Mind chain retrieval methods, devices, equipment and readable storage media

Abstract

This application discloses a method, apparatus, device, and readable storage medium for thinking chain retrieval. The method includes: obtaining keywords from a first question input by a user; searching for target nodes matching the keywords at the business scenario layer of a first knowledge graph, and filtering multiple first candidate paths based on the target nodes and user profiles; calculating the semantic similarity between each first candidate path and the first question, obtaining multiple second candidate paths with similarity higher than a first threshold, and determining K high-scoring paths from the multiple second candidate paths based on the node access frequency and time decay factor of each second candidate path; and performing variable anchor point identification and logical coherence verification based on the K high-scoring paths to obtain the final thinking chain. This invention rapidly locates multiple first candidate paths through business scenarios and user profiles, and combines semantic similarity and path optimization to achieve precise targeted retrieval, thereby improving the accuracy of thinking chain retrieval results.

Description

Technical Field

[0001] This application belongs to the field of dynamic retrieval technology, specifically relating to a mind chain retrieval method, device, equipment, and readable storage medium. Background Technology

[0002] In existing technologies, Chain-of-Thought (CoT) technology is an important means to improve the complex reasoning ability of Large Language Models (LLMs). It mainly improves the accuracy and interpretability of multi-step reasoning tasks by explicitly decomposing the problem-solving path (e.g., the step-by-step expression of "hypothesis → verification → conclusion").

[0003] The existing mind chain retrieval process relies on full-text matching or shallow semantic similarity calculation, which can only achieve coarse-grained mind chain recall and cannot meet the precise adaptation requirements in dynamic scenarios, resulting in poor retrieval results. Summary of the Invention

[0004] This application provides a method, apparatus, device, and readable storage medium for thinking chain retrieval, which solves the problem of poor retrieval results in existing thinking chain retrieval schemes.

[0005] In a first aspect, embodiments of this application provide a mind chain retrieval method, including:

[0006] Obtain keywords from the first question input by the user, where the first question is related to the first business scenario;

[0007] In the business scenario layer of the first knowledge graph, target nodes matching the keywords are searched, and based on the target nodes, multiple first candidate paths are filtered according to the user profile. The first knowledge graph is a knowledge graph composed of nodes related to business scenarios, nodes related to business decision logic, and nodes related to instance data. The nodes of the first knowledge graph have at least one edge relationship among logical order edge relationship, semantic association edge relationship, and context dependency edge relationship. The target node is a node related to the first business scenario.

[0008] Calculate the semantic similarity between each first candidate path and the first question to obtain multiple second candidate paths with similarity higher than a first threshold. Based on the node access frequency and time decay factor of each second candidate path, determine K high-scoring paths from the multiple second candidate paths, where K is an integer greater than or equal to 2.

[0009] Based on the K high-scoring paths, variable anchor point identification and logical coherence verification are performed to obtain the final thought chain.

[0010] In some embodiments, the method further includes:

[0011] The original thought chain text is segmented to obtain atomic logic nodes; the original thought chain text is the text formed by the large language model in the reasoning process of responding to user input questions in any business scenario;

[0012] The atomic logic nodes are categorized into business scenario layer, decision logic layer, and instance data layer, and multi-type edge relationships between nodes are constructed to form the first knowledge graph.

[0013] In some embodiments, the step of filtering multiple first candidate paths based on the target node and user profile includes:

[0014] Determine the decision logic layer in the first knowledge graph corresponding to the target node;

[0015] Using the user profile as a filtering condition, the paths of the decision logic layer corresponding to the target node are traversed to select multiple first candidate paths that match the user profile.

[0016] In some embodiments, determining K high-scoring paths from the plurality of second candidate paths based on the node access frequency and time decay factor of each second candidate path includes:

[0017] For each of the second candidate paths, the weight of each candidate node in the second candidate path is calculated based on the node access frequency and time decay factor of the second candidate path.

[0018] Based on the edge type, determine the edge weights between adjacent candidate nodes in the second candidate path;

[0019] The path importance score of the second candidate path is calculated based on the weights of the candidate nodes in the second candidate path and the edge weights between the adjacent candidate nodes.

[0020] The top K second candidate paths, ranked from highest to lowest score, are identified as the K high-scoring paths.

[0021] In some embodiments, calculating the path importance score of the second candidate path based on the weights of the candidate nodes in the second candidate path and the edge weights between adjacent candidate nodes includes:

[0022] Based on the weights of the candidate nodes in the second candidate path and the edge weights between the adjacent candidate nodes, the transition probability from the first supernode to the second supernode is determined, wherein a supernode is an ordered sequence of nodes composed of at least two candidate nodes in logical order;

[0023] Based on the transition probability from the first supernode to the second supernode, the path importance score of the second candidate path is calculated.

[0024] In some embodiments, the step of identifying variable anchor points and verifying logical coherence based on the K high-scoring paths to obtain the final thought chain includes:

[0025] The shared variables identified in the nodes of the two high-scoring paths are determined as anchor points, and the topology graph formed by all non-repeating nodes in the two high-scoring paths corresponding to the anchor points is determined as a splicing candidate pair. The two high-scoring paths are included in the K high-scoring paths.

[0026] For each splicing path in the candidate splicing pairs, the logical coherence of the splicing path is checked to obtain the check result;

[0027] If the verification result is unsuccessful, the splicing path is abandoned;

[0028] The splicing path with the highest path importance score among the spliced ​​paths that passed the verification and the high-scoring unspliced ​​paths is determined as the final thought chain.

[0029] In some embodiments, the logical coherence check of the splicing path includes:

[0030] Check whether there are explicit edge connections between the nodes of the splicing path;

[0031] If it exists, the verification is successful; otherwise, the semantic similarity between the nodes of the spliced ​​path is calculated.

[0032] If the semantic similarity is higher than the second threshold, virtual edges are generated between the nodes, and the path importance score of the splicing path with the newly added virtual edges is recalculated.

[0033] If the semantic similarity is lower than or equal to the second threshold, the verification is determined to fail.

[0034] Secondly, embodiments of this application also provide a mind chain retrieval device, comprising:

[0035] The acquisition module is used to acquire keywords from the first question input by the user, where the first question is related to the first business scenario;

[0036] The first processing module is used to find target nodes that match the keywords in the business scenario layer of the first knowledge graph, and based on the target nodes, filter out multiple first candidate paths according to the user profile. The first knowledge graph is a knowledge graph composed of nodes related to business scenarios, nodes related to business decision logic, and nodes related to instance data. The nodes of the first knowledge graph have at least one edge relationship among logical order edge relationship, semantic association edge relationship, and context dependency edge relationship. The target node is a node related to the first business scenario.

[0037] The second processing module is used to calculate the semantic similarity between each first candidate path and the first question, obtain multiple second candidate paths with similarity higher than a first threshold, and determine K high-scoring paths from the multiple second candidate paths based on the node access frequency and time decay factor of each second candidate path, where K is an integer greater than or equal to 2.

[0038] The third processing module is used to identify variable anchor points and verify logical coherence based on the K high-scoring paths to obtain the final thought chain.

[0039] Thirdly, embodiments of this application also provide a thought chain retrieval device, including a processor and a transceiver, wherein the transceiver receives and sends data under the control of the processor, and the processor is used to perform the following operations:

[0040] Obtain keywords from the first question input by the user, where the first question is related to the first business scenario;

[0041] In the business scenario layer of the first knowledge graph, target nodes matching the keywords are searched, and based on the target nodes, multiple first candidate paths are filtered according to the user profile. The first knowledge graph is a knowledge graph composed of nodes related to business scenarios, nodes related to business decision logic, and nodes related to instance data. The nodes of the first knowledge graph have at least one edge relationship among logical order edge relationship, semantic association edge relationship, and context dependency edge relationship. The target node is a node related to the first business scenario.

[0042] Calculate the semantic similarity between each first candidate path and the first question to obtain multiple second candidate paths with similarity higher than a first threshold. Based on the node access frequency and time decay factor of each second candidate path, determine K high-scoring paths from the multiple second candidate paths, where K is an integer greater than or equal to 2.

[0043] Based on the K high-scoring paths, variable anchor point identification and logical coherence verification are performed to obtain the final thought chain.

[0044] Fourthly, embodiments of this application also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the thought chain retrieval method described in the first aspect above.

[0045] Fifthly, embodiments of this application also provide a computer program product, including computer instructions, which, when executed by a processor, implement the steps in the mind chain retrieval method described in the first aspect above.

[0046] The above-described technical solution of the present invention has at least the following beneficial effects:

[0047] In this embodiment of the invention, keywords from a first question input by the user are obtained, and the first question is related to a first business scenario. Then, target nodes matching the keywords are searched at the business scenario layer of a first knowledge graph. Based on the target nodes, multiple first candidate paths are selected according to the user profile. The first knowledge graph is composed of nodes related to the business scenario, nodes related to business decision logic, and nodes related to instance data. The nodes in the first knowledge graph have at least one edge relationship among logical order edge relationships, semantic association edge relationships, and context dependency edge relationships. The target nodes are nodes related to the first business scenario. Next, each first candidate path is calculated. Based on the semantic similarity between the path and the first question, multiple second candidate paths with similarity higher than a first threshold are obtained. Then, based on the node access frequency and time decay factor of each second candidate path, K high-scoring paths are determined from these multiple second candidate paths, where K is an integer greater than or equal to 2. Finally, variable anchor point identification and logical coherence verification are performed based on the K high-scoring paths to obtain the final thought chain. Thus, through the above two-stage hybrid retrieval, namely the coarse-grained retrieval stage, multiple first candidate paths are quickly located based on business scenarios and user profiles. Then, the fine-grained retrieval stage, combining semantic similarity and path optimization, achieves precise targeted retrieval, thereby improving the accuracy of the thought chain retrieval results. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the mind chain retrieval method according to an embodiment of this application;

[0049] Figure 2 This is a schematic diagram of the modules of the mind chain retrieval device according to an embodiment of this application;

[0050] Figure 3 This is a schematic diagram of the hardware structure of the mind chain retrieval device according to an embodiment of this application. Detailed Implementation

[0051] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. 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.

[0052] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and are not used to describe a specified order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, not limited in number; for example, a first object can be one or more. Furthermore, in the specification and claims, "and" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0053] The existing implementation of thought chain technology generates the reasoning process as a linear text sequence and relies on natural language processing technology to output logical steps step by step. To optimize the practical application of this technology, there are typically two solutions. Solution one, the linear storage and vector retrieval scheme, involves treating the fully generated thought chain text (e.g., "User age 35 → High health insurance premium → Recommend high-coverage critical illness insurance") as a single continuous sequence. This is then encoded into a fixed-dimensional high-dimensional vector using a pre-trained language model and stored in a vector database. When a new user input arrives, the system also encodes it as a vector and matches it with historical vectors using cosine similarity or Euclidean distance calculations, recalling the most similar thought chains as candidate reuse objects.

[0054] Option 2 employs dynamic caching and reuse technology, which shortens the time overhead of model generation of thought chains by identifying high-frequency inference fragments and establishing a fast retrieval mechanism. The specific process is: cache fragment generation → cache pool management → key-value matching retrieval. Specifically, cache fragment generation involves analyzing the weight distribution of each step in the inference path during model generation of thought chains using an attention mechanism to identify frequently occurring logical units. Specifically, the weight matrix of the self-attention layer is used to statistically analyze the activation intensity of different token sequences, filtering out continuous fragments with activation values ​​higher than a threshold. Simultaneously, the fragments are standardized (e.g., variable generalization, replacing "age=35" with "age={X}") to generate reusable templated content. Cache pool management involves storing the selected fragments in the cache pool as key-value pairs, where the key is the semantic vector of the fragment, and the value is the original text fragment and its metadata (e.g., usage frequency, last access time). Key-value matching retrieval involves extracting key features from the question when a new user request arrives, encoding them into a query vector, and calculating similarity with all key vectors in the cache pool. The top-K candidate segments with similarity higher than a preset value are recalled, and their compatibility with the current context is verified by the rule engine. If the verification passes, the candidate segments are directly concatenated into the input prompt, guiding the model to skip repeated inference steps.

[0055] Existing thought chain storage formats primarily employ linear structures, limiting the expression of multi-hop logical relationships. They can only represent sequential reasoning relationships and cannot effectively represent complex, multi-level, multi-branch logical structures. For example, in insurance product recommendation scenarios, premium calculation may simultaneously rely on multiple parallel reasoning paths such as user age, occupational risk, and regional policies. Linear storage struggles to express the intertwined relationships of these paths, leading to inaccurate location of related sub-logical units during retrieval. Furthermore, linear structures lack sufficient support for cross-level semantic associations (such as the mapping between business rules and specific clauses), resulting in inefficient reuse of historical reasoning experience.

[0056] Existing retrieval processes rely on full-text matching or shallow semantic similarity calculations, which can only achieve coarse-grained recall of thought chains and cannot meet the precise adaptation needs of dynamic scenarios. For example, when a user inputs "recommendation of medical insurance for high-risk occupations," traditional methods may recall a complete thought chain containing the tag "occupational risk," but cannot remove redundant reasoning steps that are irrelevant to the current user's age or region (such as premium calculation formulas for specific age groups). At the same time, shallow semantic similarity calculations are easily affected by differences in expression (such as the difference between the terms "critical illness insurance" and "major disease insurance"), leading to a mismatch between search results and actual problem requirements.

[0057] To address the aforementioned technical problems, this application provides a method, apparatus, device, and readable storage medium for thinking chain retrieval. The method and apparatus are based on the same concept, and since the principles underlying the problems solved by the method and apparatus are similar, their implementations can be mutually referenced; repeated details will not be elaborated further.

[0058] like Figure 1 The diagram shown is a flowchart illustrating the mind chain retrieval method according to an embodiment of this application. The method may include:

[0059] Step 101: Obtain keywords from the first question input by the user, where the first question is related to the first business scenario;

[0060] Here, keywords (such as "medical insurance" and "chronic disease history") in the first question input by the user can be extracted using the Term Frequency-Inverse Document Frequency (TF-IDF) algorithm.

[0061] It should be noted that the first business scenario refers to the business scenario under the target application scenario corresponding to the large language model. Among them, the target application scenario can be financial marketing scenario, customer service and intelligent question answering scenario, content creation and copywriting scenario, education and learning scenario, medical and health scenario, travel and tourism scenario, etc.

[0062] Step 102: In the business scenario layer of the first knowledge graph, find the target node that matches the keyword, and based on the target node, filter out multiple first candidate paths according to the user profile. The first knowledge graph is a knowledge graph composed of nodes related to business scenarios, nodes related to business decision logic, and nodes related to instance data. The nodes of the first knowledge graph have at least one edge relationship among logical order edge relationship, semantic association edge relationship, and context dependency edge relationship. The target node is a node related to the first business scenario.

[0063] It should be noted that the first knowledge graph is pre-generated and is a knowledge graph specific to the target application scenario corresponding to the large language model. This first knowledge graph consists of nodes related to the business scenario, nodes related to business decision-making logic, and nodes related to instance data.

[0064] The nodes related to the business scenario can include scenario tags and general rules. Here, scenario tags can be medical insurance tags, life insurance tags, and general rules can be general recommendation logic.

[0065] Nodes related to business decision-making logic can store specific reasoning rules (such as premium calculation formulas and risk assessment models) and support parameterized variables (such as {age} and {income}).

[0066] Nodes related to instance data can record specific data such as terms and conditions and historical cases, and can be dynamically bound to the aforementioned nodes related to business decision-making logic.

[0067] The nodes of the first knowledge graph have at least one of the following edge relationships: logical sequence edge relationship, semantic association edge relationship, and context dependency edge relationship. Here, logical sequence edges describe the temporal dependencies of the reasoning process, such as "age analysis → premium assessment". Semantic association edges link similar nodes across layers, such as the "critical illness insurance recommendation" node (belonging to a node related to the business scenario) and the "cancer protection clause" node (belonging to a node related to instance data). Context dependency edges mark conditional triggering relationships, such as "age greater than 50" triggering the "premium increase" rule.

[0068] For example, the user's first input question is: "Recommend insurance for a 45-year-old diabetic patient." The keywords extracted from this first question include "diabetic patient" (corresponding to a chronic disease history) and "insurance" (corresponding to medical insurance); that is, the first business scenario involved in the first question is medical insurance. Then, the business scenario layer of the first knowledge graph is searched for target nodes that match the keywords in the first question. These target nodes are nodes related to the medical insurance scenario. After finding the target nodes, multiple first candidate paths are filtered according to the user profile (e.g., age range: "45 years old", medical history type: "diabetic patient"). Here, these multiple first candidate paths can be understood as problem-solving paths matching the user profile (also known as candidate thought chains).

[0069] By executing steps 101 and 102, coarse-grained mind chain retrieval matching user profiles was achieved. Subsequent steps 103 and 104 involve precise targeted retrieval of mind chains by combining semantic similarity and path optimization.

[0070] Step 103: Calculate the semantic similarity between each first candidate path and the first question to obtain multiple second candidate paths with similarity higher than the first threshold. Based on the node access frequency and time decay factor of each second candidate path, determine K high-scoring paths from the multiple second candidate paths, where K is an integer greater than or equal to 2.

[0071] It should be noted that among the multiple first candidate paths obtained after step 102, there may be first candidate paths that have semantic deviations from the first question input by the user. Therefore, by executing step 103, the semantic similarity between each first candidate path and the first question is calculated, and multiple first candidate paths with similarity higher than a first threshold are selected as multiple second candidate paths, while first candidate paths with significant semantic deviations from the first question are removed. Then, for each second candidate path, a path importance score is calculated using the PageRank algorithm, which incorporates node access frequency and time decay factor. Finally, K high-scoring paths are selected from the multiple second candidate paths, that is, high-frequency and most recent problem-solving paths (i.e., high-scoring paths) are selected based on node access frequency and time decay factor.

[0072] Step 104: Based on the K high-scoring paths, perform variable anchor point identification and logical coherence verification to obtain the final thought chain.

[0073] To maintain the logical rigor of the reasoning process, variable anchor point identification and logical coherence verification are performed on high-scoring paths. This allows for the logical assembly of discrete but semantically related paths to form a complete reasoning chain. This improves the reusability of the thought chain and avoids global reasoning failures caused by the failure of local segments. Finally, the path with the highest path importance score among the verified high-scoring paths is determined as the final thought chain.

[0074] Here, steps 103 and 104 involve precise targeted retrieval of thought chains by combining semantic similarity and path optimization. The method described in this application can improve the accuracy of retrieval results for thought chain searches.

[0075] The thought chain retrieval method of this application embodiment obtains keywords from a first question input by the user, the first question being related to a first business scenario; then, it searches for target nodes matching the keywords in the business scenario layer of a first knowledge graph, and based on the target nodes, filters out multiple first candidate paths according to the user profile. The first knowledge graph is a knowledge graph composed of nodes related to business scenarios, nodes related to business decision logic, and nodes related to instance data. The nodes of the first knowledge graph have at least one edge relationship among logical order edge relationship, semantic association edge relationship, and context dependency edge relationship. The target node is a node related to the first business scenario; then, it calculates each of the first... The semantic similarity between a candidate path and the first question is used to obtain multiple second candidate paths with similarity higher than a first threshold. Based on the node access frequency and time decay factor of each second candidate path, K high-scoring paths are determined from the multiple second candidate paths, where K is an integer greater than or equal to 2. Finally, variable anchor point identification and logical coherence verification are performed on the K high-scoring paths to obtain the final thought chain. In this way, through the above two-stage hybrid retrieval, i.e., the coarse-grained retrieval stage, multiple first candidate paths are quickly located through business scenarios and user profiles. Then, the fine-grained retrieval stage is entered, which combines semantic similarity and path optimization to achieve precise targeted retrieval and improve the accuracy of the retrieval results of the thought chain retrieval.

[0076] As can be seen from the above, the first knowledge graph is generated in advance. So how to obtain the first knowledge graph can be achieved in the following way.

[0077] As an optional implementation, the method of this application also includes:

[0078] The original thought chain text is segmented to obtain atomic logic nodes; the original thought chain text is the text formed by the large language model in the reasoning process of responding to user input questions in any business scenario;

[0079] It should be understood that the original thought chain text comprises multiple parts. Here, these multiple original thought chain texts include various texts generated by the large language model during the reasoning process in response to user input questions across different business scenarios.

[0080] Specifically, natural language processing (NLP) technology can be used to segment the original thought chain text into logical steps. Connectives (such as "therefore" and "however") and sentence structures (such as conditional clauses and concluding statements) can be used to identify reasoning breakpoints, breaking the text down into atomic logical nodes. For example, the statement "User age 35 → High health insurance premium → Recommendation of high-coverage critical illness insurance" can be decomposed into three independent nodes: age analysis, premium assessment, and product matching. Each node needs further annotation of contextual dependencies (such as "preconditions" and "constraints"), extraction of keywords (such as "age," "premium," and "critical illness insurance"), and generation of semantic vectors to form searchable metadata tags.

[0081] The atomic logic nodes are categorized into business scenario layer, decision logic layer, and instance data layer, and multi-type edge relationships between nodes are constructed to form the first knowledge graph.

[0082] Here, the business scenario layer can be used to define insurance product types (such as medical insurance and life insurance) and general recommendation logic. Nodes related to business scenarios can include scenario tags and general rules.

[0083] The decision logic layer, where each node can be used to store specific reasoning rules (such as premium calculation formulas and risk assessment models), supports parameterized variables (such as {age} and {income}).

[0084] The instance data layer contains nodes that can record specific data such as terms and conditions, historical cases, etc., and can be dynamically bound to the nodes related to the business decision-making logic mentioned above.

[0085] Subsequently, nodes within a layer and nodes between layers form the first knowledge graph through various types of edge relationships. These various edge relationships include logical sequence edges, semantic association edges, and context-dependent edges. Here, logical sequence edges describe the temporal dependencies in the reasoning process, such as "age analysis → premium assessment." Semantic association edges link similar nodes across layers, such as the "critical illness insurance recommendation" node, which belongs to a node related to the business scenario, and the "cancer protection clause" node, which belongs to a node related to instance data. Context-dependent edges mark conditional triggering relationships.

[0086] Here, the graph structure of the first knowledge graph can be persisted using the Neo4j graph database. Each node stores the content hash value (to prevent duplication), semantic vector, access frequency, and timestamp, supporting subsequent weight adjustments. This application achieves multi-dimensional abstraction and association of reasoning paths through a layered graph model (business scenario layer, decision logic layer, and instance data layer) and multi-type edge relationships (logical order edges, semantic association edges, and context dependency edges). For example, in the insurance recommendation scenario, the "medical insurance recommendation" node in the business scenario layer can be simultaneously associated with the "chronic disease premium calculation rules" node in the decision logic layer and the "diabetes-specific clauses" node in the instance data layer, forming a three-dimensional knowledge graph. Experiments show that compared with traditional linear storage, the multi-hop association query of the graph structure of the first knowledge graph can improve retrieval efficiency by 3-5 times and can capture implicit logical links (such as discovering the indirect association between "high-risk occupations" and "accident insurance clauses" through semantic edges), improving the recall rate of implicit relationships and the logical coverage of complex reasoning scenarios.

[0087] In some embodiments, in step 102 above, based on the target node, multiple first candidate paths are selected according to the user profile, including:

[0088] Determine the decision logic layer in the first knowledge graph corresponding to the target node;

[0089] Using the user profile as a filtering condition, the paths of the decision logic layer corresponding to the target node are traversed to select multiple first candidate paths that match the user profile.

[0090] In this embodiment, a hierarchical pruning strategy is adopted to retain only nodes that are strongly related to the first business scenario, thus avoiding the overhead of traversing the entire graph.

[0091] It should be noted that in the node attributes of the decision logic layer, label fields associated with user profiles are predefined, such as "Applicable age range: [30-50]" and "Supported medical history types: [diabetes, hypertension]". Then, when the first question carries a user profile (e.g., "diabetes history", "age 45"), the system first searches for target nodes (e.g., the "medical insurance" node) that match the keywords of the first question in the business scenario layer of the first knowledge graph. Subsequently, a path traversal is initiated to the decision logic layer corresponding to the target node. During the traversal, filtering conditions based on user profile labels can be dynamically added using Cypher query statements. The specific generation logic is as follows: the user profile carried by the first question (e.g., "diabetes history", "age 45") is converted into a node attribute matching condition expression. For example, when the user profile contains "diabetes history", the WHEREnode.disease_type CONTAINS "diabetes" condition clause is automatically generated; when an age value is carried, the node.min_age<= 45 AND node.max_age >= 45 range matching condition is generated. In this way, multiple filtering conditions are connected by the logical AND operator to form a complete attribute filtering expression embedded in the Cypher query, which ultimately filters out multiple first candidate paths that match the user profile.

[0092] The above-mentioned hierarchical pruning adopts depth-first search (DFS) and pruning threshold control: if a node in a sub-path fails to meet the label condition three times in a row, the traversal of that branch is terminated to avoid invalid search.

[0093] The above calculation of the semantic similarity between each first candidate path and the first question to obtain multiple second candidate paths with similarity higher than the first threshold may specifically include: for each first candidate path, the Faiss vector index library can be used to calculate the semantic similarity between the semantic vector of each candidate node of the first candidate path and the semantic similarity between the semantic vector of each candidate node of the first candidate path and the semantic similarity between the semantic vector of each candidate node of the first candidate path and the semantic similarity between the semantic vector and the first question are all higher than the first threshold (e.g., 0.85), then the first candidate path is determined as a second candidate path; otherwise, the first candidate path is eliminated.

[0094] As an optional implementation, in step 103 above, based on the node access frequency and time decay factor of each second candidate path, K high-scoring paths are determined from the plurality of second candidate paths, including:

[0095] For each of the second candidate paths, the weight of each candidate node in the second candidate path is calculated based on the node access frequency and time decay factor of the second candidate path.

[0096] Specifically, the dynamic weight (DW) of each candidate node is composed of the product of the base weight and the decay factor (decay_score):

[0097]

[0098] in, ;

[0099] Here, base_weight is calculated logarithmically to avoid zero-frequency / high-frequency interference; For T i The frequency of access at any given time; λ is the time decay coefficient (λ=0.01 means 1% decay per day). Timestamp unit: t current t represents the current timestamp. last_accessed This represents the timestamp of the last access. The timestamp is in days (e.g., January 1, 2024 corresponds to an integer number of days), ensuring that the business rule update cycle (such as premium adjustments, product listing / delisting) is aligned with the decay rhythm. In this application, the node access frequency (logarithmically smoothed) and the time decay factor (exponential decay in days) are combined into the node's dynamic weight, replacing the fixed weight of the traditional algorithm. This allows for responsiveness to the timeliness of business needs.

[0100] Based on the edge type, determine the edge weights between adjacent candidate nodes in the second candidate path;

[0101] The edge types include logically ordered edges, semantically related edges, and context-dependent edges. It should be noted that the edge weight (EW) between adjacent candidate nodes in the second candidate path is set according to the edge type; for example, logically ordered edges have EW=1.0, and semantically related edges have EW=0.8, reflecting the connection strength.

[0102] The path importance score of the second candidate path is calculated based on the weights of the candidate nodes in the second candidate path and the edge weights between the adjacent candidate nodes.

[0103] Specifically, the path importance score of the second candidate path can be calculated using the PageRank algorithm based on the weights of the candidate nodes in the second candidate path and the edge weights between adjacent candidate nodes.

[0104] The top K second candidate paths, ranked from highest to lowest score, are identified as the K high-scoring paths.

[0105] In an optional embodiment, calculating the path importance score of the second candidate path based on the weights of the candidate nodes in the second candidate path and the edge weights between adjacent candidate nodes may specifically include:

[0106] Based on the weights of the candidate nodes in the second candidate path and the edge weights between the adjacent candidate nodes, the transition probability from the first supernode to the second supernode is determined, wherein a supernode is an ordered sequence of nodes composed of at least two candidate nodes in logical order;

[0107] Based on the transition probability from the first supernode to the second supernode, the path importance score of the second candidate path is calculated.

[0108] In this embodiment, the logical coherence between paths is ensured by abstracting the reasoning path (i.e., the ordered sequence of nodes) into supernodes and defining transition probabilities based on the overlapping parts (shared nodes, or overlapping nodes) and their edge weights. This embodiment can use a normalized weighted formula to calculate the first supernode P. i To the second supernode P j The transition probability allows path sorting to simultaneously respond to business popularity, timeliness, and connection strength.

[0109] Specifically, the reasoning path is abstracted into supernodes. If the first supernode P i The suffix and the second supernode P j Prefix overlap, i.e., the first supernode P i The suffix and the second supernode P j They have common parts. Among them, P i The suffix is ​​the set of all subpaths containing the last node of the current path; P j The prefix is ​​the set of all subpaths that include the first node of the current path. For example, if P i =[A,B,C,D],P j If the coordinates are [C,D,E,F], then the overlapping portion is [C,D]. This means that the first supernode P has been traversed. i At that time, the last two steps are C→D, and the second supernode P j The first two steps are exactly C→D. The overlapping part (i.e., the overlapping nodes) makes it possible to start from P. i Transition to P j It became possible.

[0110] Among them, from the first supernode P i To the second supernode P j The transition probability M ij The transition probability is allocated based on the ratio of the weight of the overlapping node to the sum of the edge weights. The calculation formula is as follows:

[0111]

[0112] Among them, M ij Indicates the first supernode P i To the second supernode P j The transition probability T k Indicates the first supernode P i With the second supernode P j The k-th node in a series of overlapping nodes; Out(P i ) is from the first supernode P i The set of supernodes that can be transferred to (i.e., possibly related to P) i A set of supernodes with overlapping suffixes); T n Indicates the first supernode P i With supernode P m The nth node in a series of overlapping nodes.

[0113] In this formula, the denominator represents the value from the first supernode P. i Transfer to all possible subsequent supernodes P m The weight sum can also be understood as: starting from the first supernode P i Transfer to all possible subsequent supernodes P m The sum of connection strengths. The inner summation represents the summation for each possible subsequent supernode P. m Calculate its relationship with the first supernode P. i The connection strength of the overlapping part; the outer layer summation means that the continuous strengths of all subsequent supernodes are accumulated to obtain the total strength.

[0114] Molecular representation from the first supernode P i Transfer to the second supernode P j The connection strength is determined by the dynamic weight DW(T) of the overlapping nodes. k Edge weights of overlapping edges The sum of products is represented by DW(T). k Reflecting node T k The importance of edge weights is emphasized in this application, where the (dynamic) weight of a node incorporates its access frequency and a time decay factor, meaning nodes that are accessed frequently recently have a higher weight. Indicates the node T k To T k+1 The transition strength can be the weight of logically ordered edges or the weight of semantically related edges. The summation process corresponding to the numerator is: traversing all continuous edges in the overlapping part (i.e., each pair of adjacent nodes T) k and T k+1 Multiply the edge weight of each edge by the starting node T. k The weights are calculated and accumulated. The larger this value, the stronger the first supernode P. iWith the second supernode P j The tighter the overlap, the better the connection from the first supernode P. i Transfer to the second supernode P j The greater the likelihood, the higher the probability.

[0115] After obtaining the transition probability from the first supernode to the second supernode, the PageRank algorithm (i.e., through power iteration) is used to calculate the PageRank value of each supernode. The sum of the PageRank values ​​is then used to obtain the path importance score of the corresponding second candidate path.

[0116] In some embodiments, step 104 above, which involves identifying variable anchor points and verifying logical coherence based on the K points to obtain the final thought chain, may specifically include:

[0117] Step 1041: The shared variables identified in the nodes of the two high-scoring paths are determined as anchor points, and the topology graph formed by all non-repeating nodes in the two high-scoring paths corresponding to the anchor points is determined as a splicing candidate pair. The two high-scoring paths are included in the K high-scoring paths.

[0118] For example, if the high-scoring path P i =[A,B,C,D] and high-scoring path P j =[E,B,C,F] has an overlapping part [B,C], where [B,C] is the identified shared variable, i.e., the anchor point. After concatenation, a new topology graph G is obtained, consisting of nodes [A,B,C,D,E,F]. Topology graph G is the concatenation candidate pair. In topology graph G, besides P... i and P j In addition, new candidate splicing paths will be generated, such as splicing path P. m =[A,B,C,F], or P n =[E,B,C,D].

[0119] Here, shared variables are identified in the nodes of high-scoring paths as splicing hubs (i.e., anchor points) to generate splicing candidate pairs. Specifically, parameterized variables (such as "premium calculation result: {premium}") of the input / output nodes of candidate sub-paths are extracted, and a variable mapping table across sub-paths is established. For example, if the output node of sub-path A contains the variable {premium=5000}, and the input node of sub-path B requires {premium} as a parameter, the system will automatically create a variable binding relationship and use this parameter as an anchor point to generate splicing candidate pairs.

[0120] Step 1042: For each splicing path in the candidate splicing pair, perform a logical coherence check on the splicing path to obtain the check result;

[0121] As an optional implementation, the logical coherence of the spliced ​​path is checked to obtain the check result, including:

[0122] Check whether there are explicit edge connections between the nodes of the splicing path;

[0123] If it exists, the verification is successful; otherwise, the semantic similarity between the nodes of the spliced ​​path is calculated.

[0124] If the semantic similarity is higher than the second threshold, virtual edges are generated between the nodes, and the path importance score of the splicing path with the newly added virtual edges is recalculated.

[0125] If the semantic similarity is lower than or equal to the second threshold, the verification is determined to fail.

[0126] In this implementation, the logical coherence check is divided into three stages.

[0127] Specifically, the first stage is structural compliance verification. This can be done by using Cypher queries to traverse the connections between nodes in the concatenated path and check for explicit logical sequential edges or semantic edges. For example, verify if there is a direct edge connecting the "premium calculation" node in sub-path A and the "deductible analysis" node in sub-path B. If so, the verification passes. Otherwise, proceed to the second stage.

[0128] The second stage involves implicit relationship reasoning. When an explicit edge is missing, a semantic similarity compensation mechanism is activated, which calculates the semantic similarity between nodes in the concatenated path (e.g., calculating the cosine similarity of semantic vectors between nodes). If the semantic similarity is higher than a second threshold (which can be set to 0.85), virtual edge generation between nodes is triggered. At this time, the system will automatically create a temporary semantic association edge (weight = 0.6) and record this implicit connection to the session context cache. If the semantic similarity is lower than or equal to the second threshold, virtual edge generation is abandoned, and the concatenated path is marked as invalid, i.e., the verification fails.

[0129] The third stage involves dynamic weight correction. Newly generated virtual edges are incorporated into the path weight calculation system, and the path importance score of the spliced ​​paths connected to the virtual edges is recalculated.

[0130] Step 1043: If the verification result is unsuccessful, then the splicing path is abandoned;

[0131] Step 1044: Among the spliced ​​paths that passed the verification and the high-scoring unspliced ​​paths, the path with the highest path importance score is determined as the final thought chain.

[0132] It should be noted that after incorporating the newly generated virtual edges into the path weight calculation system and recalculating the path importance score of the spliced ​​path connected to the virtual edges, if the corrected spliced ​​path still maintains a high score ranking (e.g., remains in the top 3), the splicing operation is considered valid. If it does not maintain a high score ranking, the splicing operation is considered invalid, and the spliced ​​path is discarded. Finally, from the spliced ​​paths that have passed the logical coherence check and the high-scoring unspliced ​​paths, the path with the highest path importance score is selected as the final thought chain.

[0133] It should be noted that the final thought chain will be injected with a large model prompt, where the nodes and edges are rearranged according to the logical order, while virtual edge annotations are retained for subsequent optimization reference. This mechanism ensures improved resource reuse of high-scoring paths while maintaining the logical rigor of the reasoning process.

[0134] like Figure 2 As shown in the illustration, this application also provides a mind chain retrieval device, wherein the device may include:

[0135] The acquisition module 201 is used to acquire keywords from the first question input by the user, wherein the first question is related to the first business scenario;

[0136] The first processing module 202 is used to find target nodes that match the keywords in the business scenario layer of the first knowledge graph, and to filter out multiple first candidate paths according to the user profile based on the target nodes. The first knowledge graph is a knowledge graph composed of nodes related to business scenarios, nodes related to business decision logic, and nodes related to instance data. The nodes of the first knowledge graph have at least one edge relationship among logical order edge relationship, semantic association edge relationship, and context dependency edge relationship. The target node is a node related to the first business scenario.

[0137] The second processing module 203 is used to calculate the semantic similarity between each first candidate path and the first question, obtain multiple second candidate paths with similarity higher than a first threshold, and determine K high-scoring paths from the multiple second candidate paths based on the node access frequency and time decay factor of each second candidate path, where K is an integer greater than or equal to 2.

[0138] The third processing module 204 is used to identify variable anchor points and verify logical coherence based on the K high-scoring paths to obtain the final thought chain.

[0139] In some embodiments, the apparatus of this application further includes:

[0140] The fourth processing module is used to segment the original thought chain text to obtain atomic logic nodes; the original thought chain text is the text formed by the large language model in the reasoning process of responding to user input questions in any business scenario;

[0141] The fifth processing module is used to classify the atomic logic nodes according to the business scenario layer, decision logic layer, and instance data layer, and to construct multi-type edge relationships between nodes to form the first knowledge graph.

[0142] In some embodiments, the first processing module 202 includes:

[0143] The first processing unit is used to determine the decision logic layer in the first knowledge graph corresponding to the target node;

[0144] The second processing unit is used to traverse the paths of the decision logic layer corresponding to the target node using the user profile as a filtering condition, and filter out multiple first candidate paths that match the user profile.

[0145] In some embodiments, the second processing module 203 includes:

[0146] The first calculation unit is used to calculate the weight of each candidate node in the second candidate path based on the node access frequency and time decay factor of the second candidate path for each second candidate path.

[0147] The third processing unit is used to determine the edge weights between adjacent candidate nodes in the second candidate path based on the edge type.

[0148] The second calculation unit is used to calculate the path importance score of the second candidate path based on the weights of the candidate nodes in the second candidate path and the edge weights between the adjacent candidate nodes.

[0149] The fourth processing unit is used to determine the top K second candidate paths with the highest scores from the multiple second candidate paths as the K high-scoring paths.

[0150] In some embodiments, the second computing unit is specifically used for:

[0151] Based on the weights of the candidate nodes in the second candidate path and the edge weights between the adjacent candidate nodes, the transition probability from the first supernode to the second supernode is determined, wherein a supernode is an ordered sequence of nodes composed of at least two candidate nodes in logical order;

[0152] Based on the transition probability from the first supernode to the second supernode, the path importance score of the second candidate path is calculated.

[0153] In some embodiments, the third processing module 204 includes:

[0154] The fifth processing unit is used to identify shared variables in the nodes of two high-scoring paths as anchor points, and to identify the topology graph formed by all non-repeating nodes in the two high-scoring paths corresponding to the anchor points as splicing candidate pairs, wherein the two high-scoring paths are included in the K high-scoring paths;

[0155] The verification unit is used to perform logical coherence verification on each splicing path in the candidate splicing pair and obtain the verification result.

[0156] The sixth processing unit is used to discard the splicing path if the verification result is unsuccessful;

[0157] The seventh processing unit is used to determine the path with the highest path importance score among the spliced ​​paths that passed the verification and the high-scoring unspliced ​​paths as the final thought chain.

[0158] In some embodiments, the verification unit is specifically used for:

[0159] Check whether there are explicit edge connections between the nodes of the splicing path;

[0160] If it exists, the verification is successful; otherwise, the semantic similarity between the nodes of the spliced ​​path is calculated.

[0161] If the semantic similarity is higher than the second threshold, virtual edges are generated between the nodes, and the path importance score of the splicing path with the newly added virtual edges is recalculated.

[0162] If the semantic similarity is lower than or equal to the second threshold, the verification is determined to fail.

[0163] The thought chain retrieval device of this application embodiment obtains keywords from a first question input by a user, the first question being related to a first business scenario; then, it searches for target nodes matching the keywords in the business scenario layer of a first knowledge graph, and based on the target nodes, filters out multiple first candidate paths according to the user profile. The first knowledge graph is a knowledge graph composed of nodes related to business scenarios, nodes related to business decision logic, and nodes related to instance data. The nodes of the first knowledge graph have at least one edge relationship among logical order edge relationship, semantic association edge relationship, and context dependency edge relationship. The target node is a node related to the first business scenario; then, it calculates each of the first... The semantic similarity between a candidate path and the first question is used to obtain multiple second candidate paths with similarity higher than a first threshold. Based on the node access frequency and time decay factor of each second candidate path, K high-scoring paths are determined from the multiple second candidate paths, where K is an integer greater than or equal to 2. Finally, variable anchor point identification and logical coherence verification are performed on the K high-scoring paths to obtain the final thought chain. In this way, through the above two-stage hybrid retrieval, i.e., the coarse-grained retrieval stage, multiple first candidate paths are quickly located through business scenarios and user profiles. Then, the fine-grained retrieval stage is entered, which combines semantic similarity and path optimization to achieve precise targeted retrieval and improve the accuracy of the retrieval results of the thought chain retrieval.

[0164] like Figure 3 As shown in the illustration, this application also provides a mind chain retrieval device, including a processor 300 and a transceiver 310. The transceiver 310 receives and transmits data under the control of the processor 300, and the processor 300 is used to perform the following operations:

[0165] Obtain keywords from the first question input by the user, where the first question is related to the first business scenario;

[0166] In the business scenario layer of the first knowledge graph, target nodes matching the keywords are searched, and based on the target nodes, multiple first candidate paths are filtered according to the user profile. The first knowledge graph is a knowledge graph composed of nodes related to business scenarios, nodes related to business decision logic, and nodes related to instance data. The nodes of the first knowledge graph have at least one edge relationship among logical order edge relationship, semantic association edge relationship, and context dependency edge relationship. The target node is a node related to the first business scenario.

[0167] Calculate the semantic similarity between each first candidate path and the first question to obtain multiple second candidate paths with similarity higher than a first threshold. Based on the node access frequency and time decay factor of each second candidate path, determine K high-scoring paths from the multiple second candidate paths, where K is an integer greater than or equal to 2.

[0168] Based on the K high-scoring paths, variable anchor point identification and logical coherence verification are performed to obtain the final thought chain.

[0169] In some embodiments, the processor 300 is further configured to:

[0170] The original thought chain text is segmented to obtain atomic logic nodes; the original thought chain text is the text formed by the large language model in the reasoning process of responding to user input questions in any business scenario;

[0171] The atomic logic nodes are categorized into business scenario layer, decision logic layer, and instance data layer, and multi-type edge relationships between nodes are constructed to form the first knowledge graph.

[0172] In some embodiments, the processor 300 is further configured to:

[0173] Determine the decision logic layer in the first knowledge graph corresponding to the target node;

[0174] Using the user profile as a filtering condition, the paths of the decision logic layer corresponding to the target node are traversed to select multiple first candidate paths that match the user profile.

[0175] In some embodiments, the processor 300 is further configured to:

[0176] For each of the second candidate paths, the weight of each candidate node in the second candidate path is calculated based on the node access frequency and time decay factor of the second candidate path.

[0177] Based on the edge type, determine the edge weights between adjacent candidate nodes in the second candidate path;

[0178] The path importance score of the second candidate path is calculated based on the weights of the candidate nodes in the second candidate path and the edge weights between the adjacent candidate nodes.

[0179] The top K second candidate paths, ranked from highest to lowest score, are identified as the K high-scoring paths.

[0180] In some embodiments, the processor 300 is further configured to:

[0181] Based on the weights of the candidate nodes in the second candidate path and the edge weights between the adjacent candidate nodes, the transition probability from the first supernode to the second supernode is determined, wherein a supernode is an ordered sequence of nodes composed of at least two candidate nodes in logical order;

[0182] Based on the transition probability from the first supernode to the second supernode, the path importance score of the second candidate path is calculated.

[0183] In some embodiments, the processor 300 is further configured to:

[0184] The shared variables identified in the nodes of the two high-scoring paths are determined as anchor points, and the topology graph formed by all non-repeating nodes in the two high-scoring paths corresponding to the anchor points is determined as a splicing candidate pair. The two high-scoring paths are included in the K high-scoring paths.

[0185] For each splicing path in the candidate splicing pairs, the logical coherence of the splicing path is checked to obtain the check result;

[0186] If the verification result is unsuccessful, the splicing path is abandoned;

[0187] The path with the highest importance score among the spliced ​​paths that passed the verification and the high-scoring unspliced ​​paths is determined as the final thought chain.

[0188] In some embodiments, the processor 300 is further configured to:

[0189] Check whether there are explicit edge connections between the nodes of the splicing path;

[0190] If it exists, the verification is successful; otherwise, the semantic similarity between the nodes of the spliced ​​path is calculated.

[0191] If the semantic similarity is higher than the second threshold, virtual edges are generated between the nodes, and the path importance score of the splicing path with the newly added virtual edges is recalculated.

[0192] If the semantic similarity is lower than or equal to the second threshold, the verification is determined to fail.

[0193] The thought chain retrieval device of this application embodiment obtains keywords from a first question input by a user, the first question being related to a first business scenario; then, it searches for target nodes matching the keywords in the business scenario layer of a first knowledge graph, and based on the target nodes, filters out multiple first candidate paths according to the user profile. The first knowledge graph is a knowledge graph composed of nodes related to business scenarios, nodes related to business decision logic, and nodes related to instance data. The nodes of the first knowledge graph have at least one edge relationship among logical order edge relationship, semantic association edge relationship, and context dependency edge relationship. The target node is a node related to the first business scenario; then, it calculates each of the first... The semantic similarity between a candidate path and the first question is used to obtain multiple second candidate paths with similarity higher than a first threshold. Based on the node access frequency and time decay factor of each second candidate path, K high-scoring paths are determined from the multiple second candidate paths, where K is an integer greater than or equal to 2. Finally, variable anchor point identification and logical coherence verification are performed on the K high-scoring paths to obtain the final thought chain. In this way, through the above two-stage hybrid retrieval, i.e., the coarse-grained retrieval stage, multiple first candidate paths are quickly located through business scenarios and user profiles. Then, the fine-grained retrieval stage is entered, which combines semantic similarity and path optimization to achieve precise targeted retrieval and improve the accuracy of the retrieval results of the thought chain retrieval.

[0194] This invention also provides a thought chain retrieval device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-described functionality. Figure 1 The various processes in the illustrated thought chain retrieval method embodiment can achieve the same technical effect, and will not be described again here to avoid repetition.

[0195] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described functionality. Figure 1 The various processes in the illustrated thought chain retrieval method embodiment achieve the same technical effect, and will not be described again here to avoid repetition. The computer-readable storage medium mentioned includes, for example, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0196] This invention also provides a computer program product, including computer instructions, which, when executed by a processor, implement the above-described functionality. Figure 1 The steps in the mind chain retrieval method are shown.

[0197] It is worth noting that the technologies described in this application are not limited to Long Term Evolution (LTE) / LTE-Advanced (LTE-A) systems, but can also be used in other wireless communication systems, such as Code Division Multiple Access (CDMA), Time Division Multiple Access (TDMA), Frequency Division Multiple Access (FDMA), Orthogonal Frequency Division Multiple Access (OFDMA), Single-carrier Frequency-Division Multiple Access (SC-FDMA), and other systems. The terms "system" and "network" in this application are often used interchangeably, and the described technologies can be used with the systems and radio technologies mentioned above, as well as with other systems and radio technologies. However, the following description describes New Radio (NR) systems for illustrative purposes, and NR terminology is used in most of the following description. These technologies can also be applied to applications beyond NR systems, such as 6th Generation (6G) communication systems.

[0198] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0199] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 A device for one or more processes and / or the functions specified in one or more boxes.

[0200] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce a paper article including an instruction means, the instruction means being implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0201] These computer program instructions can also be loaded onto a computer or other programmable data processing equipment, causing the computer or other programmable equipment to perform a series of operational steps to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0202] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A mind chain retrieval method, characterized in that, include: Obtain keywords from the first question input by the user, where the first question is related to the first business scenario; In the business scenario layer of the first knowledge graph, target nodes matching the keywords are searched, and based on the target nodes, multiple first candidate paths are filtered according to the user profile. The first knowledge graph is a knowledge graph composed of nodes related to business scenarios, nodes related to business decision logic, and nodes related to instance data. The nodes of the first knowledge graph have at least one edge relationship among logical order edge relationship, semantic association edge relationship, and context dependency edge relationship. The target node is a node related to the first business scenario. Calculate the semantic similarity between each first candidate path and the first question to obtain multiple second candidate paths with similarity higher than a first threshold. Based on the node access frequency and time decay factor of each second candidate path, determine K high-scoring paths from the multiple second candidate paths, where K is an integer greater than or equal to 2. Based on the K high-scoring paths, variable anchor point identification and logical coherence verification are performed to obtain the final thought chain.

2. The method according to claim 1, characterized in that, The method further includes: The original thought chain text is segmented to obtain atomic logic nodes; the original thought chain text is the text formed by the large language model in the reasoning process of responding to user input questions in any business scenario; The atomic logic nodes are categorized into business scenario layer, decision logic layer, and instance data layer, and multi-type edge relationships between nodes are constructed to form the first knowledge graph.

3. The method according to claim 1, characterized in that, Based on the target node, multiple first candidate paths are selected according to user profiles, including: Determine the decision logic layer in the first knowledge graph corresponding to the target node; Using the user profile as a filtering condition, the paths of the decision logic layer corresponding to the target node are traversed to select multiple first candidate paths that match the user profile.

4. The method according to claim 1, characterized in that, The process of determining K high-scoring paths from the plurality of second candidate paths based on the node access frequency and time decay factor of each second candidate path includes: For each of the second candidate paths, the weight of each candidate node in the second candidate path is calculated based on the node access frequency and time decay factor of the second candidate path. Based on the edge type, determine the edge weights between adjacent candidate nodes in the second candidate path; The path importance score of the second candidate path is calculated based on the weights of the candidate nodes in the second candidate path and the edge weights between the adjacent candidate nodes. The top K second candidate paths, ranked from highest to lowest score, are identified as the K high-scoring paths.

5. The method according to claim 4, characterized in that, The step of calculating the path importance score of the second candidate path based on the weights of the candidate nodes in the second candidate path and the edge weights between adjacent candidate nodes includes: Based on the weights of the candidate nodes in the second candidate path and the edge weights between the adjacent candidate nodes, the transition probability from the first supernode to the second supernode is determined, wherein a supernode is an ordered sequence of nodes composed of at least two candidate nodes in logical order; Based on the transition probability from the first supernode to the second supernode, the path importance score of the second candidate path is calculated.

6. The method according to claim 1, characterized in that, The process of identifying variable anchor points and verifying logical coherence based on the K high-scoring paths to obtain the final thought chain includes: The shared variables identified in the nodes of the two high-scoring paths are determined as anchor points, and the topology graph formed by all non-repeating nodes in the two high-scoring paths corresponding to the anchor points is determined as a splicing candidate pair. The two high-scoring paths are included in the K high-scoring paths. For each splicing path in the candidate splicing pairs, the logical coherence of the splicing path is checked to obtain the check result; If the verification result is unsuccessful, the splicing path is abandoned; The path with the highest importance score among the spliced ​​paths that passed the verification and the high-scoring unspliced ​​paths is determined as the final thought chain.

7. The method according to claim 6, characterized in that, The logical coherence check of the spliced ​​path includes: Check whether there are explicit edge connections between the nodes of the splicing path; If it exists, the verification is successful; otherwise, the semantic similarity between the nodes of the spliced ​​path is calculated. If the semantic similarity is higher than the second threshold, virtual edges are generated between the nodes, and the path importance score of the splicing path with the newly added virtual edges is recalculated. If the semantic similarity is lower than or equal to the second threshold, the verification is determined to fail.

8. A mind chain retrieval device, characterized in that, include: The acquisition module is used to acquire keywords from the first question input by the user, where the first question is related to the first business scenario; The first processing module is used to find target nodes that match the keywords in the business scenario layer of the first knowledge graph, and based on the target nodes, filter out multiple first candidate paths according to the user profile. The first knowledge graph is a knowledge graph composed of nodes related to business scenarios, nodes related to business decision logic, and nodes related to instance data. The nodes of the first knowledge graph have at least one edge relationship among logical order edge relationship, semantic association edge relationship, and context dependency edge relationship. The target node is a node related to the first business scenario. The second processing module is used to calculate the semantic similarity between each first candidate path and the first question, obtain multiple second candidate paths with similarity higher than a first threshold, and determine K high-scoring paths from the multiple second candidate paths based on the node access frequency and time decay factor of each second candidate path, where K is an integer greater than or equal to 2. The third processing module is used to identify variable anchor points and verify logical coherence based on the K high-scoring paths to obtain the final thought chain.

9. A mind chain retrieval device, comprising a processor and a transceiver, wherein the transceiver receives and transmits data under the control of the processor, characterized in that, The processor is used to perform the following operations: Obtain keywords from the first question input by the user, where the first question is related to the first business scenario; In the business scenario layer of the first knowledge graph, target nodes matching the keywords are searched, and based on the target nodes, multiple first candidate paths are filtered according to the user profile. The first knowledge graph is a knowledge graph composed of nodes related to business scenarios, nodes related to business decision logic, and nodes related to instance data. The nodes of the first knowledge graph have at least one edge relationship among logical order edge relationship, semantic association edge relationship, and context dependency edge relationship. The target node is a node related to the first business scenario. Calculate the semantic similarity between each first candidate path and the first question to obtain multiple second candidate paths with similarity higher than a first threshold. Based on the node access frequency and time decay factor of each second candidate path, determine K high-scoring paths from the multiple second candidate paths, where K is an integer greater than or equal to 2. Based on the K high-scoring paths, variable anchor point identification and logical coherence verification are performed to obtain the final thought chain.

10. A mind chain retrieval device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the mind chain retrieval method as described in any one of claims 1 to 7.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the mind chain retrieval method as described in any one of claims 1 to 7.

12. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the steps in the mind chain retrieval method as described in any one of claims 1 to 7.

Images