogeneous mapping channel, i.e. the cache buffer physical stack, is established, the data stream begins to be format-translated and aligned along the channel. The heterogeneous feature mapping module 31 calls the hardware type alignment instructions embedded in the microprocessor (such as the data warping extension instruction set under the SIMD architecture) to force the discrete integer feature parameters extracted along the channel to be padded with bit width, and rigidly translates them with the floating-point threshold factor into 64-bit double-precision floating-point numbers (Float64) with the same byte length. Then, according to the 64-bit word length boundary, the memory address alignment arrangement is performed in the cache to complete the hardware-level comparison preparation.

[0058] In this embodiment, the rigid numerical comparison module 32 is used to perform a single-cycle read of the physically aligned memory sectors and send the aligned parameters in pairs to the arithmetic logic unit (ALU) of the microprocessor to perform machine-level Boolean comparison (CMP instruction): Specifically, in terms of intellectual property, the arithmetic logic unit extracts the number of intellectual property rights from the structured static operational characteristic parameters. (As mentioned above, 18 items were captured), and the lower limit threshold for the number of intellectual property rights was extracted from the qualification review threshold matrix. (e.g., 15 items as set in the threshold matrix); since 18≥15, the Boolean comparison of this dimension passes, and the corresponding register bit outputs a high level 1.

[0059] In terms of R&D expenses, the arithmetic logic unit extracts the total annual R&D expense from the structured static operational characteristic parameters. (e.g., 200,000 yuan) and the most recent main business revenue (For example, 10 million yuan), after division, the static book value of R&D expenditure ratio is 2%; then it is compared with the lower limit threshold of R&D expenditure ratio extracted from the threshold matrix. (5%) is compared; since 2% < 5%, the comparison of this dimension failed, and the output is low level 0.

[0060] Furthermore, based on the output of the arithmetic logic unit, a first static compliance assessment vector data stream containing a sequence of Boolean state identifiers (such as [0,1] in this scenario) is generated. Given that the target enterprise is in a period of high frequency of R&D integration and testing, for low-level signals appearing in the R&D dimension, the rigid numerical comparison module 32 does not directly throw out the final conclusion of "compliance not passed", but instead determines it as a physical lag defect in the R&D status representation caused by the static financial accounting cycle. Then, the Boolean state identifier corresponding to this dimension is forcibly configured (overwritten) as a time-delay rejection identifier in the register. Finally, the first static compliance assessment vector data stream is sent to the system memory bus as a hardware-level interrupt signal.

[0061] The system also includes a compliance generation unit 4, which calculates the edge computing power consumption drift rate based on the time-delay rejection identifier and uses the edge computing power consumption drift rate as a dynamic weight factor to compensate the total R&D expenses for feature weights, thus obtaining dynamic R&D parameters. The dynamic R&D parameters are then compared with the qualification review threshold matrix, and the comparison results and the edge computing power consumption drift rate are synchronously input into the generative language model for decoding. The generative language model then autoregressively decodes and outputs a dynamic consultation solution text data stream containing the physical computing power supporting chain and subsequent financial and tax accounting planning.

[0062] In this embodiment, the compliance generation unit 4 includes an edge computing power compensation module 41 and an offset intervention decoding module 42. The edge computing power compensation module 41 and the offset intervention decoding module 42 realize the cascading transmission of heterogeneous control flow and dynamic data flow through a serial bus. This is used to solve the secondary system defect caused by the time lag in the settlement cycle of enterprise static accounting R&D expenses, which leads to the misjudgment and rejection of ongoing high-frequency physical R&D activities in normal static compliance judgment.

[0063] Specifically, the edge computing power compensation module 41 is used to perform status polling on the system memory bus. When the rigid numerical comparison module 32 writes a time-delay rejection flag to the system memory bus because the static R&D ratio on the books does not meet the standard, the edge computing power compensation module 41 immediately responds to the system-level hardware interrupt signal transmitted by the time-delay rejection flag, suspends the normalized static compliance assessment link, and forcibly cuts into the branch control logic of underlying computing power dynamic monitoring and general hardware resource allocation.

[0064] In this embodiment, the static evaluation unit 3 is configured as a front-end first-level low-power hardware screening gateway: when the static book R&D ratio meets the qualification review threshold, it is determined that the security redundancy of the enterprise's historical financial data has covered the risk of time lag attenuation, and the system directly outputs the compliance conclusion along the conventional low-power link; only when the rigid value comparison fails, the system determines that the time lag effect has substantially caused the risk of compliance misjudgment, and then the time lag rejection flag is used as a hardware-level wake-up interrupt to activate the edge computing power compensation module 41 belonging to the high-power computing branch as needed.

[0065] Furthermore, the edge computing power compensation module 41 extracts the network topology routing identifier (in this embodiment, it is instantiated as the physical network card IP address of the edge computing node currently performing low-code cloud integration debugging in the target company) by the addressing and routing instructions of the microprocessor, and temporarily locks it in the physical register mapping space after being parsed by the preceding consultation instruction receiving module 11. The edge computing power compensation module 41 overcomes the disk read limitations of the external relational financial database and sends high-frequency network polling probe data packets to the corresponding target edge computing node through the motherboard physical network card. (To ensure real-time capture of the integration debugging status while preventing network link congestion, the sending cycle of the network polling probe data packets is rigidly constrained within a certain timeframe.) Within the physical region, in this embodiment, the preferred configuration is as follows: The hardware performance counters deployed on the motherboard firmware layer of the edge computing node are remotely activated (to monitor the internal voltage changes and instruction execution status of the microprocessor in real time, count the number of logic core registers in a high-level active state, and thus directly intercept and output the CPU thread concurrency at the physical hardware layer). The hardware gateway link layer counter (used to capture the underlying data packet throughput queue of the target node when receiving external API calls, and input the continuously sampled discrete throughput values ​​into the arithmetic logic unit of the microprocessor to perform variance calculation, thereby calculating and generating the variance of concurrent requests per second of the cloud API hardware gateway) ).

[0066] Specifically, within the time-series review sliding window parameters deconstructed and output by the state acquisition unit 1 (rigidly defined in this embodiment as the time boundary of the most recent three months composed of the absolute Unix timestamp interval [1712534400, 1720396800]), the edge computing power compensation module 41 performs sampling at a preset period. Real-time monitoring and capture of the number of concurrent CPU threads in active running state on the physical chip of the underlying edge computing node. .

[0067] It is worth noting that the sampling period The range of values ​​for is limited to In this embodiment, the preferred programming method is... Specifically, if the sampling period Below The microprocessor will frequently trigger underlying hardware interrupts, causing severe context switching and cache turbulence secondary defects in the target node, seriously dragging down the host computing power of the physical chip; if the sampling period Higher than The probe will be unable to capture sudden high-frequency thread surges that occur instantaneously in the code compilation pipeline (such as the instantaneous burst of computing power during the static linking phase of the compiler), resulting in severe distortion of dynamic feature extraction.

[0068] Synchronous, in order to translate discrete packet throughput into variance of concurrent requests per second with statistical convergence properties. The edge computing power consumption drift rate compensation calculation module 41 calls a preset local sliding statistical window (its time width) The valid value range is limited to: In this embodiment, it is preferred to lock it as Furthermore, its sliding step size is uniformly fixed as follows: The data packet throughput queue captured by the hardware gateway link layer counter is traversed; the general-purpose arithmetic logic unit (ALU) performs a traversal on the queue within each sliding step period. The variance of the discrete throughput sample set (i.e., a sample array containing 30 independent data points) collected over a time span is calculated to output the variance of concurrent requests per second of the cloud API hardware gateway in real time. By introducing this local sliding statistical window, the dilution effect of long-term time series data on instantaneous fluctuations is eliminated, ensuring... It can accurately map the actual load jitter of the current physical network card.

[0069] It is worth noting that in this embodiment, the edge computing power compensation module 41 invokes the arithmetic instruction set pipeline embedded in the microprocessor to drive the general-purpose arithmetic logic unit (ALU) to perform hardware-level pipelined multiplication, addition, and smoothing filtering operations on the data stream in the register. Specifically, in order to achieve same-dimensional normalization of heterogeneous computing power indicators while eliminating high-frequency noise interference, the general-purpose arithmetic logic unit invokes a preset computing load quantization optimization equation burned word-by-word into the read-only memory (ROM) to perform optimization and solution: ; In the formula, This represents the calculated edge computing power consumption drift rate used to characterize the actual dynamic R&D process; This indicates the number of concurrent CPU threads detected. This represents the variance of concurrent requests per second at the cloud API hardware gateway. This represents the target edge node hardware baseline concurrent capacity parameter dynamically read based on the network topology routing calibration code, used to calibrate the relative saturation of computing power consumption; This represents the computing power topology allocation weight coefficient, and its value range is... ; The nonlinear smoothing damping constant used to suppress burst traffic spikes (in this embodiment, the value ranges from 50 to 500, and in this embodiment, it is preferably burned to 100 in combination with the target node firmware configuration) is used as a denominator smoothing factor to prevent nonlinear divergence when API concurrent requests experience pulse-like drastic fluctuations.

[0070] During enterprise R&D activities, tasks such as source code compilation, model training, automated testing, and continuous integration continuously utilize the computing resources of edge computing nodes. The system monitors the running status of edge nodes associated with R&D tasks in real time and statistically analyzes the concurrent CPU threads, API call frequency, and resource usage changes corresponding to these tasks to construct the edge computing power consumption drift rate. Since the labor costs, cloud resource costs, and equipment costs associated with these R&D activities typically have a lag in financial recognition, the edge computing power consumption drift rate is used to characterize the intensity of R&D activities within the time-series review window and serves as a compensation parameter for estimating R&D investments that have not yet been financially recognized. It is integrated with static R&D expenses to calculate dynamic R&D expenses, thereby reducing static assessment errors caused by lags in financial accounting. It is worth noting that the edge computing power consumption drift rate does not directly represent R&D expenses but rather serves as a quantitative representation of the intensity of R&D activities. Based on this intensity of R&D activity, the system estimates R&D investments that have not yet been financially recognized and integrates them with static book R&D expenses to generate dynamic R&D expenses (i.e., dynamic R&D parameters).

[0071] After obtaining the unsigned floating-point number, a general data comparison instruction is called to compare it. Perform value boundary verification based on the underlying hardware benchmark.

[0072] Specifically, the drift rate of edge computing power consumption will be used to characterize this. The binary electrical signal stream, along with the lower and upper limit reference constants pre-locked in ROM, performs a single-cycle machine-level size determination. The upper and lower limit reference constants are hardware limit parameters obtained by calling hardware probes to perform physical no-load and network stress tests on the target edge computing node. Their specific engineering calibration basis is as follows: The effective physical value range of the lower limit reference constant is: In this embodiment, the noise floor characteristics of the target node (such as the Linux CentOS server core) are combined, and it is preferably calibrated to 0.08. The physical basis for this value range is that when the target node only maintains communication between the underlying operating system daemon process and the network heartbeat packet, the standby noise floor computing power consumption presented by its hardware performance counter falls within this range.

[0073] The effective physical value range of the upper limit reference constant is: In this embodiment, the value is preferably set to 0.45. The physical basis for this value range is that the maximum non-blocking throughput that the physical Ethernet card (NIC) of the node can withstand under the condition of ensuring secure communication and not triggering the NIC interrupt queue overflow falls within this range.

[0074] It is worth noting that the lower limit benchmark constant is used to characterize the standby noise floor computing power threshold of the target node operating system when maintaining communication between the underlying daemon process and network heartbeat packets; the upper limit benchmark constant is used to characterize the maximum non-blocking throughput extreme value of the node's physical Ethernet card (NIC) under the condition of ensuring secure communication.

[0075] Based on the aforementioned rigid hardware calibration benchmarks: If the machine level determines this If the calculated value is less than the lower limit benchmark constant (e.g., 0.08), it is determined that the current instruction set execution density is only the system noise floor, the target node is in a pure standby state, and the branch instruction is executed to directly block the subsequent data flow to the addressing path of the feature weight compensation equation (i.e., no compensation is given). If it is determined that If the calculated value is greater than the upper limit reference constant (e.g., 0.45), it is determined that the physical limit of the Ethernet card throughput has been exceeded, the system has encountered an abnormal traffic attack (e.g., DDoS broadcast storm), the hardware level triggers the security warning interrupt line and terminates the measurement flow. If it is determined that it is within a safe closed-loop area If the signal is detected within the specified range, it will be allowed to proceed and locked as a valid indicator of the edge computing power consumption drift rate in the R&D status. It forcibly latches the data into the specified register address space.

[0076] Furthermore, after confirming the edge computing power consumption drift rate... After successfully passing the safe value boundary verification, the floating-point unit (FPU) of the central processing unit takes over the data block, latches it as a dynamic weighting factor in the designated address space, and calls the nonlinear exponential smoothing compensation equation (i.e., the nonlinear compensation equation mentioned above) that is pre-stored in read-only memory from the system's standard function library. ; In the formula, This represents a preset engineering experience constant (in this embodiment, the engineering experience constant is used to adapt to the differences in computing power overhead of microprocessor clusters of different sizes when performing digital migration). The effective physical value range is limited to In this embodiment, it is preferably calibrated to 1.3389), a preset engineering empirical constant. As a R&D investment conversion factor for the target edge computing node group under a specific network topology routing calibration code, it is used to calibrate the virtual resource consumption conversion rate corresponding to unit computing power fluctuation at the hardware data level, thereby ensuring the dynamic R&D parameters of the solution output. ; The dynamic R&D parameters, which are strongly bound to the actual physical state of the solution output, are represented. This represents the total amount of discrete annual R&D expenses collected by the preceding data capture module 12 and locked in the first memory cache queue; This represents the weighting constant for the conversion of computing power assets, stored in read-only memory. It is used to calibrate the conversion loss rate between heterogeneous physical resources and financial elements, and its physical value range is limited to [specific value range]. ; This indicates the standardized cloud resource unit computing power rental price constant (such as the standard rental fee per hour for a single-core CPU) of the target edge node, which is synchronously captured by the system from the external cloud service provider's billing gateway or the enterprise's asset ledger via the API interface.

[0077] Specifically, based on this nonlinear exponential smoothing compensation equation, when the floating-point unit performs pipelined computation, even under extreme time-delay conditions or during the new project initiation phase, the static book value of the total annual R&D expenses read will be reduced. The absolute zero value causes the preceding multiplication expansion operator to... Even after collapsing to zero, the system can still call the additive compensation operator along the branch. The additive compensation operator forces the use of the detected real edge computing power to consume drift rate. With physical rental unit price constant By executing hardware multiply-accumulate instructions and converting them into an estimated base for R&D investment, the system avoids the feature blocking trap caused by zero-valued structured static accounts. This ensures that, in any static financial gap state, the system can quantitatively estimate the R&D resource investment that has not yet been financially confirmed within the time-series review window, based on the effective edge computing power consumption corresponding to the detected R&D tasks and the unit resource cost. This generates dynamic R&D parameters that characterize the actual scale of R&D activities. .

[0078] Subsequently, the edge computing power compensation module 41 will calculate the generated dynamic R&D parameters. The binary data stream is rigidly written into the shared buffer of the system memory, and an address ready signal is sent to the bus so that the subsequent serially cascaded bias intervention decoding module 42 can perform subsequent data reading and secondary comparison calls.

[0079] In this embodiment, the bias intervention decoding module 42 immediately performs the physical intersection of the mainline accounting data and the branchline computing power parameters based on the address ready control interrupt signal sent to the system bus by the aforementioned module 41, and drives the generative language model to perform white-box intervention decoding.

[0080] Specifically, the bias intervention decoding module 42 calls the general-purpose arithmetic logic unit (ALU) of the microprocessor via system bus instructions to read the dynamic R&D parameters that have been boosted by the underlying computing power weight from the shared buffer of memory. And compared with the most recent main business revenue pre-latched in the reference register. Perform hardware-level floating-point division to obtain a discrete floating-point number representing the dynamic R&D percentage.

[0081] Furthermore, the Arithmetic Logic Unit (ALU) uses the obtained discrete floating-point number and the lower limit threshold of R&D expense ratio extracted from the qualification review threshold matrix. The single-cycle machine-level Boolean comparison instruction (CMP instruction) is executed. If the comparison result is successful, the result of the instruction directly triggers the corresponding status register inside the microprocessor to undergo a physical level reversal, outputting a secondary comparison success status result represented by a high level, and forcibly overwriting the original low-level "time-delay rejection flag" to a high-level "secondary comparison success status code".

[0082] Simultaneously, the bias intervention decoding module 42 requests a memory space of the same length as the attention sequence of the generative language model from the system's video memory array to construct the physical bias weight matrix. While requesting video memory space, the bias intervention decoding module 42 simultaneously receives the original compliance inquiry intent transmitted through the preceding bus. Through sequence word segmentation operation of the word embedding layer, it identifies the discrete annual R&D expense aggregated total amount in the intent text that caused the preceding static assessment to fail. The location of the time sequence is hard-coded and anchored, and defined as the "failure book dimension".

[0083] The bias intervention decoding module 42 extracts the aforementioned latched edge computing power consumption drift rate and uses it as an attenuation control variable to call a preset attenuation mapping function: ; In the formula, This represents the base semantic preservation constant stored in read-only memory (its effective physical value range is limited to ). In this embodiment, the attention fault tolerance of the Transformer architecture is preferably calibrated as follows: ); This represents the preset sensitivity coefficient (in this embodiment, it is set to adapt to heterogeneous large models with different attention head dimensions). The numerical sensitivity difference is used to prevent over-penalty from causing the model to completely lose its generative ability, or under-penalty from causing time-delayed data to be unable to be effectively suppressed. The sensitivity coefficient is... The valid range of values ​​is limited to Combined with the tensor normalization property, the sensitivity coefficient is preferably 15, which is used to adjust the steepness of the penalty imposed by the computing power drift feature on the attention mechanism of large models; The floating-point attenuation penalty element representing the output of the mapping calculation is used as the physical bias weight matrix. Internal negative penalty weights applied to invalid book features; The base of the natural logarithm (natural constant) is used to drive the underlying exponential nonlinear operations; This represents the edge computing power consumption drift rate extracted from the specified register address space after the preceding safe value boundary verification.

[0084] It is worth noting that this decay mapping function introduces a basis semantics preservation constant. To achieve the physical bias weight matrix When assigning values ​​to the diagonal coordinates associated with the 'Invalid Book Dimension' label internally, through... Instead of absolute zero-value truncation, soft gating suppression constraints are constructed. This ensures that the tensor nodes corresponding to the 'failed accounting dimension' in the autoregressive decoding pipeline of the generative language model still retain basic syntactic coherence and contextual conjunction addressing capabilities, only stripping them of their high-dimensional weights in compliance logic decision-making calculations. This approach thoroughly establishes a causal tracing and state binding mechanism between the model's output character sequence and the real physical cloud resource configuration state, while avoiding the discontinuity of high-dimensional semantic vector sequences caused by excessive penalties.

[0085] Subsequently, the attenuation element Rigidly written to the physical bias weight matrix Internally, corresponding to the diagonal coordinates associated with the "failed accounting dimension" label. Through the above decay mapping transformation based on the actual computing power indicators, the abstract financial compliance rejection risk is materialized into a floating-point matrix mask array in the video memory array that can be directly called by the tensor operation unit.

[0086] Furthermore, the generative language model invoked by the bias intervention decoding module 42 has its network topology nodes fully deployed in the high-speed video memory array (VRAM) of a dedicated graphics processing unit (GPU); at the neural network topology level, the generative language model is configured as an autoregressive decoding configuration based on a one-way mask self-attention mechanism, and its physical instance allocation in the video memory space is completely independent and isolated from the aforementioned general encoder large model used for intent classification; it includes at least a word embedding layer, a multi-head self-attention layer with an autoregressive decoding architecture, a layer normalization unit, a feedforward network linear mapping layer, and a maximum likelihood probability sampling output layer.

[0087] The steps for the generative language model to implement intervention constraints are as follows: First, a specific memory addressing space is allocated through the word embedding layer, and the input consists of the secondary alignment success status code and dynamic R&D parameters. The enhanced prompt word sequence, which is formed by splicing the measured values, the qualification review threshold matrix, and the original compliance inquiry intent, is translated into high-dimensional discrete word vectors and then cascaded through positional encoding to output a high-dimensional semantic vector matrix.

[0088] The high-dimensional semantic vector matrix is ​​input to a multi-head self-attention layer with an autoregressive decoding architecture via a parallel data bus. At the hardware level, the multi-head self-attention layer consists of multiple parallel arrays of linear transformation weight matrices. As the high-dimensional semantic vector matrix passes through the GPU's multiply-accumulate instruction pipeline, it undergoes matrix multiplication with the corresponding linear transformation weight matrices, thereby decoupling and separating a query matrix that independently represents the contextual retrieval features of the current text. Key matrix AND-value matrix ; Furthermore, the physical bias weight matrix generated in the previous solution is... Forced embedding into matrix AND key matrix The attention score is calculated as follows: ; In the formula, The preset dimension representing a single attention head; This indicates that after physical bias constraints and normalization, the self-attention feature tensor is finally output by the underlying tensor operation unit, which contains contextual semantic weights that are affected by the real computing power evidence chain. Key matrix The tensor transpose matrix is ​​used to compare with the query matrix. And at the underlying level, the dot product execution unit is called to calculate the original correlation score matrix; This is a normalized exponential activation function instruction used to rigidly compress and map the biased tensor elements to... Within the probability distribution interval, ensure that the sum of the exponents of all elements equals 1.

[0089] It is worth noting that, under the intervention of the aforementioned underlying instructions, when the tensor operation unit executes the key matrix associated with the "invalid ledger dimension", At that time, encountered The negative penalty weight element, approaching zero (0.005), causes a sharp drop in the attention activation value of the corresponding tensor node due to the local Hadamard multiplication operation, rigidly suppressing its attention weight to near absolute zero. Through this physical zero-value suppression characteristic, the generative language model is forcibly deprived of computational resources from existing data, focusing its maximum likelihood sampling trajectory entirely on dynamic R&D parameters supported by a computational evidence chain. Above.

[0090] Since the underlying hardware has already generated a 'secondary comparison success status code' through an overwrite operation, the system has logically confirmed the target company's substantive compliance. However, to prevent the generative language model from being influenced by the semantic interference of the original static and lagging accounting data in the context when generating the 'dynamic consultation solution' presented to the user, thus avoiding factual illusions or negative representations, the system inputs the secondary comparison success status code and dynamic R&D parameters into the generative language model and forcibly applies a physical bias weight matrix. The zero-value masking characteristic of this matrix completely shields the large model from the semantic reading ability of the 'failed static accounting dimension' at the physical memory level, forcing the generative language model to focus entirely on the 'secondary comparison success status code' and the 'edge computing power consumption drift rate'.

[0091] Finally, the maximum likelihood sampling output layer decodes and outputs the dynamic compliance conclusion text data stream in a unidirectional autoregressive loop iteration. The specific output result of the example is: "The current physical compliance review has been passed (status code: Success). Based on the real-time monitored core node (IP: 10.24.1.100) computing power drift rate, it is proven that the company has recently carried out high-intensity digital integration and R&D activities. It is recommended that the company's finance department export the computing power load log as an auxiliary account for R&D reference, and complete the financial collection of R&D personnel salaries and cloud service fees for the current month as soon as possible, which will fully meet the lower limit standard of high-tech enterprise qualification."

[0092] Furthermore, to ensure that the generative language model can achieve low-latency parallel inference on high-speed memory arrays with limited target edge computing nodes and to prevent memory overflow (OOM) crashes, this embodiment rigidly calibrates the network hyperparameters at each level within the model based on a continuous memory allocation mechanism: The effective value range for the word embedding layer dimension and the hidden layer vector dimension is limited to: In this embodiment, the FP16 half-precision floating-point calculation specification is preferably calibrated to 4096 to ensure full-load alignment of the bus bandwidth for a single memory read and avoid wasting the idle bandwidth of the video memory read. Attention to the number of heads The range of values ​​is limited to The preferred programming dimension is 32, with a single attention head dimension rigidly bound to it. The computation converges to 128 (i.e., 4096 / 32); this 128-dimensional tensor partitioning is extremely well-suited to the throughput step size of the matrix hardware multiply-accumulate array of modern underlying tensor operation units, enabling ultra-fast parallel computation in a single clock cycle; Given the exponential expansion of memory time and space complexity inherent in self-attention mechanisms, to prevent high-frequency queries from causing instantaneous memory breakdown, the effective range of the maximum context sequence length for the autoregressive decoding is forcibly truncated to [a specific value]. In this embodiment, the hard-coded value is preferably locked to 2048; To ensure the nonlinear mapping space of the network while controlling instantaneous memory overhead, the dimensional expansion ratio of the feedforward network linear mapping layer (FFN) is preferably locked at 4, which corresponds to a physical mapping dimension of 16384 for the hidden layer.

[0093] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent question-and-answer system for enterprise consultation based on a large language model, characterized in that: include: Status acquisition unit (1), the status acquisition unit (1) receives natural language consultation instruction messages based on the time-constrained gateway, and extracts structured static operation feature parameters by triggering physical network card addressing to call the underlying hardware database; Among them, the natural language consultation instruction message contains the target company identification information and the intent of compliance inquiry; Threshold parsing unit (2) is used to receive the natural language consultation instruction message, determine the target qualification category label, recall the unstructured policy text that matches the target qualification category label from the distributed vector database, perform semantic segmentation extraction on the unstructured policy text, and construct the qualification review threshold matrix; Static evaluation unit (3) is used to receive the structured static operation feature parameters and qualification review threshold matrix, and generate a first static compliance evaluation vector data stream containing a Boolean state identifier sequence based on the comparison result by triggering the underlying hardware comparison instruction. In the case of any numerical comparison failing, the corresponding Boolean state identifier is configured as a time-delay rejection identifier. The compliance generation unit (4) calculates the edge computing power consumption drift rate based on the time-delay rejection identifier, and uses the edge computing power consumption drift rate as a dynamic weight factor to compensate the total R&D expenses in the structured static operating characteristic parameters to obtain dynamic R&D parameters; it performs a secondary comparison between the dynamic R&D parameters and the qualification review threshold matrix, and inputs the comparison result and the edge computing power consumption drift rate into the generative language model for decoding, and outputs a dynamic compliance conclusion text data stream.

2. The intelligent question-and-answer system for enterprise consultation based on a large language model according to claim 1, characterized in that, The status acquisition unit (1) includes an inquiry instruction receiving module (11) and a data capture module (12). The consultation instruction receiving module (11) is used to receive the natural language consultation instruction message and parse out the timing review sliding window parameters; The data capture module (12) is used to generate an SQL query message with hard time constraints based on the time-series review sliding window parameters, send a query request to the underlying hardware database through the API gateway, capture the corresponding financial general ledger account table and R&D project initiation ledger data, and extract structured static operating characteristic parameters; wherein, the structured static operating characteristic parameters include the total annual R&D expenses. Number of intellectual property rights and the most recent main business revenue .

3. The intelligent question-and-answer system for enterprise consultation based on a large language model according to claim 1, characterized in that, The threshold parsing unit (2) includes an intent classification routing module (21) and a policy text regular expression extraction module (22). The intent classification routing module (21) is used to send the natural language consultation instruction message into a preset large language model, extract the query entity features through the attention mechanism, and output the target qualification category label. The policy text regular expression extraction module (22) is used to parse unstructured policy texts retrieved from the distributed vector database that match the target qualification category labels, and extract the lower limit threshold of R&D expense ratio. With the lower limit threshold of the number of intellectual property rights Construct a qualification review threshold matrix.

4. The intelligent question-and-answer system for enterprise consultation based on a large language model according to claim 3, characterized in that, The neural network architecture of the pre-set large language model includes a word embedding layer, a multi-head self-attention encoder, and a fully connected classification layer; then the intent classification routing module (21) inputs the natural language consultation instruction message into the word embedding layer, outputs the corresponding intent category and matches and maps it to the target qualification category label, and the specific data flow steps involved are as follows: The word embedding layer receives the natural language consultation instruction message, calls a preset word segmentation dictionary to parse it into a discrete token sequence, and maps the discrete token sequence into a continuous high-dimensional input vector sequence; The multi-head self-attention encoder receives the high-dimensional input vector sequence, constructs a query matrix, a key matrix, and a value matrix based on the high-dimensional input vector sequence, calculates the attention weight distribution, and uses the attention weight distribution to perform a weighted summation on the value matrix to extract the context semantic feature matrix. The fully connected classification layer performs a linear transformation on the context semantic feature matrix, inputs a normalized exponential function for probabilistic activation, and outputs a probability distribution vector corresponding to the predetermined compliant business space. Extract the classification node corresponding to the maximum likelihood probability in the probability distribution vector as the corresponding intent category, and trigger the addressing instruction to perform key-value matching in the intent mapping hash table of static random access memory to obtain the target qualification category label.

5. The intelligent question-and-answer system for enterprise consultation based on a large language model according to claim 3, characterized in that, The policy text regular expression extraction module (22) extracts the lower limit threshold and constructs the qualification review threshold matrix. The specific steps involved are as follows: Based on unstructured policy texts retrieved from a distributed vector database that match the target qualification category tags, and segmented into multiple discrete syntactic logic blocks based on preset punctuation identifiers; Feature traversal is performed on each of the syntactic logic blocks based on a pre-set multimodal regular expression constraint template; wherein, the multimodal regular expression constraint template includes a first regular expression template and a second regular expression template, and each of the first regular expression template and the second regular expression template contains a text prefix anchor and a regular expression capture group; The first regular expression template performs character traversal on each of the syntactic logic blocks to extract the lower limit threshold of R&D expense ratio. ; The second regular expression template performs character traversal on each of the syntactic logic blocks to extract the lower limit threshold for the number of intellectual property rights. ; Extract a contiguous physical address space from the system memory bus and set the lower limit threshold for the R&D expense ratio. With respect to the lower limit threshold of the number of intellectual property rights The qualified review threshold matrix is ​​constructed by sequentially pushing the memory offsets into the continuous physical address space.

6. The intelligent question-and-answer system for enterprise consultation based on a large language model according to claim 1, characterized in that, The static evaluation unit (3) includes a heterogeneous feature mapping module (31) and a rigid numerical comparison module (32). The heterogeneous feature mapping module (31) is used to establish a data type alignment and physical addressing mapping channel between the structured static operational feature parameters and the qualification review threshold matrix, namely the heterogeneous mapping channel; The rigid numerical comparison module (32) is used to extract the corresponding feature parameters and threshold factors along the heterogeneous mapping channel, perform machine-level Boolean comparison, and force the corresponding Boolean state identifier to be configured as the time-delay rejection identifier.

7. The intelligent question-and-answer system for enterprise consultation based on a large language model according to claim 6, characterized in that, The rigid numerical comparison module (32) performs machine-level Boolean comparison along the heterogeneous mapping channel, and the specific steps involved are as follows: The total annual R&D expenses extracted from the structured static operational characteristic parameters along the heterogeneous mapping channel. Compared with the most recent main business revenue And the lower limit threshold of R&D expense ratio extracted from the qualification review threshold matrix. Determine the total amount of annual R&D expenses collected. Compared with the most recent main business revenue Is the ratio greater than or equal to the lower limit threshold for the proportion of R&D expenses? ; Simultaneously, the number of intellectual property rights is extracted from the structured static operational feature parameters along the heterogeneous mapping channel. And extract the lower limit threshold for the number of intellectual property rights from the qualification review threshold matrix. Determine the number of intellectual property rights. Is it greater than or equal to the lower limit threshold for the number of intellectual property rights? ; If the total annual R&D expenditure is determined to be discrete Compared with the most recent main business revenue The ratio is lower than the lower limit threshold for the proportion of R&D expenses. If the historical static financial accounting data has a physical lag defect in representing the R&D status, then the corresponding Boolean status identifier is used as the time-lag rejection identifier to generate the first static compliance assessment vector data stream.

8. The intelligent question-and-answer system for enterprise consultation based on a large language model according to claim 2, characterized in that, The compliance generation unit (4) includes an edge computing power compensation module (41) and an offset intervention decoding module (42). Among them, the edge computing power compensation module (41) listens to the physical computing power load status of the target node in real time to construct the edge computing power consumption drift rate, and uses the edge computing power consumption drift rate as a dynamic weight factor to perform nonlinear feature weight compensation on the total R&D expenses in the structured static operation feature parameters, and outputs dynamic R&D parameters. The bias intervention decoding module (42) is used to receive the dynamic R&D parameters, re-compare them with the qualification review threshold matrix, generate a physical bias weight matrix based on the edge computing power consumption drift rate mapping, apply bias intervention constraints to the autoregressive decoding calculation of the preset generative language model, and finally output the dynamic compliance conclusion text data stream.

9. The intelligent question-and-answer system for enterprise consultation based on a large language model according to claim 8, characterized in that, The edge computing power compensation module (41) outputs dynamic R&D parameters bound to the physical state, and the specific steps involved are as follows: Based on the hardware interrupt signal triggered by the time-delay rejection flag, the network topology routing calibration code corresponding to the target enterprise identification information is extracted, and the hardware performance counter on the corresponding underlying edge computing node motherboard is activated. Within the time interval defined by the timing review sliding window parameter, capture the number of concurrent CPU threads that are actively running on the underlying edge computing node. Variance of concurrent requests per second with cloud API hardware gateway ; The number of concurrent CPU threads The variance of concurrent requests per second compared to the cloud API hardware gateway The optimization solution is performed to construct the edge computing power consumption drift rate. ; The edge computing power consumption drift rate As the dynamic weighting factor, and by invoking a pre-set nonlinear exponential smoothing compensation equation, the extracted total annual R&D expenditure is aggregated. Perform numerical compensation and output the dynamic R&D parameters. .

10. The intelligent question-and-answer system for enterprise consultation based on a large language model according to claim 9, characterized in that, The specific steps involved in the biased intervention decoding module (42) performing a secondary comparison and outputting the dynamic compliance conclusion text data stream are as follows: For the aforementioned dynamic R&D parameters Compared with the most recent main business revenue Perform a floating-point division operation to generate a dynamic R&D ratio, and then compare the dynamic R&D ratio with the lower limit threshold of R&D expense ratio in the qualification review threshold matrix. Execute machine-level Boolean comparison instructions; Based on the logical judgment result of the machine-level Boolean comparison instruction, the status register is triggered to overwrite the corresponding status flag bit, so as to output the secondary comparison pass status result represented by the Boolean logic status value, and overwrite the original time-delay rejection flag as the secondary comparison success status code. The pre-set generative language model is invoked, with the edge computing power consuming drift rate. As a decay control variable, it is mapped to generate the physical bias weight matrix. ; The successful secondary comparison status code is compared with the dynamic R&D parameters. The input is fed into the generative language model, and the physical bias weight matrix is ​​adjusted accordingly. Tensor multiplication is performed, and the generative language model is autoregressively decoded to output a dynamic compliance conclusion text data stream.

Images