Data acquisition method, inference method and device for enhancing large model inference
By constructing a conflict graph and comprehensively evaluating source confidence, lifecycle status, and interactive feedback, and using confidence-based text filtering, the limitations of traditional RAG systems in handling dynamic, multi-source, and conflicting information are overcome, thereby improving the accuracy and reliability of large model inference.
Patent Information
- Application Number
- CN202511983962.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-26
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2045-12-26
AI Technical Summary
Traditional RAG systems struggle to effectively distinguish the authority of information sources when dealing with dynamic, multi-source, and conflicting information. They are unable to handle conflicts between different versions of content, and existing credibility assessment methods lack the ability to infer deep structures and complex relationships, making them susceptible to being circumvented by carefully crafted content, leading to inaccurate or misleading results.
By obtaining related raw texts from different retrieval sources, breaking them down into text blocks, constructing conflict graphs, evaluating source confidence, lifecycle status, and interaction feedback credibility, comprehensively analyzing credibility, and selecting texts with high comprehensive analysis credibility as target data.
It improves the authenticity, accuracy, and consistency of large model inference results, reduces the generation of erroneous information, ensures that the model uses the latest and most effective information, and reduces potential risks in practical applications.
Smart Images

Figure CN121809677B_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the field of artificial intelligence information retrieval enhancement technology, and in particular to a data acquisition method, reasoning method, and apparatus for enhancing large model reasoning. Background Technology
[0002] Traditional Retrieval Augmentation (RAG) systems typically rely heavily on vector similarity-based recall and ranking mechanisms when processing information. While this approach performs reasonably well with structured, static knowledge bases, it faces significant challenges as the information ecosystem becomes increasingly complex, particularly with the introduction of rapidly iterating technical documents, multi-source heterogeneous knowledge bases, and highly diverse community-generated content. Vector similarity struggles to effectively distinguish the authority of information sources, cannot handle conflicts between different versions of content, and is prone to misuse of semantic fragments that only hold true in specific contexts.
[0003] Furthermore, valuable validation evidence such as likes and adoptions in community content is often difficult to utilize in a structured way, leading to the amplification of erroneous information by large language models (LLM), generating inaccurate or misleading results.
[0004] To address these challenges, some existing technologies attempt to directly calculate credibility scores for documents or paragraphs. However, these methods are essentially simple linear combinations of various weights, lacking the ability to infer the deep structure and complex relationships behind the information. This linear combination strategy can easily be "circumvented" by carefully crafted content—that is, by artificially manipulating textual features or metadata to obtain high credibility scores, when the content itself may not actually be reliable. More importantly, this shallow evaluation mechanism is insufficient to effectively support the various conflicts existing in complex knowledge ecosystems, failing to deeply understand the relativity and contextual dependence of information, thus making it difficult to achieve effective conflict resolution. Summary of the Invention
[0005] In view of this, the present disclosure provides a data acquisition method, reasoning method, and apparatus for enhancing large model reasoning, which can solve the limitations of traditional RAG methods in handling dynamic, multi-source, and conflicting information, as well as the shortcomings of current credibility assessment methods, and can achieve credibility enhancement through multi-source conflict resolution.
[0006] In a first aspect, embodiments of this disclosure provide a data acquisition method for enhancing large model inference, including: Retrieve several related original texts from different search sources based on the target requirements; Each of the original text entries is split into several text blocks; A conflict graph is constructed based on all the text blocks; the conflict graph includes the relationship edges between each pair of the original texts and their strength weights, and each node in the conflict graph is one of the original texts; Determine the source confidence level of the retrieval source for each of the original texts; Based on the conflict map and all the source confidence scores, determine the conflict confidence score for each node; Determine the lifecycle status credibility of each node based on the stated target requirements; The credibility of the interaction feedback of each node is determined based on the interaction metadata corresponding to each node and the source confidence level. Based on the conflict credibility, the lifecycle state credibility, and the interaction feedback credibility, the comprehensive analysis credibility of each node is determined. The original text corresponding to all nodes whose comprehensive analysis credibility is greater than a preset credibility threshold is used as the target data for enhancing the large model inference.
[0007] Secondly, embodiments of this disclosure also provide an enhanced RAG large model inference method, including: Retrieve several related original texts from different search sources based on the target requirements; Each of the original text entries is split into several text blocks; Construct a conflict graph based on all the aforementioned text blocks; The conflict graph includes the relationship edges between each pair of the original texts and their strength weights, and each node in the conflict graph is one of the original texts. Determine the source confidence level of the retrieval source for each of the original texts; Based on the conflict map and all the source confidence scores, determine the conflict confidence score for each node; Determine the lifecycle status credibility of each node based on the stated target requirements; The credibility of the interaction feedback of each node is determined based on the interaction metadata corresponding to each node and the source confidence level. Based on the conflict credibility, the lifecycle state credibility, and the interaction feedback credibility, the comprehensive analysis credibility of each node is determined. The original text corresponding to all nodes whose comprehensive analysis credibility is greater than the preset credibility threshold is used as the target data; The target data is then output as the reasoning result through a large model.
[0008] Thirdly, this disclosure also provides a computer device, which adopts the following technical solution: The computer device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which enable the at least one processor to perform either the data acquisition method for enhanced large model inference or the large model inference method for enhanced RAG described above.
[0009] Fourthly, embodiments of this disclosure also provide a computer-readable storage medium storing computer instructions for causing a computer to perform any of the above-described data acquisition methods for enhanced large model inference or the enhanced RAG large model inference methods.
[0010] Fifthly, embodiments of this disclosure also provide a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.
[0011] The data acquisition method disclosed in this application for enhancing large model inference first obtains several related original texts from different retrieval sources according to the target requirements, ensuring that the acquired corpus is centered around the target requirements rather than randomly scraped. By obtaining data from different retrieval sources, a wider range of information perspectives and knowledge from different sources can be covered, avoiding information being limited to a single dimension. Then, each original text is divided into several text blocks, allowing the analysis to delve into the local level of the text. Next, a conflict graph is constructed based on all text blocks, systematically quantifying the interrelationships between different original texts, especially potential conflicts. Then, the source confidence of the retrieval source for each original text is determined, introducing an assessment of the trust in the information source itself. Then, based on the conflict graph and the confidence of all sources, the lifecycle state confidence of each node is determined according to the target requirements. Based on the interaction metadata and source confidence of each node, the interaction feedback confidence of each node is determined. Based on the conflict confidence, lifecycle state confidence, and interaction feedback confidence, the comprehensive analysis confidence of each node is determined. Finally, The original texts corresponding to all nodes whose comprehensive analysis credibility exceeds a preset credibility threshold are used as the corpus for enhancing the large model's inference. Interactive feedback provides a qualitative assessment from the perspective of user feedback, compensating for the shortcomings of purely technical analysis. By assessing the lifecycle status of the text, outdated, invalid, or updated information can be identified, ensuring that the model uses the latest and most effective information. The conflict credibility of each node is determined. By combining the structured conflict relationships between texts (conflict graph) and the reliability of the source itself (source confidence), the conflict credibility of an original text in the overall information set can be assessed more accurately. The three independent credibility dimensions (conflict, lifecycle, and interactive feedback) are organically combined to form a more comprehensive and robust "comprehensive analysis credibility". By integrating multiple assessment dimensions, the true value and reliability of the text can be reflected more accurately, improving the accuracy and robustness of the assessment. Finally, through strict threshold screening, it is ensured that every original text entering the corpus has undergone multi-dimensional and refined credibility verification. When large models use such rigorously selected corpora for reasoning, the authenticity, accuracy, and consistency of their output are greatly improved, effectively reducing the generation of illusions and misinformation. This provides reliable information increments for large models, enabling them to generate more persuasive and trustworthy answers, thereby reducing the potential risks of the model in practical applications. It effectively solves the limitations of traditional RAG methods in handling dynamic, multi-source, and conflicting information, as well as the shortcomings of current credibility assessment methods.
[0012] The above description is merely an overview of the technical solution disclosed herein. In order to better understand the technical means of this disclosure and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description
[0013] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0014] Figure 1 This is a flowchart illustrating a data acquisition method for enhancing large model inference provided in an embodiment of this disclosure.
[0015] Figure 2 This is a flowchart illustrating a method for retrieving several related original texts from different retrieval sources according to target requirements, as provided in an embodiment of this disclosure.
[0016] Figure 3 This is a flowchart illustrating a method for constructing a conflict graph based on all text blocks, as provided in an embodiment of this disclosure.
[0017] Figure 4 A flowchart illustrating the method for determining the conflict credibility of each node provided in this embodiment of the disclosure.
[0018] Figure 5 This is a flowchart illustrating the method for obtaining the lifecycle state reliability of each node provided in this embodiment of the disclosure.
[0019] Figure 6 This is a flowchart illustrating the method for obtaining the credibility of interactive feedback for each node as provided in the embodiments of this disclosure.
[0020] Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present disclosure. Detailed Implementation
[0021] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.
[0022] It should be understood that the following specific examples illustrate the implementation of this disclosure, and those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.
[0023] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.
[0024] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The drawings only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.
[0025] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.
[0026] Reference Figure 1 This application discloses a data acquisition method for enhancing large model inference, which can enhance the credibility of multi-source conflict resolution. The method specifically includes: S100: Retrieve several related original texts from different retrieval sources according to the target requirements.
[0027] S200 breaks each original text into several text blocks.
[0028] S300 constructs a conflict graph based on all text blocks.
[0029] The conflict graph includes the relationship edges between each pair of original texts and their strength weights. Each node in the conflict graph is an original text.
[0030] S400 determines the source confidence of the retrieval source for each original text.
[0031] S500 determines the conflict confidence of each node based on the conflict graph and the confidence of all sources.
[0032] S600 determines the lifecycle status reliability of each node based on target requirements.
[0033] S700 determines the credibility of each node's interaction feedback based on the interaction metadata and source confidence level corresponding to each node.
[0034] S800 determines the overall analysis credibility of each node based on conflict credibility, lifecycle status credibility, and interaction feedback credibility.
[0035] S900 uses the original text corresponding to all nodes whose comprehensive analysis credibility is greater than the preset credibility threshold as the corpus (i.e., target data) for enhancing large model inference.
[0036] Reference Figure 2 The method for S100, "obtaining several related original texts from different retrieval sources according to target requirements," specifically includes: S110, Determine the intent type of the target requirement.
[0037] Specifically, lightweight classification models (such as BERT-Tiny or rule-based keyword matching) can be used to classify user queries and obtain the corresponding intent type.
[0038] The intent type can be any one of the following: API usage, version change, troubleshooting, or best practice.
[0039] API usage class refers to a specific type of information requirement or user query category, which can include: querying function parameters, return values, and configuration items. The characteristic words corresponding to API usage class include: config, parameter, how to use, argument, etc. That is, when a user's query contains one or more of these words, it can be determined that the query belongs to the "API usage class".
[0040] Version change categories can include: inquiries about new features, deprecated features, and migration guidelines; the corresponding characteristic words for this category are: deprecated, new in v5, migration, difference, etc. That is, when a user's query contains one or more of these words, it can be determined that the query belongs to the "version change category".
[0041] Troubleshooting categories include: querying error messages and bug fixing; the corresponding characteristic words for this category include: error, exception, fail to start, bug, etc. That is, when a user's query contains one or more of these words, it can be determined that the query belongs to the "troubleshooting category".
[0042] Best practice categories include queries about architecture design and performance optimization. The corresponding keywords for this category include: bestpractice, optimization, architecture, etc. That is, when a user's query contains one or more of these keywords, it can be determined that the query belongs to the "best practice category".
[0043] S120, determine the list of target sources corresponding to the intent type according to the preset routing mapping table.
[0044] Table 1 Preset Route Mapping Table Referring to Table 1, the preferred search source refers to the authoritative core source, the secondary search source refers to the code / engineering source, and the excluded / degraded source refers to the community / experience source; the target source list includes summary information of the preferred search source, secondary search source, and excluded / degraded source.
[0045] S130, Identify key technology entities in the target requirements.
[0046] Specifically, a pre-trained NER model can be used, specifically trained for the technical documentation domain, and then the key technical entities in the target requirements can be identified based on the trained NER model.
[0047] S140, based on the target source list, retrieve several original texts associated with the key technology entities from the corresponding retrieval sources.
[0048] Furthermore, if the entity is an Open Source Project (such as React): automatically append the GitHub repository address of the project; if the entity is a Cloud Service (such as AWS): automatically append the address of the official SDK documentation for the service.
[0049] For example, in Example 1 (Neo4j Configuration), the user query is: "How do I configure the n10s plugin in Neo4j version 5.26?" Intent recognition includes: keyword configuration, which is determined to be an API usage class; keyword 5.26, which is determined to contain version constraints.
[0050] Source routing decisions include: prioritizing the official Neo4j documentation (Tier 1), especially the Operations Manual; and secondarily selecting the GitHub repository neo4j-contrib / neosemantics (Tier 2) to verify specific plugin version behaviors. Further supplementary information includes the Neo4j community forums (Tier 3) to find the latest user troubleshooting records. Exclusion criteria include: CSDN / Medium blogs from before 2022 (because the query specifies the very recent version 5.26, older blogs represent high-risk noise).
[0051] In this embodiment, the three-layer strategy of "intent recognition, route mapping, and entity expansion" effectively avoids blindly searching the entire network. Instead, it can accurately identify the range of data sources most likely to contain the correct answer. This not only improves retrieval efficiency but also reduces the probability of low-quality information (such as outdated blogs) entering the system from the source.
[0052] In this embodiment, S100 ensures that the acquired corpus is centered around the target requirement, rather than being randomly crawled; by acquiring it from different retrieval sources, it can cover a wider range of information perspectives and knowledge from different sources, avoiding information being limited to a single dimension, laying the foundation for more refined analysis later, emphasizing the relevant original text, indicating that preliminary information screening has been carried out, excluding text that is irrelevant to the target requirement, which can effectively improve the efficiency and effectiveness of subsequent processing.
[0053] S200, "Divide each original text into several text blocks," specifically includes: S210: Perform parsing processing on each piece of original text, extract the plain text fragment containing the text content, remove embedded formatting tags and non-text elements, and obtain the target text.
[0054] S220 divides each target text into several discrete text blocks.
[0055] Specifically, each target text is broken down into independent, verifiable, core information claims containing factual statements—that is, divided into several discrete text blocks. This step enables a leap from "unstructured documents" to "structured knowledge points," facilitating subsequent comparisons.
[0056] Suppose that in this step we retrieve the following two original document fragments, which discuss the same technical point (i.e., correspond to the same entity), but the content conflicts. We perform parsing on each original text, extract the plain text fragment containing the text content, remove embedded formatting tags and non-text elements, and obtain the target text.
[0057] Target text A (from the 2021 Neo4j 4.x official documentation); Text: "To configure the then10s extension in Neo4j, you must add the following line to your neo4j.conffile: dbms.unmanaged_extension_classes=n10s.endpoint= / rdf. After saving the file, a server restart is required." Meta: {id: "doc_A", ver: "4.2", source: "official"}.
[0058] Target text B (from the Neo4j 5.x migration guide of 2024); Text: "In Neo4j 5, thedbms.unmanaged_extension_classes setting has been removed. You should now register n10s using server.dynamic_setting.allow_list or via the Cypherprocedure CALL n10s.graph.config(...)." Meta: {id: "doc_B", ver: "5.26",source: "official"}.
[0059] Then, LLM can be called to decompose the two text segments into independent, atomic claim nodes, that is, to divide each target text into several discrete text blocks.
[0060] During the LLM call, the issued instructions may include "Please break down the following text into several independent atomic facts, requiring 1) each claim must be a complete sentence containing a subject; 2) the version restrictions, configuration parameter names, and operation conditions in the original text must be retained; 3) the format is a JSON List".
[0061] The segmentation results of target text A include: Prop A1 : "In Neo4j 4.x, configuring n10s requires editing theneo4j.conf file."Prop A2 : "In Neo4j 4.x, the configuration parameter isdbms.unmanaged_extension_classes."Prop A3 : "Configuring n10s in Neo4j 4.xrequires a server restart." The segmentation results of target text B include: Prop B1 : "In Neo4j 5, the dbms.unmanaged_extension_classes setting isremoved."Prop B2 : "In Neo4j 5, n10s should be registered using server.dynamic_setting.allow_list."Prop B3 "In Neo4j 5, n10s can be configured via Cypherprocedure CALL n10s.graph.config." This step yields six independent claim nodes, i.e., six text blocks, containing a mixture of correct and incorrect information.
[0062] In this embodiment, S200 breaks down the original text into smaller text blocks, enabling analysis to delve into the local level of the text. This is crucial for identifying subtle information, potential conflicts, or differing arguments within the text, laying the foundation for constructing a detailed conflict map. This embodiment allows for independent evaluation of each part of the text, thereby enabling more effective mining and utilization of scattered but valuable information within the text.
[0063] Reference Figure 3 The S300 method of "constructing a conflict map based on all text blocks" specifically includes: S310, obtain the comparison results of the relationship between each pair of original texts based on all text blocks.
[0064] The relationship comparison results are contradictory, competitive, or parallel.
[0065] In this embodiment, several text blocks corresponding to each original text constitute an information claim. Specifically, a pre-trained general NLI model (such as BER, RoBERTa, DeBERTa, etc., and fine-tuned) or an NLI model for a specific technical field can be used to determine the relationship between two information claims and obtain the comparison results of the relationship between the pairs of original texts.
[0066] If the relationship between two information claims is determined to be mutually exclusive or contradictory, the relationship comparison result is determined to be contradictory; if the relationship between two information claims is determined to be mutually exclusive solutions to the same problem, the relationship comparison result is determined to be competitive; if the two information claims are both feasible paths to complete a certain task and there is no direct conflict or substitution relationship between them, the relationship comparison result is determined to be parallel.
[0067] It should be noted that this application retrieves several related original texts from different search sources based on the target requirements, therefore the entities for the information claims corresponding to all original texts are the same.
[0068] S320 maps the information claim corresponding to each original text to a node in the graph database.
[0069] The graph database can be Neo4j, ArangoDB, Amazon Neptune, etc.
[0070] S330 generates the relationship edges between each pair of nodes and their strength weights based on the relationship comparison results.
[0071] Among them, the relation edges include negative edges, weak negative edges, or positive edges, and the strength weight corresponding to the negative edge is less than the strength weight corresponding to the weak negative edge, and the strength weight corresponding to the weak negative edge is less than the strength weight corresponding to the positive edge.
[0072] In this step, the generated relation edges (i.e., edge types) and strength weights are mapped to the relationships between nodes. In this embodiment, all generated relations (edges) and nodes are imported into the graph database in batches to obtain the conflict graph.
[0073] Furthermore, the strength weight corresponding to a negative edge is preferably -1.0, indicating a strong conflict and mutual negation. The strength weight corresponding to a weak negative edge is preferably -0.8, indicating the existence of conflict or mutual exclusion, but it may not be an absolute negation, and there is a degree of weakening relationship. The strength weight corresponding to a positive edge is preferably 1.0 (indicating support, consistency, or complementarity) or 0 (indicating no direct relationship, and no edge is added).
[0074] For example, comparison group 1: Prop A2 vs Prop B1The NLI inference result is: Contradiction. The entity is: dbms.unmanaged_extension_classes. Operation: Create a Prop in the graph database. A2 Node (if it does not exist) and Prop B1 Node (if it does not exist). In Prop A2 and Prop B1 Establish a CONTRADICTS relationship between the two nodes and set the strength weight of the relationship to -1.0.
[0075] Comparison group 2: Prop A2 vs Prop B2 Its NLI inference result is: Competition. The entity is: dbms.unmanaged_extension_classes(Prop A2 )vs server.dynamic_setting (PProp B2 Note: The entities here are different, but the system needs to determine whether they are competing entities under the intent of "how to configure a similar function in Neo4j". A2 A specific parameter was mentioned, Prop B2 A list of allowed settings was mentioned. Action: In Prop A2 and Prop B2 Establish a COMPETES_WITH relationship between the two nodes. Set the strength weight of this relationship to -0.8.
[0076] Comparison group 3: Prop B2 vs Prop B3 Its NLI inference result is: parallel. Entity: server.dynamic_setting(Prop B2 ) vs Cypher CALL (Prop B3 ). Context: Both could be the recommended approach for implementing a certain functionality in Neo4j version 5.x. Operation (assuming strategy A is used, creating a positive edge): In Prop B2 and Prop B3 Establish a SUPPORTS relationship between the two nodes. Set the strength weight of this relationship to +1.0.
[0077] Operation (assuming strategy B is adopted, no edge is created): No operation is performed, and these two nodes are currently not directly connected in the graph.
[0078] Furthermore, we can visualize JSON as a graph: such as Prop. B1(5.x) Sends a red arrow (attack / negation) pointing at Prop. A2 (4.x). Prop B2 (5.x) Standing on Prop B1 Beside it, as a new alternative, this structure clearly tells subsequent algorithms: Prop A2 Since it is the target of the attack, its credibility should be reduced; while Prop B1 It is the initiator of the attack (and from a new version), so its credibility should be very high. This is the basis of "inferring credibility from the structure".
[0079] The implementation of the conflict graph in this application is a multi-stage process involving information extraction, NLI reasoning, relation mapping, and graph database operations. The key is how to effectively train the NLI model to capture the conflict and competition relationships specific to the technology field, and how to design a flexible representation to store these relationships and their strength.
[0080] Furthermore, the reliability of the information source can be considered when assigning weights. For example, claims from official documents should have higher weights than those from personal blogs, and this can be adjusted by introducing a "reputation score" outside the NLI model.
[0081] In this embodiment, S300 constructs a conflict graph, which for the first time systematically quantifies the relationships between different original texts, especially potential conflicts. The relationship edges and their strength weights can identify contradictions that were originally hidden in massive amounts of information. This is crucial for processing dynamic, multi-source, and conflicting information, and avoids inference errors caused by the model directly adopting contradictory information. Using original texts as nodes and relationships as edges provides a structured way to understand the complex connections between information, which facilitates subsequent quantitative analysis.
[0082] Regarding S400's "determining the source confidence of each original text retrieval source," the official source confidence is greater than the community source confidence.
[0083] Specifically, the source confidence of old official sources is lower than that of new official statements / announcements / blogs; the source confidence of ordinary technical blogs / forums is lower than that of old official sources; and the source confidence of well-known community platforms is lower than that of new official statements / announcements / blogs but higher than that of ordinary technical blogs / forums.
[0084] Specifically, suppose there are 4 typical nodes. Node V1 (old official version) contains the following message: "Configuring n10s requires modifying dbms.unmanaged_extension_classes." Source: Neo4j v4.2 official documentation; source confidence: 0.9 (official documentation usually has a high initial score).
[0085] Node V2 (New Official Statement / Announcement / Blog), content: "The dbms.unmanaged_extension_classes configuration item has been removed." Source: Neo4j v5.26 Migration Guide; Source confidence: 0.95.
[0086] Node V3 (new official documentation), content: "Please use server.dynamic_setting for configuration." Source: Neo4j v5.26 User Manual; Source confidence level: 1.
[0087] Node V4 (General Technical Blog / Forum), content: "No configuration required, just copy the JAR file and restart." Source: A personal blog (2020); Source confidence: 0.4 (low initial score for personal blogs).
[0088] In this embodiment, S400 introduces a trust assessment of the information source itself. Different search sources (e.g., official documents, news reports, personal blogs, academic papers, etc.) have different levels of trust. This step can quantify this trust level, providing an important basis for subsequent comprehensive trust assessment; it helps to distinguish texts from more reliable sources, thus tending to accept information with higher source trust when there is information conflict.
[0089] Reference Figure 4 The S500 method of "determining the conflict confidence of each node based on the conflict map and the confidence levels of all sources" specifically includes the following methods for determining the conflict confidence of each node: S510 configures an initial authority score for each node based on its origin.
[0090] Specifically, the initial authority score can be preset based on the source metadata (Source Meta) when the original text is entered into the database. In a preferred embodiment, when the source corresponding to the node is official documentation (basic Tier 1), it is set to 0.9~0.95; when the source corresponding to the node is code / project source (i.e., Tier 2), it is set to 0.8; and when the source corresponding to the node is community blogs, etc. (i.e., Tier 3), it is set to 0.4.
[0091] S520: Determine the relationship edges and their strength weights of the nodes to be analyzed from the conflict graph.
[0092] S530, based on the initial authority score and strength weight of the node corresponding to the relation edge of the node to be analyzed, determine the degree of conflict between the node corresponding to the relation edge of the node to be analyzed and the node to be analyzed.
[0093] S540 obtains the conflict credibility of the node to be analyzed based on all conflict levels and the initial authority score of the node to be analyzed.
[0094] No. The confidence level of the conflict of each node is : .
[0095] in, For the first The initial authority score of each node, The influence coefficient, In order to be with the first The node with the associated relationship The initial authority score of each node, For the first The node and the first Strength weights between nodes For the first The node is the first The degree of conflict of each node (i.e., the node to be analyzed).
[0096] Furthermore, This is used to control the influence strength of neighbor node conflict information with semantic overlap. If A larger value indicates that the credibility of the claim relies more heavily on the "support" or "refutation" of other evidence; if A lower score indicates that the credibility of the claim depends more on its own basic authority score.
[0097] In this embodiment, S500 can more accurately assess the conflict credibility of an original text in the overall information set by combining the structured conflict relationship (conflict map) between texts and the reliability of the source itself (source confidence). If a text conflicts with many other sources with lower credibility, its conflict credibility will decrease. Conversely, if a text itself has high source credibility and is consistent with other high-credibility texts, its conflict credibility will remain at a high level even if there are a few conflicts.
[0098] Reference Figure 5 The method for S600 to "determine the lifecycle state credibility of each node according to target requirements," specifically the method for obtaining the lifecycle state credibility of each node, includes: S610, determine the lifecycle status of each node. The lifecycle status includes the existing version time and the corresponding existing version description features.
[0099] S620: Obtain the current time and current version description features corresponding to the target requirement.
[0100] S630 determines the validity type of each node's status based on the current version description features, current time, and lifecycle status.
[0101] S640, determine the corresponding state weight based on the state validity type.
[0102] In this embodiment, if an existing version description feature is explicitly removed from the current version description feature, the state validity type of the corresponding node is determined to be the version removal class, and the corresponding state weight is determined to be 0.01 to prevent complete discarding, but to make its ranking extremely low, so a very small value is assigned.
[0103] If an existing version description feature is explicitly deprecated in the current version description feature, the state validity type of the corresponding node is determined to be version deprecation removal, and the corresponding state weight is determined to be 0.05.
[0104] If the existing version description features are not explicitly defined in the current version description features, and the existing version time is not later than the current time, the status validity type of the corresponding node is determined to be active, and the corresponding status weight is determined to be 1.0.
[0105] S650 determines the lifecycle state reliability of the node based on the state weight and a preset formula.
[0106] The reliability of the lifecycle state corresponding to each node is: : Where D is the version difference between the current time and the existing version time. These are the state weights.
[0107] Specifically, for example, if the target requirement is "How to configure the n10s plugin in Neo4j 5.26?", the current time corresponding to the target requirement is obtained as follows: (i.e., target version), current version description features.
[0108] The nodes to be evaluated include: 1) Prop A (Old method): "Use dbms.unmanaged_extension_classes." Corresponding source metadata: Neo4j v4.2 documentation (released in 2021). 2) Prop B(New method): "Use server.dynamic_setting." Corresponding source metadata: Neo4j v5.26 documentation (released in 2024). 3) Prop C (Intermediate state): "dbms.unmanaged... is marked as Deprecated, migration is recommended." Corresponding source metadata: Neo4j v4.4 changelog (released in 2022).
[0109] For Prop A (v4.2) Prediction: and It crosses a major version number; prediction rule: crossing a major version usually means a breaking change; inferred status: Removed or High_Risk. For Prop C Prediction for (v4.4): The original text mentions "v4.4 Deprecated". Lifecycle trajectory: Active (v4.0) -> Deprecated (v4.4) -> Removed (v5.0+); Prediction conclusion: In v5.26, this feature should have been completely removed; Inferred status: Removed. Regarding Prop... B (v5.26) Prediction: Inferring state: Active (active / valid). Then, the reliability of the lifecycle state for each node can be determined.
[0110] In this embodiment, S600 can identify outdated, invalid, or updated information by evaluating the lifecycle status of the text, ensuring that the model uses the latest and most effective information, filtering out old information that no longer has practical significance, and preventing the model from reasoning based on outdated knowledge, thereby improving the timeliness and accuracy of the response.
[0111] Reference Figure 6 The S700 method of "determining the credibility of each node's interaction feedback based on the interaction metadata and source confidence of each node" specifically includes the following methods for obtaining the credibility of each node's interaction feedback: S710, obtain the interaction metadata corresponding to each node; among which, the interaction metadata includes the number of approvals and the number of confirmations.
[0112] Recognition includes one or more of the following: likes, quantity, and quantity. The quantity of recognition refers to the total number of recognitions. Confirmation includes one or more of the following: positive feedback, being cited, and being adopted. The quantity of confirmation refers to the total number of confirmations.
[0113] S720 determines the initial energy information of each node based on the source confidence level corresponding to each node.
[0114] In this embodiment, the higher the source confidence, the larger the initial energy information is configured. For example, if the source is an official developer / maintainer reply, the initial energy information is preset to the highest (e.g., 0.9); if the source is a high-reputation platform (e.g., StackOverflow accepted answers), the initial energy information is preset to the second highest (e.g., 0.8); and if the source is a regular community post / personal blog, the initial energy information is preset to a lower level (e.g., 0.4).
[0115] The S730 determines the credibility of each node's interactive feedback based on the number of acknowledgements, confirmations, and initial energy information.
[0116] The credibility of interactive feedback is : .
[0117] in, As the recognition coefficient, The number of people who have given their approval (this is the most basic signal of group approval). To confirm the quantity, To confirm the coefficients, This refers to the initial energy information of the corresponding node. In this embodiment, Greater than This step transforms feedback from the community and other sources into structured information, taking into full account user interaction behavior and further improving credibility.
[0118] In this step, for conduct Logarithmic processing is used to prevent the "Matthew effect" in social data, that is, to prevent a viral post with a few thousand likes from having such a high impact that it overshadows all other evidence. The processing slows down the energy growth as the number of likes increases, ensuring the stability of the system.
[0119] In this step, The corresponding confirmation can be obtained through explicit verification feedback extracted from the comments section using LLM (Large Model) or semantic parsing. For example, the number of replies containing phrases such as "personally tested and effective", "solved my problem", and "confirmed to be true".
[0120] in, The preferred value is 0.5. The preferred value is 0.2, that is The weight (0.5) is significantly higher than (0.2) This is because the system believes that "explicit confirmation with textual feedback" has higher technical reference value than "casual likes". This weighting design ensures that even if the number of likes is not high, as long as multiple experts confirm its validity, the energy field of the knowledge point will be rapidly enhanced. At the same time, through the setting of these two weights, it is possible to identify invalid content that "gains high likes simply because it is funny or has a clickbait title" but lacks technical confirmation.
[0121] The formula is set according to the official documentation ( (High) has not yet been updated, but the community issue (although) Low, but When the new configuration method is pointed out (extremely high), the final energy value of the issue will surpass that of the official documentation, effectively solving the drawbacks of the traditional RAG.
[0122] In this embodiment, the interactive metadata in S700 (such as likes, comments, citations, number of adoptions, etc.) reflects the feedback and recognition the text receives in actual applications; combined with source confidence, it can more comprehensively evaluate the actual value and influence of a text; high interactive feedback usually means that the text has high practicality or is widely accepted, which is an independent and important credibility indicator. Interactive feedback provides a qualitative assessment from the perspective of user feedback, making up for the shortcomings of purely technical analysis.
[0123] The S800 method of "determining the comprehensive analysis credibility of each node based on conflict credibility, lifecycle state credibility, and interaction feedback credibility" specifically includes: weighted summation of conflict credibility, lifecycle state credibility, and interaction feedback credibility to obtain the score corresponding to the comprehensive analysis credibility of each node.
[0124] In S800, the three independent credibility dimensions (conflict, lifecycle, and interaction feedback) are organically combined to form a more comprehensive and robust "comprehensive analysis credibility". Any single credibility indicator has limitations. This multi-dimensional assessment method is more comprehensive and in-depth than a single source confidence or simple verification method. It can more accurately reflect the true value of the text. By integrating multiple assessment dimensions, it can more accurately reflect the true value and reliability of the text, and improve the accuracy and robustness of the assessment.
[0125] The S900's provision of "using the original text corresponding to all nodes whose comprehensive analysis credibility exceeds the preset credibility threshold as the corpus for enhancing large-scale model inference" includes: directly outputting the original text corresponding to nodes whose comprehensive analysis credibility exceeds the preset credibility threshold, and deleting other original texts that do not meet the conditions. Through rigorous threshold screening, it ensures that every original text entering the corpus undergoes multi-dimensional and refined credibility checks. When the large model uses such a rigorously screened corpus for inference, the authenticity, accuracy, and consistency of its output results are greatly improved, effectively reducing the generation of illusions and misinformation. This provides reliable information increments for the large model, enabling it to generate more persuasive and trustworthy answers, thereby reducing the potential risks of the model in practical applications.
[0126] In this embodiment, the introduction of lifecycle state credibility can identify and filter out outdated and no longer valid information, ensuring that the model uses the latest and most relevant knowledge. The combined effect of source confidence and conflict graphs can assess the original reliability of information from different retrieval sources and identify potential contradictions between different sources, thus preventing the model from mistakenly accepting unreliable original text. The conflict graph, a core innovation, quantifies and identifies textual conflicts by analyzing the strength of relationships between pairs of original texts. This allows the model to handle information conflicts more intelligently, for example, by selecting texts with lower conflict levels or marking conflicting parts, rather than mechanically mixing all information and causing chaotic inference results. Interactive feedback credibility utilizes interactive metadata, effectively introducing social proof or user feedback, which can serve as an effective indicator of text acceptance and actual utility—more grounded in reality than purely technical evaluation.
[0127] Finally, the corpus is filtered by comprehensively analyzing texts whose credibility exceeds a preset credibility threshold. This ensures that only high-quality, highly credible texts are used to enhance model inference, thereby improving the credibility of the model's output from the source. Through this meticulous corpus filtering, the model can obtain higher-quality and more accurate retrieval results when performing RAG operations, significantly reducing the occurrence of illusions and improving the authenticity and consistency of generated answers.
[0128] For fields such as finance, healthcare, law, and news analysis, which are characterized by complex information, high real-time requirements, and frequent information conflicts, this method can provide more reliable corpus support, enabling large models to function more safely and effectively in these fields.
[0129] This application discloses a data acquisition method for enhancing large-scale model inference, aiming to address the numerous challenges encountered by traditional RAG (Retrieval-Augmented Generation) in handling complex information and to compensate for the shortcomings of existing credibility assessment methods. This method provides high-quality, high-credibility corpora, significantly improving the performance of large-scale model inference. It offers a streamlined corpus construction mechanism that automates the processing of massive, heterogeneous data, greatly reducing manual intervention and data preprocessing workload. By intelligently filtering the corpus, the amount of data that the model needs to retrieve and process during inference can be effectively reduced, potentially lowering real-time latency and computational costs, and improving overall engineering efficiency.
[0130] Secondly, this application discloses a method for enhancing large-scale model inference in RAG, including: S10, Based on the target requirements, retrieve several related original texts from different retrieval sources; S20: Each original text is split into several text blocks, and a conflict graph is constructed based on all text blocks.
[0131] The conflict graph includes the relationship edges between each pair of original texts and their strength weights. Each node in the conflict graph is an original text. S30, determine the source confidence of the retrieval source for each original text; S40, Based on the conflict graph and the confidence of all sources, determine the conflict confidence of each node; S50 determines the lifecycle status reliability of each node based on target requirements; S60, based on the interaction metadata and source confidence of each node, determine the credibility of the interaction feedback of each node; S70, based on the credibility of conflict, the credibility of life cycle status, and the credibility of interactive feedback, determine the comprehensive analysis credibility of each node; S80, the original texts corresponding to all nodes whose comprehensive analysis credibility is greater than the preset credibility threshold are used as the target corpus; S90 outputs the target corpus as the reasoning result through a large model.
[0132] It should be noted that the schemes protected by S10-S80 are consistent with the data acquisition method for enhancing large model inference disclosed in the first aspect of this application. For details of the specific implementation scheme, please refer to the detailed content disclosed in the first aspect of this application, so they will not be repeated here.
[0133] The method disclosed in this application enables enhanced credibility-aware retrieval, allowing large models to automatically identify credible claims, accurately separate authoritative from low-quality sources, and proactively resolve version conflicts in scenarios involving knowledge conflicts, multiple versions, and mixed low-quality sources. This effectively prevents misinformation from being amplified and significantly improves factual consistency, interpretability, and robustness.
[0134] Traditional RAG methods have limitations in handling dynamic, multi-source, and conflicting information. This application constructs a complex knowledge ecosystem that can effectively incorporate advanced technologies in real time. This ecosystem not only requires the ability to retrieve relevant content from massive amounts of information, but more importantly, it must be able to understand the source, version, and contextual dependencies of information, and evaluate community consensus and verification clues, i.e., perform deep structural inference. This ensures the accuracy and reliability of the information retrieval and generation process, and ultimately improves the application effect of large-scale language models in complex knowledge environments.
[0135] A computer device according to embodiments of the present disclosure includes a memory and a processor. The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.
[0136] The processor may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the computer device to perform desired functions. In one embodiment of this disclosure, the processor is used to execute computer-readable instructions stored in the memory, causing the computer device to perform all or part of the steps of the data acquisition method for enhanced large model inference or the enhanced RAG large model inference method described in the foregoing embodiments of this disclosure.
[0137] Those skilled in the art will understand that, in order to solve the technical problem of how to achieve a good user experience, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included within the protection scope of this disclosure.
[0138] like Figure 7 This is a schematic diagram of a computer device provided for an embodiment of the present disclosure. It illustrates a structural schematic diagram suitable for implementing the computer device in the embodiments of the present disclosure. Figure 7 The computer device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.
[0139] like Figure 7 As shown, a computer device may include a processor (such as a central processing unit, graphics processing unit, etc.), which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) or programs loaded from storage devices into random access memory (RAM). The RAM also stores various programs and data required for the operation of the computer device. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.
[0140] Typically, the following devices can be connected to the I / O interface: input devices, such as sensors or visual information acquisition devices; output devices, such as displays; storage devices, such as magnetic tapes or hard drives; and communication devices. Communication devices allow the computer device to communicate wirelessly or wiredly with other devices (such as edge computing devices) to exchange data. Although Figure 7 A computer apparatus with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.
[0141] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from ROM. When the computer program is executed by a processor, all or part of the steps of the data acquisition method for enhanced large model inference or the enhanced RAG large model inference method of the embodiments of this disclosure are performed.
[0142] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.
[0143] A computer-readable storage medium according to embodiments of the present disclosure stores non-transitory computer-readable instructions. When these non-transitory computer-readable instructions are executed by a processor, all or part of the steps of the data acquisition method for enhanced large model inference or the large model inference method for enhanced RAG described in the foregoing embodiments of the present disclosure are performed.
[0144] The aforementioned computer-readable storage media include, but are not limited to: optical storage media (e.g., CD-ROM and DVD), magneto-optical storage media (e.g., MO), magnetic storage media (e.g., magnetic tape or portable hard drive), media with built-in rewritable non-volatile memory (e.g., memory card), and media with built-in ROM (e.g., ROM cartridge).
[0145] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.
[0146] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.
[0147] In this disclosure, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The block diagrams of devices, apparatuses, devices, and systems involved in this disclosure are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as "comprising," "including," "having," etc., are open-ended terms meaning "including but not limited to," and are used interchangeably with them. The terms "or" and "and" as used herein refer to the terms "and / or," and are used interchangeably with them unless the context clearly indicates otherwise. The term "such as" as used herein refers to the phrase "such as but not limited to," and is used interchangeably with it.
[0148] Additionally, as used herein, the "or" used in a list of items beginning with "at least one" indicates a separate list, such that a list of, for example, "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not imply that the described example is preferred or better than other examples.
[0149] It should also be noted that in the systems and methods of this disclosure, the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions to this disclosure.
[0150] Various changes, substitutions, and modifications can be made to the technology described herein without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of this disclosure is not limited to the specific aspects of the processes, machines, manufactures, events, means, methods, and actions described above. Currently existing or later-developed processes, machines, manufactures, events, means, methods, or actions that perform substantially the same function or achieve substantially the same result as the corresponding aspects described herein can be utilized. Therefore, the appended claims include such processes, machines, manufactures, events, means, methods, or actions within their scope.
[0151] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.
[0152] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.
Claims
1. A data acquisition method for enhancing large model inference, characterized in that, include: Retrieve several related original texts from different search sources based on the target requirements; Each of the original text entries is split into several text blocks; A conflict graph is constructed based on all the text blocks; the conflict graph includes the relationship edges between each pair of the original texts and their strength weights, and each node in the conflict graph is one of the original texts; Determine the source confidence level of the retrieval source for each of the original texts; Based on the conflict map and all the source confidence scores, determine the conflict confidence score for each node; Determine the lifecycle status credibility of each node based on the stated target requirements; The credibility of the interaction feedback of each node is determined based on the interaction metadata corresponding to each node and the source confidence level. Based on the conflict credibility, the lifecycle state credibility, and the interaction feedback credibility, the comprehensive analysis credibility of each node is determined. The original text corresponding to all nodes whose comprehensive analysis credibility is greater than a preset credibility threshold is used as the target data for enhancing the large model inference; The step of determining the lifecycle state credibility of each node based on the target requirement includes: determining the lifecycle state corresponding to each node, wherein the lifecycle state includes the existing version time and the corresponding existing version description features; obtaining the current time and current version description features corresponding to the target requirement; determining the state validity type of each node based on the current version description features, the current time, and the lifecycle state; determining the corresponding state weight based on the state validity type; and determining the lifecycle state credibility of the node based on the state weight. The step of determining the credibility of the interaction feedback of each node based on the interaction metadata corresponding to each node and the source confidence score includes: obtaining the interaction metadata corresponding to each node, wherein the interaction metadata includes the number of approvals and the number of confirmations; determining the initial energy information of each node based on the source confidence score; and determining the credibility of the interaction feedback of each node based on the number of approvals, the number of confirmations, and the initial energy information.
2. The data acquisition method for enhancing large model inference according to claim 1, characterized in that, The process of retrieving several related original texts from different retrieval sources according to the target requirements includes: Determine the type of intent for the target need; The list of target sources corresponding to the intent type is determined according to the preset routing mapping table; Identify the key technical entities in the target requirements; Based on the target source list, retrieve several original texts associated with the key technology entity from the corresponding retrieval sources; The number of original text entries is not less than the number of sources of the retrieval source; Each original text entry includes at least source information, version information, publication time information, and interaction records.
3. The data acquisition method for enhancing large model inference according to claim 2, characterized in that, The step of splitting each original text into several text blocks includes: Each piece of original text is parsed to extract the plain text fragment containing the text content, remove embedded formatting tags and non-text elements, and obtain the target text; Each target text is divided into several discrete text blocks; Several discrete text blocks include at least version information and interaction records.
4. The data acquisition method for enhancing large model inference according to claim 1, characterized in that, The construction of the conflict graph based on all the text blocks includes: Based on all the text blocks, obtain the relationship comparison results between each pair of the original texts; the relationship comparison results are contradictory, competing, or parallel. Map the information claim corresponding to each original text to a node in the graph database; Based on the relationship comparison results, generate the relationship edges between each pair of nodes and their strength weights; The relation edges include negative edges, weak negative edges, or positive edges, and the strength weight corresponding to the negative edge is less than the strength weight corresponding to the weak negative edge, and the strength weight corresponding to the weak negative edge is less than the strength weight corresponding to the positive edge.
5. The data acquisition method for enhancing large model inference according to claim 4, characterized in that, The step of determining the conflict confidence level of each node based on the conflict map and all the source confidence levels includes: Configure an initial authority score for each node based on its origin. The relationship edges and their strength weights of the nodes to be analyzed are determined from the conflict graph. Based on the initial authority score and strength weight of the node corresponding to the relation edge of the node to be analyzed, determine the degree of conflict between the node corresponding to the relation edge of the node to be analyzed and the node to be analyzed. The conflict credibility of the node to be analyzed is obtained based on all the conflict levels and the initial authority score of the node to be analyzed.
6. A method for enhancing large-scale RAG inference, characterized in that, include: Retrieve several related original texts from different search sources based on the target requirements; Each of the original text entries is split into several text blocks; Construct a conflict graph based on all the aforementioned text blocks; The conflict graph includes the relationship edges between each pair of the original texts and their strength weights, and each node in the conflict graph is one of the original texts. Determine the source confidence level of the retrieval source for each of the original texts; Based on the conflict map and all the source confidence scores, determine the conflict confidence score for each node; Determine the lifecycle status credibility of each node based on the stated target requirements; The credibility of the interaction feedback of each node is determined based on the interaction metadata corresponding to each node and the source confidence level. Based on the conflict credibility, the lifecycle state credibility, and the interaction feedback credibility, the comprehensive analysis credibility of each node is determined. The original text corresponding to all nodes whose comprehensive analysis credibility is greater than the preset credibility threshold is used as the target data; The target data is output as the reasoning result through a large model; The step of determining the lifecycle state credibility of each node based on the target requirement includes: determining the lifecycle state corresponding to each node, wherein the lifecycle state includes the existing version time and the corresponding existing version description features; obtaining the current time and current version description features corresponding to the target requirement; determining the state validity type of each node based on the current version description features, the current time, and the lifecycle state; determining the corresponding state weight based on the state validity type; and determining the lifecycle state credibility of the node based on the state weight. The step of determining the credibility of the interaction feedback of each node based on the interaction metadata corresponding to each node and the source confidence score includes: obtaining the interaction metadata corresponding to each node, wherein the interaction metadata includes the number of approvals and the number of confirmations; determining the initial energy information of each node based on the source confidence score; and determining the credibility of the interaction feedback of each node based on the number of approvals, the number of confirmations, and the initial energy information.
7. A computer device, characterized in that, The computer device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to perform the data acquisition method for enhanced large model inference as described in any one of claims 1-5 or the large model inference method for enhanced RAG as described in claim 6.
8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a computer to perform the data acquisition method for enhanced large model inference as described in any one of claims 1-5 or the large model inference method for enhanced RAG as described in claim 6.
9. A computer program product comprising computer instructions, characterized in that, When executed by a processor, the computer instructions implement the data acquisition method for enhanced large model inference as described in any one of claims 1-5 or the large model inference method for enhanced RAG as described in claim 6.
Citation Information
Patent Citations
Software development data management method and system based on cloud service
CN120540692A
Semantic enhancement processing system based on large model agent RAG database
CN121072782A
