A method and system for multimodal information extraction from multi-page image documents

By employing a multimodal information extraction method that combines page-by-page classification, interval aggregation, and field-level rolling scheduling, the problems of redundant reasoning, visual information loss, and system instability in multi-page image document processing are solved, achieving efficient and accurate multi-page document information extraction and segmentation classification.

CN123072045APending Publication Date: 2026-11-10浙江实在智能科技有限公司 +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611571954.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-12
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing multi-modal extraction technologies for multi-page image documents suffer from problems such as redundant reasoning and uncontrolled costs, loss of visual semantic information, amplification of OCR cascade errors, slow location of key fields in long documents, lack of entity-level error correction capabilities, and coarse concurrent scheduling, making it difficult to meet the efficiency and accuracy requirements of enterprise-level large-scale processing.

Method used

The method employs page-by-page classification and aggregation of continuous similar intervals, category-based differentiated extraction, cost-driven orchestration, field-level scrolling skipping, high-concurrency closed-loop control, and result correction. It directly processes image data through a multimodal large model, and combines field-level scrolling scheduling and a unified control plane to optimize task orchestration and concurrency management, thereby achieving field-level scrolling scheduling and result correction.

Benefits of technology

Significantly reduces the number of calls and economic costs, improves processing efficiency and accuracy, enhances system stability and operability, adapts to multi-document segmentation and classification scenarios, and ensures rapid detection of key fields and traceability of results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123072045A_ABST
    Figure CN123072045A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of intelligent document processing technology, specifically relating to a method and system for multimodal information extraction from multi-page image documents. The method includes: uniformly converting images from each page into image data adapted to a large multimodal model input; constructing a single-label classification task for each page image, determining the document category page by page, and merging consecutive pages of the same type into classification intervals; selecting a set of field rules and prompt words according to the classification intervals; combining candidate extraction tasks based on the number of pages and the number of field rules, supporting multi-image merging requests; performing field-level rolling scheduling based on field completion status, with fields that have been hit and configured to be skipped not being scheduled in subsequent pages; executing the extraction task through a unified control plane, which includes token bucket shaping, PID closed-loop speed regulation, circuit breaking, and exponential backoff retries, with classification and extraction sharing the same control plane; collecting extraction results, backfilling the source page numbers, normalizing and automatically correcting specified fields, and outputting structured results and classification intervals.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of intelligent document processing technology, specifically relating to a method and system for extracting multimodal information from multi-page image documents. Background Technology

[0002] Intelligent document processing aims to transform unstructured or semi-structured documents such as scanned documents, image-based PDFs, invoices, contracts, and archives into structured data. Traditional solutions typically employ a cascaded path of "image preprocessing—OCR—layout analysis—rule or model extraction," which offers some engineering usability in scenarios with fixed layouts, good scan quality, and stable field positions. With the development of multimodal large-scale models, these models can directly receive page images and simultaneously understand text, layout, and visual cues, providing new possibilities for end-to-end information extraction. Existing technologies mainly focus on OCR cascaded extraction, template / layout location extraction, document classification and classification-driven extraction, multimodal large-scale model document understanding, multi-page document visual question answering, and interface rate limiting and circuit breaking.

[0003] However, existing multimodal extraction techniques for multi-page image documents still have the following prominent problems: 1. The full combination of "page × field" leads to redundant reasoning and uncontrolled costs: The same field often appears only once or a few times in a multi-page document, but the full field solution per page forces it to be repeatedly reasoned. The number of API calls and costs increase exponentially with the number of pages and fields, making it difficult to support large-scale enterprise-level processing.

[0004] 2. Lack of cost awareness and strategic adaptability in task scheduling: Different combinations of "page / field / image count" result in significant differences in fixed overhead, image token overhead, and field token overhead per request. Existing methods typically employ a fixed organization method and cannot adaptively select a lower-cost orchestration strategy based on the number of pages, fields, page group size, and rule group size. They also struggle to amortize fixed call overhead by combining multiple images into a single request.

[0005] 3. Loss of visual semantic information and amplification of OCR cascade errors. OCR-based cascaded solutions lose visual information such as layout, seals, watermarks, and table borders in the "image → text" step, and OCR recognition errors are directly amplified into extraction errors. In low-resolution scanning, seal obstruction, tilted photography, and complex table scenarios, a single character error in key fields such as amount, account number, and invoice number can cause the field-level accuracy to drop to zero.

[0006] 4. Slow location of key fields in long documents: Key fields in contracts, documents, and other files are often located on the first and last pages. For example, the client's name is usually on the first page, while the date and signature are usually on the last page. A sequential page-by-page scan requires traversing to the very end to find the relevant fields, resulting in numerous invalid calls on intermediate pages, extended inference time, and an inability to meet the high-efficiency business requirements for the first valid result.

[0007] 5. Lack of entity-level error correction and normalization capabilities: For long entities such as company and organization names, the model or OCR output often has problems such as typos due to similar character shapes, missing characters, and inconsistencies between simplified / traditional and full-width characters. Existing solutions lack configurable automatic correction and normalization mechanisms with ambiguous criteria, which means that downstream businesses cannot directly use the extracted results and still require a lot of manual review.

[0008] 6. Inefficient concurrency scheduling and flow control: Simple concurrency lacks token bucket shaping, PID rate closed-loop, dynamic concurrency scaling, slow start, circuit breaking, and rate reduction mechanisms, making it prone to triggering upstream rate limiting and causing failures and retries. Static rate limiting faces a dilemma when upstream capacity fluctuates: too small a limit leads to insufficient throughput, while too large a limit leads to 429 errors, timeouts, and retry storms. The system is prone to avalanche under high load or abnormal scenarios.

[0009] 7. Lack of multi-document segmentation and classification capabilities: In real-world scenarios, multi-page documents often contain a mix of invoices, contracts, bank statements, itineraries, and other document types. This necessitates the ability to "classify page by page and merge consecutive pages of the same type into intervals" so that different extraction templates can be applied to each interval. Existing single-category classification schemes for the entire document cannot meet the needs of mixed-format scenarios. Furthermore, classification and extraction tasks often employ two separate scheduling, rate limiting, and retry mechanisms, resulting in a complex architecture and inconsistent governance strategies.

[0010] In summary, existing technologies cover individual dimensions such as OCR recognition, template extraction, invoice processing, multimodal model training, document question answering, or general rate limiting. However, none of them explicitly or implicitly suggest integrating the entire process of "page-by-page classification → aggregation of consecutive pages of the same type → category-driven extraction → multi-image unified cost planning → field-level rolling ablation → high-concurrency closed-loop control → result fusion and entity correction" into a single multimodal document processing pipeline. Therefore, in scenarios involving multimodal information extraction and document segmentation and classification of multi-page image documents, the challenge of simultaneously achieving high accuracy, processing efficiency, economic cost, and system stability remains. Summary of the Invention

[0011] This invention aims to overcome the problems of existing multi-page image document processing methods based on OCR cascading or template matching, such as OCR error cascading and visual information loss, high cost due to full combination of page-by-page and field-by-field, lack of segmentation and classification of mixed documents, slow location of key fields in long documents, lack of entity error correction, and coarse concurrency control. It provides a multimodal information extraction method and system for multi-page image documents that can classify and aggregate continuous similar intervals page-by-page, extract information differentially by category, use cost-driven arrangement, allow field-level scrolling and skipping, implement high-concurrency closed-loop control, and correct results.

[0012] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: A method for extracting multimodal information from multi-page image documents includes the following steps; S1, convert the images from each page of a multi-page image document into image data that is adapted to the input of a multimodal large model; S2 constructs a single-label classification task for each page of images, uses a multimodal large model to determine the document category page by page, and merges consecutive pages of the same category into a classification interval according to page order; S3, select the corresponding set of field rules and prompt words according to the document category of the classification range; S4 combines multi-page images with field rules into candidate extraction tasks based on the number of pages and the number of field rules, and supports merging multi-page images corresponding to the same rule into a multi-image merging request; S5. For the candidate extraction tasks generated in step S4, perform field-level rolling scheduling based on the field completion status: maintain the field completion status at the field level, and for fields configured to be skipped after a hit, if the field has been successfully extracted in any page of images, skip the field in the task scheduling of subsequent pages, and only add the unhit fields to the task to be scheduled; the field completion status is updated in real time according to the extraction results. S6, the extraction task is executed through a unified control plane, which includes a global semaphore, asynchronous queue, token bucket shaping, PID closed-loop speed regulation based on queue length, circuit breaking and exponential backoff retry, and the classification task and the extraction task share the same worker thread pool, token bucket and circuit breaking retry strategy. S7 collects the extraction results, backfills the source page number for the field values, performs normalization and automatic correction on the specified fields, and outputs the structured extraction results and classification intervals.

[0013] Preferably, in step S1, image access supports dataURL passthrough, HTTP / HTTPS image download, and object storage relative path reading, and all accessed images are uniformly converted into Base64 encoded image data; wherein, HTTP / HTTPS image download is executed through an asynchronous client, and object storage reading encapsulates synchronous storage I / O in a thread pool for execution, and download concurrency is limited by semaphores during the image conversion stage.

[0014] Preferably, in step S2, the classification task is uniformly modeled as a single-field task of document category. The classification prompt includes a list of candidate categories, category descriptions, and strict JSON output constraints. When parsing the classification response, only predefined categories are accepted, and illegal categories or empty results are marked as unclassified. The classification task and the extraction task reuse the same response parsing and page number backfilling mechanism.

[0015] Preferably, in step S3, the prompt words are constructed using a system-level hard constraint and field three-element separation method; wherein, each field is described by three elements: display name, internal code, and semantic description, the display name is used as the JSON key of the multimodal large model output, and the internal code is used for internal system association; the prompt words require that the multimodal large model output is pure JSON, all fields must appear, and fields that have not been extracted with valid values ​​are represented by empty arrays; when cleaning the response of the multimodal large model, it is compatible with the message response structure returned by different upstream models, stripping code block markers and thought chain text, and extracting objects that conform to the JSON format.

[0016] Preferably, in step S4, the candidate extraction task is generated in the following way: cost estimation is performed based on fixed cost, image cost, and rule cost, and an orchestration method is selected between page-first strategy and rule-first strategy according to the estimated total cost; wherein, the page-first strategy takes a page or page group as the outer dimension, and the rule-first strategy takes a field or field group as the outer dimension; when the maximum number of images requested in a single request is less than or equal to 1, the page-first strategy is automatically adopted.

[0017] Preferably, in step S4, the multi-image merging request submits multiple pages of images corresponding to the same rule together in a single request, and requests the multimodal large model to return an array organized by page, the array including multiple page objects, each page object containing a page number and fields; if the multi-image merging request fails due to the interface limitations of the multimodal large model, it is automatically downgraded to a page group or single-page task with fewer pages.

[0018] Preferably, in step S5, the completion status of a field is uniquely identified by its internal code. A field can be configured to be skipped after a single extraction or hit. Once a field configured to be skipped after a single extraction or hit is successfully extracted, it will not be scheduled for subsequent pages. Page access order strategies include strict page-by-page, look-ahead scrolling, edge-first, and grouped scrolling. Edge-first scrolling accesses the document alternately from both ends towards the middle, following the order of first page, last page, second first page, and second last page. Grouped scrolling organizes field rules and pages into a two-dimensional task matrix, generating parallel batches along the anti-diagonal line, where the anti-diagonal line is the line in the two-dimensional task matrix whose sum is equal to the sum of the row containing the rule and the column containing the page. Before each batch begins, triangular ablation is performed on field rules that have been hit and configured to be skipped. Triangular ablation refers to removing candidate nodes of the field rule in the subsequent page direction from the batch, creating skipped areas in the task matrix.

[0019] Preferably, in step S6, the unified control surface performs speed regulation and protection in the following manner: The PID closed-loop speed control mechanism adjusts the token generation rate based on the error between the current queue length and the target queue length; the circuit breaker mechanism includes closed, open, and half-open states, and enters the half-open state for detection after cooling down in the open state; when a rate limiting response is received, the current token generation rate is immediately halved, and exponential backoff retries are performed for timeouts, temporary network anomalies, and retryable upstream errors; after service startup or circuit breaker recovery, slow start is performed, and the number of concurrent worker threads is dynamically adjusted based on the average response time and queue depth.

[0020] Preferably, in step S7, multi-level normalization, length-based bucketing candidate retrieval, and edit distance calculation are performed on the specified field, which includes the company or organization name; Whether to perform automatic correction is determined based on ambiguity judgment conditions, which include: For short names, the original value is retained without correction; the difference in edit distance between the best and second-best candidates is not less than the configuration margin; the similarity of the correction results is not lower than the threshold. Automatic correction is performed only when the optimal candidate is unique and meets the aforementioned ambiguity judgment conditions, and the original value, edit distance, and similarity annotation are retained during correction; empty strings, empty arrays, and invalid objects are filtered during result collection, and the source page number is written for each valid match.

[0021] The present invention also provides a multimodal information extraction system for multi-page image documents, comprising: The image access and preprocessing module is used to uniformly convert the images on each page into image data that is adapted to the input of a multimodal large model; The page-by-page multimodal classification and interval aggregation module is used to determine the document category page by page and merge consecutive pages of the same type into a category interval; The category-driven extraction configuration module is used to select the set of field rules and prompt words according to the category range; The cost-driven task planning module is used to generate candidate extraction tasks based on cost estimates and supports multi-graph merging requests. The field-level incremental scrolling scheduling module is used to maintain the completion status of fields at the field level. For fields configured to be skipped after being hit, they will not be scheduled again on subsequent pages after successful extraction. Only fields that were not hit will be scheduled. A high-concurrency multimodal inference engine is used to execute extraction tasks through a unified control plane, which includes global semaphores, asynchronous queues, token bucket shaping, PID closed-loop speed regulation, circuit breaking, and retrying. The classification task and the extraction task share the same control plane. The results fusion and correction module is used to backfill the source page number for field values, normalize and automatically correct specified fields, and output structured results and classification intervals.

[0022] Compared with the prior art, the beneficial effects of this invention are: (1) significantly reducing the number of calls and economic costs: the classification pre-processing mechanism enables the system to activate the corresponding fields and prompt words only in the corresponding category interval, cutting off the invalid reasoning of irrelevant fields on irrelevant pages from the source; the field-level incremental scrolling scheduling tracks the field hit status with the field internal code (tag_code), and the hit fields are automatically skipped across pages, so that the number of API calls is reduced from the product complexity of "number of pages × number of fields" to an approximately linear scale related to the actual occurrence position of the field; the cost-driven planner automatically selects the best option between page-major and rule-major according to the cost estimation model, and the multi-graph fusion request further reduces the fixed overhead of a single call. The three work together to systematically compress the total call cost of processing large batches of multi-page documents: (2) Significantly improve processing efficiency and response speed: The edge-rolling strategy alternates between the beginning and end of the document and the middle, so that high-frequency key fields such as the name of the party on the first page of the contract and the signature date on the last page can be hit in the first round, significantly shortening the time to return the first effective result (TTFR, Time To First Result); Combined with the high-concurrency asynchronous worker pool (WorkerPool), token bucket smoothing, proportional-integral-derivative (PID) closed-loop speed regulation and dynamic concurrent scaling, the overall processing latency is further narrowed in batch scenarios; (3) Fundamentally eliminate OCR error cascading and improve extraction accuracy: This invention uses page images as the direct input of multimodal large models, and completes page understanding, text recognition and field extraction end-to-end, fully preserving and utilizing visual semantic information such as page layout, table borders, seals, watermarks, and handwritten annotations, avoiding the recognition errors in OCR cascading schemes. Error propagation amplifies the extraction process; the company / organization name directory correction mechanism further corrects the long entity recognition results through multi-level normalization, length bucket pruning, editing distance and multiple ambiguity criteria, effectively improving the output quality of key fields; (4) Enhance system robustness and stability: This invention constructs a unified high-concurrency call control plane, organically combining global semaphores, token bucket shaping, PID queue closed-loop speed regulation, dynamic worker scaling, slow start warm-up, progressive circuit breaking (CLOSED / OPEN / HALF_OPEN), exponential backoff retry and 429 fast deceleration.This control plane enables the system to have adaptive adjustment capabilities in upstream flow limiting, network jitter or abnormal recovery scenarios, significantly reducing the probability of 429 errors, avalanches and invalid retries, and maintaining stable processing performance in high-throughput batch scenarios; (5) Automated task orchestration and scale adaptation: The cost-driven task planner does not require manual parameter tuning when the number of pages / rules changes, and automatically selects the lower-cost orchestration method between page-major or rule-major, avoiding the cost degradation caused by fixed orchestration strategies when the task scale changes. This mechanism enables the solution to smoothly adapt to a wide range of task scales, from single-page single-field to hundreds of pages and hundreds of fields; (6) Native support for multi-document segmentation classification: This invention sinks the classification granularity to the "page-by-page" level, and through page-by-page multimodal classification and continuous similar page interval aggregation, it naturally adapts to real business scenarios where multiple document types such as invoices, contracts, and bank receipts are mixed in a single document; The classification interval serves as prior information in the extraction stage, so that the application of subsequent differentiated templates / prompt words has a clear scope boundary, rather than relying on the single category assumption of the entire document; (7) Enhanced operability and traceability: The classification interval and field-level extraction results are associated with the source page number, supporting downstream page-by-page segmentation, verification, and transfer; The field definition adopts a three-element separation design of "display name | internal code | semantic description", taking into account both model readability and system relevance; The directory correction process retains the original value, edit distance, and similarity annotation, and each correction can be traced, which is convenient for manual review, auditing, and rollback; The unified authentication encapsulation and response cleaning layer further reduces the complexity of operation and maintenance. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall process of the multimodal information extraction method for multi-page image documents in this invention; Figure 2 This is a flowchart of the pre-classification and category prompt word-driven extraction process in this invention; Figure 3 This is a flowchart illustrating the high-concurrency scheduling control logic in this invention; Figure 4 This is a schematic diagram illustrating the anti-diagonal parallel batch and triangular ablation process in the group rolling mode of this invention; Figure 5 This is a schematic diagram of a process for segmenting, classifying, and differentially extracting mixed ticket packages according to an embodiment of the present invention. Detailed Implementation

[0024] To more clearly illustrate the embodiments of the present invention, specific implementation methods will be described below with reference to the accompanying drawings. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without any creative effort.

[0025] like Figure 1 As shown, this invention proposes a method and system for large-scale multimodal information extraction and document segmentation and classification of multi-page image documents. The overall process is as follows: adaptive image source access -> page-by-page multimodal classification -> aggregation of consecutive pages of the same type -> selection of field rules and prompt words by category -> cost-driven task planning -> field-level rolling scheduling -> high-concurrency multimodal inference -> result fusion, page number backfilling, and entity correction.

[0026] The method specifically includes the following steps: 1. Convert the images on each page of a multi-page image document into image data that is adapted to the input of a multimodal large model; 2. Construct a single-label classification task for each page of images, use a multimodal large model to determine the document category page by page, and merge consecutive pages of the same category into classification intervals according to page order; 3. Select the corresponding set of field rules and prompt words according to the document category of the classification range; 4. Based on the number of pages and the number of field rules, combine multi-page images with field rules into candidate extraction tasks, and support merging multi-page images corresponding to the same rule into a multi-image merging request; 5. For the candidate extraction tasks generated in step 4, perform field-level rolling scheduling based on the field completion status: maintain the field completion status at the field level, and for fields configured to be skipped after a hit, if the field has been successfully extracted in any page of images, skip the field in the task scheduling of subsequent pages, and only add the fields that were not hit to the task to be scheduled; the field completion status is updated in real time according to the extraction results. 6. The extraction task is executed through a unified control plane, which includes a global semaphore, asynchronous queue, token bucket shaping, PID closed-loop speed regulation based on queue length, circuit breaking and exponential backoff retries, and the classification task and the extraction task share the same worker thread pool, token bucket and circuit breaking retry strategy. 7. Collect the extraction results, backfill the source page number for the field values, perform normalization and automatic correction on the specified fields, and output the structured extraction results and classification intervals.

[0027] The following provides a detailed description of each core module in the system of this invention: (I) Image Source Adaptive Access and Preprocessing Module Unified Input Access: The access layer supports three types of input: dataURL passthrough, HTTP / HTTPS image download, and object storage relative path reading. These inputs are uniformly converted into a multimodal model-consumable format: data:image / <specific image subtype>;Base64, <Base64 encoded data of the image>.

[0028] Asynchronous and threaded models: HTTP downloads obtain image content through asynchronous clients and parse the Content-Type from the response header; object storage reads encapsulate synchronous storage I / O in a thread pool to avoid blocking the event loop.

[0029] Concurrency control and observability: During the image conversion stage, download concurrency is limited by semaphores, and request identifier request_id, image sequence number, source type, time consumption and exception information are recorded, so that the document classification and information extraction links share the same set of image preparation capabilities.

[0030] (II) Pre-classification and Category Differentiation Hint Module like Figure 2 As shown, this invention uses document classification as a pre-extraction step. The system constructs a single-label classification task for each page, writing a list of categories and their descriptions into classification prompts, requiring the model to output only JSON with the specified category names. The system constructs a document category field (document_category) classification rule for each page, with classification prompts including a list of candidate categories, category descriptions, and strict JSON output constraints. The classification task and extraction task use the same worker pool, the same token bucket, and the same circuit breaker retry strategy, naturally protecting the classification phase from production-level concurrency control. During classification response parsing, the system extracts candidate values ​​from categories, values, or similar keys, accepting only predefined categories; illegal categories or empty results are marked as unclassified.

[0031] After classification, the system scans the classification results page by page, merging consecutive pages of the same category into intervals. For example, if P1-P8 are invoices, P9-P16 are contracts, P17-P20 are bank receipts, and P21-P30 are itineraries, the output would be {Invoice:[[1,8]], Contract:[[9,16]], Bank Receipt:[[17,20]], Itinerary:[[21,30]]}. During the extraction phase, the system selects the field set, sets rules, and provides rule group prompts based on the interval to avoid ineffective tasks such as extracting invoice tax amounts from the contract page or contract signature dates from the invoice page.

[0032] (III) Cost-driven multimodal task planner (task splitting and orchestration strategy selection) This module is one of the core components of this invention. Given N pages of images and M extraction rules (fields), the system does not blindly perform a full combination of "page × rule". Instead, it first divides the N pages into several page groups according to the "maximum number of images per request" and divides the M rules into several rule groups according to the "maximum number of rules per request" (to avoid excessively long prompts per request). Then, it uniformly estimates and compares the costs of the two arrangement strategies and selects the optimal one.

[0033] Task unit definition: Page-major strategy: using "page × rule group" as the task unit; Rule-major strategy: The task unit is "single rule × page group". It supports carrying multiple pages of images in a single request and requires the model to return field results page by page according to the page array.

[0034] Cost estimation model: The overhead of a single multimodal call is broken down into three components: fixed cost, image cost, and rule cost. Cost per request = C_fixed + C_image × number of images + C_rule × number of rules; Where C_fixed represents the inherent cost (including latency) of each request, C_image represents the cost of each page image, and C_rule represents the cost of each rule. The total cost of the two strategies can then be estimated separately. Page-major total cost ∝ number of pages × number of rule groups × (C_fixed + C_image + C_rule × number of rules within a group); Total cost of rule-major rule is calculated as follows: (Number of rules × Number of page groups × (C_fixed + C_image × Number of pages within a group + C_rule)).

[0035] Adaptive strategy selection: The system defaults to auto mode, calculates and compares the estimated total cost of the two strategies, and automatically selects the one with the lower cost; it also supports forcibly specifying page-major or rule-major through configuration; when "maximum number of images per request ≤ 1" (i.e., multi-image merging is not supported), it automatically degenerates into the page-first strategy.

[0036] Multi-image merging request: When the rule priority strategy is selected, the system submits multiple pages of images corresponding to the same rule together in a single request. The model returns a structured result organized by page at once, thereby reducing the fixed overhead of each call and significantly reducing the total number of calls.

[0037] (iv) High-concurrency processing and scheduling control module (refined flow control) like Figure 3As shown, the high-concurrency control plane consists of four layers: the scheduling layer is responsible for classifying and extracting subtasks and putting them into the asynchronous queue; the shaping layer controls the instantaneous concurrency and request rate through global semaphores and token buckets; the feedback layer adjusts the token generation rate according to the queue length through PID; and the protection layer achieves fault recovery through circuit breaking, half-open detection, slow start, retry, and 429 fast deceleration.

[0038] Global semaphore: The specific implementation method is process-level sharing. Both classification and extraction require acquiring the semaphore first to prevent multiple business requests from overlapping and impacting the upstream.

[0039] Token bucket: The specific implementation is that the worker thread obtains a token before initiating a model call; the token is replenished at the current rate to smooth out sudden traffic.

[0040] PID controller: The specific implementation method is to calculate the error e every 1 second = current queue length - target queue length, and adjust the rate according to Kp*e + Ki*cumulative error + Kd*error change, and then use upper and lower limits for pruning.

[0041] PID tuning: The default parameters are P first, then I, then D; the specific implementation is to first set Ki=Kd=0, then gradually increase Kp until the queue length shows slight oscillations; add Ki to eliminate long-term backlog; add Kd to suppress overshoot. If 429 increases or the average response time increases, then decrease Kp or the upper limit rate.

[0042] Slow start: Specifically, after the service starts or the circuit breaker recovers, the rate gradually increases until the PID recommended value or the maximum rate is reached.

[0043] Circuit breaking: The specific implementation is as follows: in the CLOSED state, failures accumulate; in the OPEN state, new requests are rejected or fail quickly; after cooling down, in the HALF_OPEN state, a small number of probes are allowed to pass, and if the probe is successful, it is closed, and if it fails, it is reopened.

[0044] Retry: Specifically, it involves exponential backoff for timeouts, temporary network anomalies, and retryable upstream errors; and immediately halving the current rate upon detecting a 429 error or a rate limit response.

[0045] The above mechanisms together achieve closed-loop adaptive governance of multimodal calls, which is the underlying capability that can be reused for text and multimodal, extraction and classification.

[0046] (v) Field-level incremental rolling scheduling module (skip if field is already selected) This module represents the most targeted innovation of this invention in terms of "cost reduction and speed improvement." Its core idea is that a field in a multi-page document typically only needs to be successfully extracted once. Once a field has been matched on a page, subsequent pages do not need to initiate inference for that field again, as detailed below: (1) The field completion tracker system uses tag_code as the unique identifier for each field and maintains the field completion tracker. Each field can declare whether to "skip after a hit" through extract_once (this parameter controls whether all fields are extracted at once) or skip_after_found (this parameter controls whether to skip after a hit). If not explicitly declared, the global default value controls this. Before each round of scheduling, the system only takes the fields that have not yet been completed to form a list of pending fields to be extracted, pending_rules. When a field returns a non-empty valid value, it is marked as completed, and subsequent pages will no longer initiate reasoning for it. When all fields have been hit or all pages have been traversed, the rolling scheduling automatically terminates to ensure that the process will inevitably converge.

[0047] (2) Page access order strategy: In addition to the above field-level incremental mechanism, the system further provides a variety of configurable page access order strategies: strict (strict page-by-page): The window size is 1, and it proceeds strictly in page order; rolling (look-ahead scrolling): Advances multiple pages at once according to the configured look-ahead page number, taking into account the continuous information of adjacent pages; edge_rolling (edge ​​priority): The access order is 0 (first page), N-1 (last page), 1 (secondary first page), N-2 (secondary last page)..., squeezing from both ends of the document towards the middle, prioritizing common fields on the first / last page (such as the name of the party A on the first page of the contract, and the date and signature on the last page). group_rolling: Organizes the rule dimension and image dimension into a two-dimensional task matrix, forming parallel batches along the anti-diagonal.

[0048] (3) Two-dimensional matrix batching and triangular ablation in group_rolling: The group_rolling mode abstracts candidate tasks into a two-dimensional matrix of "rule R_i × image P_j": the horizontal axis represents the image page, and the vertical axis represents the rule (or rule group). Each cell represents a candidate subtask that can be submitted to the multimodal large model once. For example Figure 4As shown, the horizontal axis represents images P1-P5, and the vertical axis represents rules R1-R5; dark lines (■) indicate tasks actually executed, and light lines (□) indicate tasks skipped because the rules have already been extracted. The system generates batches according to the anti-diagonal (i+j is a constant) – B1 contains (R1, P1), B2 contains (R1, P2) and (R2, P1), and so on until the last batch. The same batch markers B1-B9 indicate the same anti-diagonal where i+j is a constant, and these can be executed concurrently within the same batch. Since tasks on the same anti-diagonal only depend on the field completion status of earlier batches before execution, there is no sequential dependency within the batch, allowing for simultaneous concurrent submission. There is no sequential dependency of "field hit" between nodes within a batch, so they can be enqueued simultaneously; batches advance from B1 to B9.

[0049] Before each batch begins, the scheduler reads the field completion tracker and performs triangular ablation on the hit rules: if a rule has already extracted a valid value on the previous image and the rule is configured as extract_once / skip_after_found, then all candidate nodes for that rule in the subsequent image direction are removed from the batch, so that the upper right region of the matrix forms a "skip triangle", that is, a shallow textured area of ​​the upper right triangle, to prevent the same rule from continuing to generate invalid inference in subsequent pages.

[0050] When the matrix degenerates into a single row or column, anti-diagonal grouping no longer introduces additional dependency constraints: a single row represents one rule facing multiple pages of images, and a single column represents multiple rules facing one page of images. Both can be directly executed concurrently as a dependency-free batch. In the degenerate scenario, there is no need to wait by anti-diagonal grouping; all subtasks can be launched at once. For example, a single row represents 1 rule × 5 pages of images, with task nodes R1P1, R1P2, R1P3, R1P4, and R1P5; a single column represents 5 rules × 1 page of images, with task nodes R1P1, R2P1, R3P1, R4P1, and R5P1; both can be executed in parallel within the same batch. This degenerate logic ensures that `group_rolling` does not introduce unnecessary waiting in small-scale scenarios.

[0051] (vi) Multimodal Prompt Construction and Inference Execution Module This module is responsible for constructing multimodal prompts with strong constraints and performing inference, which is a key step in ensuring that the extraction results are structured, parsable, and associative.

[0052] (1) The system-level hard constraints and field three-element separation extraction prompt words adopt the design of "system-level hard constraints + field three-element separation". The following hard constraints are fixed in the system prompt: the output must be pure JSON, all fields must appear, fields without values ​​are placed in empty arrays, fabrication and speculation are prohibited, multiple values ​​are arranged in reading order (from left to right, from top to bottom), duplicate values ​​are not deduplicated and must be recorded, and table data is extracted by row and column.

[0053] Each field to be extracted is described by three elements: "Field Name | Internal Code | Meaning," and it is explicitly agreed that the key of the model's output JSON must use the field name; the internal code is only used for internal system associations and is prohibited from being output as a JSON key. This decouples "model-oriented readability" from "system-oriented associativity."

[0054] (2) Differentiated prompts for the three tasks Single-page task: Inject the current page number into the prompt ("This is page N of the document") to guide the model to understand the page number context; Multi-image merging task: The prompt lists the set of pages included in this request and requires the model to return an array of pages (the pages array), with each page object containing two fields: page number (page_num) and fields (fields); Classification task: Use independent classification prompts, requiring that the category value must belong to one of the given category sets (single label, single selection), and strictly constrain the output format.

[0055] (3) Unified upstream call and response cleaning: Upstream model calls are encapsulated through a unified large language model client (LLMClient), which can adaptively adapt to various authentication methods such as bearer tokens, Azure API keys, generative artificial intelligence (AIGC) authentication, and hash-based message authentication code (HMAC) signatures. The response cleaning module is compatible with two response structures: the message content field (message.content) and the message content field of the first element of the option array (choices[0].message.content). It removes code block markers marked with triple backticks and json and redundant thought chain text, and extracts the outermost JSON object that can be directly parsed by the JSON parsing function (json.loads), ensuring that the result can directly enter the structured parsing process.

[0056] (vii) Results integration, entity correction and segmented classification aggregation module This module is responsible for processing the scattered, raw results returned by concurrent inference into structured outputs that can be directly stored in the database. It includes two sub-steps: result collection and null value filtering, and company / organization name directory correction.

[0057] (1) Results collection and null filtering After the concurrent subtasks are completed, the system performs thread-safe accumulation of the results of each subtask according to the field's internal code (tag_code); uniformly filters invalid matches such as empty strings, empty arrays, and invalid objects; and writes a page number (page_num) for each valid match so that the field value can be traced back to the source page.

[0058] (2) Company / Institution Name Directory Correction For the extracted values ​​of specified fields (defined by a configurable list of fields), the system reads a hot-loadable list of fields and a company directory, and performs low-risk automatic correction, mainly including: Multi-level normalization: Unicode compatibility standardization (NFKC) + case folding + removal of punctuation / spaces; Length-based binning of candidate retrieval: The directory is binned according to the normalized length. During retrieval, the distance is only calculated for candidates with similar lengths to avoid full table scanning. Edit distance with pruning (Levenshtein distance): Prune in advance with a maximum allowable distance + margin as the upper limit; Multiple ambiguity criteria: Short names are preserved without correction; the "distance difference" between the best and second-best candidates must not be less than the configuration margin, otherwise the ambiguity is determined and not corrected; the corrected "similarity" must not be lower than the threshold—that is, automatic correction is only performed when the best candidate is unique and the risk is sufficiently low; in this embodiment, the similarity is not based on semantic vector similarity, but on normalized string similarity based on Levenshtein edit distance, to ensure that the company name correction process is interpretable and reproducible. The similarity threshold, the maximum allowed edit distance, and the candidate distance difference margin are all configurable.

[0059] Traceability: The original value is preserved during correction, and edit distance and similarity annotations are generated to facilitate auditing and rollback; Hot loading: Changes to directory and field list files are detected by signature with "modification time + file size" and take effect without restarting.

[0060] Through the in-depth processing of this module, scattered, raw, and potentially error-prone model outputs are transformed into clean, consistent, traceable, and directly integrateable structured data.

[0061] In the above embodiments, the image source adaptive access and preprocessing module (i.e., the image access and preprocessing module) is used to execute method step 1; the classification pre-processing and category differentiation prompt module (i.e., the page-by-page multimodal classification and interval aggregation module and the category-driven extraction configuration module) is used to execute method steps 2 and 3; the cost-driven task planning module (i.e., the cost-driven task planning module) is used to execute method step 4; the field-level incremental rolling scheduling module (i.e., the field-level incremental rolling scheduling module) is used to execute method step 5; the high-concurrency processing and scheduling control module (i.e., the high-concurrency multimodal inference engine) is used to execute method step 6; the strongly constrained multimodal prompt module (i.e., the category-driven extraction configuration module) is used to execute method step 3; and the result integration, entity correction, and segmented classification aggregation module (i.e., the result fusion and correction module) is used to execute method step 7.

[0062] Based on the technical solution of this invention, the following case scenario illustrates the implementation process of this invention in practical applications. The specific application implementation scheme is as follows: like Figure 5 As shown, this embodiment uses a 30-page mixed financial scan package as input data, with a page resolution of approximately 150-200 dpi, including 8 pages of invoices, 8 pages of contracts, 4 pages of bank receipts, and 10 pages of travel itineraries. Candidate categories are set as five: invoices, contracts, bank receipts, travel itineraries, and others. The field sets corresponding to each category are pre-configured according to business needs, totaling 5 fields for invoices, 6 fields for contracts, 5 fields for bank receipts, and 4 fields for travel itineraries.

[0063] The classification and interval aggregation results are as follows: P1-P8: The classification result is invoice, and the range aggregation is {"invoice": [[1, 8]]}. The extracted fields include: invoice number, buyer's name, seller's name, total price including tax, and invoice date.

[0064] P9-P16: The classification result is "Contract", and the interval aggregation is {"Contract": [[9, 16]]}. The extracted fields include: Contract Number, Party A, Party B, Signing Date, Contract Amount, and Signature Page.

[0065] P17-P20: The classification result is bank receipts, and the range aggregation is {"bank receipts": [[17, 20]]}. The extracted fields include: payer, payee, account number, transaction number, and amount.

[0066] P21-P30: The classification result is a travel itinerary, and the interval aggregation is {"travel itinerary": [[21, 30]]}. The extracted fields include: passenger name, flight number, departure point, and arrival point.

[0067] Under the same input conditions, using C_fixed=1, C_image=8, and C_rule=0.2 as indicative weights, the costs of different orchestration strategies are compared, and the results are as follows: Table 1. Cost Comparison of Different Orchestration Strategies

[0068] The above costs are relative comparison examples under illustrative weights. In actual deployment, corresponding coefficients can be configured according to the billing model of the model service provider. During system operation, the cost is calculated in real time based on the current page number N, the number of rules M, the maximum number of images per request Imax, the maximum number of rules per request Rmax, and the field completion status, and the system selects the best option for execution. If the merging of multiple images fails due to limitations on the model side, the system automatically degrades to a smaller page group or a single-page task.

[0069] The following is an example of a key execution process log: Category-based access: task_id=cls_P09, rule=document_category, page=9, result={"category": "contract"}; Range aggregation: {"Invoice": [[1,8]], "Contract": [[9,16]], "Bank Receipt": [[17,20]], "Itinerary": [[21,30]]}; Field tracking: After buyer_name is hit in P2, it is marked as completed=true, and no further inference will be initiated for this field in P3-P8; group_rolling scheduling: Batch B3 contains (R1,P3), (R2,P2), and (R3,P1), and is executed concurrently within the batch; if R2 has already been hit on a previous page, then (R2,P2) and all nodes with the same rule on subsequent pages are triangulated; Current limiting control: The current queue length queue_len=37, the target setpoint=20, and the error e=17. The PID controller increases the token generation rate accordingly. When the upstream returns a 429 current limiting response, the current rate is immediately halved. Fuse recovery: After 10 consecutive failures, the fuse enters the OPEN state, cools down for 30 seconds, switches to the HALF_OPEN state, allows 5 probe requests, and returns to the CLOSED state after all are successful.

[0070] Examples of partial field extraction and correction results are shown below: Table 2. Example of partial field extraction and correction results.

[0071] This embodiment fully verifies the feasibility of the present invention from image access, page-by-page classification, interval aggregation, cost-driven orchestration, field-level incremental scrolling, high-concurrency rate limiting scheduling to result fusion and entity correction. In a 30-page mixed document scenario, compared with the full single-page single-field solution, the actual number of tasks executed is reduced from 600 to about 88, a reduction of more than 85%.

[0072] This invention proposes a method and system for multimodal large-model information extraction and document segmentation and classification of multi-page image documents, effectively solving several core problems in multimodal large-model extraction scenarios for multi-page image documents: 1) Redundant inference and cost overrun caused by the full combination of "page × field"; 2) Lack of cost awareness and scale adaptability in task orchestration; 3) OCR cascade errors and loss of visual information; 4) Slow location of key fields in long documents; 5) Difficulty in correcting misidentification of entities such as company / organization names; 6) Difficulty in segmenting and classifying mixed documents of various types; 7) Slow and unstable inference efficiency under high concurrency scenarios.

[0073] The innovative aspects of this invention are as follows: The classification-first, differential extraction mechanism first classifies each page into multiple modal categories and merges consecutive pages of the same type. Then, based on the category range, it selects the corresponding field set, role_setting, or group_prompt for differential extraction. This reduces invalid reasoning for irrelevant fields on irrelevant pages and is the general principle for cost reduction and efficiency improvement of the "classify first, extract later" pipeline.

[0074] Classification task extraction mechanism: Classification is uniformly modeled as a single-field (single-label) task of document_category, and the WorkerPool, rate limiting, circuit breaking, retry, response parsing and page number backfilling capabilities of the multimodal extraction framework are reused, so that classification and extraction share the same set of scheduling and governance infrastructure.

[0075] Cost-driven multimodal task planning mechanism: Establish a unified cost estimation model for fixed cost, image cost, and rule cost, and estimate the page-major and rule-major orchestration methods respectively. In auto mode, the lower cost is automatically selected to achieve scale-adaptive cost reduction of differentiated orchestration.

[0076] Multi-image merging request and page-based structured return mechanism: A single request carries multiple page images, requiring the model to return the results of each page according to the pages / page_num / fields structure, and the fields of each page are filled accordingly, thus reducing the fixed call overhead.

[0077] Field-level incremental rolling scheduling mechanism: Tracks the completion status of fields with tag_code, supports extract_once / skip_after_found configuration, automatically skips hit fields across pages, and continues reasoning only for hit fields, and converges the number of calls from the product level of "page number × number of fields" to the scale related to the actual occurrence position of the field.

[0078] Edge-first and group-rolling page access strategies: multiple strategies coexist, including strict, rolling, edge_rolling, and group_rolling; edge_rolling squeezes from both ends of the document towards the middle, prioritizing high-frequency key fields on the first and last pages; group_rolling optimizes multiple field groups in parallel, collectively shortening the return time of the first valid result.

[0079] High-concurrency closed-loop flow control engine: It combines global semaphores, asynchronous queues, token bucket shaping, PID queue closed-loop speed regulation, dynamic worker scaling, slow start warm-up, progressive circuit breaking, exponential backoff retries and 429 fast deceleration into a unified control surface to ensure high throughput and low jitter under batch processing.

[0080] Strongly constrained multimodal hint mechanism: separation of the three elements of the field (display name | internal encoding | semantic description), pure JSON hard constraints, page number context injection, multi-image unified hints and category-specific hints, combined with response cleaning and parsing (stripping code blocks and thought chains, and extracting JSON that can be parsed by json.loads).

[0081] Company / organization name directory correction mechanism: Based on NFKC normalization, length binning, pruning edit distance and multiple ambiguity criteria of "candidate distance margin + similarity threshold + short name fidelity", it achieves low-risk automatic correction; it supports hot loading of the directory and retention of original values ​​(original_value / edit_distance / similarity), and is fully traceable.

[0082] A unified page number traceability mechanism for classification and extraction results: both classification intervals and field match are associated with page numbers, enabling downstream processes to be segmented, verified, and transferred by page, significantly improving the interpretability and auditability of the results.

[0083] The above description is merely a detailed explanation of preferred embodiments and principles of the present invention. For those skilled in the art, there may be changes in specific implementation methods based on the ideas provided by the present invention, and these changes should also be considered within the scope of protection of the present invention.

Claims

1. A method for extracting multimodal information from multi-page image documents, characterized in that, Includes the following steps; S1, convert the images from each page of a multi-page image document into image data that is adapted to the input of a multimodal large model; S2 constructs a single-label classification task for each page of images, uses a multimodal large model to determine the document category page by page, and merges consecutive pages of the same category into a classification interval according to page order; S3, select the corresponding set of field rules and prompt words according to the document category of the classification range; S4 combines multi-page images with field rules into candidate extraction tasks based on the number of pages and the number of field rules, and supports merging multi-page images corresponding to the same rule into a multi-image merging request; S5. For the candidate extraction tasks generated in step S4, perform field-level rolling scheduling based on the field completion status: maintain the field completion status at the field level. For fields configured to be skipped after a hit, if the field has been successfully extracted in any page of images, skip the field in the task scheduling of subsequent pages and only add the unhit fields to the task to be scheduled. The completion status of the field is updated in real time based on the extraction results; S6, the extraction task is executed through a unified control plane, which includes a global semaphore, asynchronous queue, token bucket shaping, PID closed-loop speed regulation based on queue length, circuit breaking and exponential backoff retry, and the classification task and the extraction task share the same worker thread pool, token bucket and circuit breaking retry strategy. S7 collects the extraction results, backfills the source page number for the field values, performs normalization and automatic correction on the specified fields, and outputs the structured extraction results and classification intervals.

2. The method for multimodal information extraction from multi-page image documents according to claim 1, characterized in that, In step S1, image access supports dataURL passthrough, HTTP / HTTPS image download, and object storage relative path reading, and converts all accessed images into Base64 encoded image data. Among them, HTTP / HTTPS image download is executed through an asynchronous client, and object storage reading encapsulates synchronous storage I / O in a thread pool for execution. During the image conversion phase, download concurrency is limited by semaphores.

3. The method for multimodal information extraction from multi-page image documents according to claim 1, characterized in that, In step S2, the classification task is uniformly modeled as a single-field task for document categories. The classification prompt includes a list of candidate categories, category descriptions, and strict JSON output constraints. When parsing the classification response, only predefined categories are accepted, and illegal categories or empty results are marked as unclassified. The classification task and the extraction task reuse the same response parsing and page number backfilling mechanism.

4. The method for multimodal information extraction from multi-page image documents according to claim 1, characterized in that, In step S3, the prompt words are constructed using a system-level hard constraint and field three-element separation method. Each field is described by three elements: display name, internal code, and semantic description. The display name is used as the JSON key in the output of the multimodal large model, and the internal code is used for internal system association. The prompt words require that the output of the multimodal large model be pure JSON, all fields must appear, and fields without valid values ​​are represented by empty arrays. When cleaning the response of the multimodal large model, it is compatible with the message response structure returned by different upstream models, stripping code block markers and thought chain text, and extracting objects that conform to the JSON format.

5. The method for multimodal information extraction from multi-page image documents according to claim 1, characterized in that, In step S4, the candidate extraction task is generated in the following way: cost estimation is performed based on fixed cost, image cost, and rule cost, and an orchestration method is selected between page-first strategy and rule-first strategy according to the estimated total cost; wherein, the page-first strategy takes a page or page group as the outer dimension, and the rule-first strategy takes a field or field group as the outer dimension; when the maximum number of images in a single request is less than or equal to 1, the page-first strategy is automatically adopted.

6. The method for multimodal information extraction from multi-page image documents according to claim 1, characterized in that, In step S4, the multi-image merging request submits multiple pages of images corresponding to the same rule together in a single request, and requests the multimodal large model to return an array organized by page. The array includes multiple page objects, each containing a page number and fields. If the multi-image merging request fails due to the interface limitations of the multimodal large model, it is automatically downgraded to a page group or single-page task with fewer pages.

7. The method for multimodal information extraction from multi-page image documents according to claim 1, characterized in that, In step S5, the completion status of a field is uniquely identified by its internal code. A field can be configured to be extracted once or skipped after a successful extraction. Once a field configured to be extracted once or skipped after a successful extraction, it will not be scheduled for subsequent pages. Page access order strategies include strict page-by-page, look-ahead scrolling, edge-first, and grouped scrolling. Edge-first scrolling accesses the document alternately from both ends towards the middle, following the order of first page, last page, second first page, and second last page. Grouped scrolling organizes field rules and pages into a two-dimensional task matrix, generating parallel batches along the anti-diagonal line, where the anti-diagonal line is the line in the two-dimensional task matrix whose sum is the row containing the rule and the column containing the page. Before each batch begins, triangular ablation is performed on field rules that have been hit and configured to be skipped. Triangular ablation refers to removing candidate nodes of the field rule in the subsequent page direction from the batch, creating skipped areas in the task matrix.

8. The method for multimodal information extraction from multi-page image documents according to claim 1, characterized in that, In step S6, the unified control surface performs speed regulation and protection in the following manner: The PID closed-loop speed control mechanism adjusts the token generation rate based on the error between the current queue length and the target queue length; the circuit breaker mechanism includes closed, open, and half-open states, and enters the half-open state for detection after cooling down in the open state; when a rate limiting response is received, the current token generation rate is immediately halved, and exponential backoff retries are performed for timeouts, temporary network anomalies, and retryable upstream errors; after service startup or circuit breaker recovery, slow start is performed, and the number of concurrent worker threads is dynamically adjusted based on the average response time and queue depth.

9. The method for multimodal information extraction from multi-page image documents according to claim 1, characterized in that, In step S7, multi-level normalization, length-based bucketing candidate retrieval, and edit distance calculation are performed on the specified field, which includes the company or organization name; Whether to perform automatic correction is determined based on ambiguity judgment conditions, which include: For short names, retain the original value and do not perform any correction; The difference in edit distance between the best and second-best candidates is not less than the configuration margin; the similarity of the correction results is not less than the threshold. Automatic correction is performed only when the optimal candidate is unique and meets the aforementioned ambiguity judgment conditions, and the original value, edit distance, and similarity annotation are retained during correction; empty strings, empty arrays, and invalid objects are filtered during result collection, and the source page number is written for each valid match.

10. A multimodal information extraction system for multi-page image documents, used to implement the multimodal information extraction method for multi-page image documents as described in any one of claims 1-9, characterized in that, The multimodal information extraction system for multi-page image documents includes: The image access and preprocessing module is used to uniformly convert the images on each page into image data that is adapted to the input of a multimodal large model; The page-by-page multimodal classification and interval aggregation module is used to determine the document category page by page and merge consecutive pages of the same type into a category interval; The category-driven extraction configuration module is used to select the set of field rules and prompt words according to the category range; The cost-driven task planning module is used to generate candidate extraction tasks based on cost estimates and supports multi-graph merging requests. The field-level incremental scrolling scheduling module is used to maintain the completion status of fields at the field level. For fields configured to be skipped after being hit, they will not be scheduled again on subsequent pages after successful extraction. Only fields that were not hit will be scheduled. A high-concurrency multimodal inference engine is used to execute extraction tasks through a unified control plane, which includes global semaphores, asynchronous queues, token bucket shaping, PID closed-loop speed regulation, circuit breaking, and retrying. The classification task and the extraction task share the same control plane. The results fusion and correction module is used to backfill the source page number for field values, normalize and automatically correct specified fields, and output structured results and classification intervals.