Multi-step reasoning information acquisition method, system and equipment and storage medium

Abstract

The invention discloses a multi-step reasoning information acquisition method, system and device and a storage medium, which are corresponding schemes: performing semantic analysis and dynamic correction on task query by using a large language model to realize multi-step reasoning and adaptive optimization of query semantics; moreover, an incremental multi-path recall system fused by heterogeneous associated signals is designed, and keyword, semantic and structured features are fully fused through a multi-path, multi-view and multi-round progressive recall joint mechanism, so that high-coverage and high-precision candidate information recall is realized; in addition, a fine ranking scheme based on deep thinking (thinking chain) and reasoning expansion is provided, and interpretable logic analysis and accurate ranking are carried out on candidate information. Generally speaking, the information understanding, reasoning depth and retrieval accuracy of the model under the complex incidence relation can be remarkably improved while the calculation efficiency is kept, and universal technical support is provided for a high-reliability knowledge acquisition and scientific intelligent system.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, system, device and storage medium for multi-step reasoning information acquisition. Background Technology

[0002] Information retrieval aims to help users quickly and accurately find relevant information that meets their needs from massive amounts of data. With the rapid development of information technology, the volume of data is exploding, and users' information needs are becoming increasingly complex and specialized. This makes efficient and accurate information retrieval technology play a crucial role in scientific research, education, software development, and daily life. Through information retrieval, users can effectively manage massive amounts of information, thereby significantly improving the efficiency and quality of knowledge acquisition and supporting the continuous development of scientific research, decision analysis, and innovative applications.

[0003] However, in many cutting-edge and specialized fields, information no longer exists in the form of simple, isolated facts, but is interconnected through complex relationships, forming structured and logically robust knowledge systems. These complex relationships refer to the indirect, multi-layered, and conditional logical or semantic connections between pieces of information. Its core characteristic lies in the indirectness of these connections; that is, relevant information is not connected through superficial word overlap or simple semantic similarity, but requires multi-step reasoning, logical deduction, or the introduction of domain knowledge to establish a link. This characteristic is particularly prominent in specialized fields such as mathematical deduction, coding, and scientific research. Faced with these scenarios, information acquisition technologies must possess the ability to handle complex relationships, integrating information from different sources through multi-step reasoning to reveal the deep logical connections between knowledge, thereby accurately meeting users' complex information needs.

[0004] To address the ever-increasing demand for information retrieval, existing technologies typically employ a multi-path recall followed by refined re-ranking. During the recall phase, to cover as many potential relevant documents as possible, the system uses a multi-path recall strategy, including keyword recall based on word frequency matching and vector recall based on semantic space similarity. Subsequently, in the refined re-ranking phase, the system interacts with the query and candidate documents, calculates their relevance scores, and optimizes the ranking of the recall results, thereby improving the accuracy of the final outcome.

[0005] While the aforementioned multi-path recall followed by refined reordering schemes has achieved significant results in general information retrieval tasks, it still has many shortcomings when dealing with complex relationships. Existing research and applications mainly rely on general internet corpora, but the proportion of high-quality, structured professional knowledge in such data is extremely low. Therefore, existing models often lack the ability to deeply understand and logically model symbolic language (such as mathematical formulas and program code) in professional fields, making it difficult to analyze their abstract concepts and complex reasoning processes, and thus unable to accurately capture related information in complex scenarios. This makes it difficult for both vector recall and keyword recall to effectively identify truly relevant content due to their inability to understand the complex meaning behind the information, resulting in limited information coverage and accuracy in the recall stage.

[0006] More importantly, existing solutions generally focus on synonym matching or semantic modeling of questions and answers, rather than deeply mining the complex logical relationships and reasoning paths hidden behind the information. This keeps the core capabilities of existing models at a shallow semantic level, lacking the ability to model complex relationships and making them unsuitable for information retrieval tasks that require multi-step logical reasoning. In addition, information retrieval tasks oriented towards complex relationships are often verbose and semantically ambiguous, containing multiple constraints and lacking explicit keyword matching or direct semantic correspondence with the target information, thus limiting the information discrimination ability and ranking accuracy in the fine-grained re-ranking stage.

[0007] These challenges make it difficult for existing information retrieval technologies to achieve efficient and accurate knowledge acquisition in cutting-edge and specialized fields. Therefore, there is an urgent need to fully leverage the powerful reasoning and knowledge understanding capabilities of large language models to design novel information retrieval systems capable of handling complex relationships. These systems should be able to integrate multi-source knowledge based on multi-step logical reasoning, uncover deep connections between information, and thus achieve more intelligent and reliable acquisition of complex information. Summary of the Invention

[0008] The purpose of this invention is to provide a multi-step reasoning information acquisition method, system, device, and storage medium that, in information acquisition tasks involving complex relationships, utilizes the reasoning capabilities of a large language model to achieve stronger comprehension and higher retrieval accuracy.

[0009] The objective of this invention is achieved through the following technical solution: A multi-step reasoning information acquisition method includes: Through a multi-step reasoning mechanism and dynamic retrieval enhancement strategy, the original query of the task is semantically understood and structurally decomposed based on the first major language model, and retrieval-enhanced query is gradually generated. The query enhancement based on retrieval uses multi-path recall to obtain multiple candidate result sets, and performs a comprehensive evaluation of all candidate result sets from multiple dimensions to obtain an initial comprehensive evaluation result. Based on the initial comprehensive evaluation result, a multi-round progressive recall strategy is adopted to gradually expand the search scope and obtain the final search information. Based on the thought chain reasoning mechanism and the reasoning-time expansion strategy, the final retrieved information is semantically understood and relevance-ranked to obtain the rearranged information.

[0010] A multi-step reasoning information acquisition system, used to implement the aforementioned method, includes: The iterative query understanding module based on retrieval enhancement is used to perform semantic understanding and structured decomposition of the original query of the task based on the first major language model through a multi-step reasoning mechanism and dynamic retrieval enhancement strategy, and gradually generate retrieval-enhanced queries. The heterogeneous correlation signal fusion incremental multipath recall module is used to recall multiple candidate result sets through multipath based on search enhancement query, and to comprehensively evaluate all candidate result sets from multiple dimensions to obtain an initial comprehensive evaluation result; based on the initial comprehensive evaluation result, a multi-round progressive recall strategy is adopted to gradually expand the search scope and obtain the final search information. The fine-ranking module based on deep thinking and reasoning extension is used to perform semantic understanding and relevance ranking on the final retrieved information based on the thinking chain reasoning mechanism and reasoning extension strategy, so as to obtain the reordered information.

[0011] A processing device includes: one or more processors; and a memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0012] A readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0013] As can be seen from the technical solution provided by the present invention, it addresses the challenges of modeling multi-source heterogeneous signals, cross-semantic associations, and logical dependencies in complex relational information requirements. A unified framework integrating retrieval enhancement, reasoning extension, and deep ranking is proposed. First, a large language model is used to perform semantic parsing and dynamic correction of the task query, achieving multi-step reasoning and adaptive optimization of the query semantics. Second, an incremental multi-path recall system for heterogeneous relational signal fusion is designed. Through a joint mechanism of multi-path, multi-perspective, and multi-round progressive recall, keywords, semantics, and structured features are fully integrated to achieve high-coverage and high-precision candidate information recall. Finally, a fine-ranking scheme based on deep thinking (thought chain) and reasoning-time extension is proposed to perform interpretable logical analysis and precise ranking of candidate information. Overall, the present invention can significantly improve the model's information understanding, reasoning depth, and retrieval accuracy under complex relational relationships while maintaining computational efficiency, providing general technical support for highly reliable knowledge acquisition and scientific intelligent systems. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of a multi-step reasoning information acquisition method provided in an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram illustrating the overall working process of a multi-step reasoning information acquisition method provided in an embodiment of the present invention.

[0017] Figure 3 This is a schematic diagram of an iterative query understanding scheme based on retrieval enhancement provided in an embodiment of the present invention.

[0018] Figure 4 This is a schematic diagram of an incremental multipath recall system for heterogeneous correlation signal fusion provided in an embodiment of the present invention.

[0019] Figure 5 This is a schematic diagram of a fine-sorting scheme based on deep thinking and reasoning extension provided in an embodiment of the present invention.

[0020] Figure 6 This is a schematic diagram of a multi-step reasoning information acquisition system provided in an embodiment of the present invention.

[0021] Figure 7 This is a schematic diagram of a processing device provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0023] First, the following explanations are provided for the terms that may be used in this article: The terms "comprising," "including," "containing," "having," or other similar semantic descriptions should be interpreted as non-exclusive inclusion. For example, including a technical feature element (such as raw material, component, ingredient, carrier, dosage form, material, size, part, component, mechanism, device, step, process, method, reaction conditions, processing conditions, parameter, algorithm, signal, data, product or article of manufacture, etc.) should be interpreted as including not only the expressly listed technical feature element, but also other technical feature elements that are not expressly listed and are well-known in the art.

[0024] The following provides a detailed description of a multi-step reasoning information acquisition method, system, device, and storage medium provided by this invention. Contents not described in detail in the embodiments of this invention are prior art known to those skilled in the art. Where specific conditions are not specified in the embodiments of this invention, conventional conditions or conditions recommended by the manufacturer in the art shall apply. Reagents or instruments used in the embodiments of this invention, unless otherwise specified by the manufacturer, are all commercially available conventional products.

[0025] Example 1 This invention provides a method for obtaining multi-step reasoning information, such as... Figure 1 As shown, it mainly includes the following steps: Step 1: Iterative query understanding based on retrieval enhancement.

[0026] This step aims to achieve multi-step reasoning and adaptive query optimization for complex information needs by combining semantic analysis, task understanding, and dynamic retrieval feedback. Specifically, in this embodiment of the invention, through a multi-step reasoning mechanism and a dynamic retrieval enhancement strategy, the original query of the task is semantically understood and structurally decomposed based on the first major language model to gradually generate a retrieval-enhanced query.

[0027] The preferred implementation of this step is as follows: The first language model is used to perform semantic understanding and structured parsing of the original query of the task, identifying the query intent, information demand type, and logical dependencies as analysis results. For tasks with multiple constraints or multiple semantic layers, automatic demand decomposition is performed through reasoning, dividing the overall task into several sub-tasks or sub-demands. Semantic understanding and structured parsing are then performed on each sub-task or sub-demand to obtain corresponding analysis results. Subsequently, an initial query is generated based on the analysis results. An external retrieval engine is used to execute the initial query and obtain preliminary candidate results. These preliminary candidate results are then used to modify or expand the initial query, regenerating a new query, and the external retrieval engine continues to obtain corresponding candidate results. This process is iterated continuously, with the first language model evaluating in each iteration whether the current query fully reflects the intent of the original query and has sufficient retrieval effectiveness. When the first language model determines that the query has been rewritten well enough, the iteration terminates, resulting in a retrieval-enhanced query.

[0028] In this embodiment of the invention, the step of modifying or expanding the initial query using preliminary candidate results and regenerating a new query includes: performing semantic verification and reasoning analysis on the preliminary candidate results using a first major language model, optimizing the expression or retrieval scope of the initial query based on the verification and analysis results, and generating a new query.

[0029] Preferably, a strategy optimization mechanism based on reinforcement learning is also introduced. During the iteration process, a reward signal is calculated based on the retrieved candidate results to guide the first language model to optimize its high-quality query generation strategy. The reward function is set according to the ranking accuracy, relevance improvement or information coverage improvement index of the candidate results to positively incentivize the query generation strategy of the first language model. In some scenarios, user feedback or task completion is also used as additional rewards.

[0030] In this embodiment of the invention, the query generation strategy optimization based on reinforcement learning uses a reward function to adjust the probability of the strategy model generating an answer, thereby correcting different query rewrites.

[0031] Step 2: Incremental multipath recall through heterogeneous correlation signal fusion.

[0032] In this embodiment of the invention, the search-enhanced query obtains multiple candidate result sets through multi-path recall, and comprehensively evaluates all candidate result sets from multiple dimensions; based on the comprehensive evaluation results, a multi-round progressive recall strategy is adopted to gradually expand the search scope and obtain the final search information. Its preferred implementation is as follows: (1) Multi-path recall: Utilizing the generation and rewriting capabilities of the first language model, multiple sampling is performed on the retrieval-enhanced query to generate several semantically diverse query rewriting paths; each rewriting path independently performs recall operations under different keyword representations, semantic subspaces or contextual concerns to obtain multiple candidate result sets.

[0033] (2) Multi-perspective fusion: A multi-perspective fusion mechanism is introduced to conduct unified correlation evaluation and fusion of all candidate result sets; among them, the candidate result sets are comprehensively judged from multiple dimensions, including keyword matching score, semantic similarity score and contextual relevance; the evaluation results of each dimension are weighted and fused to obtain the initial comprehensive evaluation result. The weight coefficients involved are set according to the task type, retrieval target or expert experience, or dynamically and adaptively adjusted through the performance of the validation set.

[0034] (3) Multi-round progressive recall: Based on the initial comprehensive evaluation results, multiple threshold intervals are set. In the first round, the candidate result set of the highest threshold interval is selected, and the query for enhanced retrieval is rewritten based on the first language model. Then, multi-path recall and multi-dimensional comprehensive evaluation are carried out in sequence, and some candidate results are recalled based on the comprehensive evaluation results. In subsequent rounds, based on the comprehensive evaluation results of the previous round, the candidate results that are semantically related, causally related, or contextually extended but not selected are selected from the comprehensive evaluation results of the initial comprehensive evaluation results. The query for enhanced retrieval is rewritten based on the first language model, and multi-path recall and multi-dimensional comprehensive evaluation are carried out in sequence, and some candidate results are recalled based on the comprehensive evaluation results. The candidate results recalled in each round are screened and verified using the reasoning ability of the first language model or external knowledge index. Through the multi-round progressive recall mechanism, all recalled candidate results constitute the final retrieval information.

[0035] Step 3: Fine-tuning based on deep thinking and reasoning.

[0036] In this embodiment of the invention, based on the thought chain reasoning mechanism and the reasoning-time expansion strategy, the final retrieved information is semantically understood and relevance-ranked to obtain the rearranged information.

[0037] The above-mentioned solution provided by the embodiments of the present invention has the following main advantages: It proposes a unified framework scheme integrating retrieval enhancement, reasoning expansion, and deep ranking; firstly, it utilizes a large language model to perform semantic parsing and dynamic correction of task queries, realizing multi-step reasoning and adaptive optimization of query semantics; secondly, it designs an incremental multi-path recall system based on heterogeneous association signal fusion, fully integrating keywords, semantics, and structured features through a joint mechanism of multi-path, multi-perspective, and multi-round progressive recall to achieve high-coverage and high-precision candidate information recall; finally, it proposes a fine-ranking scheme based on deep thinking (thought chain) and reasoning-time expansion, performing interpretable logical analysis and precise ranking of candidate information. Overall, the present invention can significantly improve the model's information understanding, reasoning depth, and retrieval accuracy under complex association relationships while maintaining computational efficiency, providing general technical support for highly reliable knowledge acquisition and scientific intelligent systems.

[0038] To more clearly demonstrate the technical solution and its effects provided by the present invention, the method provided by the embodiments of the present invention will be described in detail below with reference to specific examples.

[0039] I. Overall Introduction to the Plan.

[0040] The solution provided by the embodiments of the present invention mainly includes three parts: iterative query understanding based on retrieval enhancement, incremental multipath recall based on heterogeneous correlation signal fusion, and fine ranking based on deep thinking and reasoning extension.

[0041] 1. Iterative query understanding based on retrieval enhancement aims to perform deep semantic understanding and structured decomposition of task queries by introducing a multi-step reasoning mechanism and dynamic retrieval enhancement strategy, thereby gradually generating retrieval queries that better meet the needs of complex information. Its workflow mainly includes the following steps: (1) Task reasoning: Based on the original query of the task, semantic analysis and task intent recognition are performed using the first language model to deduce the potential information demand types and their logical structures. (2) Demand decomposition: For complex or multi-level query tasks, the first language model is used to decompose them into several sub-demands that can be solved independently, so as to gradually aggregate and reason. (3) Query generation: Based on the reasoning results and the semantics of the sub-demands, the first language model is used to generate the corresponding structured retrieval query, covering information such as keywords, semantic expressions and logical constraints. (4) Retrieval enhancement and further iteration: Candidate information is obtained by using external retrieval systems (such as vector databases or document retrieval engines), and the retrieval results are fed back to the first language model for semantic verification and adaptive updates, iteratively optimizing the query expression and retrieval scope until the preset information demand is met. Through the aforementioned enhanced iterative query understanding process, this invention enables dynamic interactive optimization between query and knowledge, allowing the information acquisition process to possess multi-step reasoning and semantic self-correction capabilities, thereby significantly improving the accuracy of information retrieval and the efficiency of knowledge acquisition under complex relationships.

[0042] 2. Heterogeneous Relational Signal Fusion and Incremental Multipath Recall: This system aims to construct a retrieval framework that balances recall diversity and accuracy in response to the multi-source heterogeneous signals and cross-semantic association problems in complex information needs. By fusing multi-dimensional relational signals such as keywords, semantics, and structured features, this system achieves hierarchical exploration and gradual expansion of the knowledge space. Its main working mechanism includes the following three levels: (1) Multi-path recall: Multiple query rewriting paths are generated by sampling multiple times using the first language model, and independent recall is performed under different keyword representations and semantic spaces. This mechanism can jointly utilize keyword features and semantic information, thereby effectively improving the diversity and coverage of retrieval results and avoiding the omission of relevant information due to a single query expression. (2) Multi-perspective fusion: Keyword matching scores and semantic similarity scores are weighted and fused, and the two types of information sources, namely word similarity and semantic relevance, are comprehensively considered to achieve more accurate relevance assessment. This fusion strategy can establish complementary relationships between multi-dimensional signals, thereby maintaining the stability and reliability of information acquisition in complex relational contexts. (3) Multi-round progressive recall: A multi-round progressive recall strategy is adopted, first using high-confidence recall results as the core for initial matching, and then gradually expanding to low-confidence or indirectly related information areas. Through this layer-by-layer progressive recall method, the system can improve the completeness and robustness of knowledge retrieval while ensuring high relevance. Through the above-mentioned synergistic mechanism of multi-path, multi-perspective and multi-round progressive recall, the heterogeneous correlation signal fusion progressive multi-path recall system of the present invention can achieve high coverage and high-precision retrieval of complex information space, providing a more diverse and reliable candidate information foundation for subsequent multi-step reasoning and result integration modules.

[0043] 3. Based on deep thinking and reasoning-time extension, the fine ranking aims to achieve deep semantic understanding and high-precision relevance ranking of candidate information by introducing a thinking chain reasoning mechanism and reasoning-time extension strategy. Compared with the traditional ranking method that relies on shallow semantic matching, this module can perform multi-step logical deduction and criterion-driven decision optimization in the reasoning stage, thereby effectively improving the accuracy and reliability of complex information acquisition tasks. Its core process includes the following four technologies: (1) Criterion-based relevance point estimation: First, based on the preset semantic and logical criteria, the second language model is used to estimate the relevance points of the query and candidate documents one by one. This process combines explicit semantic similarity and implicit logical consistency evaluation to achieve relevance judgment and provide interpretable scoring basis. (2) Thinking chain reasoning: In the process of performing criterion-based relevance point estimation, a thinking chain reasoning mechanism is introduced to enhance the logical consistency and semantic interpretation ability of the second language model in scoring. This mechanism guides the second language model to generate a multi-step reasoning process, enabling it to conduct structured thinking before giving a score, and to deduce the semantic association and logical dependency between candidate documents and query tasks layer by layer, thereby obtaining more accurate and interpretable relevance evaluation results. (3) Reasoning-time expansion: The module dynamically expands the number of thought chains and reasoning depth during the reasoning stage, and selects the most logically reasonable and consistent conclusions from multiple reasoning paths through a consistency mechanism, thereby enhancing the reasoning depth and discrimination accuracy of the second language model in complex semantic tasks. (4) Reinforcement learning: To further optimize the reasoning and ranking performance of the second language model, a training mechanism based on reinforcement learning is introduced. By using user feedback, ranking accuracy, or logical consistency as reward signals, the second language model can continuously adjust its reasoning strategy and decision criteria, and gradually form an adaptive optimization capability for complex reasoning tasks. Through the above mechanisms, the deep thinking and reasoning-time expansion fine ranking of this invention realizes the leap from shallow semantic matching to deep logical reasoning, significantly improving the accuracy, interpretability, and robustness of information ranking, and providing efficient and reliable reasoning support for the acquisition of information with complex relationships.

[0044] II. Detailed introduction of the plan.

[0045] like Figure 2 The diagram shown is a schematic representation of the overall working process of the above-described solution of the present invention. The three parts of the present invention will be further described below with reference to the accompanying drawings.

[0046] 1. Iterative query understanding based on retrieval enhancement.

[0047] This section aims to achieve multi-step reasoning and adaptive query optimization for complex information needs by combining semantic analysis, task understanding, and dynamic retrieval feedback. For example... Figure 3The diagram illustrates an iterative query understanding scheme based on retrieval enhancement. First, the primary language model is used to perform semantic understanding and structured parsing of the input original task query, identifying the query intent, information need type, and potential logical dependencies. For complex tasks containing multiple constraints or multi-layered semantics, automatic requirement decomposition is performed through reasoning, dividing the overall query into several sub-tasks or sub-requirements for subsequent step-by-step reasoning and aggregation. Then, based on the analysis results, an initial retrieval query is generated, integrating keywords, semantic features, and logical constraints to form a structured retrieval expression.

[0048] Building upon this foundation, the core innovation of this invention lies in its enhanced retrieval and iterative retrieval mechanism. Specifically, an external retrieval engine (e.g., a vector database, a hybrid retrieval system, or a document index) can be used to execute queries and obtain preliminary candidate results. After receiving the retrieval results, the first language model not only assesses their surface relevance but also performs semantic verification and reasoning analysis to identify insufficient information coverage, expression biases, or logical gaps. Based on this, the first language model modifies or expands the query semantics, for example, by adding synonyms, adjusting keyword weights, introducing hypernyms or hyponyms, or supplementing reasoning conditions, thereby regenerating new query expressions.

[0049] Preferably, in each iteration, the query content and retrieval scope are adaptively optimized by combining the retrieval results from the previous round with the inference output of the primary language model, gradually improving the relevance and completeness of the results. Simultaneously, a reinforcement learning-based strategy optimization mechanism is introduced. During iteration, reward signals are calculated based on retrieval feedback results to guide the primary language model in optimizing its query generation and correction rules. Specifically, reward functions can be set based on indicators such as the ranking accuracy of candidate results, the magnitude of relevance improvement, or the improvement in information coverage, positively incentivizing the primary language model's query generation strategy, thereby prompting the primary language model to automatically learn more efficient query construction and expansion methods. In some scenarios, user feedback or task completion can also be used as additional reward signals to further enhance the adaptability of the query strategy to real-world task requirements.

[0050] Based on the dynamic cyclic mechanism of retrieval-feedback-correction-re-retrieval provided in the embodiments of the present invention, combined with reinforcement learning-driven policy updates, semantic self-correction and intelligent optimization of complex tasks can be achieved without significantly increasing computational costs. This enables the information acquisition process to have continuous learning and evolution capabilities, thereby significantly improving the accuracy of information retrieval and task satisfaction in complex relational scenarios.

[0051] 2. Incremental multipath recall through heterogeneous correlation signal fusion.

[0052] like Figure 4The diagram illustrates a multi-path recall system based on heterogeneous correlation signal fusion. It mainly comprises the following three stages: (1) Multi-path recall.

[0053] In this embodiment of the invention, the generation and rewriting capabilities of the first large language model are utilized to sample the input query multiple times, generating several semantically diverse query rewriting paths. Each rewriting path independently performs a recall operation under different keyword representations, semantic subspaces, or contextual concerns, forming multiple candidate result sets.

[0054] For example, multiple recall channels can be employed in parallel to enhance the diversity and comprehensiveness of the results: on the one hand, keyword-level retrieval is performed using traditional inverted index algorithms such as BM25 and TF-IDF to ensure high-precision matching of explicitly relevant text; on the other hand, dense vector retrieval is performed by combining semantic retrieval models based on vector representation (such as BGE-M3, GTE, Contriever, etc.) to capture deep semantic associations and implicit contextual relationships. Through this multi-path, multi-channel joint mechanism, the risk of information omission can be significantly reduced while ensuring recall diversity, providing a rich candidate base for subsequent multi-perspective fusion and multi-tiered recall.

[0055] (2) Multi-perspective fusion.

[0056] In this embodiment of the invention, in order to take into account the complementarity and consistency between different retrieval signals, a multi-perspective fusion mechanism is introduced on the basis of the candidate results obtained by multi-path recall to perform unified correlation evaluation and fusion ranking of the results from different retrieval channels.

[0057] Specifically, candidate documents are comprehensively evaluated from multiple dimensions, including keyword matching score, semantic similarity score, and contextual relevance. These multi-dimensional signals are then weighted and fused to obtain a comprehensive relevance index. ; Where R is the comprehensive correlation index. To score the keyword matching score Score semantic similarity. For contextual relevance, These are weighting coefficients, which can be set according to task type, retrieval target, or expert experience, or dynamically and adaptively adjusted based on validation set performance. In more complex implementations, a learning-based ranking fusion model can be introduced to jointly model keyword matching features, semantic similarity features, and contextual relevance, thereby achieving adaptive nonlinear weighted fusion and fine-grained ranking optimization.

[0058] Through the above multi-perspective fusion mechanism, relevance and coverage can be taken into account at the lexical, semantic and contextual levels, effectively improving the matching accuracy and robustness under complex semantic tasks, and providing a high-quality candidate foundation for subsequent multi-tiered progressive recall and fine ranking modules.

[0059] (3) Multi-stage progressive recall.

[0060] In this embodiment of the invention, to ensure high relevance of search results while also considering knowledge coverage and robustness, a multi-tiered progressive recall mechanism is further introduced based on multi-path recall and multi-perspective fusion. This mechanism takes high-confidence results as the core and expands layer by layer to low-confidence or indirectly related candidate regions, thereby realizing a progressive information acquisition process from core matching to semantic expansion.

[0061] Specifically, firstly, based on the comprehensive relevance score obtained from multi-perspective fusion, multi-level threshold intervals (e.g., high confidence, medium confidence, and low confidence intervals) are set. In the first round, results from the high confidence interval are prioritized as the primary basis for precise matching and verification. In subsequent rounds, based on the semantic features or contextual clues of the previous round's results, the search scope is dynamically expanded, incorporating relevant information from semantically adjacent documents, potentially causally related documents, or citation chains. Furthermore, during each round of expansion, the reasoning capabilities of the primary language model or external knowledge indexes can be utilized to screen and verify newly added candidates, ensuring that the introduced information is logically consistent with the original query. Through this multi-round progressive recall mechanism, the recall scope can be gradually expanded while maintaining high accuracy, thereby effectively improving the completeness, robustness, and information association depth of knowledge retrieval in complex semantic tasks or cross-domain association scenarios.

[0062] 3. Refined sorting based on deep thinking and reasoning.

[0063] like Figure 5 The diagram illustrates a refined ranking scheme based on deep thinking and reasoning. The core components involved are described below.

[0064] (1) Criterion-based correlation point estimation.

[0065] In this embodiment of the invention, a set of relevance evaluation criteria is first established based on the task type and retrieval objective, such as information relevance, information completeness, timeliness, source authority, and logical consistency. In implementation, a point estimation mechanism can be used: that is, relevance scoring is performed independently for each candidate document. The inputs include the task query (Q), candidate document content (D), and the set of criteria (C). The task query can be the original query or a retrieval-enhanced query obtained through an iterative query understanding method based on retrieval enhancement. The candidate document content is the candidate result set obtained through the aforementioned heterogeneous association signal fusion incremental multipath recall system. These three elements are used as input prompts to the second language model, which generates corresponding relevance score vectors based on multidimensional criteria. Each component corresponds to a score for a criterion. Finally, a comprehensive score R1 for each candidate document is obtained through a weighted aggregation function R1=f(S), which serves as its relevance estimation result. Here, f(.) is the weighted aggregation function. This mechanism can explicitly quantify the fit between candidate information and query requirements and provide a visual evaluation basis.

[0066] (2) Reasoning by thought chain.

[0067] In criterion-based relevance estimation, a thought chain reasoning mechanism can be further introduced to perform multi-step logical expansion of candidate information before final evaluation. The second language model, during generation, not only judges the surface relevance of candidate documents but also infers their implicit logical dependencies and semantic support relationships through multi-step thought chains. For example, for scientific literature tasks, the second language model can deduce whether the document provides implicit elements such as experimental evidence, thereby generating a structured reasoning path. During this process, the second language model can continuously reflect and verify, examining whether the preceding reasoning is sufficient and the evidence is reasonable, achieving deep thinking and dynamic correction. Each thought chain can include four stages: premise identification, evidence extraction, logical judgment, and conclusion generation. The second language model can continuously repeat this process within the thought chain, reflecting and verifying, and revising preceding hypotheses or supplementing evidence at any time, thus completing the reasoning and evaluation of complex related information in an interpretable and traceable manner.

[0068] (3) Expand during reasoning.

[0069] During the reasoning phase, the number and depth of generated thought chains can be expanded in various ways to enhance the second language model's ability to analyze and judge complex semantic tasks. Specifically, multiple sampling, different prompt templates, or multi-model collaboration can be used to generate multiple thought chains with diverse reasoning paths. Simultaneously, by inserting suggestive transition words (such as "consider further," "another possibility," etc.) or setting a reasoning budget, the second language model can be guided to extend its thinking depth within a single chain. After obtaining multiple thought chains, a consistency mechanism is used to compare and aggregate the conclusions and reasoning paths of different chains, selecting the most logically sound and consistent result. This mechanism effectively improves the reasoning depth and discrimination accuracy of the second language model in tasks such as multi-layered semantic associations and complex causal judgments, achieving a balance between reasoning breadth and reliability.

[0070] (4) Strengthen learning.

[0071] To further optimize the overall performance of the second language model in the inference and ranking stages, this invention introduces an adaptive optimization mechanism based on reinforcement learning. User feedback, ranking accuracy, logical consistency, and inference stability are used as reward signals to continuously update and optimize the model's inference strategy and decision-making criteria. Specifically, during reinforcement learning, when the policy model samples and generates candidate outputs, it evaluates each candidate document output based on the aforementioned metrics. For example, ranking accuracy scores can be calculated based on the ranking results of positive and negative sample documents and used as reward signals to incentivize the policy model to generate better candidate ranking results. Simultaneously, user feedback and logical consistency evaluations can be combined to assign differentiated reward values ​​to different sampled outputs of the same candidate document, thereby guiding the second language model to gradually optimize its inference path and decision-making mode. Furthermore, when the second language model's confidence is low or the inference results are uncertain, it can further interact with the external environment to obtain supplementary information from existing knowledge bases or retrieval systems, forming new contextual comparison signals to enrich the model's judgment basis. Based on this, the second language model continuously adjusts its thought chain generation strategy, weight aggregation method, and inference expansion depth, gradually forming a self-evolving capability for complex inference tasks. This mechanism not only improves the model's generalization performance in diverse task scenarios, but also ensures the interpretability and consistency of the inference path and relevance evaluation process.

[0072] Those skilled in the art will understand that the first and second mainly serve to distinguish them, that is, to indicate that each of the above three parts is configured with a corresponding large language model and will be optimized according to the corresponding task; the specific structure of the large language model involved can be implemented by the user by selecting an existing model according to the actual situation, and the present invention does not impose specific limitations.

[0073] Through the above description of the embodiments, those skilled in the art can clearly understand that the above embodiments can be implemented by software, or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of the above embodiments can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.), including several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0074] Example 2 This invention also provides a multi-step reasoning information acquisition system, which is mainly used to implement the methods provided in the foregoing embodiments, such as... Figure 6 As shown, the system mainly includes: The iterative query understanding module based on retrieval enhancement is used to perform semantic understanding and structured decomposition of the original query of the task based on the first major language model through a multi-step reasoning mechanism and dynamic retrieval enhancement strategy, and gradually generate retrieval-enhanced queries. The heterogeneous correlation signal fusion incremental multipath recall module is used to recall multiple candidate result sets through multipath based on search enhancement query, and to comprehensively evaluate all candidate result sets from multiple dimensions to obtain an initial comprehensive evaluation result; based on the initial comprehensive evaluation result, a multi-round progressive recall strategy is adopted to gradually expand the search scope and obtain the final search information. The fine-ranking module based on deep thinking and reasoning extension is used to perform semantic understanding and relevance ranking on the final retrieved information based on the thinking chain reasoning mechanism and reasoning extension strategy, so as to obtain the reordered information.

[0075] Since the main technical details of this system have been described in previous embodiments, they will not be repeated here.

[0076] Those skilled in the art will understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above.

[0077] Example 3 The present invention also provides a processing device, such as Figure 7 As shown, it mainly includes: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method provided in the foregoing embodiments.

[0078] Furthermore, the processing device also includes at least one input device and at least one output device; in the processing device, the processor, memory, input device, and output device are connected via a bus.

[0079] In this embodiment of the invention, the specific types of the memory, input device, and output device are not limited; for example: Input devices can be touchscreens, image acquisition devices, physical buttons, or mice, etc. The output device can be a display terminal; The memory can be random access memory (RAM) or non-volatile memory, such as disk storage.

[0080] Example 4 The present invention also provides a readable storage medium storing a computer program that, when executed by a processor, implements the method provided in the foregoing embodiments.

[0081] In this embodiment of the invention, the readable storage medium is a computer-readable storage medium and can be disposed in the aforementioned processing device, for example, as a memory in the processing device. Furthermore, the readable storage medium can also be any medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0082] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. The information disclosed in the background section is intended only to enhance the understanding of the overall background technology of the present invention and should not be construed as an admission or implication in any way that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for acquiring information through multi-step reasoning, characterized in that, include: Through a multi-step reasoning mechanism and dynamic retrieval enhancement strategy, the original query of the task is semantically understood and structurally decomposed based on the first major language model, and retrieval-enhanced query is gradually generated. The query enhancement based on retrieval retrieves multiple candidate result sets through multi-path recall, and comprehensively evaluates all candidate result sets from multiple dimensions to obtain an initial comprehensive evaluation result; Based on the initial comprehensive evaluation results, a multi-round progressive recall strategy is adopted to gradually expand the search scope and obtain the final search information; Based on the thought chain reasoning mechanism and the reasoning-time expansion strategy, the final retrieved information is semantically understood and relevance-ranked to obtain the rearranged information.

2. The multi-step reasoning information acquisition method according to claim 1, characterized in that, The process involves using a multi-step reasoning mechanism and a dynamic retrieval enhancement strategy to perform semantic understanding and structured decomposition of the original query based on the first major language model, gradually generating retrieval-enhanced queries, including: The first major language model is used to perform semantic understanding and structured parsing on the original query of the task, and the query intent, information demand type and logical dependency relationship are identified as the analysis results. For tasks with multiple constraints or multiple semantic layers, automatic demand decomposition is performed through reasoning, dividing the overall task into several sub-tasks or sub-demands. Semantic understanding and structured parsing are then performed on each sub-task or sub-demand to obtain the corresponding analysis results. Finally, the initial query is generated based on the analysis results. The process involves executing an initial query using an external retrieval engine to obtain preliminary candidate results. These preliminary candidate results are then used to modify or expand the initial query, generating a new query. The external retrieval engine is then used again to obtain the corresponding candidate results. This process iterates continuously, with the primary language model evaluating in each iteration whether the current query reflects the intent of the original query and has retrieval effectiveness. If so, the iteration terminates, resulting in a retrieval-enhanced query.

3. The multi-step reasoning information acquisition method according to claim 2, characterized in that, The step of using preliminary candidate results to modify or expand the initial query and regenerate a new query includes: The first major language model is used to perform semantic verification and reasoning analysis on the preliminary candidate results. Based on the verification and analysis results, the expression or retrieval scope of the initial query is optimized, and a new query is generated.

4. A multi-step reasoning information acquisition method according to claim 2 or 3, characterized in that, Also includes: introduction Based on the reinforcement learning-based policy optimization mechanism, a reward signal is calculated based on the retrieved candidate results during the iteration process, which is used to guide the first language model to optimize its query generation strategy. Among them, a reward function is set based on the ranking accuracy of candidate results, the improvement of relevance, or the improvement of information coverage to positively incentivize the query generation strategy of the first language model. In some scenarios, user feedback or task completion is also used as additional rewards.

5. The multi-step reasoning information acquisition method according to claim 1, characterized in that, The retrieval-enhanced query obtains multiple candidate result sets through multi-path recall, including: Leveraging the generation and rewriting capabilities of the largest language model, multiple samplings are performed on the retrieval-enhanced query to generate several semantically diverse query rewriting paths. Each rewriting path independently performs a recall operation under different keyword representations, semantic subspaces, or contextual concerns to obtain multiple candidate result sets.

6. The multi-step reasoning information acquisition method according to claim 1, characterized in that, The comprehensive evaluation of all candidate result sets from multiple dimensions includes: A multi-perspective fusion mechanism is introduced to perform unified correlation evaluation and fusion of all candidate result sets; The candidate result set is comprehensively judged from multiple dimensions, including keyword matching score, semantic similarity score and contextual relevance; the evaluation results of each dimension are weighted and fused to obtain an initial comprehensive evaluation result. The weight coefficients involved are set according to the task type, retrieval target or expert experience, or dynamically and adaptively adjusted through the performance of the validation set.

7. The multi-step reasoning information acquisition method according to claim 1, characterized in that, The initial comprehensive evaluation results are used to employ a multi-round progressive recall strategy to gradually expand the search scope and obtain the final search information, including: Based on the initial comprehensive evaluation results, multi-level threshold intervals are set. In the first round, the candidate result set of the highest threshold interval is selected, and the query for enhanced retrieval is rewritten based on the first large language model. Then, multi-path recall and comprehensive evaluation of multiple dimensions are carried out in sequence, and some candidate results are recalled based on the comprehensive evaluation results. Subsequent rounds, based on the comprehensive evaluation results of the previous round, select candidate results that are semantically related, causally related, or contextually extended from the comprehensive evaluation results that are not selected in the range of the initial comprehensive evaluation results (excluding the highest threshold range). Continue to rewrite the search-enhanced query based on the first language model, and then perform multi-path recall and multi-dimensional comprehensive evaluation in sequence. Based on the comprehensive evaluation results, recall some candidate results. Each round of recall candidate results is screened and verified using the reasoning capabilities of the largest language model or external knowledge indexes; through a multi-round progressive recall mechanism, all recalled candidate results constitute the final retrieval information.

8. A multi-step reasoning information acquisition system, characterized in that, To implement the method according to any one of claims 1 to 7, comprising: The iterative query understanding module based on retrieval enhancement is used to perform semantic understanding and structured decomposition of the original query of the task based on the first major language model through a multi-step reasoning mechanism and dynamic retrieval enhancement strategy, and gradually generate retrieval-enhanced queries. The heterogeneous correlation signal fusion incremental multipath recall module is used to recall multiple candidate result sets through multipath based on retrieval enhancement query, and to comprehensively evaluate all candidate result sets from multiple dimensions to obtain an initial comprehensive evaluation result; based on the initial comprehensive evaluation result, a multi-round progressive recall strategy is adopted to gradually expand the retrieval scope and obtain the final retrieval information. The fine-ranking module based on deep thinking and reasoning extension is used to perform semantic understanding and relevance ranking on the final retrieved information based on the thinking chain reasoning mechanism and reasoning extension strategy, so as to obtain the reordered information.

9. A processing device, characterized in that, include: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A readable storage medium storing a computer program, characterized in that, When a computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images