An academic literature intelligent retrieval system and method based on a large model drive

Abstract

The application discloses a kind of academic literature intelligent retrieval system and method based on large model driving, belong to academic literature information retrieval technical field.System uses three-layer architecture, including: data access layer, is equipped with logical analysis construction module, heuristic query detection module, cross-library semantic mapping reconstruction engine and elastic starting architecture;Business logic layer, is equipped with heuristic complexity analysis pre-judgment module, database perception semantic alignment module, hardware perception capture distribution architecture and intelligent download strategy management module;User interaction layer, is equipped with adaptive semantic routing module, multistage cascade acquisition pipeline, asynchronous non-blocking GUI architecture and robustness exception handling module.The application realizes the full-process automation from natural language query to the full text acquisition of multi-source heterogeneous database by three-layer decoupling architecture;The full text acquisition uses three-stage cascade acquisition pipeline (first level PMC API, second level publisher LinkOut page HTML analysis, third level legal open access backup source), only for legal open access literature and institution subscribed resources, does not include any technical means to bypass copyright wall.

Description

Technical Field

[0001] This invention belongs to the field of academic literature information retrieval technology, and in particular relates to an intelligent academic literature retrieval system and method based on a large model. Background Technology

[0002] With the explosive growth in the number of academic documents, researchers face numerous technical challenges in the process of literature retrieval and acquisition.

[0003] Traditional Boolean logic retrieval systems suffer from a significant semantic gap when handling complex natural language queries. User-inputted search queries often contain multiple layers of nested logic, cross-field constraints, and domain-specific terms. Traditional rule-based or string replacement-based parsing methods struggle to accurately understand user intent, easily leading to logical misalignments or search results that deviate from expectations.

[0004] Existing literature retrieval systems also have significant shortcomings in the full-text acquisition stage. Different academic databases use different syntax rules and field identifiers, and the migration of retrieval queries between heterogeneous data sources requires a large amount of manual adaptation. At the same time, paywall restrictions prevent many documents from being accessed in full text directly through official channels, and existing systems lack effective alternative acquisition mechanisms.

[0005] Furthermore, existing retrieval systems often experience lag or even freezes in the user interface when handling high-concurrency network requests, severely impacting user experience. System configuration parameters cannot be persistently saved, requiring users to repeatedly configure API keys, search preferences, and other parameters each time they use the system. Facing varying computing power environments, existing systems lack resilience and often crash directly rather than smoothly degrade when dependency libraries are missing.

[0006] To address the aforementioned problems, this invention proposes an intelligent academic literature retrieval system and method based on a large model. Summary of the Invention

[0007] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention employs a three-layer decoupled architecture consisting of a data access layer, a business logic layer, and a user interaction layer. This architecture automates the entire process from natural language query input to full-text retrieval from multi-source heterogeneous databases. The responsibilities of each layer are clearly defined, and the interfaces are standardized, facilitating system expansion and maintenance.

[0008] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: an intelligent academic literature retrieval system based on a large model, comprising: The data access layer includes a logical parsing and construction module, a heuristic query detection module, a cross-database semantic mapping reconstruction engine, and an elastic startup architecture. The data access layer is used to transform unstructured natural language or semi-structured retrieval expressions into logically rigorous abstract syntax trees, perform multi-dimensional complex quantification and intelligent decision-making based on query features, dynamically generate standardized target retrieval expressions based on the grammatical constraints of the target database, and achieve smooth degradation and elastic startup of components under different computing power environments. The business logic layer includes a heuristic complexity analysis and prediction module, a database-aware semantic alignment module, a hardware-aware crawling and distribution architecture, and an intelligent download strategy management module. The business logic layer is used to perform structural scanning on the original query string to quantify the task difficulty and guide differentiated parsing strategies, automatically detect the target database based on query characteristics and load the corresponding expert model prompts, automatically match the corresponding crawling algorithm based on the top-level domain of the document URL to perform full-text crawling, and make intelligent decisions between traditional acquisition methods and alternative resources. The user interaction layer includes an adaptive semantic routing module, a multi-level cascaded acquisition pipeline, an asynchronous non-blocking GUI architecture, and a robust exception handling module. This layer transforms unstructured natural language or mixed-syntax queries into standardized database retrieval expressions, dynamically schedules network I / O resources to acquire full-text documents based on data source availability and compliance priorities, maintains user interface responsiveness during high-concurrency network requests, and performs exception capture and automatic recovery for unstable network environments and diverse document data.

[0009] As a preferred embodiment, in the data access layer: The logical parsing construction module includes a recursive descent parsing unit and a field semantic binding unit. The recursive descent parsing unit is used to transform the input search string into a token stream with semantic annotation through lexical analysis, and to construct a hierarchical abstract syntax tree composed of query nodes through a recursive function. Each node contains node type, semantic content, and child node references. The field semantic binding unit is used to automatically propagate field attributes to all term nodes or phrase nodes in the logical subtree through a depth-first traversal algorithm when a database field identifier is identified. The heuristic query detection module includes a multi-dimensional complexity quantification unit and a MeSH semantic awareness unit for the biomedical field. The multi-dimensional complexity quantification unit is used to extract features from the input text in real time through the query syntax detection module, including multi-dimensional grammatical features such as Boolean operator existence, field label existence, and wildcard existence, and calculates a normalized confidence score based on the extracted features. When the complexity score exceeds a set threshold, it automatically switches to the LLM-first mode. The MeSH semantic awareness unit for the biomedical field is used to identify key terms in the search query for the PubMed database and recommend MeSH mapping to the user through LLM.

[0010] Furthermore, in the data access layer, the cross-database semantic mapping reconstruction engine includes: Recursive string projection unit, used to follow the priority rules of the target database when converting abstract syntax trees back to strings and automatically fill in parentheses where necessary to prevent semantic drift; The field mapping table unit is used to automatically remap the field identifiers of the source database to the compatible tags of the target database based on the built-in heterogeneous database field lookup table; The special character escaping normalization unit is used to automatically perform character normalization based on the differences in how special symbols are handled in different databases.

[0011] The elastic boot architecture includes: The smooth degradation strategy unit is used to automatically scan the dependency libraries in the runtime environment during the launcher initialization phase, and trigger an automatic degradation mechanism to seamlessly switch to the basic mode when the missing dependency of the LLM enhancement module is detected. The multi-mode interactive distribution unit is used to manually specify the running mode via command-line parameters.

[0012] As a preferred embodiment, in the business logic layer: The heuristic complexity analysis and prediction module includes: The multi-factor feature extraction unit is used to dynamically detect the depth of nested parentheses in the query statement, count the number and combination types of Boolean operators, match the proprietary syntax of a specific database through regular expressions, and calculate the number of core concepts contained in the query through the analysis of word frequency and logical partitions. Complexity quantification representation unit, used to calculate a comprehensive score based on a weighted formula and automatically classify queries into multiple complexity levels.

[0013] Furthermore, the database-aware semantic alignment module includes: The database-aware dynamic suggestion unit is used to automatically detect the target database based on query features and load the corresponding expert model suggestion words, perform dual mapping of intent and syntax, and force the model to return structured data containing fields of parsed query, confidence, and optimization suggestions through JSONSchema constraints. The robust response cleaning and self-healing unit is used to perform atomic-level cleaning before deserialization to filter out invisible control characters and automatically close missing parentheses, complete unclosed quotation marks, or remove redundant commas through a heuristic repair engine.

[0014] Furthermore, the hardware-aware crawling and distribution architecture includes a publisher-specific crawling and distribution engine and a dynamic content rendering unit. The publisher-specific crawling and distribution engine is used to automatically match the corresponding crawling algorithm based on the top-level domain of the document URL, design exclusive CSS selector sequences and waiting strategies for multiple mainstream publishers, activate the intelligent document awareness algorithm to extract the full text using text density indicators for unidentified publisher websites, and integrate a DOI automatic redirection parsing module to track the redirection path by simulating HEAD requests. The dynamic content rendering unit is used to bypass JavaScript rendering obstacles and handle asynchronously loaded content using headless browser technology. The intelligent download strategy management module includes a historical success rate decision unit, an ordered legitimate open access source list maintenance unit, and a PDF integrity verification unit. The historical success rate decision unit records the success rate statistics of traditional links and legitimate open access backup sources in real time, forcing the traditional link to be used preferentially for open access journals, and automatically triggering a DOI-based legitimate open access backup source search link for documents that fail to be crawled using traditional methods. The ordered legitimate open access source list maintenance unit maintains a priority-sorted list of legitimate open access sources, attempting them in priority order before the download task starts. The first source with a normal response is used for downloading; if it fails, it is marked and switched to the next legitimate source. The PDF integrity verification unit performs multiple verifications on the byte stream after download completion, including header signature detection, error text scanning, and end-of-file identifier verification. If verification fails, the current source is automatically marked as unavailable and switched to the next legitimate open access source for re-acquisition.

[0015] As a preferred embodiment, in the user interaction layer: The adaptive semantic routing module includes: The query parsing template dispatching unit is used to select the corresponding database-specific parsing template based on the detected target database type branch. The target database type is determined by detecting the database-specific field tags (PubMed square bracket field tags / WoS equal sign field tags / Scopus parentheses field tags) in the query string. The database-specific parsing templates include four sets: PubMed, Web of Science, Scopus, and a general compatible set. The multimodal LLM service orchestration failover unit is used to implement a unified calling interface for APIs of multiple large model service providers and support the configuration of primary and backup service providers. It also uses a cleaning function to recursively traverse the data structure returned by the LLM to remove phantom text and extract clean search expressions. When the primary service response times out or fails to parse, it automatically and seamlessly switches to the backup API.

[0016] Furthermore, in the user interaction layer, the multi-level cascaded acquisition pipeline includes: The metadata retrieval and ID mapping unit is used for authentication via a single NCBI API Key and configurable request intervals, and after obtaining the PubMedID list, it calls the ID conversion service to build a mapping dictionary from PMID to PMCFullTextID; The three-level full-text acquisition unit includes a first-level acquisition subunit, a second-level acquisition subunit, and a third-level acquisition subunit. The first-level acquisition subunit is used to download structured XML or official PDFs of documents that match the PMCID mapping via PMCAPI and extract the clean text using an XML parser. The second-level acquisition subunit is used to acquire documents that do not match the PMCID mapping by parsing the full-text link block HTML in the PubMed document details page or parsing the publisher's original link using a DOI parser. The third-level acquisition subunit is used to query the best_oa_location open access PDF direct link corresponding to the document DOI via the Unpaywall API when the first two levels of acquisition fail, and uses the institutional knowledge base configured by the deployer (DSpace-style DOI retrieval) as a supplementary source. After acquiring the legal open access PDF file stream based on the DOI, multiple checks are performed on the file header, error text, and file footer identifiers.

[0017] Furthermore, in the user interaction layer: The asynchronous non-blocking GUI architecture includes a thread-safe log synchronization unit, a collapsible layout unit, and an input debouncing feedback verification unit. The thread-safe log synchronization unit constructs a queue-based message bus. A background fetching thread encapsulates log information, progress updates, and state changes into message objects and pushes them onto the message bus. The main UI thread polls the message bus at a fixed period. The collapsible layout unit implements the expansion and collapse operations of various interface blocks through a custom collapsible container component and automatically saves the collapsible state to a state file. When the system is closed, it automatically serializes configuration parameters and stores them in a local file. When the system starts, it automatically deserializes and restores the context configuration. The input debouncing feedback verification unit binds keyboard events to the intelligent search input box and triggers background syntax detection and complexity analysis after the user stops inputting for a preset time. Based on the verification results, it changes the color state of UI elements in real time and dynamically displays the inferred database syntax type. The robust exception handling module includes a filename disinfection unit, a dual-stream capture unit, and an exponential backoff circuit breaker unit. The filename disinfection unit is used to automatically remove illegal characters from filenames using regular expression filters and to intelligently truncate excessively long titles. The dual-stream capture unit is used to redirect standard output and standard error and simultaneously distribute them to the GUI log panel and the local log file. The exponential backoff circuit breaker unit is used to execute an exponential backoff retry strategy when encountering connection errors or timeouts during interaction with the server, and to trigger a local circuit breaker mechanism to skip the current entry and record the error log when the number of consecutive failures exceeds a threshold.

[0018] This invention also proposes an intelligent academic literature retrieval method based on a large model, comprising the following steps: Step 1: Perform lexical analysis and recursive descent parsing on the input search string to construct a hierarchical abstract syntax tree, bind field attributes to syntax tree nodes, perform complex metric quantization based on the syntax detection module and switch to LLM enhanced parsing when the threshold is exceeded, and recursively project the abstract syntax tree into a compatible target search expression according to the grammatical constraints of the target database. Step 2: Perform multi-factor feature extraction and weighted scoring on the query statement to achieve complexity grading; load corresponding expert hint words according to database features and constrain the model output structured data; perform cleaning and format repair on the model response; schedule the corresponding crawling algorithm according to the document URL domain name, and extract the full text by text density when it is not recognized, and use a headless browser to render dynamic content; make link decisions based on download success rate statistics and try legal open acquisition sources in priority order (the first normal response is used, and the failure is marked to switch to the next legal source); perform multiple verifications on the downloaded byte stream and switch to the next legal source when verification fails. Step 3: Based on the detected target database type branch, select the corresponding dedicated parsing template (PubMed / WoS / Scopus / four general compatible sets). Extract the retrieval expression from the LLM response through the cleaning function and switch to the backup API when the main service fails. After authentication with a single NCBI API Key and configurable request intervals, and after building an ID mapping dictionary, perform three-level full-text retrieval. Implement asynchronous communication between the background thread and the main UI thread based on a message queue and achieve interface state persistence through a collapsible container. Bind debouncing logic to the input box and update the UI feedback in real time according to the verification results. Perform sanitization processing on the file name and distribute standard output and standard error to the GUI panel and log file simultaneously. When encountering connection anomalies, perform exponential backoff retries and trigger partial circuit breaking when the number of failures exceeds the limit.

[0019] (III) Beneficial Effects Compared with existing technologies, this invention provides an intelligent academic literature retrieval system and method based on a large model, which has the following beneficial effects: I. This invention transforms unstructured or semi-structured retrieval queries into a logically rigorous tree-like topology by constructing a hierarchical abstract syntax tree. This fundamentally solves the logical misalignment problem easily caused by traditional string replacement methods, ensuring lossless expression and cross-platform migration of complex nested Boolean logic. Simultaneously, through multi-factor complex metric analysis and an adaptive routing mechanism, it achieves differentiated processing for simple and complex queries. While ensuring millisecond-level response speeds for simple queries, it leverages the semantic understanding capabilities of a large language model to solve the parsing problem of complex natural language queries, achieving an optimal balance between computational cost and parsing accuracy. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the system module structure of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention. Figure 1 ; Figure 3 This is a schematic diagram of the method flow of the present invention. Figure 2 ; Figure 4 This is a schematic diagram of the method flow of the present invention. Figure 3 . Detailed Implementation

[0021] To better understand the purpose, structure, and function of this invention, the following will further describe a large-model-driven intelligent academic literature retrieval system and method based on specific embodiments in conjunction with the accompanying drawings.

[0022] Example 1: Data Access Layer I. Logical parsing construction module It includes: 1. Recursive descent parsing mechanism In this embodiment, the system performs a deep scan on the input search string using a preset regular expression engine. This engine can accurately identify phrases (delimited by quotation marks), Boolean operators (AND / OR / NOT), field qualifiers, and nested bracket logic, and convert them into a semantically annotated token stream.

[0023] The parser employs a depth-first strategy, constructing a hierarchical AST composed of QueryNodes through recursive functions such as `_parse_or` and `_parse_and`. Each node contains a node type (NodeType), semantic content (Value), and references to child nodes, ensuring accurate decoupling of complex logic (such as multi-level nested parentheses). This hierarchical parsing architecture fundamentally solves the logical misalignment problem easily caused by traditional string replacement methods, providing a structured logical foundation for subsequent cross-database syntax conversion.

[0024] 2. Field semantic binding and transmission When the parser recognizes a specific database field identifier (such as TS= in WoS), the system initiates the field binding mechanism. Using a depth-first traversal algorithm, the field attributes are automatically propagated to all term or phrase nodes in its logical subtree.

[0025] II. Heuristic Query Detection Module It includes: 1. Multi-dimensional and complex metrics The system uses the QuerySyntaxDetector module to extract features from the input text in real time, including Boolean operator density (the ratio of the number of Boolean operators to the number of query words), operator diversity (the number of Boolean operator types after deduplication), field label coverage (the ratio of the number of field labels to the number of query words), and field label count, etc., which are multi-dimensional continuous quantitative indicators (corresponding to query_parser.py:257-262).

[0026] The system calculates a normalized confidence score based on the extracted features. For example, queries containing multi-level nested logic or MeSH lexical potential will receive a higher LLM processing priority. When the complexity score exceeds a set threshold, the system automatically switches to "LLM priority mode," leveraging the semantic understanding capabilities of the large model for synonym expansion and logical reconstruction, rather than simply relying on hard-coded rules.

[0027] 2. MeSH Semantic Awareness for the Biomedical Field For biomedical databases such as PubMed, the PubMedParser module integrates the ability to identify key terms (such as Neoplasms, Immunology, etc.). The system can identify candidate words with high search value in the search query and recommend high-precision MeSH mapping through LLM to the user, significantly improving the recall and precision of the search.

[0028] III. Cross-library semantic mapping refactoring engine It includes: 1. Recursive projection of nodes to strings During the process of converting AST back to string, QueryTransformer strictly follows the priority rules of the target database (such as AND over OR) and automatically completes parentheses where necessary to prevent semantic drift.

[0029] 2. Special character escaping and standardization The system incorporates a heterogeneous database field mapping table (SchemaMapping). For example, it automatically remaps the "TS (Topic)" field in WoS to the PubMed-compatible "[Title / Abstract]" tag, achieving a seamless migration of search logic. Addressing the differences in how different databases handle special characters such as quotation marks and asterisks, the reconstruction engine automatically performs character normalization to ensure that the generated search queries do not trigger syntax errors in the target API or web interface.

[0030] IV. Flexible Startup Architecture 1. Smooth degradation strategy based on dynamic loading During the initialization phase, the launcher detects dependency libraries in the runtime environment through exception handling. If a complex dependency required by the LLM enhancement module is detected as missing, the system will trigger an automatic degradation mechanism, seamlessly switching to "Basic Mode." By decoupling the EnhancedGUI from the basic logic, the system ensures the atomicity of the core document retrieval function, maintaining high availability of the basic retrieval function even if the user environment cannot support the LLM components.

[0031] 2. Multi-mode interaction layer distribution The system allows users to manually specify the operating mode via command-line parameters (such as `--modeenhanced`). For high-performance workstations, the enhanced version is enabled to obtain LLM accessibility features; for lightweight office equipment, the basic version is used to minimize resource consumption (memory / CPU). The system has established a complete traceback capture mechanism, which can automatically fall back to the console to print error details when the GUI fails to load, and guide the user to fix dependencies.

[0032] Example 2: Business Logic Layer I. Heuristic Complexity Analysis and Prediction Module It includes: 1. Multi-factor feature extraction mechanism The system dynamically detects the depth of nested parentheses in query statements. If the nesting level exceeds a preset threshold (e.g., 3 levels), the system will identify it as "deep nesting" and significantly increase the complexity weight to capture complex logical inclusion relationships. The system statistically analyzes the number and combination types of Boolean operators (AND, OR, NOT), considering high-density operator combinations as strong logical correlation signals requiring higher-precision parsing support. It matches proprietary syntax of specific databases (e.g., PubMed tags, WoS TS= fields) using regular expressions; once mixed syntax or non-standard abbreviations are detected, the module will automatically increase the "special syntax" score. Through analysis of word frequency and logical partitioning, the system calculates the number of core concepts contained in the query; queries with multiple concepts (>3) are marked as "complex concept combinations."

[0033] 2. Complex metric representation The analysis engine calculates a comprehensive score S=Σ(Factor_i×Weight_i) based on a weighted formula and automatically categorizes queries into four levels: "simple," "medium," "complex," or "very complex." This pre-evaluation mechanism avoids excessive consumption of LLM computing power for simple queries while ensuring that complex queries can achieve deep semantic understanding, thus realizing efficient allocation of computing resources.

[0034] II. Database-Aware Semantic Alignment Module It includes: 1. Database-aware dynamic suggestions The parser automatically detects the target database based on query features. If MeSH terminology features are detected, the system loads the "PubMed Expert Model" suggestion, requiring the model to incorporate MeSH subtopic term expansion during optimization. The system not only optimizes query syntax but also performs a dual mapping between intent and syntax. For example, it converts unstructured natural language research intent into precise expressions conforming to WoSNEAR operator logic. By embedding JSON templates in the suggestion, the system guides the model to return structured data containing fields such as `parsed_query`, `confidence`, and `optimization_suggestions`, ensuring seamless parsing by the business logic layer.

[0035] 2. Robust response cleaning and format self-healing For invisible control characters (such as \x00-\x1f) that may be carried in LLM responses, the module performs atomic-level cleaning before deserialization. A heuristic repair engine is integrated, which can automatically close missing parentheses, complete unclosed quotation marks, or remove redundant commas, greatly improving the fault tolerance of the automated pipeline in cases of network instability or model illusion.

[0036] III. Hardware-Aware Fetching and Distribution Architecture It includes: 1. Publisher-specific crawler distribution engine The system automatically matches the corresponding crawling algorithm based on the top-level domain of the document URL. For major publishers such as Nature, Science, Cell, and Elsevier, exclusive CSS selector sequences and waiting strategies are designed. For unidentified publisher websites, the system activates the "intelligent document awareness" algorithm, traversing high-text-density tags such as `article` and `main` on the page, and using a text / HTML ratio algorithm to extract the full text of unknown page structures. The system integrates a transparent DOI parsing module, simulating HEAD requests to trace redirection paths and transforming static DOI links into actionable, real-time publisher URLs.

[0037] 2. Dynamic content rendering based on Selenium By utilizing HeadlessChrome technology, the system can bypass JavaScript rendering barriers, handle asynchronous content loading common on modern publisher websites, and ensure the integrity of the crawled content.

[0038] IV. Intelligent Download Strategy Management Module It includes: 1. Intelligent decision-making based on historical success rates The policy manager records the success rate statistics of traditional search links and legitimate open access alternative sources in real time. Traditional search links are forcibly prioritized for open access journals; for documents that fail to be crawled using traditional methods, the search link based on DOI-based legitimate open access alternative sources is automatically triggered.

[0039] 2. Legally open access to backup sources via sequential polling. The system maintains a predefined list of mirror domain names. The downloader attempts to download from each mirror in a predefined order. The first mirror node that responds normally is used for downloading, and failed mirror domain names are marked as invalid, which significantly reduces the failure rate of download requests.

[0040] 3. Atomic-level integrity verification of PDF files After downloading, the system does not submit the results directly. Instead, it performs multiple verifications on the byte stream: including checking the %PDF file header signature, scanning for HTML tags that represent errors (such as the text "AccessDenied"), and verifying the end-of-file (EOF) marker. If verification fails, the system automatically marks the current source as unavailable and switches to the next legitimate open-access source to re-acquire the document, ensuring that the final product delivered to the user is a genuine and readable PDF document.

[0041] Example 3: User Interaction Layer I. Adaptive Semantic Routing Module It includes: 1. Dynamic query complexity analysis mechanism The system integrates a lightweight QueryComplexityAnalyzer. During the real-time process of user input of query commands, the analyzer calculates a normalized complexity score S_complex∈[0,10] based on the depth of the abstract syntax tree (AST), the density of logical operators (AND / OR / NOT), and the ambiguity of domain-specific terms.

[0042] The system selects the corresponding database-specific parsing template (PubMed, Web of Science, Scopus, and a general-compatible template) based on the detected target database type. The target database type is determined by detecting database-specific field tags in the query string (PubMed square bracket field tags / Web of Science equal sign field tags / Scops parentheses field tags), achieving precise assignment of the target database syntax (corresponding to llm_query_parser.py:710-740 + :742 / :816 / :845 / :874). The system calculates a complexity score and appends this score as metadata to the large language model's prompt words to guide parsing. 2. Multimodal LLM Service Orchestration and Failover The module internally implements abstract encapsulation of multi-source large model APIs (such as SiliconFlow and OpenRouter). The system supports configuring a primary service provider and a fallback service provider.

[0043] To address the uncertainty in the output of large models, a `safe_extract_query_string` cleaning function was designed. This function recursively traverses the JSON, dictionary, or non-standard string structures returned by the LLM, removing hallucination text and extracting clean search expressions. If the main service times out or fails to parse, the system automatically and seamlessly switches to the backup API to ensure the continuity of search tasks.

[0044] II. Multi-level Cascaded Acquisition Pipeline It includes: 1. Metadata retrieval and ID mapping mechanism The system implements authentication for the NCBI E-utilities interface through a single `ncbi_api_key` configuration item. For requests with a configured API key, the system employs a configurable request interval, significantly improving the throughput of metadata retrieval. After obtaining the PubMed ID (PMID) list, the system calls the `id_converter` service to construct a mapping dictionary from PMIDs to PMC Full Text IDs (PMCIDs). This step is a key filter distinguishing between "summary-only articles" and "open access full-text articles."

[0045] 2. Three-level full-text acquisition strategy Level 1 (PMC Direct): For documents that match the PMCID mapping, the system prioritizes downloading structured XML or official PDF via the PMC API. The system has a built-in XML parser that can remove metadata tags and extract the clean text.

[0046] Level 2 (Publisher Original Link Resolution): For documents that do not match the PMCID mapping, the system obtains the publisher's original link through HTML parsing or DOI redirection parsing of the full-text link block on the PubMed document details page. The system also parses the publisher's original link through PubMed LinkOut page Hypertext Markup Language parsing or DOI Resolver. Level 3 (Legitimate Open Access Backup Source): As a last resort, if the first two levels fail, the system queries the best_oa_location open access PDF direct link corresponding to the document DOI via the Unpaywall API, and uses the institutional knowledge base configured by the deployer (DSpace-style DOI search) as a supplementary source. After obtaining the legitimate open access PDF file stream based on the DOI, it performs multiple quality checks on the file header, error text, and file footer identifiers. It also performs file header and footer identifier checks on the PDF byte stream. III. Asynchronous Non-blocking GUI Architecture It includes: 1. Thread-safe logging and state synchronization mechanisms The system employs a cross-thread safe scheduling mechanism based on a window event loop. The background fetching thread uses this mechanism to asynchronously submit log information, progress updates, and state changes to the main UI thread. This design completely avoids race conditions and UI freezes caused by cross-thread GUI operations, achieving a smooth projection from background logic to foreground visualization.

[0047] 2. Foldable modular layout The interface uses custom components called CollapsibleContainer and CollapsibleSection. These components support dynamic expand / collapse operations and automatically save the collapsed state of each block to ui_state.json. Upon closing, the system first cancels the suspended debouncing timer handle via a close event hook to coordinate callback timings under a single-threaded event loop, then automatically serializes all current configuration parameters (including API key, search preferences, and window geometry) and stores them in a local JSON file, before destroying the window. Upon startup, it prioritizes reading the saved window geometry, clamps it to the screen boundary, and then automatically deserializes and restores the context. This provides users with a seamless operating experience, eliminating the need to repeatedly configure complex search parameters.

[0048] 3. Input stabilization and real-time feedback verification In the "Smart Search Input" box, the system is linked to... <keyrelease>The system introduces debouncing logic: when a new input event arrives, the debouncing timer first cancels the previously suspended timer handle and sets the local handle reference to null. This handle is then canceled again in the window close event hook to coordinate with the timing of the configuration serialization. Background syntax checking and complexity analysis are only triggered if the user stops inputting for more than a preset time (e.g., 500ms). Based on the verification results, the system changes the color state of UI elements in real time (e.g., green for valid syntax, red for errors) and dynamically displays the inferred database syntax type (e.g., PubMed / WoS / Scopus), achieving a "what you see is what you get" interactive feedback.

[0049] IV. Robust Exception Handling Module It includes: 1. Atomic-level file writing and filename cleaning To address filename restrictions imposed by different operating systems, the system incorporates a regular expression filter that automatically removes illegal characters (such as \ / :*?"<>|) and intelligently truncates excessively long headers to prevent OSError. The system redirects standard output (stdout) and standard error (stderr), distributing them simultaneously to both the GUI log panel and local log files via the EnhancedLogCapture class. Even when running in headless mode, it can fully trace any exception stack trace.

[0050] 2. Exponential backoff and automatic circuit breaking in network I / O When interacting with the NCBI E-utilities interface or a legitimate open-access source server, if a ConnectionError or Timeout occurs, the system does not immediately report an error, but instead executes an exponential backoff retry strategy. If the number of consecutive failures exceeds a threshold, the system triggers a partial circuit breaker mechanism, skipping the current entry and logging the error, ensuring that the failure to download a single document does not cause the entire batch processing task to be interrupted (Fail-Safe).

[0051] This invention achieves full-process automation from natural language query to full-text retrieval from multi-source heterogeneous databases through a three-layer decoupled architecture; the full-text retrieval adopts a three-level cascaded retrieval pipeline, which is only used for legally open access literature and institutionally subscribed resources, and does not contain any technical means of bypassing the copyright holder's paywall.

[0052] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.< / keyrelease>

Claims

1. An intelligent academic literature retrieval system based on a large model, characterized in that, include: The data access layer includes a logical parsing and construction module, a heuristic query detection module, a cross-database semantic mapping reconstruction engine, and an elastic startup architecture. The business logic layer includes a heuristic complexity analysis and prediction module, a database-aware semantic alignment module, a hardware-aware crawling and distribution architecture, and an intelligent download strategy management module. The user interaction layer includes an adaptive semantic routing module, a multi-level cascaded acquisition pipeline module, an asynchronous non-blocking GUI architecture, and a robust exception handling module. The data access layer is used to transform the search query into an abstract syntax tree and perform cross-database syntax mapping. The business logic layer is used for query complexity prediction, database-aware prompt word construction, and full-text crawling and distribution. The user interaction layer is used for query semantic routing, multi-level full-text retrieval, and asynchronous interface response.

2. The intelligent academic literature retrieval system based on a large model as described in claim 1, characterized in that, In the data access layer: The logical parsing construction module includes a recursive descent parsing unit and a field semantic binding unit. The recursive descent parsing unit is used to convert the retrieval string into a token stream and construct a hierarchical abstract syntax tree through lexical analysis. The field semantic binding unit is used to propagate the database field identifiers to each node in the syntax tree. The heuristic query detection module includes a multi-dimensional complex metric unit and a MeSH semantic awareness unit. The multi-dimensional complex metric unit is used to extract features from the input text in real time and calculate a confidence score based on the features. When the score exceeds a set threshold, it switches to the LLM priority mode. The MeSH semantic awareness unit is used to identify key terms in the PubMed database search query and recommend MeSH mapping through LLM.

3. The intelligent academic literature retrieval system based on a large model as described in claim 2, characterized in that, In the data access layer: The cross-database semantic mapping reconstruction engine includes a recursive string projection unit, a field mapping table unit, and a special character escaping normalization unit; the recursive string projection unit is used to follow the priority rules of the target database and automatically complete parentheses when converting the abstract syntax tree into a string; The field mapping table unit is used to remap the source database field identifier to the target database compatible label according to the heterogeneous database field lookup table; The special character escaping normalization unit is used to perform character normalization on the differences in special symbol processing in different databases; The elastic startup architecture includes a smooth degradation strategy unit and a multi-mode interactive distribution unit; The smooth degradation strategy unit is used to scan the runtime environment dependency libraries at startup and automatically downgrade to the basic mode when the missing LLM enhancement module dependency is detected. The multi-mode interactive distribution unit is used to specify the running mode through command line parameters.

4. The intelligent academic literature retrieval system based on a large model as described in claim 1, characterized in that, In the business logic layer: the heuristic complexity analysis and prediction module includes a multi-factor feature extraction unit and a complexity quantification representation unit; The multi-factor feature extraction unit is used to extract features of the nested bracket depth, Boolean operator density and diversity, syntax specificity and semantic breadth of the query statement; The complexity quantification representation unit is used to calculate a comprehensive score based on a weighted formula and classify queries into multiple complexity levels. The database-aware semantic alignment module includes a database-aware dynamic prompting unit and a robust response cleaning and self-healing unit. The database-aware dynamic prompting unit is used to detect the target database based on query features and load the corresponding expert model prompt words, and force the model to return structured data through JSONSchema constraints; The robust response cleaning and self-healing unit is used to filter out invisible control characters before deserialization and automatically repair format errors through a heuristic repair engine.

5. The intelligent academic literature retrieval system and method based on a large model as described in claim 4, characterized in that, In the business logic layer, the hardware-aware crawling and distribution architecture includes a publisher-dedicated crawling and distribution engine and a dynamic content rendering unit; The publisher-specific crawling and distribution engine is used to match the corresponding crawling algorithm based on the top-level domain of the document URL, extract the full text of unidentified publisher websites using the text density index, and integrates a DOI automatic redirection parsing module to track redirection paths. The dynamic content rendering unit is used to process asynchronously loaded content using headless browser technology; The intelligent download strategy management module includes a historical success rate decision-making unit, an orderly and legal open access source list maintenance unit, and a PDF integrity verification unit. The historical success rate decision unit is used to make download link decisions based on the real-time recorded statistics of the success rates of traditional downloads and alternative downloads. The ordered legal open access source list maintenance unit is used to maintain a legal open access source list sorted by priority. It tries sources in priority order, and the first source with a normal response is used for downloading. If it fails, it is marked to switch to the next legal source. The PDF integrity verification unit is used to perform multiple verifications on the downloaded byte stream. If the verification fails, it automatically switches to an alternative image and reacquires the data.

6. The intelligent academic literature retrieval system based on a large model as described in claim 1, characterized in that, In the user interaction layer: the adaptive semantic routing module includes a query parsing template dispatch unit and a multimodal LLM service orchestration failover unit; The query parsing template dispatching unit selects the corresponding database-specific parsing template based on the detected target database type branch. The target database type is determined by detecting database-specific field tags in the query string; the database-specific parsing templates include four sets: PubMed, Web of Science, Scopus, and a general compatible set. The multimodal LLM service orchestration failover unit is used to realize unified invocation and master / slave switching of multi-source large model APIs, and to extract clean search expressions through a cleaning function.

7. The intelligent academic literature retrieval system based on a large model as described in claim 6, characterized in that... The multi-level cascaded acquisition pipeline module includes a metadata retrieval and ID mapping unit and a three-level full-text acquisition unit; The metadata retrieval and ID mapping unit is used to authenticate with a single NCBI API Key and configure the request interval to build a mapping dictionary from PMID to PMCID; The three-level full-text acquisition unit includes a first-level acquisition subunit, a second-level acquisition subunit, and a third-level acquisition subunit, which are used to acquire the full text through three channels: PMCAPI, publisher LinkOut, and legal open access backup sources.

8. The intelligent academic literature retrieval system based on a large model as described in claim 1, characterized in that, In the user interaction layer: the asynchronous non-blocking GUI architecture includes a thread-safe log synchronization unit, a collapsible layout unit, and an input debouncing feedback verification unit; The thread-safe log synchronization unit is used to achieve asynchronous communication between the background fetching thread and the main UI thread through a message queue; The collapsible layout unit is used to save and restore the collapsible state of each block of the interface and to serialize and persist the configuration parameters. The input debounce feedback verification unit is used to bind debounce logic to the input box and change the state of UI elements in real time according to the verification result.

9. The intelligent academic literature retrieval system based on a large model as described in claim 8, characterized in that, The robust anomaly handling module includes a filename disinfection unit, a dual-stream capture unit, and an exponential backoff circuit breaker unit. The file name disinfection unit is used to remove illegal characters from file names and truncate excessively long titles; The dual-stream capture unit is used to simultaneously distribute standard output and standard error to the GUI log panel and the local log file; The exponential backoff circuit breaker unit is used to perform exponential backoff retries when there is a connection error or timeout, and to trigger a partial circuit breaker when the number of consecutive failures exceeds a threshold.

10. A method for intelligent academic literature retrieval based on a large model driven by the system described in any one of claims 1-9, characterized in that, Includes the following steps: Step 1: Perform lexical analysis and recursive descent parsing on the input search string to construct a hierarchical abstract syntax tree, bind field attributes to syntax tree nodes, perform complex metric quantization based on the syntax detection module and switch to LLM enhanced parsing when the threshold is exceeded, and recursively project the abstract syntax tree into a compatible target search expression according to the grammatical constraints of the target database. Step 2: Perform multi-factor feature extraction and weighted scoring on the query statement to achieve complexity classification; load corresponding expert prompts based on database features and constrain the model output structured data; perform cleaning and format repair on the model response; schedule the corresponding crawling algorithm based on the document URL domain name, extract the full text by text density when it is not recognized, and render dynamic content using a headless browser; make link decisions based on download success rate statistics and try legitimate open acquisition sources in priority order; perform multiple verifications on the downloaded byte stream and switch to the next legitimate source when verification fails; Step 3: Select the corresponding dedicated parsing template based on the detected target database type branch, extract the retrieval expression from the LLM response through the cleaning function, switch the backup API when the main service fails, perform three-level full-text retrieval after single NCBI API Key authentication and configurable request interval and ID mapping dictionary construction, realize asynchronous communication between background thread and main UI thread based on message queue and realize interface state persistence through collapsible container, bind debouncing logic to input box and update UI feedback in real time according to the verification result, perform sanitization processing on file name and distribute standard output and standard error to GUI panel and log file at the same time, perform exponential backoff retries when encountering connection abnormality and trigger partial circuit breaker when the failure limit is exceeded.

Images