Information retrieval method, system and equipment and computer readable storage medium

Abstract

The invention discloses an information retrieval method, system and device and a computer readable storage medium, and relates to the technical field of intelligent retrieval. Determining a core judgment standard and an exclusion rule matched with the query information; according to the core judgment standard and the exclusion rule, processing the query information to obtain a system cue word; performing information retrieval based on the system cue word to obtain a target retrieval result of the query information; wherein the core judgment standard comprises a function matching standard, a field consistency standard and a context-related standard; the exclusion rules comprise a general description exclusion rule, a system boundary exclusion rule, a vocabulary overlapping exclusion rule, an indirect correlation exclusion rule and a historical version exclusion rule. According to the method, the system cue word which is matched with the novelty search information function, consistent with the field, associated with the context, free of general description, cross-system boundaries, free of overlapping words, directly associated and consistent with the latest version can be obtained, and the search precision is improved.

Description

Technical Field

[0001] This application relates to the field of intelligent retrieval technology, and more specifically, to an information retrieval method, system, device, and computer-readable storage medium. Background Technology

[0002] With the development of deep learning technology, information retrieval based on vector similarity can be performed according to user needs in processes such as intelligent question answering. This involves converting the user-input text into a high-dimensional vector representation and determining text relevance by calculating metrics such as cosine similarity between vectors.

[0003] While vector similarity retrieval has improved semantic understanding, it still has many shortcomings. For example, vectors have limited representational power, making it difficult to capture complex semantic relationships and contextual information. Furthermore, vector retrieval lacks sufficient understanding of contextual information, hindering deep semantic reasoning and judgment. When queries involve complex technical logic relationships, simple vector similarity calculations often fail to accurately determine relevance.

[0004] In conclusion, improving retrieval accuracy is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this application is to provide an information retrieval method that can, to a certain extent, solve the technical problem of how to improve retrieval accuracy. This application also provides an information retrieval system, an electronic device, and a computer-readable storage medium.

[0006] To achieve the above objectives, this application provides the following technical solution:

[0007] An information retrieval method, comprising:

[0008] Retrieve query information to be processed;

[0009] Determine the core judgment criteria and exclusion rules for matching the query information;

[0010] The query information is processed according to the core judgment criteria and the exclusion rules to obtain system prompts.

[0011] Information retrieval is performed based on the system prompts to obtain the target retrieval results for the query information.

[0012] The core judgment criteria include functional matching criteria, domain consistency criteria, and context-related criteria; the exclusion rules include general description exclusion rules, system boundary exclusion rules, word overlap exclusion rules, indirect relevance exclusion rules, and historical version exclusion rules.

[0013] Preferably, the process of processing the query information according to the core judgment criteria and the exclusion rules to obtain system prompts includes:

[0014] The query information is flattened to obtain the atomic entries corresponding to the query information;

[0015] The atomic entries are grouped to obtain the entry grouping results;

[0016] Based on the core judgment criteria and the exclusion rules, the grouping results of the entries are processed in batches to obtain system prompt words.

[0017] Preferably, the flattening process of the query information to obtain the atomic entries corresponding to the query information includes:

[0018] Context information for generating the query information;

[0019] The query information is processed to obtain initial information;

[0020] Evidence extraction and index enhancement are performed on the initial information to obtain candidate information;

[0021] Feature extraction is performed on the candidate information to obtain feature data;

[0022] The feature data is deduplicated and weakly similar data is merged to obtain atomic entries;

[0023] Based on the context information, the metadata of the atomic entries is generated.

[0024] Preferably, the process of processing the query information to obtain initial information includes:

[0025] The paragraphs in the query information are subjected to noise reduction, whitespace normalization, line break merging, list merging, unit standardization, and terminology standardization to obtain the natural paragraph processing results;

[0026] The tables and images in the query information are keyed in the header and the text is extracted to obtain structured placeholder entries and position pointers.

[0027] According to semantic boundaries, the statements that span paragraphs in the query information are divided into the smallest semantic units, and the parent-child relationship and sequential index between the smallest semantic units and the corresponding paragraphs are generated;

[0028] The natural segment processing result, the structured placeholder entry, the position pointer, the smallest semantic unit, the parent-child relationship, and the sequence index are used as initial information.

[0029] Preferably, generating the metadata of the atomic entry based on the context information includes:

[0030] Based on the context information, metadata for the atomic entries is generated, including source file, paragraph location results, topic candidate results, evidence fragment summary, length, and quality score.

[0031] Preferably, the step of performing information retrieval based on the system prompts to obtain the target retrieval results for the query information includes:

[0032] Retrieve cached pre-search results;

[0033] Detect whether the system prompts match the pre-search results;

[0034] In response to a system prompt matching the pre-search results, the target search result for the query information is extracted from the pre-search results.

[0035] Preferably, after detecting whether the system prompt word matches the pre-search result, the method further includes:

[0036] If the system prompt does not match the pre-search result, the system prompt is aligned and constrained according to a unified semantic template to obtain the semantic processing result.

[0037] The trained large language model is invoked to retrieve the semantic processing results, and initial retrieval results with values ​​of temperature coefficient and numbers matching the maximum number of tokens are obtained.

[0038] The initial search results are subjected to consistency verification and confidence aggregation to obtain the target search results for the query information.

[0039] An information retrieval system, comprising:

[0040] The query information acquisition module is used to acquire query information to be processed.

[0041] The filtering criteria determination module is used to determine the core judgment criteria and exclusion rules that match the query information;

[0042] The query processing module is used to process the query information according to the core judgment criteria and the exclusion rules to obtain system prompt words;

[0043] The retrieval module is used to perform information retrieval based on the system prompts and obtain the target retrieval results of the query information;

[0044] The core judgment criteria include functional matching criteria, domain consistency criteria, and context-related criteria; the exclusion rules include general description exclusion rules, system boundary exclusion rules, word overlap exclusion rules, indirect relevance exclusion rules, and historical version exclusion rules.

[0045] An electronic device, comprising:

[0046] Memory, used to store computer programs;

[0047] A processor for executing the computer program to implement the steps of any of the above-described information retrieval methods.

[0048] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above-described information retrieval methods.

[0049] This application provides an information retrieval method that involves: acquiring query information to be processed; determining core judgment criteria and exclusion rules for matching the query information; processing the query information according to the core judgment criteria and exclusion rules to obtain system suggestion words; and performing information retrieval based on the system suggestion words to obtain the target retrieval results for the query information. The core judgment criteria include functional matching criteria, domain consistency criteria, and context relevance criteria; the exclusion rules include general description exclusion rules, system boundary exclusion rules, word overlap exclusion rules, indirect relevance exclusion rules, and historical version exclusion rules. In this application, after acquiring the query information, it is necessary to process the query information according to the functional matching criteria, domain consistency criteria, and context relevance criteria, and also according to the general description exclusion rules, system boundary exclusion rules, word overlap exclusion rules, indirect relevance exclusion rules, and historical version exclusion rules. This process yields system suggestion words that are functionally matched to the novelty information, domain consistent, context-related and without general descriptions, cross system boundaries, do not contain overlapping words, are directly related, and are consistent with the latest version. This improves the semantic understanding capability during the information retrieval process, accurately understands the true intent and needs of the query, enhances domain adaptability, reduces the false positive rate, and improves retrieval accuracy. The information retrieval system, electronic device, and computer-readable storage medium provided in this application also solve the corresponding technical problems. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0051] Figure 1 A flowchart illustrating an information retrieval method provided in this application embodiment;

[0052] Figure 2This is a flowchart for information retrieval based on entries;

[0053] Figure 3 Diagram of the optimization mechanism for batch processing;

[0054] Figure 4 This is a flowchart illustrating information retrieval based on pre-search and full-search.

[0055] Figure 5 This is an architecture diagram for the pre-retrieval cache.

[0056] Figure 6 System architecture diagram for information retrieval;

[0057] Figure 7 Flowchart for intelligent relevance judgment;

[0058] Figure 8 This is a schematic diagram of the structure of an information retrieval system provided in an embodiment of this application;

[0059] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application;

[0060] Figure 10 This is another structural schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0061] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0062] With the development of deep learning technology, information retrieval is required in processes such as intelligent question answering, and in complex technical fields such as automotive, manufacturing, and software engineering, including system requirements analysis, document management, and knowledge retrieval applications.

[0063] Traditional search primarily relies on keyword matching techniques, including TF-IDF (Term Frequency-Inverse Document Frequency) and BM25 algorithms. These techniques determine relevance by calculating the degree of matching between the query term and words in the document. For example, when a user searches for "Park Switch wake-up function," the system searches the relevant document for entries containing keywords such as "Park," "Switch," and "wake-up." However, traditional keyword search techniques have significant limitations. First, they cannot understand semantic meaning and only perform literal matching, resulting in the omission of a large amount of semantically related but worded content. For example, when a document uses different expressions such as "parking switch" and "P gear button" to describe the same function, traditional search methods cannot identify their relevance. Second, keyword search is prone to false positives; when a document contains the query keyword but is semantically unrelated, it will still be incorrectly marked as relevant. Furthermore, traditional methods cannot handle semantically related but worded cases such as synonyms and near-synonyms, such as the semantic association between technical terms like "wake-up" and "activation," or "switch" and "button."

[0064] With the rise of vector similarity retrieval technology, text can be converted into high-dimensional vector representations, and text relevance can be determined by calculating metrics such as cosine similarity between vectors. For example, word vector models such as Word2Vec and GloVe, or pre-trained language models such as BERT and GPT, can be used to convert text into vector representations. Although vector similarity retrieval has improved semantic understanding, it still has many shortcomings: vector representation capabilities are limited, making it difficult to capture complex semantic relationships and contextual information. For example, in automotive system requirements, "E_CSW wake-up" and "electronic control switch activation" are highly related semantically, but vector representations may not accurately capture this association; furthermore, vector retrieval lacks sufficient understanding of contextual information and cannot perform deep semantic reasoning and judgment. When queries involve complex technical logical relationships, simple vector similarity calculations often fail to accurately determine relevance.

[0065] Another common retrieval method is based on predefined rules and pattern matching. These methods identify relevant requirements by defining specific rule patterns, such as defining regular expression patterns to match requirement descriptions in a specific format. However, rule-based retrieval techniques have significant drawbacks: high rule maintenance costs, poor scalability, and difficulty adapting to new query patterns and changing requirements. When business needs change or new technical terms emerge, rules need to be manually updated, resulting in extremely high maintenance costs. Furthermore, rule conflict handling is complex; when multiple rules match simultaneously, it's difficult to determine priority and the final result. Most importantly, rule-based methods lack intelligent adaptability and cannot learn from and improve upon historical queries.

[0066] According to relevant technical literature, such as the traditional retrieval methods mentioned in *Introduction to Information Retrieval* (Christopher D. Manning et al.) and the vector retrieval techniques discussed in *Deep Learning in Information Retrieval* (ACM Computing Surveys, 2019), existing retrieval methods generally suffer from insufficient accuracy when handling complex semantic queries. In the field of automotive industry requirements analysis, relevant technical literature such as *Requirements Engineering for Automotive Electronic Systems* (SAE International, 2020) points out that traditional retrieval methods are ineffective in handling technical terminology and complex technical relationships. The literature states that automotive system requirements documents typically contain a large number of technical terms, technical parameters, and complex logical relationships, making it difficult for traditional keyword- or vector-based retrieval methods to accurately understand this complex technical content.

[0067] After comprehensive analysis, existing retrieval methods suffer from the following core problems: insufficient semantic understanding capabilities, failing to accurately understand the true intent and needs of users' queries; low retrieval accuracy, high false positive rate, and inaccurate relevance judgment; poor domain adaptability, struggling to handle complex terminology and concepts in specialized fields; lack of intelligent judgment capabilities, unable to perform in-depth semantic analysis and relevance judgment; and limited scalability, making it difficult to adapt to new query patterns and business needs. These technical deficiencies severely restrict the practicality and accuracy of customer demand retrieval systems, especially in fields with high technical requirements such as the automotive industry, where traditional retrieval methods can no longer meet increasingly complex business needs. The information retrieval solution provided in this application can improve retrieval accuracy.

[0068] Please see Figure 1 , Figure 1 This is a flowchart of an information retrieval method provided in an embodiment of this application.

[0069] An information retrieval method provided in this application embodiment may include the following steps:

[0070] Step S101: Obtain the query information to be processed.

[0071] In practical applications, the query information to be processed can be obtained first, for example, by using the llm_relevance_judgment() function. The query information can be generated by the user or by the human-computer interaction device, and the type of query information can be flexibly determined according to the application scenario. For example, the query information can be information for analyzing vehicles or information for managing software, etc. This application does not make any specific limitations here.

[0072] Step S102: Determine the core judgment criteria and exclusion rules for matching the query information; the core judgment criteria include the function matching criteria, the domain consistency criteria, and the context-related criteria; the exclusion rules include the general description exclusion rules, the system boundary exclusion rules, the word overlap exclusion rules, the indirect relevance exclusion rules, and the historical version exclusion rules.

[0073] Step S103: Process the query information according to the core judgment criteria and exclusion rules to obtain system prompts.

[0074] In practical applications, existing solutions often fail to accurately understand the true intent and content of queries, leading to significant discrepancies between search results and expected outcomes. Therefore, to improve search accuracy, this application enhances the deep semantic understanding capabilities of the large language model to accurately understand query intent, including complex technical terms, professional concepts, and contextual relationships. This requires defining core judgment criteria and exclusion rules for matching the query information. For example, core judgment criteria and exclusion rules can be determined based on the type of query information. These criteria include functional matching, domain consistency, and contextual relevance. Exclusion rules include general description exclusion, system boundary exclusion, word overlap exclusion, indirect relevance exclusion, and historical version exclusion. The query information is then processed according to these core judgment criteria and exclusion rules to obtain system prompts. For ease of understanding, assuming a user's query is "Park Switch wake-up related needs," this application accurately understands that this is not merely a keyword matching problem but a comprehensive query involving multiple technical aspects such as the wake-up mechanism of the parking switch, related hardware inputs, and communication, and thus obtains relevant system prompts.

[0075] It should be noted that the function matching standard is used to limit system prompts to explicitly describe the core functions or features of the query information; the domain consistency standard is used to limit system prompts to belong to the technical field, system, or component involved in the query information; the context relevance standard is used to limit system prompts to be related to the operating environment and usage scenario involved in the query information; the general description exclusion rule is used to exclude prompts that only contain general descriptions but have no specific function implementation; the system boundary exclusion rule is used to exclude prompts that belong to other systems or components and are unrelated to the query information; the lexical overlap exclusion rule is used to exclude prompts that only overlap with the query information in terms of lexical meaning but have no substantial connection; the indirect relevance exclusion rule is used for prompts that require multiple steps to be sorted to be related to the issues involved in the query information; and the historical version exclusion rule is used to exclude obsolete or outdated prompts.

[0076] Step S104: Perform information retrieval based on system prompts to obtain the target retrieval results for the query information.

[0077] In practical applications, after receiving system suggestions, information retrieval can be performed based on them. For example, the `call_llm_api()` function can be used to call a pre-trained Large Language Model (LLM) to process the system suggestions in batches, obtaining the target retrieval results. These results can include filenames, data content, etc. Furthermore, retrieval coverage can be calculated based on the target retrieval results and query information. Adjustments to the information retrieval process can then be made based on this coverage to ensure effective information retrieval.

[0078] It should be noted that information retrieval methods can be selected or combined based on specific business needs, technical conditions, and cost budgets. For example, in scenarios with high computational efficiency requirements, a hybrid retrieval scheme based on vector databases can be adopted. This scheme constructs a vector representation of the requirement document, combines vector similarity calculation with traditional keyword matching to achieve the requirement retrieval function. Specific technical implementations include: using a pre-trained language model to convert the requirement document into a high-dimensional vector representation, storing it in a vector database, calculating the similarity between the query vector and the document vector using algorithms such as cosine similarity, performing keyword matching using methods such as TF-IDF, and combining the results of both methods to determine relevance. The advantages are high computational efficiency, fast vector retrieval speed, relatively simple technical implementation, and low development and maintenance costs due to its mature vector retrieval technology.

[0079] In scenarios requiring high retrieval accuracy, an intelligent retrieval scheme based on multi-model fusion can be adopted. This scheme integrates various retrieval models, including keyword matching models, vector similarity models, and semantic understanding models, and combines the results of each model through a model fusion algorithm. Specific technical implementations include: training multiple specialized retrieval models, each responsible for different retrieval tasks; designing model fusion algorithms, such as weighted averaging, voting mechanisms, and learning ranking; and optimizing the weights and fusion strategies of each model through machine learning methods. This scheme can comprehensively utilize the advantages of different models, improving the robustness and accuracy of retrieval. Through model fusion, the limitations of a single model can be reduced, improving the overall retrieval effect. The technical implementation has good scalability, allowing new models to be added or the fusion strategy to be adjusted as needed.

[0080] In large-scale data processing scenarios, a parallel retrieval scheme based on a distributed architecture can be adopted. This scheme distributes retrieval tasks across multiple computing nodes for parallel processing by constructing a distributed retrieval architecture. Specifically, it includes: designing a distributed retrieval architecture, including a master node and multiple worker nodes; implementing a task distribution mechanism to assign retrieval tasks to different worker nodes; using technologies such as message queues to achieve communication and coordination between nodes; and designing a result aggregation algorithm to integrate the retrieval results from each node. This scheme can fully utilize multi-core CPUs and distributed computing resources, significantly improving the processing capacity of large-scale data. Through parallel processing, it can greatly shorten retrieval time, increase system throughput, and the technical architecture has good scalability, allowing the number of nodes to be dynamically adjusted according to data scale and processing needs.

[0081] In scenarios with high interpretability requirements, a rule-based intelligent retrieval solution can be adopted. This solution achieves intelligent demand retrieval functionality by constructing a domain knowledge base and a rule engine. Specific technical implementations include: building a knowledge base for the automotive industry, containing professional terminology, technical concepts, and relationship graphs; designing a rule engine, defining various retrieval rules and judgment logic; implementing a reasoning mechanism for intelligent reasoning based on the knowledge base and rules; and designing a rule update mechanism to support the dynamic addition and modification of rules. This solution offers good interpretability, with relatively transparent retrieval results and judgment processes, facilitating user understanding and verification. Furthermore, the rule engine provides good controllability, allowing for flexible adjustments to retrieval strategies based on business needs. The technical implementation is relatively stable, does not rely on external API services, and exhibits good reliability.

[0082] In scenarios requiring high semantic understanding, an end-to-end retrieval solution based on deep learning can be adopted. This solution trains an end-to-end deep learning model to directly learn relevance judgments from user queries and requirement documents. Specifically, this involves: constructing a large-scale training dataset containing user queries, requirement documents, and relevance tags; designing an end-to-end neural network architecture, such as pre-trained models like BERT or GPT; implementing model training and optimization; optimizing model parameters through backpropagation; and designing a model inference mechanism to achieve rapid relevance judgment. This solution can learn complex semantic relationships end-to-end, possesses powerful representation learning capabilities, and through large-scale data training, the model can learn rich semantic knowledge and domain features. The technical solution has good generalization ability and can adapt to different query patterns and document types.

[0083] This application provides an information retrieval method that involves: acquiring query information to be processed; determining core judgment criteria and exclusion rules for matching the query information; processing the query information according to the core judgment criteria and exclusion rules to obtain system suggestion words; and performing information retrieval based on the system suggestion words to obtain the target retrieval results for the query information. The core judgment criteria include functional matching criteria, domain consistency criteria, and context relevance criteria; the exclusion rules include general description exclusion rules, system boundary exclusion rules, word overlap exclusion rules, indirect relevance exclusion rules, and historical version exclusion rules. In this application, after acquiring the query information, it is necessary to process the query information according to the functional matching criteria, domain consistency criteria, and context relevance criteria, and also according to the general description exclusion rules, system boundary exclusion rules, word overlap exclusion rules, indirect relevance exclusion rules, and historical version exclusion rules. This process yields system suggestion words that are functionally matched to the novelty information, domain consistent, context-related and without general descriptions, cross system boundaries, do not contain overlapping words, are directly related, and are consistent with the latest version. This improves the semantic understanding capability during the information retrieval process, accurately understands the true intent and needs of the query, enhances domain adaptability, reduces the false positive rate, and improves retrieval accuracy. Furthermore, in terms of efficiency, by improving the accuracy of semantic understanding, the accuracy of demand retrieval can be increased from 60-70% of traditional methods to over 90%, significantly reducing the workload of manual secondary screening. In terms of cost control, accurate semantic understanding can reduce the cost of repetitive work and rework caused by retrieval errors, and is expected to reduce labor costs by 30-40%. In terms of quantitative effects, the retrieval accuracy can be increased from 65% of traditional methods to over 92%, and the false positive rate can be reduced from 35% to below 8%. In terms of qualitative effects, through strict exclusion rules and intelligent judgment criteria, false positives such as general descriptions, system boundary mismatches, and overlapping words without substantial relevance can be effectively filtered, greatly improving the quality and credibility of retrieval results.

[0084] Based on the above embodiments, information retrieval schemes suffer from low retrieval efficiency when processing large-scale data. To avoid this situation, please refer to... Figure 2 The information retrieval method provided in this application embodiment may include the following steps:

[0085] Step S201: Obtain the query information to be processed.

[0086] Step S202: Determine the core judgment criteria and exclusion rules for matching the query information; the core judgment criteria include the function matching criteria, the domain consistency criteria, and the context-related criteria; the exclusion rules include the general description exclusion rules, the system boundary exclusion rules, the word overlap exclusion rules, the indirect relevance exclusion rules, and the historical version exclusion rules.

[0087] Step S203: Flatten the query information to obtain the atomic entries corresponding to the query information.

[0088] Step S204: Group the atomic entries to obtain the entry grouping results.

[0089] Step S205: Process the grouped results of the entries in batches according to the core judgment criteria and exclusion rules to obtain system prompt words.

[0090] In practical applications, during the process of processing query information according to core judgment criteria and exclusion rules to obtain system prompts, the query information can be flattened to obtain atomic entries corresponding to the query information. This involves organizing content at hierarchical levels such as file / chapter / paragraph / table / image, converting the query information into independently identifiable atomic entries. These entries allow for fine-grained reflection of the query information's characteristics. Atomic entries can then be grouped, for example, using the `search_and_print()` function to group them according to the `BATCH_SIZE` configuration. During grouping, entries with the same topic, device (e.g., ECU), or file can be prioritized to improve contextual consistency. The grouped results can then be processed in batches according to the core judgment criteria and exclusion rules, such as calling the `llm_relevance_judgment()` function to process the grouped results and obtain system prompts. To facilitate user awareness of the processing, a progress display mechanism can be used to show the processing progress, allowing users to understand the current batch number, total batch count, and current batch entry count, etc. Figure 3 As shown. In this way, the features in the query information can be reflected in a fine-grained manner by using entries, and the grouped results of the entries can be processed in parallel by batching, which can improve the processing efficiency by 3-5 times, reduce the number of calls to the large language model API by more than 80%, shorten the processing time by more than 60%, and support the efficient processing of large-scale data, thereby improving the efficiency of system prompt word generation and thus improving information retrieval efficiency.

[0091] In an exemplary embodiment, during the process of flattening the query information to obtain the atomic entries corresponding to the query information, context information of the query information can be generated. The context information may include file name, chapter path, REQ_ID, page number, coordinates, etc.; the query information is processed to obtain initial information; evidence extraction and index enhancement are performed on the initial information, such as by signal name, message ID, state machine terminology, scene words and ECU entities, to obtain candidate information; feature extraction is performed on the candidate information to obtain feature data, which may be inverted / vector features, etc.; the feature data is deduplicated and weakly similar is merged to retain the strongest evidence version and record synonyms / aliases to obtain atomic entries; and metadata of the atomic entries is generated based on the context information.

[0092] In specific application scenarios, during the processing of query information to obtain initial information, the paragraphs in the query information can be denoised, whitespace standardized, line breaks merged, list merged, unit standardized, and terminology standardized to obtain the natural paragraph processing results; the tables and images in the query information can be keyed with table headers and text extracted, including OCR / title extraction, to obtain structured placeholder entries and position pointers, including external links / coordinate pointers; according to semantic boundaries, the sentences spanning paragraphs in the query information are divided into the smallest semantic units, generating parent-child relationships and sequential indexes between the smallest semantic units and corresponding paragraphs; the natural paragraph processing results, structured placeholder entries, position pointers, smallest semantic units, parent-child relationships, and sequential indexes are used as the initial information.

[0093] In specific application scenarios, during the process of generating metadata for atomic entries based on contextual information, the metadata for atomic entries can be generated based on contextual information. The metadata includes source file, paragraph location results, topic candidate results, evidence fragment summary, length and quality score.

[0094] Step S206: Perform information retrieval based on system prompts to obtain the target retrieval results for the query information.

[0095] To further improve retrieval efficiency based on the above embodiments, please refer to [link / reference]. Figure 4 The information retrieval method provided in this application embodiment may include the following steps:

[0096] Step S301: Obtain the query information to be processed.

[0097] Step S302: Determine the core judgment criteria and exclusion rules for matching the query information; the core judgment criteria include the function matching criteria, the domain consistency criteria, and the context-related criteria; the exclusion rules include the general description exclusion rules, the system boundary exclusion rules, the word overlap exclusion rules, the indirect relevance exclusion rules, and the historical version exclusion rules.

[0098] Step S303: Process the query information according to the core judgment criteria and exclusion rules to obtain system prompts.

[0099] Step S304: Obtain the cached pre-retrieval results.

[0100] Step S305: Detect whether the system prompt word matches the pre-search result; if the system prompt word matches the pre-search result, proceed to step S306; if the system prompt word does not match the pre-search result, proceed to step S307.

[0101] Step S306: Extract the target search results from the pre-search results to obtain the query information.

[0102] Step S307: Align and constrain the system prompt words according to a unified semantic coordinate system to obtain a semantic template.

[0103] Step S308: Call the trained large language model to retrieve semantic templates and obtain initial retrieval results with values ​​of temperature coefficient and quantities matching the maximum number of tokens.

[0104] Step S309: Perform consistency verification and confidence aggregation on the initial search results to obtain the target search results for the query information.

[0105] In practical applications, considering that different users may search for the same information, caching search results allows for the rapid determination of the target search result based on pre-search results. Specifically, during the process of retrieving information based on system prompts and obtaining the target search result, cached pre-search results can be retrieved; the system prompts and pre-search results can be checked for match; if they match, the target search result is extracted from the pre-search results; if they do not match, a full search of the system prompts can be performed to obtain the corresponding target search result. In terms of cost control, the pre-search result caching mechanism can significantly reduce redundant computation costs, potentially reducing computational costs by more than 50%. Regarding energy conservation, by optimizing the processing flow and reducing invalid computations, overall energy consumption can be reduced by more than 40%.

[0106] In an exemplary embodiment, the topic name and subquery of the pre-search results can be received as input parameters through the search_from_preindex() function, and intelligent secondary retrieval can be performed on the cached pre-search results to realize the pre-search function. Specifically, the cached pre-search results are loaded using the `load_preindex_entries()` function, the pre-search base directory in the pre-search results is parsed using the `resolve_preindex_base_dir()` function, the topic names in the pre-search results are converted to a file system-friendly format using the `slugify()` function, the path to the `entries.json` file is constructed, the cached entry data in the pre-search results is loaded, and then the `search_strict_in_entries()` function is called to perform intelligent searching on the cached entries. For example, `load_preindex_entries()` loads the cached entries in the pre-search results, `list_preindex_topics()` lists the available pre-search topics, `find_best_topic_for_query()` finds the best matching pre-search topic for the system suggestion, and `resolve_topic_name()` checks whether the system suggestion matches the pre-search topic. When the system suggestion matches the pre-search topic, the search result in the pre-search results is used as the target search result; when the system suggestion does not match the pre-search topic, it automatically falls back to full search mode to ensure the completeness of the search results. Figure 5 As shown. During this process, a multi-level retrieval strategy can be implemented through the main() function, prioritizing the use of pre-retrieval results and automatically switching to full retrieval mode when the cache misses.

[0107] In specific application scenarios, pre-retrieval results can be organized using a "topic-subquery" secondary index. After obtaining the query information to be processed, pre-retrieval can be performed directly based on the query information. During this process, the topic index maps high-frequency retrieval results to entries.json under data / pre_index / topic / . After the subquery index finds a topic with new information, it performs secondary filtering based on specific intents (such as "wake-up threshold" or "CAN NM status"). Strict matching and evidence alignment can be completed in memory first, and only a small number of high-purity candidates enter LLM adjudication. When the topic is not hit or the meta-data quality is below the threshold, it automatically reverts to full flattening recall before entering the subsequent three-dimensional judgment and batch process of this application. Figure 6 As shown.

[0108] In an exemplary embodiment, during the full retrieval process, system prompts can be aligned and constrained according to a unified semantic template to obtain semantic processing results with a clear "yes / no" conclusion. The semantic template can be a three-dimensional semantic template. For example, in the automotive industry, a three-dimensional semantic template can be constructed based on functions such as wake-up / start-stop / anti-accidental touch, domains such as body control / network management / switching systems, and scenarios such as parking / power-on / sleep boundary conditions. Constraints can be imposed based on expert prompts to unify the judgments to the same semantic coordinate system and achieve high consistency. The trained large language model is then used to retrieve the semantic processing results, obtaining values ​​that are temperature coefficients and whose quantity matches the maximum token. The initial retrieval results of number matching are used to ensure stable output of the large language model by employing a temperature coefficient, and the data volume of the initial retrieval results is limited by the maximum number of tokens, where the temperature coefficient can be a low temperature coefficient, and the maximum number of tokens can be 1024, etc. Consistency checks and confidence aggregation are performed on the initial retrieval results to obtain the target retrieval results for the query information. For example, the initial retrieval results are only used as the target retrieval results for the query information when the mechanism evidence for the same system prompt words is sufficient and the decisions are consistent. In specific application scenarios, the target retrieval results output by the large language model can be a judgment result composed of yes and no, and can be converted into a list of boolean values ​​and returned to the caller, such as... Figure 7 As shown, this is for the caller to process and then display to the user.

[0109] To facilitate understanding of the information retrieval scheme provided in this application, let's assume a search is conducted in the automotive industry, using the methods and effects shown in Table 1. The following steps would be included:

[0110] Obtaining vehicle-related query information input by users can be used for new project requirements analysis, impact assessment of requirement changes, requirement consistency checks, and solution design.

[0111] Determine the core judgment criteria and exclusion rules for matching the query information; the core judgment criteria include functional matching criteria, domain consistency criteria, and context relevance criteria; the exclusion rules include general description exclusion rules, system boundary exclusion rules, word overlap exclusion rules, indirect relevance exclusion rules, and historical version exclusion rules;

[0112] Contextual information for generating query information;

[0113] The paragraphs in the query information are processed by denoising, whitespace normalization, line break merging, list merging, unit standardization, and terminology standardization to obtain the natural paragraph processing results.

[0114] The table and image data in the query information are keyed in the header and the text is extracted to obtain structured placeholder entries and position pointers;

[0115] Based on semantic boundaries, statements that span paragraphs in the query information are divided into the smallest semantic units, and parent-child relationships and sequential indexes between the smallest semantic units and their corresponding paragraphs are generated.

[0116] The initial information includes the natural paragraph processing results, structured placeholder entries, position pointers, minimum semantic units, parent-child relationships, and sequential indexes.

[0117] Evidence extraction and index enhancement are performed on the initial information to obtain candidate information;

[0118] Feature extraction is performed on the candidate information to obtain feature data;

[0119] The feature data is deduplicated and weakly similarized to obtain atomic entries;

[0120] Based on contextual information, metadata for atomic entries is generated, including source file, paragraph location results, topic candidate results, evidence fragment summary, length and quality score;

[0121] Group the atomic entries to obtain the entry grouping results;

[0122] Based on the core judgment criteria and exclusion rules, the batch grouping results of the entries are processed to obtain system prompt words;

[0123] Retrieve cached pre-search results;

[0124] The system detects whether the suggested words match the pre-search results.

[0125] If the system prompt matches the pre-search results, the target search result containing the query information is extracted from the pre-search results.

[0126] If the system prompts do not match the pre-search results, the system prompts are aligned and constrained according to a unified semantic coordinate system to obtain a semantic template.

[0127] The trained large language model is called to retrieve semantic templates, and initial retrieval results are obtained with values ​​of temperature coefficient and numbers matching the maximum number of tokens.

[0128] The initial search results are subjected to consistency verification and confidence aggregation to obtain the target search results for the query information.

[0129] Table 1. Implementation methods and effects of this solution

[0130]

[0131] The results of comparing this application with existing solutions are shown in Table 2. After testing and verifying the solution of this application, including functional testing, performance testing, stability testing, and compatibility testing, the test results show that the solution of this application has achieved the expected goals in terms of accuracy, efficiency, and stability, and has practical application value.

[0132] Table 2 Comparison of this application with existing proposals

[0133]

[0134] Please see Figure 8 , Figure 8 This is a schematic diagram of the structure of an information retrieval system provided in an embodiment of this application.

[0135] An information retrieval system provided in this application embodiment may include:

[0136] The query information acquisition module 101 is used to acquire query information to be processed.

[0137] The filtering criteria determination module 102 is used to determine the core judgment criteria and exclusion rules that match the query information;

[0138] The query processing module 103 is used to process the query information according to the core judgment criteria and exclusion rules to obtain system prompt words;

[0139] The retrieval module 104 is used to perform information retrieval based on system prompts and obtain the target retrieval results for the query information.

[0140] The core judgment criteria include functional matching criteria, domain consistency criteria, and context-related criteria; the exclusion rules include general description exclusion rules, system boundary exclusion rules, word overlap exclusion rules, indirect relevance exclusion rules, and historical version exclusion rules.

[0141] This application provides an information retrieval system, wherein the query processing module may include:

[0142] Flattening units are used to flatten query information to obtain atomic entries corresponding to the query information;

[0143] Grouping units are used to group atomic entries to obtain entry grouping results;

[0144] The processing unit is used to process the grouped results of entries in batches according to the core judgment criteria and exclusion rules to obtain system prompt words.

[0145] This application provides an information retrieval system in which a flattened unit is used to: generate context information for query information; process the query information to obtain initial information; extract evidence and enhance the index of the initial information to obtain candidate information; extract features from the candidate information to obtain feature data; deduplicate and merge weakly similar features to obtain atomic entries; and generate metadata for the atomic entries based on the context information.

[0146] This application provides an information retrieval system in which a flattening unit is used to: denoise, standardize whitespace, merge line breaks, merge lists, standardize units, and standardize terminology in paragraphs of the query information to obtain natural paragraph processing results; perform header keying and text extraction on tables and images in the query information to obtain structured placeholder entries and position pointers; divide sentences spanning paragraphs in the query information into the smallest semantic units according to semantic boundaries, and generate parent-child relationships and sequential indexes between the smallest semantic units and corresponding paragraphs; and use the natural paragraph processing results, structured placeholder entries, position pointers, smallest semantic units, parent-child relationships, and sequential indexes as initial information.

[0147] This application provides an information retrieval system in which a flattened unit is used to generate metadata for atomic entries based on contextual information. The metadata includes source files, paragraph location results, topic candidate results, evidence fragment summaries, length and quality scores.

[0148] This application provides an information retrieval system, wherein the retrieval module may include:

[0149] The pre-retrieval result acquisition unit is used to acquire cached pre-retrieval results;

[0150] The detection unit is used to detect whether the system prompts match the pre-search results;

[0151] The pre-retrieval unit is used to extract the target retrieval result from the pre-retrieval result if the system prompt word matches the pre-retrieval result.

[0152] An information retrieval system provided in this application embodiment may further include:

[0153] The full retrieval unit is used to detect whether the system prompts match the pre-retrieval results. If the system prompts do not match the pre-retrieval results, the system prompts are aligned and constrained according to a unified semantic coordinate system to obtain a semantic template. The trained large language model is called to retrieve the semantic template to obtain initial retrieval results with values ​​of temperature coefficient and quantities matching the maximum number of tokens. The initial retrieval results are then subjected to consistency verification and confidence aggregation to obtain the target retrieval results for the query information.

[0154] This application also provides an electronic device and a computer-readable storage medium, both of which have the corresponding effects of the information retrieval method provided in the embodiments of this application. Please refer to... Figure 9 , Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0155] An electronic device provided in this application includes a memory 201 and a processor 202. The memory 201 stores a computer program, and the processor 202 executes the computer program to implement the steps of the information retrieval method described in any of the above embodiments.

[0156] Please see Figure 10 Another electronic device provided in this application embodiment may further include: an input port 203 connected to the processor 202 for transmitting commands input from the outside to the processor 202; a display unit 204 connected to the processor 202 for displaying the processing results of the processor 202 to the outside; and a communication module 205 connected to the processor 202 for enabling communication between the electronic device and the outside. The display unit 204 may be a display panel, a laser scanning display, etc.; the communication method adopted by the communication module 205 includes, but is not limited to, Mobile High-Definition Link (MHL), Universal Serial Bus (USB), High-Definition Multimedia Interface (HDMI), wireless connection: Wireless Fidelity (WiFi), Bluetooth communication technology, Bluetooth Low Energy communication technology, and communication technology based on IEEE 802.11s.

[0157] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the information retrieval method described in any of the above embodiments.

[0158] The computer-readable storage media involved in this application include random access memory (RAM), memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs (compact disc read-only memory), or any other form of storage media known in the art.

[0159] This application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the information retrieval method described in any of the above embodiments.

[0160] The technical implementation of this application can be based on the Python programming language and adopts a modular design architecture. The core code file is function_search_agent.py, containing 714 lines of code, which mainly implements the intelligent relevance judgment function based on a large language model. The code structure is clear, including core functions such as llm_relevance_judgment(), search_and_print(), and search_from_preindex(), each with clearly defined input and output parameters and functional descriptions. Taking the automotive industry as an example, the code implementation uses professional automotive system requirements analysis terminology, such as "E_CSW", "Park Switch", and "wake-up mechanism", reflecting the professional application background of this solution in the automotive industry. The code comments are detailed, including function descriptions, parameter explanations, and return value descriptions, facilitating user understanding of the technical implementation details.

[0161] This application utilizes a YAML configuration file, settings.yaml, for parameter management. This file contains key information such as API provider configuration, model parameter settings, and file path configuration. The system supports two API providers: OpenAI and LiteLLM, allowing for flexible API switching through configuration parameters. Key configuration parameters include: BATCH_SIZE=10 (batch size), temperature=0.1 (LLM output randomness control), max_tokens=1024 (output length limit), and debug_mode=false (debug mode control). These parameter settings directly affect the system's performance and retrieval results and are integral parts of the solution.

[0162] The data processed in this application can be in JSON format, containing fields such as ID, REQ_ID, TEXT, IMG_URL, and TABLE. The data source is automotive industry requirement documents, such as OEM_REQ.json, which contain a large amount of automotive electronic system requirement information. The pre-retrieval result cache can adopt a thematic storage structure, with each theme corresponding to a directory containing two files: meta.json (metadata) and entries.json (cache entries). The cache directory structure is data / pre_index / theme_name / , facilitating theme management and cache maintenance.

[0163] Furthermore, this application's solution boasts excellent scalability, supporting adjustments to system behavior through configuration parameters and the addition of new API providers and model types. The code employs a modular design, with each functional module relatively independent, facilitating feature expansion and maintenance. The system supports detailed logging and debugging information output, aiding in problem diagnosis and performance optimization. Through configuration file management, the system can adapt to different deployment environments and performance requirements.

[0164] For descriptions of relevant parts in the information retrieval system, electronic device, and computer-readable storage medium provided in this application's embodiments, please refer to the detailed description of the corresponding parts in the information retrieval method provided in this application's embodiments; they will not be repeated here. Furthermore, parts of the technical solutions provided in this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0165] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0166] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An information retrieval method, characterized in that, include: Retrieve query information to be processed; Determine the core judgment criteria and exclusion rules for matching the query information; The query information is processed according to the core judgment criteria and the exclusion rules to obtain system prompts. Information retrieval is performed based on the system prompts to obtain the target retrieval results for the query information. The core judgment criteria include functional matching criteria, domain consistency criteria, and context-related criteria; the exclusion rules include general description exclusion rules, system boundary exclusion rules, word overlap exclusion rules, indirect relevance exclusion rules, and historical version exclusion rules.

2. The method according to claim 1, characterized in that, The query information is processed according to the core judgment criteria and the exclusion rules to obtain system prompts, including: The query information is flattened to obtain the atomic entries corresponding to the query information; The atomic entries are grouped to obtain the entry grouping results; Based on the core judgment criteria and the exclusion rules, the grouping results of the entries are processed in batches to obtain system prompt words.

3. The method according to claim 2, characterized in that, The flattening process of the query information to obtain the atomic entries corresponding to the query information includes: Context information for generating the query information; The query information is processed to obtain initial information; Evidence extraction and index enhancement are performed on the initial information to obtain candidate information; Feature extraction is performed on the candidate information to obtain feature data; The feature data is deduplicated and weakly similar data is merged to obtain atomic entries; Based on the context information, the metadata of the atomic entries is generated.

4. The method according to claim 3, characterized in that, The process of processing the query information to obtain initial information includes: The paragraphs in the query information are subjected to noise reduction, whitespace normalization, line break merging, list merging, unit standardization, and terminology standardization to obtain the natural paragraph processing results; The tables and images in the query information are keyed in the header and the text is extracted to obtain structured placeholder entries and position pointers. According to semantic boundaries, the statements that span paragraphs in the query information are divided into the smallest semantic units, and the parent-child relationship and sequential index between the smallest semantic units and the corresponding paragraphs are generated; The natural segment processing result, the structured placeholder entry, the position pointer, the smallest semantic unit, the parent-child relationship, and the sequence index are used as initial information.

5. The method according to claim 4, characterized in that, The process of generating the metadata for the atomic entries based on the context information includes: Based on the context information, metadata for the atomic entries is generated, including source file, paragraph location results, topic candidate results, evidence fragment summary, length, and quality score.

6. The method according to claim 5, characterized in that, The process of retrieving information based on the system prompts to obtain the target retrieval results for the query information includes: Retrieve cached pre-search results; Detect whether the system prompts match the pre-search results; In response to a system prompt matching the pre-search results, the target search result for the query information is extracted from the pre-search results.

7. The method according to claim 6, characterized in that, After detecting whether the system prompt words match the pre-search results, the method further includes: If the system prompt does not match the pre-search result, the system prompt is aligned and constrained according to a unified semantic template to obtain the semantic processing result. The trained large language model is invoked to retrieve the semantic processing results, and initial retrieval results with values ​​of temperature coefficient and numbers matching the maximum number of tokens are obtained. The initial search results are subjected to consistency verification and confidence aggregation to obtain the target search results for the query information.

8. An information retrieval system, characterized in that, include: The query information acquisition module is used to acquire query information to be processed. The filtering criteria determination module is used to determine the core judgment criteria and exclusion rules that match the query information; The query processing module is used to process the query information according to the core judgment criteria and the exclusion rules to obtain system prompt words; The retrieval module is used to perform information retrieval based on the system prompts and obtain the target retrieval results of the query information; The core judgment criteria include functional matching criteria, domain consistency criteria, and context-related criteria; the exclusion rules include general description exclusion rules, system boundary exclusion rules, word overlap exclusion rules, indirect relevance exclusion rules, and historical version exclusion rules.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the information retrieval method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the information retrieval method as described in any one of claims 1 to 7.

Images