Retrieval enhancement type generation method and system, electronic equipment and storage medium

Abstract

The invention provides a retrieval enhancement type generation method and system, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence, the retrieval enhancement type generation method comprises the following steps: obtaining a current input text, and determining text information of the current input text and context information of the current input text in a historical input text; based on the text information and the context information, determining a retrieval path corresponding to the current input text; retrieving in a preset knowledge base based on the retrieval paths, and fusing the candidate documents retrieved under each retrieval path to obtain a document set corresponding to the current input text; and inputting the current input text and the document set corresponding to the current input text into the pre-training language model to obtain an output text generated by the pre-training language model. According to the method and the system provided by the invention, the accuracy, the response speed and the document matching quality of the large language model in a complex query scene are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a retrieval-enhanced generation method, system, electronic device, and storage medium. Background Art

[0002] With the widespread application of large language models (LLMs), they have demonstrated powerful natural language understanding and generation capabilities in areas such as question-answering systems, writing assistance, and automatic summarization. However, because pre-trained models contain fixed knowledge and are difficult to update in real time, they are prone to hallucinations and are difficult to adapt to application scenarios that require the latest information or specialized domain knowledge.

[0003] To address this issue, the Retrieval-Augmented Generation (RAG) method emerged. This method significantly improves the accuracy and credibility of the output by matching user input with relevant documents in a pre-set knowledge base and then incorporating the retrieval results into the model-assisted content generation. Related retrieval-augmented generation methods typically employ a single strategy, either keyword information inverted indexing or vector similarity retrieval. These strategies struggle to balance retrieval efficiency and semantic accuracy, especially in situations with dense unstructured text or ambiguous cross-domain semantics. This results in poor document matching, retrieval results that deviate from the question's semantics, and slow response times.

[0004] Therefore, how to improve the accuracy, response speed and document matching quality of large language models in complex query scenarios has become a technical problem that needs to be solved urgently in the industry. Summary of the Invention

[0005] The present invention provides a retrieval-enhanced generation method, system, electronic device, and storage medium for solving the technical problem of how to improve the accuracy, response speed, and document matching quality of large language models in complex query scenarios.

[0006] The present invention provides a retrieval-enhanced generation method, comprising: Acquire current input text, and determine text information of the current input text and context information of the current input text in historical input texts; Determining a search path corresponding to the current input text based on the text information and the context information; the search path includes at least one of a keyword search path, a topic cluster search path, and a semantic vector search path; Searching in a preset knowledge base based on the search path, fusing candidate documents retrieved under each search path to obtain a document set corresponding to the current input text; The current input text and a document set corresponding to the current input text are input into a pre-trained language model to obtain an output text generated by the pre-trained language model.

[0007] In some embodiments, determining the text information of the current input text and the context information of the current input text in the historical input text includes: Performing word segmentation and keyword extraction on the current input text to determine keyword information of the current input text; Performing semantic vectorization on the current input text to determine a semantic feature vector of the current input text; Performing topic modeling on the current input text to determine a topic distribution vector of the current input text; Determining text information of the current input text based on the keyword information, the semantic feature vector, and the topic distribution vector; Information is extracted from historical input texts associated with the current input text to generate context information of the current input text.

[0008] In some embodiments, determining the search path corresponding to the current input text based on the text information and the context information includes: Determining a query type of the current input text based on the text information and the context information; Based on the query type of the current input text, a search path and a search path scheduling strategy corresponding to the current input text are determined; the search path scheduling strategy includes a search path combination, a search path calling sequence, a search path concurrent execution relationship, and hardware resource configuration information.

[0009] In some embodiments, searching in a preset knowledge base based on the search path and fusing candidate documents retrieved under each search path to obtain a document set corresponding to the current input text includes: In the case where the search path includes a keyword search path, performing a reverse search in the preset knowledge base based on keyword information of the current input text to determine a first candidate document; In the case where the search path includes a topic cluster search path, performing similarity matching on the topic distribution vector of the current input text and the topic distribution vectors of each topic cluster in the preset knowledge base, and determining the document in the topic cluster with the highest topic distribution vector similarity as the second candidate document; In the case where the search path includes a semantic vector search path, performing similarity matching between the semantic feature vector of the current input text and the semantic feature vectors of each document in the preset knowledge base, and determining the document with the highest similarity as the third candidate document; The first candidate document, the second candidate document and the third candidate document are merged to obtain a document set corresponding to the current input text.

[0010] In some embodiments, the fusing of the first candidate document, the second candidate document, and the third candidate document to obtain a document set corresponding to the current input text includes: Determining a comprehensive quality score for each candidate document based on the semantic similarity, keyword coverage, and topic consistency between each candidate document and the current input text, the semantic consistency between each candidate document and the context information, and the document authority and timeliness of each candidate document; Based on the candidate documents whose comprehensive quality scores are greater than a preset score, a document set corresponding to the current input text is determined.

[0011] In some embodiments, the method further comprises: Obtaining user feedback information on the output text; the user feedback information includes at least one of user click behavior, page dwell time, user usage rating, secondary question behavior, and user error correction behavior; Determining the validity of each search path based on the user feedback information; When the validity of the keyword search path is invalid, updating the posting list of the preset knowledge base; In the case that the validity of the topic clustering retrieval path is invalid, each document in the preset knowledge base is re-clustered to obtain an updated topic cluster.

[0012] In some embodiments, the preset knowledge base is updated based on the following steps: Get the target document; Determining keyword information, semantic feature vectors, and topic distribution vectors of the target document; Based on the keyword information of the target document, generating an inverted index of the keyword information of the target document; Generating a semantic vector index of the target document based on the semantic feature vector of the target document; Generating a topic clustering index of the target document based on the topic distribution vector of the target document; The target document and the keyword information inverted index, semantic vector index and topic clustering index corresponding to the target document are added to the preset knowledge base.

[0013] The present invention provides a retrieval-enhanced generation system, comprising: An acquisition module, configured to acquire a current input text, determine text information of the current input text and context information of the current input text in historical input texts; a determination module, configured to determine a search path corresponding to the current input text based on the text information and the context information; the search path comprising at least one of a keyword search path, a topic cluster search path, and a semantic vector search path; A retrieval module, configured to search a preset knowledge base based on the retrieval path, merge candidate documents retrieved under each retrieval path, and obtain a document set corresponding to the current input text; The generation module is used to input the current input text and the document set corresponding to the current input text into the pre-trained language model to obtain the output text generated by the pre-trained language model.

[0014] The present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and operable on the processor, wherein the processor implements the retrieval-enhanced generation method when executing the computer program.

[0015] The present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the retrieval-enhanced generation method is implemented.

[0016] The retrieval-enhanced generation method, system, electronic device and storage medium provided by the present invention obtain the current input text, determine the text information of the current input text and the context information of the current input text in the historical input text; based on the text information and the context information, determine the retrieval path corresponding to the current input text; the retrieval path includes at least one of a keyword retrieval path, a topic clustering retrieval path and a semantic vector retrieval path; based on the retrieval path, search is performed in a preset knowledge base, and the candidate documents retrieved under each retrieval path are merged to obtain a document set corresponding to the current input text; the current input text and the document set corresponding to the current input text are input into a pre-trained language model to obtain an output text generated by the pre-trained language model; on the one hand, multiple retrieval paths are adopted, which can be dynamically selected according to the text information and context information of the input text, thereby realizing efficient document screening of a large-scale knowledge base; on the other hand, the context information of the input text is taken into consideration, and the user's intention can be fully understood, so that the document retrieval results are consistent with the semantics of the input text, significantly improving the relevance of document matching and the system response speed; and improving the accuracy, response speed and document matching quality of the large language model in complex query scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 This is one of the flow charts of the retrieval-enhanced generation method provided by the present invention.

[0020] Figure 2 This is one of the structural diagrams of the retrieval-enhanced generation system provided by the present invention.

[0021] Figure 3 This is the second structural diagram of the retrieval-enhanced generation system provided by the present invention.

[0022] Figure 4 This is the second flow chart of the retrieval-enhanced generation method provided by the present invention.

[0023] Figure 5 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0024] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0025] It should be noted that the terms "first," "second," and the like in the present invention are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having," as well as any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps, units, or modules is not necessarily limited to those steps, units, or modules that are explicitly listed, but may include other steps, units, or modules that are not explicitly listed or that are inherent to these processes, methods, products, or apparatuses.

[0026] The retrieval-enhanced generation method in related technologies has the following problems in large-scale knowledge base scenarios: (1) It is difficult for a single index structure to take into account both retrieval efficiency and semantic accuracy at the same time, especially when unstructured text is dense or cross-domain semantics are ambiguous, the document matching effect is poor; (2) It lacks context-awareness and cannot fully understand user intentions and historical conversation context, resulting in a deviation between retrieval results and question semantics; (3) The retrieval path is fixed and cannot dynamically select the optimal retrieval path based on input features, affecting response speed; (4) The system lacks an index self-evolution mechanism. When faced with incremental updates to the knowledge base or changes in hot spots, the index structure becomes redundant or outdated, and the maintenance cost is high; (5) The efficiency of data flow and computing resource scheduling is low, making it difficult to meet the service stability requirements in high-concurrency scenarios.

[0027] In order to solve the above technical problems, Figure 1 This is one of the flow charts of the search enhancement generation method provided by the present invention, such as Figure 1 As shown, the method includes step 110 , step 120 , step 130 and step 140 .

[0028] Step 110: Acquire the current input text, determine the text information of the current input text and the context information of the current input text in the historical input text.

[0029] Specifically, the retrieval-enhanced formula generation method provided in the embodiments of the present invention is implemented as a retrieval-enhanced formula generation system. This system can be implemented in software, such as a retrieval-enhanced formula generation program running on a computer, or in hardware, such as a computer or server that executes the retrieval-enhanced formula generation method.

[0030] The retrieval-enhanced generation method provided in this embodiment is suitable for vertical domain question-answering systems with complex semantics, high timeliness requirements, and high concurrent access, such as government decision support, legal compliance review, and medical knowledge assistance. The system can include a retrieval-enhanced generation system, a preset knowledge base, and a pretrained language model.

[0031] A pretrained language model is a large language model pre-trained using large amounts of text data. It can generate natural language text based on the input question text. A pre-set knowledge base is a database storing structured or semi-structured information, providing additional background knowledge and data support for the pre-trained language model. This helps the model better understand complex concepts, recent facts, and domain-specific information. The retrieval-enhanced generation system matches user input text with relevant documents in the pre-set knowledge base and then incorporates the search results to assist the pre-trained language model in generating content, significantly improving the accuracy and credibility of the output.

[0032] The current input text is the text content entered by the user at the current moment, which can be a query, a question, a message, or any other form of text input. The user can submit the input text in natural language form through the interface or interface.

[0033] Text information refers to the information obtained after parsing the current input text, which may include keywords, topics, named entities, intentions, emotional tendencies, etc.

[0034] Users may have multiple rounds of conversations around a topic in the system, that is, they will enter multiple texts in chronological order. Historical input text is the text entered by the user before the current moment.

[0035] Contextual information refers to the background information related to the current input text. It includes both the position and relationship of the current input text within the previous input text, as well as the textual information in the previous input text that is related to the current input text. Contextual information helps understand the meaning, intent, and relevance of the current input text.

[0036] After obtaining the current input text, on the one hand, natural language processing can be performed on the current input text to obtain text information; on the other hand, the similarity between the current input text and the historical input text can be calculated through the context window and embedding vector to determine the context relationship, and the context information can be determined based on the historical input text with the context relationship.

[0037] Step 120: Determine a search path corresponding to the current input text based on the text information and context information; the search path includes at least one of a keyword search path, a topic cluster search path, and a semantic vector search path.

[0038] Specifically, a search path refers to the search method and path selected by the system during the information retrieval process based on the input text and contextual information. Different search paths are suitable for different scenarios and needs, and can improve the accuracy and efficiency of retrieval.

[0039] In the embodiment of the present invention, the search path includes a keyword search path, a topic cluster search path, and a semantic vector search path.

[0040] Keyword search involves extracting keywords from input text and using them to search within a knowledge base or document collection. Keyword search is simple and fast, suitable for scenarios requiring precise matching and rapid document recall. Document recall is a core task in information retrieval, which involves retrieving documents relevant to a user's query from a large document collection.

[0041] The topic clustering retrieval approach analyzes the topic distribution of input text, categorizes it into predefined topic clusters, and then searches within those clusters. This approach captures semantic similarity and is suitable for coarse-grained matching based on complex topics.

[0042] The semantic vector retrieval path refers to converting the input text into a semantic vector and using vector similarity calculation to search in a knowledge base or document collection. It can capture semantic similarity and is suitable for scenarios that process complex semantics. It is used for fine-grained matching based on semantic understanding. Among them, the vector index in the semantic vector retrieval path can include semi-dense vector indexing and high-dimensional dense vector indexing. Semi-dense vector indexing is a retrieval method that combines dense vectors and sparse vectors. It improves the accuracy and coverage of search results by simultaneously utilizing the semantic matching capabilities of dense vectors and the keyword matching capabilities of sparse vectors. High-dimensional dense vector indexing is a technology for indexing and retrieving dense vectors of high-dimensional data. Dense vectors are usually generated by deep learning models and can capture the semantic features of text, images or other data.

[0043] In an embodiment of the present invention, a multi-dimensional index collaboration mechanism can be constructed in the retrieval-enhanced generation system to construct and integrate an index structure including a keyword retrieval path, a topic clustering retrieval path, and a semantic vector retrieval path (including semi-dense vectors and high-dimensional dense vectors).

[0044] Based on the text and context, the query type of the current input text can be determined. Query types can include explicit searches (e.g., legal document number searches), fuzzy queries (e.g., trending questions), and exploratory queries (e.g., open-ended recommendations). Based on the query type, at least one or more search paths are selected.

[0045] Step 130: Search in a preset knowledge base based on the search path, merge the candidate documents retrieved under each search path, and obtain a document set corresponding to the current input text.

[0046] Specifically, a search is performed in the preset knowledge base according to the selected search path. According to these search paths, the document recall task can be performed in parallel in the preset knowledge base to obtain the candidate documents that are most semantically relevant to the current input text.

[0047] The candidate documents retrieved under each search path are deduplicated, aggregated, scored, and sorted, and finally a document set corresponding to the current input text is obtained.

[0048] Step 140: Input the current input text and the document set corresponding to the current input text into the pre-trained language model to obtain the output text generated by the pre-trained language model.

[0049] Specifically, the current input text and the document collection can be input into a pre-trained language model, and the pre-trained language model generates the output text through reasoning.

[0050] The pre-trained language model receives the current input text and a document collection, uses the current input text as a prompt, performs semantic understanding on each document in the document collection, uses the Chain-of-Thought mechanism to perform multi-step reasoning, and gradually generates output text.

[0051] The pre-trained language model supports multiple rounds of generation and candidate answer evaluation mechanisms. If the quality of the output text generated for the first time is poor, the pre-trained language model will trigger a regeneration process, using different ideas and multi-prompt word sampling to generate multiple candidate texts, and then filter the optimal result through model evaluation and user preference scoring.

[0052] The retrieval-enhanced generation method provided by the embodiment of the present invention obtains the current input text, determines the text information of the current input text and the context information of the current input text in the historical input text; based on the text information and the context information, determines the retrieval path corresponding to the current input text; the retrieval path includes at least one of a keyword retrieval path, a topic clustering retrieval path and a semantic vector retrieval path; based on the retrieval path, searches are performed in a preset knowledge base, and the candidate documents retrieved under each retrieval path are merged to obtain a document set corresponding to the current input text; the current input text and the document set corresponding to the current input text are input into a pre-trained language model to obtain an output text generated by the pre-trained language model; on the one hand, multiple retrieval paths are adopted, which can be dynamically selected according to the text information and context information of the input text, thereby realizing efficient document screening of a large-scale knowledge base; on the other hand, the context information of the input text is taken into consideration, which can fully understand the user's intention, so that the document retrieval results are consistent with the semantics of the input text, significantly improving the relevance of document matching and the system response speed; and improving the accuracy, response speed and document matching quality of the large language model in complex query scenarios.

[0053] It should be noted that each embodiment of the present invention can be freely combined, the order can be changed, or it can be executed separately, and does not need to rely on or depend on a fixed execution order.

[0054] In some embodiments, determining text information of the current input text and context information of the current input text in historical input texts includes: Perform word segmentation and keyword extraction on the current input text to determine the keyword information of the current input text; Perform semantic vectorization on the current input text to determine the semantic feature vector of the current input text; Perform topic modeling on the current input text to determine the topic distribution vector of the current input text; Determine the text information of the current input text based on keyword information, semantic feature vectors, and topic distribution vectors; Information is extracted from historical input texts associated with the current input text to generate context information of the current input text.

[0055] Specifically, the current input text can be processed, with the main purpose of converting the user input from the original text form into a structured semantic representation. At the same time, it can also integrate multi-source information such as user portraits and current session context to provide a semantic basis for subsequent retrieval path planning.

[0056] Perform word segmentation and keyword extraction on the current input text to obtain keyword information for the current input text. Keyword extraction can include steps such as part-of-speech tagging, named entity recognition, and syntactic dependency analysis to identify high-information vocabulary, proper nouns, quantitative expressions, temporal and spatial scopes, and other language units in the input, and extract them from the input text as keywords.

[0057] You can call an embedding model (such as the BERT model) to perform semantic vectorization on the current input text to obtain a high-dimensional representation vector, which is used as the semantic feature vector of the current input text.

[0058] Topic modeling is used to discover hidden thematic structures in text data. It analyzes the distribution of words in a text and divides it into multiple topics. Each topic consists of a set of words with a high probability of appearing within that topic. Topic modeling can be performed on the input text using a generative model based on Bayesian statistics (Latent Dirichlet Allocation (LDA)) to determine the topic distribution vector for the input text. The topic distribution vector represents the probability distribution of the input text across different topics. It is a vector in which each element represents the probability of the input text being on a specific topic, allowing for inference of the topic areas or knowledge areas that the user may be interested in.

[0059] The keyword information, semantic feature vector and topic distribution vector can be used as the text information of the current input text.

[0060] It is also possible to check whether the current input text is in a multi-round dialogue. If so, the previous round of question and answer segments are extracted from the dialogue record and used as historical input text. A context summary vector is constructed through a summary algorithm or semantic compression technology and used as context information.

[0061] Furthermore, user profiles, including past query preferences, industry, and frequently visited topics, can be obtained to generate personalized search tags. These processing results will be used uniformly in subsequent route planning to achieve contextual awareness and personalized scheduling.

[0062] The retrieval-enhanced generation method provided by the embodiment of the present invention determines the text information of the current input text and the context information of the current input text in the historical input text, and can fully understand the user's intention so that the document retrieval results are consistent with the semantics of the input text.

[0063] In some embodiments, determining a search path corresponding to the current input text based on the text information and context information includes: Determine the query type of the current input text based on the text information and context information; Based on the query type of the current input text, the search path and search path scheduling strategy corresponding to the current input text are determined; the search path scheduling strategy includes search path combination, search path calling order, search path concurrent execution relationship and hardware resource configuration information.

[0064] Specifically, based on text information and contextual information, including user profiles, the query type of the current input text can be determined through rule matching and model discrimination. Query types include explicit retrieval, fuzzy questioning, and exploratory queries.

[0065] According to the query type of the current input text, combined with the topic span, keyword density and sentence length characteristics of the current input text, the search path corresponding to the current input text is determined, that is, at least one of the keyword search path, topic cluster search path and semantic vector search path is selected.

[0066] At the same time, the search path scheduling strategy is also determined. The search path scheduling strategy includes the search path combination, the search path calling sequence, the search path concurrent execution relationship and the hardware resource configuration information.

[0067] The search path combination refers to the type and number of search paths selected. The search path call order refers to the order in which the search paths are used. The search path concurrent execution relationship refers to whether the search paths execute document recall tasks concurrently. Hardware resource configuration information refers to the configuration of hardware resources, including whether to use cache indexes and whether to call for graphics processing unit (GPU) or field programmable gate array (FPGA) acceleration.

[0068] Historical query logs and user behavior feedback can also be used to conduct risk assessment and scoring of each search path. If historical data shows that a path has poor recall effect or high latency, the system will downgrade or replace it to ensure the dynamic rationality of path selection.

[0069] The search path scheduling strategy can be used as a structured scheduling instruction and handed over to the scheduling engine and multi-level indexing module for execution, entering the parallel search process.

[0070] The retrieval-enhanced generation method provided by the embodiment of the present invention dynamically selects the most appropriate retrieval path based on the text information, contextual information and user profile of the input text, thereby maximizing the accuracy and relevance of the recalled documents; it can not only identify different types of query tasks (explicit, fuzzy, exploratory), but also perform personalized scheduling based on user domain preferences, effectively improving the contextual relevance of the recalled documents and reducing interference from irrelevant information.

[0071] In some embodiments, a search is performed in a preset knowledge base based on a search path, and candidate documents retrieved under each search path are merged to obtain a document set corresponding to the current input text, including: In the case where the search path includes a keyword search path, performing a reverse search in a preset knowledge base based on keyword information of the current input text to determine a first candidate document; In the case where the search path includes a topic cluster search path, the topic distribution vector of the current input text is matched with the topic distribution vectors of each topic cluster in the preset knowledge base for similarity, and the document in the topic cluster with the highest topic distribution vector similarity is determined as the second candidate document; In the case where the search path includes a semantic vector search path, the semantic feature vector of the current input text is matched with the semantic feature vectors of each document in the preset knowledge base for similarity, and the document with the highest similarity is determined as the third candidate document; The first candidate document, the second candidate document and the third candidate document are fused to obtain a document set corresponding to the current input text.

[0072] Specifically, a hybrid retrieval module and a scheduling engine can be constructed. The hybrid retrieval module is used to support keyword retrieval paths, topic clustering retrieval paths, and semantic vector retrieval paths; the scheduling engine executes the retrieval path scheduling strategy and performs document recall tasks in parallel in the hybrid retrieval module to obtain candidate documents that are most relevant to the user's query semantics.

[0073] When the search path includes a keyword search path, a reverse index search is performed within the pre-set knowledge base based on the keyword information in the current input text to identify the top candidate document. For example, the system will perform a reverse index search on the core keywords extracted from the current input text, using a pre-built reverse index to locate relevant document sets and rank them based on keyword hit count, word frequency, and location information. Keyword search paths offer extremely high query efficiency and are suitable for precise questions with well-defined named entities and clear structures.

[0074] When the search path includes a topic cluster search path, the topic distribution vector of the current input text is matched against the topic distribution vectors of each topic cluster in the preset knowledge base. The document in the topic cluster with the highest topic distribution vector similarity is identified as the second candidate document. For example, the system uses the topic distribution vector of the input text to perform a similarity match against the topic distribution vectors of multiple predefined topic clusters in the preset knowledge base, searches for the one or more topic clusters with the highest similarity, and then quickly filters documents within the topic cluster. This approach is suitable for scenarios where a conceptual understanding of the problem is somewhat limited but keywords are weak, and can effectively improve semantic consistency.

[0075] When the search path includes a semantic vector search path, the semantic feature vector of the current input text is matched against the semantic feature vectors of each document in the preset knowledge base for similarity, and the document with the highest similarity is determined as the third candidate document. For example, for questions with ambiguous semantics or flexible expressions, the system can use the Approximate Nearest Neighbor (ANN) algorithm to match the semantic feature vector of the current input text with the semantic feature vectors of each document in the preset knowledge base for similarity, searching for the most semantically similar document in a sparse or dense high-dimensional vector space. If the query complexity is high or the recall accuracy requirements are strict, the system will automatically call on hardware resources such as processors for vector calculations and acceleration to reduce response latency and improve throughput.

[0076] The results of all search paths are returned to the scheduling engine cache in a streaming manner, waiting for subsequent fusion processing.

[0077] The retrieval enhancement generation method provided by the embodiment of the present invention constructs an index structure including a keyword retrieval path, a topic cluster retrieval path, and a semantic vector retrieval path (including semi-dense vectors and high-dimensional dense vectors), and flexibly schedules each retrieval path according to the input features, so that the system can accurately locate relevant documents and improve the semantic matching coverage when faced with complex semantic expressions, fuzzy descriptions or cross-domain problems. This design significantly makes up for the defects of traditional retrieval enhancement systems that rely only on a single vector index and have low recall rates when queries are ambiguous or mixed with hyponymous and hyponymous concepts. In addition, the combination of hardware acceleration devices to perform vector calculations and high-dimensional retrieval greatly improves the system throughput in high-load scenarios. This solution breaks through the problems of serial query and low hardware resource utilization of traditional retrieval enhancement generation methods, and achieves low latency, high concurrency, and high-precision knowledge retrieval capabilities, making it suitable for response-sensitive applications such as real-time question answering and intelligent search.

[0078] In some embodiments, the first candidate document, the second candidate document, and the third candidate document are merged to obtain a document set corresponding to the current input text, including: Determine the comprehensive quality score of each candidate document based on the semantic similarity, keyword coverage, and topic consistency between each candidate document and the current input text, the semantic consistency between each candidate document and the context information, and the document authority and timeliness of each candidate document; Based on the candidate documents whose comprehensive quality scores are greater than the preset scores, a document set corresponding to the current input text is determined.

[0079] Specifically, after completing multi-way document recall (i.e., obtaining the first, second, and third candidate documents), these candidate documents can be fused. The main tasks of this stage are to deduplicate, aggregate, score, and rank the candidate documents, selecting the most semantically relevant and contextually consistent document set for the final output text generation.

[0080] First, the system uniformly merges the first candidate document, the second candidate document, and the third candidate document, and uses document identification, summary signature, or semantic hash algorithm to remove duplicate content.

[0081] Next, the system activates a comprehensive scoring model to calculate the overall quality score for each document, taking into account multiple dimensions, including semantic similarity, keyword coverage, and topic consistency between each candidate document and the current input text; semantic consistency between each candidate document and contextual information; and each candidate document's authority and timeliness. This scoring process features adjustable weighting, allowing authority or timeliness to be prioritized in specific scenarios (such as legal and medical) to better meet business needs.

[0082] Semantic consistency between candidate documents and contextual information refers to the semantic similarity between each candidate document and the contextual information of the input text. This can be determined by comparing the similarity of semantic feature vectors. If the similarity exceeds a preset threshold, semantic consistency is considered. Especially in conversational, multi-round queries, the system compares candidate documents with contextual information to identify potential logical contradictions, conflicting positions, or documents with contextual drift. These documents are then removed or downgraded accordingly to avoid semantic misleading.

[0083] Finally, based on the candidate documents whose comprehensive quality scores are greater than a preset score, a document set corresponding to the current input text is determined. For example, multiple candidate documents with scores greater than a preset score can be selected based on the comprehensive quality scores and packaged in a structured form to generate a document set.

[0084] The retrieval-enhanced generation method provided by the embodiment of the present invention can screen out a document set with the greatest semantic relevance and contextual consistency for final generation by fusing candidate documents.

[0085] In some embodiments, the method further comprises: Obtaining user feedback information on the output text; the user feedback information includes at least one of user click behavior, page dwell time, user usage rating, secondary question behavior, and user error correction behavior; Determine the effectiveness of each search path based on user feedback information; When the validity of the keyword search path is invalid, the inverted list of the preset knowledge base is updated; When the validity of the topic clustering retrieval path is invalid, each document in the preset knowledge base is re-clustered to obtain an updated topic cluster.

[0086] Specifically, in order to achieve closed-loop optimization and long-term evolution capabilities of the system, the system continues to collect interactive behaviors and drive retrieval path updates and strategy adjustments after the user receives the generated content (output text).

[0087] After sending the output text to the user, user feedback information can be collected, including user click behavior, page stay time, user usage rating, secondary question behavior and user error correction behavior.

[0088] Based on user feedback, each search path can be evaluated, and the recall efficiency, document quality, generation satisfaction and other indicators of each search path can be analyzed to evaluate the effectiveness of each current search path.

[0089] If the keyword search path fails, the system will re-extract keyword information from each document in the pre-set knowledge base and update the postings list. The postings list is a crucial data structure in information retrieval systems, widely used in search engines, database indexing, and other scenarios. It enables fast keyword retrieval by mapping terms in a document to a list of documents containing those terms.

[0090] When the validity of the topic clustering retrieval path is invalid, the system will automatically trigger the clustering repartitioning or document reallocation process to re-cluster the documents in the preset knowledge base to obtain the updated topic clusters.

[0091] When the validity of the semantic vector retrieval path is invalid, the system will automatically trigger the re-extraction of semantic vectors for each document in the preset knowledge base.

[0092] The retrieval-enhanced generation method provided by the present invention incorporates a user-behavior-driven feedback learning and index evolution mechanism. During operation, the system continuously collects data metrics such as user satisfaction, click feedback, and bounce rate, and uses this data to automatically evaluate the effectiveness of search paths and strategies. For underperforming topic clusters, inefficient keywords, or invalid vector spaces, the system automatically performs cluster reconstruction, index updates, and weight adjustments. The system also possesses online optimization capabilities, enabling long-term, efficient operation and self-evolution.

[0093] In some embodiments, the preset knowledge base is updated based on the following steps: Get the target document; Determine the keyword information, semantic feature vector and topic distribution vector of the target document; Generate an inverted index of keyword information of the target document based on the keyword information of the target document; Generate a semantic vector index of the target document based on the semantic feature vector of the target document; Generate a topic clustering index of the target document based on the topic distribution vector of the target document; The target document and the keyword information inverted index, semantic vector index and topic clustering index corresponding to the target document are added to the preset knowledge base.

[0094] Specifically, the embodiment of the present invention also supports updating of a preset knowledge base. The target document can be a new document in a hot topic.

[0095] A target document can be obtained, and the keyword information, semantic feature vector, and topic distribution vector of the target document can be determined according to the method provided in the above embodiment. Then, a keyword information inverted index, a semantic vector index, and a topic cluster index can be generated respectively.

[0096] The keyword information inverted index is used for high-speed coarse screening. The topic cluster index is used for topic understanding and semantic classification. The semantic vector index includes a semi-dense vector index and a high-dimensional dense vector index. The semi-dense vector index is used in low-resource, high-recall scenarios, while the high-dimensional dense vector index is used for fine-grained semantic matching. Consistency is maintained between indexes through mapping tables or vector co-reference mechanisms. The index supports distributed deployment and asynchronous updates, adapting to the scalability needs of large-scale knowledge bases. It also supports index maintenance functions such as online updates and local reconstruction.

[0097] Finally, the target document and the keyword information inverted index, semantic vector index and topic clustering index corresponding to the target document are added to the preset knowledge base to update the preset knowledge base.

[0098] The retrieval-enhanced generation method provided by the embodiment of the present invention improves the document matching quality of a large language model in complex query scenarios by updating a preset knowledge base.

[0099] The following describes a system provided by an embodiment of the present invention. The system described below and the method described above can refer to each other.

[0100] Figure 2 This is one of the structural diagrams of the search enhancement generation system provided by the present invention. Figure 2 As shown, the system includes: An acquisition module 210 is configured to acquire a current input text and determine text information of the current input text and context information of the current input text in historical input texts; Determination module 220, for determining a search path corresponding to the current input text based on the text information and context information; the search path includes at least one of a keyword search path, a topic cluster search path, and a semantic vector search path; The retrieval module 230 is used to search the preset knowledge base based on the search path, merge the candidate documents retrieved under each search path, and obtain a document set corresponding to the current input text; The generation module 240 is configured to input the current input text and the document set corresponding to the current input text into the pre-trained language model to obtain the output text generated by the pre-trained language model.

[0101] The retrieval-enhanced generation system provided by the embodiment of the present invention obtains the current input text, determines the text information of the current input text and the context information of the current input text in the historical input text; determines the retrieval path corresponding to the current input text based on the text information and the context information; the retrieval path includes at least one of a keyword retrieval path, a topic clustering retrieval path and a semantic vector retrieval path; searches in a preset knowledge base based on the retrieval path, merges the candidate documents retrieved under each retrieval path, and obtains a document set corresponding to the current input text; inputs the current input text and the document set corresponding to the current input text into a pre-trained language model to obtain an output text generated by the pre-trained language model; on the one hand, multiple retrieval paths are adopted, which can be dynamically selected according to the text information and context information of the input text, thereby realizing efficient document screening of a large-scale knowledge base; on the other hand, the context information of the input text is taken into consideration, which can fully understand the user's intention, so that the document retrieval results are consistent with the semantics of the input text, significantly improving the relevance of document matching and the system response speed; and improving the accuracy, response speed and document matching quality of the large language model in complex query scenarios.

[0102] Figure 3 This is the second structural diagram of the search enhancement generation system provided by the present invention, such as Figure 3 As shown, the system includes an input processing module 310 , a context-aware routing module 320 , a multi-level hybrid indexing module 330 , a parallel retrieval scheduling module 340 , a document aggregation and generation module 350 , and a feedback learning and index evolution module 360 ​​.

[0103] The input processing module receives natural language input from users and performs multi-dimensional semantic and structural analysis. This module can include subunits such as a lexical analyzer, a named entity recognizer, a contextual summarizer, and a query intent discriminator. These subunits are responsible for extracting keywords, topic intent, semantic feature vectors, and historical conversation context from user input. The extracted semantic features are fed into the subsequent context-aware routing module to determine the optimal search path and serve as semantic cues in the final generation phase. This module can also integrate with user identity information, behavior logs, domain information, and other information to enable cross-session context modeling.

[0104] The context-aware routing module is the core decision-making unit of the entire system. It is responsible for dynamically planning the optimal search path and strategy based on the semantic features, contextual information, and historical behavior data output by the input processing module. It includes a feature amalgamator, query classifier, path selector, and strategy controller. It determines whether the current input is a named entity query, a fuzzy topic query, or a cross-domain comprehensive query, and then selects the path to enter the keyword indexing layer, the topic clustering layer, or the high-dimensional vector indexing layer. This module also has self-learning capabilities, continuously optimizing the path selection strategy based on search results and user feedback.

[0105] The multi-level hybrid indexing module is the indexing core of the knowledge base, supporting a combination of keyword inverted indexes, topic clustering indexes, semi-dense vector indexes, and high-dimensional dense vector indexes. When documents are stored, the system synchronously generates multiple levels of indexes: a keyword layer for high-speed coarse screening, a clustering layer for topic understanding and semantic classification, a semi-dense layer for low-resource, high-recall scenarios, and a high-dimensional dense layer for fine-grained semantic matching. Consistency is maintained between indexes through mapping tables or vector co-reference mechanisms. The module supports distributed deployment and asynchronous updates to accommodate the expansion needs of large-scale knowledge bases. It also supports index maintenance functions such as online updates and local reconstruction.

[0106] The parallel search scheduling module provides high-concurrency scheduling and resource management capabilities. It integrates a parallel query scheduler, a processor acceleration interface, a task dispatcher, and an asynchronous data pipeline engine. Based on the routing module's policy decisions, this module dispatches query tasks to the corresponding index subsystems, executing retrieval tasks in parallel to maximize computing resource utilization. Retrieval results are returned via non-blocking queues or caches, ensuring a low-latency, multi-path, and efficient recall process across index layers. Dynamic scheduling and reconfiguration are supported when the index structure or hardware resources change.

[0107] The document aggregation and generation module receives candidate document collections from each index layer and performs operations such as document deduplication, aggregation, sorting, and similarity ranking to ensure the accuracy, relevance, and high-quality of the resulting contextual information. The module includes components such as a candidate document aggregator, a semantic consistency checker, a document scorer, and generation triggers. The aggregated documents, along with the user's original input, are fed into a language model (such as a large language generation model) to generate answers. The module supports enhanced output methods such as thought chain generation, reference document embedding, and visual explanations. Different generation templates (such as question-and-answer, summary, and conversational) can be selected based on the content type.

[0108] The feedback learning and index evolution module is responsible for collecting user interaction behaviors (click-through rate, bounce rate, satisfaction rating, feedback correction, etc.) and performance indicators in the retrieval / generation process, and adjusting the index structure and routing strategy based on this feedback. The module includes a user behavior analyzer, a feedback collector, an index evolution controller, and an online clusterer. For documents with poor recall or high-frequency invalid documents, the system can automatically re-cluster, rebuild the index, or adjust the index weight; for popular topics and semantically drifting content, the system can dynamically optimize the topic cluster to ensure that the index structure maintains high performance in the long term. This module also supports grayscale index updates and version rollback mechanisms to ensure the stability and controllability of the system's online evolution.

[0109] Figure 4 This is the second flow chart of the search-enhanced generation method provided by the present invention, such as Figure 4 As shown, the modules in the above embodiment cooperate with each other to perform the following steps: Step 410: Construct a multi-level hybrid index structure, including: a keyword inverted index layer, a topic clustering layer, a semi-dense vector index layer, and a high-precision dense vector index layer; Step 420: Based on the query statement input by the user, its semantic features, keyword structure and historical context are extracted and input into the context-aware routing module to determine the adapted search path and candidate index layer combination; Step 430: Utilize the index structure to execute the search task in parallel in multiple index layers, and optimize the search execution efficiency through a hardware acceleration module (such as a GPU or FPGA); Step 440: De-duplicate, aggregate, and sort the candidate documents returned by each search channel based on similarity to obtain a document set that best matches the user's query semantics. Step 450: The document set and the user input are jointly input into the language generation model to generate the final reply content; Step 460: Collect user feedback behavior, and incrementally optimize the topic clustering structure and vector index weights based on the retrieval effect evaluation mechanism to achieve index self-repair and structure evolution.

[0110] This method achieves efficient document screening in large-scale knowledge bases by introducing a multi-level index structure that combines sparse and dense vectors, coupled with online topic clustering and index evolution mechanisms. By building a context-aware routing and scheduling module, it adaptively selects the optimal retrieval path based on the semantic characteristics of user queries, significantly improving the relevance of document matching and system response speed. Furthermore, it utilizes asynchronous parallel retrieval and hardware acceleration to achieve low-latency document retrieval in high-concurrency scenarios. This approach is particularly suitable for question-and-answer systems in vertical fields with complex semantics, high timeliness requirements, and high concurrent access, such as government decision support, legal compliance review, and medical knowledge assistance, and possesses high engineering application value and innovation.

[0111] Figure 5 Schematic diagram of the structure of the electronic device provided by the present invention, such as Figure 5 As shown, the electronic device may include: a processor (Processor) 510, a communication interface (Communications Interface) 520, a memory (Memory) 530 and a communication bus (Communications Bus) 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other via the communication bus 540. The processor 510 may call the logic commands in the memory 530 to execute the method described in the above embodiments, for example: Obtain the current input text, determine the text information of the current input text and the context information of the current input text in the historical input text; based on the text information and the context information, determine the retrieval path corresponding to the current input text; the retrieval path includes at least one of a keyword retrieval path, a topic clustering retrieval path and a semantic vector retrieval path; based on the retrieval path, search in a preset knowledge base, fuse the candidate documents retrieved under each retrieval path, and obtain a document set corresponding to the current input text; input the current input text and the document set corresponding to the current input text into a pre-trained language model to obtain an output text generated by the pre-trained language model.

[0112] Furthermore, the logical commands in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several commands for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0113] The processor in the electronic device provided by the embodiment of the present invention can call the logic instructions in the memory to implement the above method. Its specific implementation method is consistent with the implementation method of the above method and can achieve the same beneficial effects, which will not be repeated here.

[0114] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method provided in the above embodiments is implemented.

[0115] Its specific implementation is consistent with the aforementioned method implementation and can achieve the same beneficial effects, so it will not be repeated here.

[0116] An embodiment of the present invention provides a computer program product, including a computer program. When the computer program is executed by a processor, the method described above is implemented.

[0117] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0118] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A search-enhanced generation method, characterized in that: include: Acquire current input text, and determine text information of the current input text and context information of the current input text in historical input texts; Determining a search path corresponding to the current input text based on the text information and the context information; the search path includes at least one of a keyword search path, a topic cluster search path, and a semantic vector search path; Searching in a preset knowledge base based on the search path, fusing candidate documents retrieved under each search path to obtain a document set corresponding to the current input text; The current input text and a document set corresponding to the current input text are input into a pre-trained language model to obtain an output text generated by the pre-trained language model.

2. The search-enhanced generation method according to claim 1, characterized in that: The determining of the text information of the current input text and the context information of the current input text in the historical input texts includes: Performing word segmentation and keyword extraction on the current input text to determine keyword information of the current input text; Performing semantic vectorization on the current input text to determine a semantic feature vector of the current input text; Performing topic modeling on the current input text to determine a topic distribution vector of the current input text; Determining text information of the current input text based on the keyword information, the semantic feature vector, and the topic distribution vector; Information is extracted from historical input texts associated with the current input text to generate context information of the current input text.

3. The search-enhanced generation method according to claim 1, characterized in that: The determining, based on the text information and the context information, a search path corresponding to the current input text includes: Determining a query type of the current input text based on the text information and the context information; Based on the query type of the current input text, a search path and a search path scheduling strategy corresponding to the current input text are determined; the search path scheduling strategy includes a search path combination, a search path calling sequence, a search path concurrent execution relationship, and hardware resource configuration information.

4. The search-enhanced generation method according to claim 1, wherein: The searching in the preset knowledge base based on the search path and fusing the candidate documents retrieved under each search path to obtain a document set corresponding to the current input text include: In the case where the search path includes a keyword search path, performing a reverse search in the preset knowledge base based on keyword information of the current input text to determine a first candidate document; In the case where the search path includes a topic cluster search path, performing similarity matching on the topic distribution vector of the current input text and the topic distribution vectors of each topic cluster in the preset knowledge base, and determining the document in the topic cluster with the highest topic distribution vector similarity as the second candidate document; In the case where the search path includes a semantic vector search path, performing similarity matching between the semantic feature vector of the current input text and the semantic feature vectors of each document in the preset knowledge base, and determining the document with the highest similarity as the third candidate document; The first candidate document, the second candidate document and the third candidate document are merged to obtain a document set corresponding to the current input text.

5. The search-enhanced generation method according to claim 4, characterized in that: The fusing of the first candidate document, the second candidate document, and the third candidate document to obtain a document set corresponding to the current input text includes: Determining a comprehensive quality score for each candidate document based on the semantic similarity, keyword coverage, and topic consistency between each candidate document and the current input text, the semantic consistency between each candidate document and the context information, and the document authority and timeliness of each candidate document; Based on the candidate documents whose comprehensive quality scores are greater than a preset score, a document set corresponding to the current input text is determined.

6. The search-enhanced generation method according to any one of claims 1 to 5, characterized in that: The method further comprises: Obtaining user feedback information on the output text; the user feedback information includes at least one of user click behavior, page dwell time, user usage rating, secondary question behavior, and user error correction behavior; Determining the validity of each search path based on the user feedback information; When the validity of the keyword search path is invalid, updating the posting list of the preset knowledge base; In the case that the validity of the topic clustering retrieval path is invalid, each document in the preset knowledge base is re-clustered to obtain an updated topic cluster.

7. The search-enhanced generation method according to any one of claims 1 to 5, characterized in that: The preset knowledge base is updated based on the following steps: Get the target document; Determining keyword information, semantic feature vectors, and topic distribution vectors of the target document; Based on the keyword information of the target document, generating an inverted index of the keyword information of the target document; Generating a semantic vector index of the target document based on the semantic feature vector of the target document; Generating a topic clustering index of the target document based on the topic distribution vector of the target document; The target document and the keyword information inverted index, semantic vector index and topic clustering index corresponding to the target document are added to the preset knowledge base.

8. A search-enhanced generation system, characterized in that: include: An acquisition module, configured to acquire a current input text, determine text information of the current input text and context information of the current input text in historical input texts; a determination module, configured to determine a search path corresponding to the current input text based on the text information and the context information; The search path includes at least one of a keyword search path, a topic cluster search path, and a semantic vector search path; A retrieval module, configured to search a preset knowledge base based on the retrieval path, merge candidate documents retrieved under each retrieval path, and obtain a document set corresponding to the current input text; The generation module is used to input the current input text and the document set corresponding to the current input text into the pre-trained language model to obtain the output text generated by the pre-trained language model.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the retrieval-enhanced generation method according to any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the retrieval-enhanced generation method according to any one of claims 1 to 7 is implemented.

Images