An Adaptive Hybrid Retrieval Method and System for Multidimensional Spatial Data
By using an intent-aware adaptive hybrid retrieval method, the fusion weights are dynamically adjusted and computational resources are optimized, which solves the problems of rigid fusion strategies and redundant fine-ranking computation in existing technologies, thereby improving retrieval efficiency and accuracy.
Patent Information
- Application Number
- CN202611204719.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-10
- Publication Date
- 2026-11-10
AI Technical Summary
Existing multi-path hybrid retrieval schemes cannot adaptively adjust fusion weights when faced with different types of queries, resulting in a decrease in recall rate; fine-ranking computation is costly and lacks a lightweight, interpretable dynamic adjustment mechanism, resulting in a waste of computing resources.
By introducing an intent-aware adaptive hybrid retrieval method, the intent score is calculated using query feature extraction, the fusion weight is dynamically adjusted, and a threshold truncation mechanism based on intent confidence is introduced to optimize coarse and fine ranking calculations and reduce unnecessary inference overhead.
It implements a refined adaptive retrieval strategy based on query intent, which improves the accuracy and recall of retrieval results, while reducing computational costs and latency. It also features lightweight and interpretable feature extraction and computation methods.
Smart Images

Figure CN123068596A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of information retrieval and natural language processing, specifically to an adaptive hybrid retrieval method and system for multidimensional spatial data. This invention is particularly applicable to scenarios involving multi-path hybrid retrieval of large-scale knowledge bases in Retrieval-Augmented Generation (RAG) systems. It aims to precisely control retrieval strategies and improve the accuracy, recall, and computational efficiency of retrieval results by using adaptive fusion weights based on query intent awareness and a threshold truncation mechanism based on intent confidence. Background Technology
[0002] With the rise of Large Language Models (LLM), Relational Anchoring (RAG) technology has become a key solution to address model illusion and knowledge gaps. In a RAG system, knowledge base documents are segmented into text chunks, transformed into high-dimensional vectors using an embedding model, and stored in a vector database. The system retrieves relevant documents from the knowledge base to assist the Large Language Model in generating more accurate and traceable answers. The quality of the retrieval stage directly determines the reliability and accuracy of the generated answers.
[0003] Current mainstream RAG retrieval schemes have evolved from "single-path vector retrieval" to "multi-path hybrid retrieval." Single-path vector retrieval maps queries and documents to a unified semantic space through an embedding model, and performs matching through cosine similarity or dot product operations. This scheme performs well in semantic generalization, capturing deep conceptual relationships between queries and documents. However, when faced with queries that require precise matching, such as technical terms, specific numbers, or filenames, it is prone to "semantic drift"—documents that should be precisely matched are not recalled because their semantic distance in the vector space is too large.
[0004] Multi-path hybrid retrieval technology was proposed to address the problems of "semantic drift" and "missing exact matches." It achieves more comprehensive recall coverage by parallelly executing dense retrieval based on vector similarity (Dense Retrieval) and sparse retrieval based on keyword matching (Sparse Retrieval, such as the BM25 algorithm). However, with the increasing complexity of application scenarios and the diversification of user query types, existing multi-path hybrid retrieval solutions face the following three serious challenges in practical engineering implementation:
[0005] Challenge 1: Coarse fusion and ranking, neglecting query intent, and static, fixed retrieval strategies. The fusion of multi-path recall results often relies on static weights or simple linear interpolation, failing to adaptively adjust based on query intent type. Specifically, different types of queries (e.g., exact match queries and semantic understanding queries) have drastically different degrees of dependence on dense and sparse retrieval: exact match queries containing explicit filenames and numbers are more suitable for keyword retrieval (BM25), while open-ended, conceptual semantic understanding queries are more suitable for semantic vector retrieval (FAISS). Existing systems use a "one-size-fits-all" fixed fusion parameter (e.g., fixed α=0.5), failing to adaptively adjust the weight distribution of the two retrieval paths in the fusion based on query features, leading to a decrease in query recall for specific intents. For example, for exact matches containing filenames, if a fixed vector retrieval is applied...
[0006] Excessive weighting may lead to the omission of target files due to insufficient semantic matching, resulting in a serious degradation of search quality.
[0007] Challenge 2: Fine-grained ranking inference is computationally expensive, and there is a lack of computational cost control mechanisms based on intent confidence. While cross-encoder-based re-ranking models offer high accuracy, their inference costs are exorbitant—each inference requires concatenating the query and candidate documents and performing forward computation through a Transformer network. Existing solutions perform full fine-grained ranking on a fixed number of candidate documents (e.g., Top-50) regardless of the query intent. For highly specific, exact-match queries (e.g., direct queries containing complete filenames), the top documents in the coarse-grained ranking stage already possess extremely high relevance confidence, making full fine-grained ranking of a large number of candidate documents a significant waste of computational resources.
[0008] Challenge 3: Dynamic weight adjustment schemes based on Large Language Models (LLMs) are computationally expensive, and there is a lack of lightweight, interpretable alternatives. Some existing solutions attempt to dynamically adjust the fusion weights by evaluating the matching degree between queries and retrieval results using LLMs. While this achieves high accuracy, each query introduces additional LLM inference overhead (typically hundreds of milliseconds to several seconds), leading to uncontrollable latency in high-concurrency scenarios. Furthermore, the black-box nature of LLMs results in a lack of interpretability and debuggability in the weight adjustment process. Existing technologies lack a lightweight, parameterized, interpretable dynamic weight adjustment mechanism that does not rely on online large model inference.
[0009] To address the above challenges, there is an urgent need for an adaptive retrieval strategy based on query intent awareness. This strategy should employ lightweight and interpretable feature extraction and numerical computation methods to achieve continuous adaptive adjustment of fusion weights according to query intent. Simultaneously, a coarse-ranking threshold truncation mechanism based on intent confidence should be introduced to automatically match the computational cost of fine-ranking with the query intent confidence, thereby improving computational efficiency while ensuring retrieval accuracy. Summary of the Invention
[0010] This invention provides an adaptive hybrid retrieval method and system for multidimensional spatial data to solve the problems of rigid fusion strategies, redundant fine-ranking calculations, and high costs of dynamic adjustment in existing hybrid retrieval methods.
[0011] To achieve the above objectives, the present invention adopts the following technical solution:
[0012] An adaptive hybrid retrieval method for multidimensional spatial data, characterized by the following steps:
[0013] S1. Obtain the query text input by the user, extract features from the query text, and obtain a query feature set containing at least one feature;
[0014] S2. Calculate the intent score based on the set of query features, whereby the intent score is used to characterize the probability that the user query tends to be an exact match;
[0015] S3. Calculate the fusion weight dynamically based on the intent score. The fusion weight includes dense retrieval weight and sparse retrieval weight.
[0016] S4. Perform dense retrieval based on vector similarity and sparse retrieval based on keyword matching in parallel to obtain the dense retrieval score set and the sparse retrieval score set respectively;
[0017] S5. Normalize the dense retrieval score set and the sparse retrieval score set respectively to obtain the normalized dense score set and the normalized sparse score set.
[0018] S6. Based on the fusion weight, the normalized dense retrieval score and the normalized sparse retrieval score are weighted and fused to obtain the fusion ranking score of each document, and the documents are sorted from high to low according to the fusion ranking score to determine the coarse ranking candidate set.
[0019] S7. Based on the comparison result between the intent score and the preset precision threshold, perform dynamic truncation based on intent to determine the number of candidate documents to be sent for fine ranking. Based on the number of candidate documents in the fine ranking stage, truncate the coarse ranking candidate set to obtain the fine ranking candidate set.
[0020] S8. Based on the refined candidate set, the candidate documents are sent to the cross-encoder model for refined ranking and reasoning one by one, and the final retrieval results are output in descending order of the refined ranking score.
[0021] Furthermore, in step S1, the query feature set includes one or more of the following features:
[0022] feature (Query Length Ratio): Based on the characteristics of query text length calculation, a smooth mapping function is used to map the number of characters in the query text to the [0,1] interval. The input of the mapping function includes the number of characters in the query text and a preset center length parameter.
[0023] feature (Keyword density): Calculates the proportion of content words in the query to the total number of words, mapped to the [0,1] interval. The content words are words excluding the preset set of functional words;
[0024] feature (File search signal strength): Calculated based on the combination of specific symbols, file extensions or specific verbs contained in the query. Multiple signals are accumulated and scored, and the value range is mapped to [0,1].
[0025] feature (Interrogative Sentence Signal): Calculated based on the combination of interrogative words, interrogative symbols, or interrogative tone words contained in the query. Multiple signals are accumulated and scored, with the value range mapped to [0,1].
[0026] feature (Embedded semantic similarity difference): Calculate the first similarity between the query vector and the preset precise prototype vector, and the second similarity between the query vector and the preset semantic prototype vector, based on the difference between the first similarity and the second similarity.
[0027] feature (Proportion of English technical terms): Calculated based on the ratio of the number of English words to the number of Chinese characters in the query.
[0028] Furthermore, the calculation of the intent score based on the query feature set specifically includes: intent score. The calculation uses a linear weighted clipping formula:
[0029]
[0030] In the formula, For the first Preset weight coefficients for each feature; It is a bias term used to adjust the baseline offset of the overall intention score. This is used to limit the results to a preset range. The range of is a continuous interval (0, 1). →0 indicates that the query tends to be based on pure semantic understanding. →1 indicates that the query tends to be a pure exact match.
[0031] Furthermore, the dynamically calculated dense retrieval weights The formula is:
[0032]
[0033]
[0034] In the formula, The base weight represents the initial trust ratio assigned to the two search paths when the intent is unclear. To adjust the amplitude factor, the value range is greater than 0, which is used to control the intensity of the influence of the intent score on the final weight; For the boundary truncation function, ensure The value range is always between [0, 1].
[0035] Further, S5 specifically includes: performing Min-Max normalization on the dense retrieval score and the sparse retrieval score respectively, mapping the scores to the (0,1) interval to eliminate the difference in units. Normalizing the dense retrieval score The calculation formula is:
[0036]
[0037] In the formula, d represents the specific document for which the score is to be calculated; Represents any document in the candidate set C; C is the set of candidate documents in the current coarse ranking stage; This is a very small positive number, used to prevent calculation anomalies caused by a zero denominator when all candidate documents in the set have the same score. Normalized sparse search score. The formula is completely symmetrical to the one described above.
[0038] Furthermore, S6 specifically includes: calculating the fusion ranking score according to the formula:
[0039]
[0040] In the formula, and These are the normalized dense retrieval score and the normalized sparse retrieval score for document d, respectively. This refers to the intensive retrieval weights dynamically calculated in S3. This corresponds to the sparse retrieval weights. The system assigns weights to all candidate documents based on... Sort the documents from highest to lowest quality, and select the top N documents after sorting to form a coarse candidate set. N is the preset upper limit of the capacity of the coarse-ranked candidate set.
[0041] Furthermore, in step S7, the dynamic truncation based on intent scoring specifically includes the following steps:
[0042] S7a, Determine the intent score Is it greater than the preset precision threshold? ;in The preset intent confidence threshold value for the system indicates the level of confidence the system determines in the current query. The above probabilities are for triggering the truncation strategy during exact match queries.
[0043] S7b, when When the query is determined to be a high-confidence exact match, the number of candidates M for fine ranking is determined according to the following formula:
[0044]
[0045] In the formula, The default number of candidates for fine-grained ranking, representing the standard fine-grained ranking window size for semantically or ambiguous queries; This is a lower bound on the number of candidates for fine-grained ranking, used to ensure that even for the most confident exact query, the fine-grained ranking window still contains at least [number missing]. 10 candidate documents to avoid outputting empty results in extreme cases; This is the floor function.
[0046] S7c, when When (i.e., semantically understandable or ambiguous intent queries), the default number of fine-ranking candidates remains unchanged. .
[0047] From the coarse-ranked candidate set The first M documents are extracted to form a candidate set for fine-grained ranking. .
[0048] The technical effect of this threshold truncation mechanism is as follows: For high-confidence, exact-match queries (such as queries containing explicit filenames and file extensions), the top-ranked documents in the coarse-rank fusion sort already have extremely high relevance confidence, eliminating the need for expensive cross-encoder inference on a large number of candidate documents. The M-value is automatically reduced through a formula to decrease computational overhead. For semantically understanding queries, the complete fine-ranking window is maintained to fully explore semantic relationships. Thus, an adaptive match is achieved between query intent confidence and fine-ranking computational overhead.
[0049] Furthermore, in step S8, the step of feeding candidate documents into the cross-encoder model for fine-grained ranking and inference one by one specifically includes: processing the fine-grained candidate set... Candidate documents are fed one by one into a cross-encoder model (such as BGE-Reranker) for deep inference scoring to generate a refined ranking score. The calculation formula is as follows:
[0050] ,
[0051] In the formula, This means concatenating the user query q with the candidate document d into a unified input sequence; W represents the hidden layer vector of the [CLS] token output by the cross-encoder model after processing the input sequence; W and B are the weight and bias parameters of the fully connected output layer of the cross-encoder model. These parameters are fixed parameters learned during the model pre-training and fine-tuning stages and remain unchanged during the inference stage.
[0052] Score by ranking The candidate documents are sorted in descending order, and the final Top-K documents are extracted as the search results.
[0053] This invention also provides an adaptive hybrid retrieval system for multidimensional spatial data, characterized by comprising a unified retrieval interface layer, a dynamic routing scheduling layer, a multi-path recall execution layer, and a coarse and fine ranking processing layer, wherein:
[0054] The unified retrieval interface layer is used to receive user queries, encapsulate the query text into a retrieval context object, and send the retrieval context object to the dynamic routing scheduling layer.
[0055] The dynamic routing scheduling layer includes an intent parsing module and a dynamic weight calculation module: the intent parsing module is configured to obtain query text from the retrieval context object, execute the steps described in S1 and S2 of claim 1, extract a set of query features, and calculate an intent score. The dynamic weight calculation module is configured to perform the step S3 in claim 1, according to the... Preset basic weights Dynamically calculate dense retrieval weights using the adjustment amplitude factor λ. ;
[0056] The multi-path recall execution layer includes a dense retrieval engine and a sparse retrieval engine: the dense retrieval engine is configured to encode the query text into a high-dimensional vector and perform an approximate nearest neighbor search; the sparse retrieval engine is configured to perform a keyword matching algorithm based on an inverted index; the dense retrieval engine and the sparse retrieval engine are configured to be invoked in parallel by the dynamic routing scheduling layer through a concurrency mechanism;
[0057] The coarse and fine ranking processing layer includes a coarse ranking fusion module, a threshold truncation module, and a fine ranking engine. The coarse ranking fusion module is configured to execute the steps described in S5 and S6 of claim 1, normalizing and weighting the two retrieval scores to generate a coarse ranking candidate set. The threshold truncation module is configured to execute the intent-based truncation step described in claim 7, based on the intent score. With preset precise threshold The comparison results determine the candidate set for fine ranking; the fine ranking engine is configured to load the cross-encoder model, perform deep reasoning scoring on the candidate set for fine ranking, and output the final retrieval result.
[0058] Furthermore, the system also includes a configuration center module, which establishes communication connections with the intent parsing module, the dynamic weight calculation module, and the threshold truncation module, respectively. The configuration center module is configured to store and manage the set of parameters required for the operation of the intent parsing module, the dynamic weight calculation module, and the threshold truncation module. The set of parameters includes feature weight coefficients, bias terms, basic weights, adjustment amplitude factors, precise thresholds, and truncation quantity parameters.
[0059] The intent parsing module, the dynamic weight calculation module, and the threshold truncation module are all configured to read the latest parameter values from the parameter set in real time from the configuration center module, so as to achieve hot updates of algorithm strategies without restarting the service.
[0060] Furthermore, the system includes a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the method as described in any one of claims 1 to 8.
[0061] Furthermore, as described in S2, the calculation of the embedding semantic similarity difference of f5 specifically includes:
[0062] A precise prototype vector set and a semantic prototype vector set are pre-constructed; the precise prototype vector set contains several texts used to represent precise retrieval intent, and the semantic prototype set contains several texts used to represent semantic understanding intent.
[0063] The precise prototype set and the semantic prototype set are encoded into vectors using an embedding model, and the center vectors are calculated to obtain the precise matching feature vectors. and semantic understanding feature vectors ;
[0064] Encode the user query text q into a query vector. Calculate the query vector respectively The exact matching feature vector The semantic understanding feature vector The cosine similarity is used to obtain the first and second similarities;
[0065] Calculate the original difference based on the difference between the first similarity and the second similarity. ;
[0066] By normalization function Mapped to the [0,1] interval, 0 corresponds to exact matching, and 1 corresponds to semantic understanding. This normalized value is used as a feature. Participating in the intent scoring The calculation.
[0067] Compared with the prior art, the present invention has the following beneficial effects:
[0068] The intent parsing module introduces an intent evaluation function. The query intent is transformed from a discrete "rule classification" into a continuous numerical control within the (0, 1) interval; the dynamic weight calculation module uses a dynamic weight calculation formula. The intent score is smoothly mapped to the fusion weight; the retrieval strategy is finely and adaptively adjusted according to the query intent, which effectively solves the problems of accuracy fluctuation and semantic drift when the fixed weight strategy is used to handle different types of queries.
[0069] In the coarse and fine ranking processing layer, the threshold truncation module performs a coarse ranking threshold truncation mechanism based on the intent confidence, so that the fine ranking computation cost is automatically matched with the query intent confidence. For high-confidence precise queries, the system automatically reduces the fine ranking window to significantly reduce computation costs. For semantic queries, the full fine ranking scale is maintained. The system significantly improves computational efficiency and overall throughput while ensuring accuracy.
[0070] In the computation link of the dynamic routing and scheduling layer, a rule-driven mechanism based on closed mathematical operations is adopted. The weight calculation can be completed without relying on online large language model inference, avoiding the risks of high latency and uninterpretability. The configuration center centrally manages all parameters and supports hot updates, A / B testing and fast rollback of algorithm strategies. The system can flexibly adapt to the business needs of different vertical fields and has engineering portability. Attached Figure Description
[0071] Figure 1 This is a flowchart illustrating the steps of the adaptive hybrid retrieval method and system for multidimensional spatial data according to the present invention.
[0072] Figure 2 This is a schematic diagram of the link of the adaptive hybrid retrieval system for multidimensional spatial data according to the present invention;
[0073] Figure 3This is a schematic diagram of the architecture of the adaptive hybrid retrieval system for multidimensional spatial data according to the present invention. Detailed Implementation
[0074] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0075] Example 1
[0076] This embodiment applies to an enterprise-level knowledge base retrieval scenario, specifically describing the complete implementation process of an adaptive hybrid retrieval method for multi-dimensional spatial data. The core of this embodiment lies in constructing an intent-aware model, dynamically adjusting the fusion weights of the hybrid retrieval using query features, and adaptively controlling the fine-ranking scale based on intent confidence, thereby addressing the accuracy loss and computational resource waste issues caused by fixed-weight strategies.
[0077] The system receives the user-input query text q through a unified retrieval interface layer, sends it to the dynamic routing and scheduling layer via a retrieval context object, and extracts features from the query text through the intent parsing module. The feature query set includes one or more of the following features:
[0078] feature (Query length ratio): The query length is mapped to the [0,1] interval using the Sigmoid smoothing mapping function. The calculation formula is as follows:
[0079]
[0080] In the formula, L is the number of characters in the query text (including Chinese and English), r is the number of characters at the center point, and u is the width of the temperature parameter control transition zone. In this embodiment, r=30 and u=8 are preferably set. Physical meaning: Short queries are mostly exact searches in the form of keywords / filenames, while long queries are mostly semantic queries in the form of natural language. f1 takes the value [0,1], with 0 corresponding to long queries (partially semantic) and 1 corresponding to short queries (partially exact).
[0081] feature (Keyword density): Extracting lexical units (Chinese and English words) from the query using regular expressions, and calculating the proportion of non-functional words (content words) among them. Physical meaning: High-density queries primarily consist of independent nouns / terms (e.g., "Article 52 of the Contract Law is invalid"), leaning towards exact matching; low-density queries primarily consist of natural language sentences (e.g., "Under what circumstances are contracts invalid according to the Contract Law"), leaning towards semantic understanding. The value ranges from 0 to 1. The function words include a total of about 85 lexical items, covering interrogative words (24 in total such as what / how / why, etc.), structural auxiliary words (de / shi / zai, etc.), modal particles (ma / ne / ba), and punctuation marks (? . , 、 ;, etc.).
[0082] Feature (File search signal strength): It is the strongest positive distinguishing feature for distinguishing exact queries from semantic queries. A multi-signal cumulative scoring mechanism is adopted. Physical meaning: the book title marks, file suffixes, quotation marks, and lookup verbs are all strong signals of exact search intent, and superposition of multiple signals can further improve the confidence. Detection signals and scores: In this embodiment, it is preferably set that (1) the occurrence of the book title marks 《》 adds 0.35; (2) the occurrence of file suffixes (.pdf / .docx / .xlsx / .pptx / .txt / .md / .csv / .json / .xml / .html) adds 0.35; (3) being enclosed by quotation marks "" adds 0.25; (4) the occurrence of lookup verbs (look up / retrieve / search / read / open / extract / query / download / acquire / consult / refer to / refer to / file, etc.) adds 0.20; (5) an additional 0.15 is added when multiple signals appear at the same time. The total score is clipped to [0,1].
[0083] Feature (Interrogative sentence signal): It is the strongest negative distinguishing feature for distinguishing semantic queries from exact queries. A multi-signal cumulative scoring mechanism is adopted. Physical meaning: interrogative words, question marks, modal particles and semantic verbs are all strong signals of semantic question-answering intent, which are negatively correlated with exact matching intent. Detection signals and scores: In this embodiment, it is preferably set that (1) the occurrence of interrogative words (18 in total such as what / how / why / why / how / which / who / when / where / how many / which / what kind of / how about / whether / can / whether / is there / is it, etc.) adds 0.40; (2) ending with a question mark (? or ?) adds 0.30; (3) ending with an interrogative modal particle (ma / ne / ba) adds 0.25; (4) the occurrence of semantic verbs (12 in total such as explain / elaborate / discuss / analyze / inquire / illustrate / describe / introduce / summarize / generalize / evaluate / discuss, etc.) adds 0.15. The total score is clipped to [0,1].
[0084] Feature (Embedding semantic similarity difference): An embedding model (such as BGE-M3, 1024-dimensional) is used to encode each group of texts into vectors, and the arithmetic mean is taken respectively to obtain the exact matching feature vector and the semantic understanding feature vector . The query text is encoded into a query vector , and the calculations are performed respectively and , the cosine similarity of and Original difference The value range is [-1, 1] (positive values are semantically more favorable, negative values are more precise). This is achieved through a normalization function. Will Mapped to the [0,1] interval, 0 corresponds to exact matching intent, and 1 corresponds to semantic understanding intent. In this embodiment, four sets of exact prototype phrases and four sets of semantic prototype phrases are predefined. Exact prototype phrases: {filename\document\search\retrieve file}, {number\clause\chapter\regulation\standard}, {pdf\doc\txt\file format\attachment}, {view\materials\data\records\report}. Semantic prototype phrases: {what\how\why\reason\mechanism\impact}, {analysis\explanation\elaborate\discussion\exploration\research}, {principle\concept\theory\method\process\relationship}, {comparison\distinction\feature\characteristic\classification\summary}.
[0085] feature (Proportion of English technical terms): The number of English words with ≥3 letters in the query. and the number of Chinese characters The calculation formula is: English words are weighted 3 times to reflect their information density (the information content of one English term such as "RAG" is approximately equal to 3 Chinese characters). Physical meaning: a high proportion indicates that the query contains a large number of English technical terms (such as "RAG system evaluation benchmark"), indicating a more exact match; a low proportion indicates a query using pure Chinese natural language.
[0086] Calculate the intent score based on the set of query features. This score represents the probability of a user's query tendency and an exact match. The calculation formula is:
[0087]
[0088] In the formula, The preset weight coefficient for the i-th feature reflects the contribution of that feature to the "exact match" intent. This indicates that the feature increases the probability of an exact match intent (such as a file extension). This means that the feature reduces the probability of an exact match intent, which in turn increases the probability of semantic understanding intent (such as interrogative words). It is a bias term used to adjust the baseline offset of the overall intention score. The range of values is a continuous interval (0, 1). ->1 indicates that the query tends to be an exact match. ->0 indicates that the query tends to be semantically understood.
[0089] For this enterprise scenario, the preferred setting is to set the weight coefficients for all features { to ={0.10,0.15,0.45,-0.45,-0.40,0.05}, bias term b=0.20. All feature weights. and bias terms Instead of being hard-coded, it is stored in an external configuration center (such as Apollo or Nacos), allowing operations and maintenance personnel to perform hot updates according to different business areas.
[0090] Scoring based on intent Preset basic weights and adjustment magnitude factor Dynamically calculate the fusion weights.
[0091] The system dynamically calculates the intensive retrieval weights based on intent scoring. The core logic is: when a query tends to be an exact match ( When the value is close to 1), reduce the semantic retrieval weight; conversely, reduce it.
[0092] The calculation formula is as follows:
[0093]
[0094]
[0095] In the formula, This is the basic weight, preferably set to 0.5, which means that the two retrieval strategies are given equal trust when the intent is unclear; To adjust the amplitude factor, used to control the intensity of the influence of the intent score on the final weight; in this embodiment, For example, take 0.6. In this embodiment, the truncation function is used. ,make sure As P_exact changes linearly across the entire range [0.05, 0.95], a continuous transition from "pure semantic" to "pure precision" retrieval strategy is achieved.
[0096] The system performs dense retrieval based on vector similarity and sparse retrieval based on keyword matching in parallel, obtaining a combined score set for dense retrieval and a combined score set for sparse retrieval, respectively. In this embodiment, the dense retrieval specifically uses the FAISS index for approximate nearest neighbor search, and the sparse retrieval specifically uses the BM25 algorithm for inverted index matching. Since the scores of dense and sparse retrieval are on different scales and have vastly different distributions, the system performs Min-Max normalization on both scores, mapping them to the (0,1) interval. In this embodiment, the system retrieves the top-N documents (e.g., N=30) with the highest scores from both recall results, merges and deduplicates the two results, and retains all candidates for Min-Max normalization. Weighted fusion sorting.
[0097] Score based on dense retrieval For example, the Min-Max normalization formula is:
[0098]
[0099] In the formula, The document representing the current document whose score is yet to be calculated; Represents any one document in the candidate set C; This is the collection of candidate documents after merging and deduplicating the two recall paths during the coarse-sorting stage (before fusion). It is a very small positive number, preferably set as This is used to prevent the denominator from being zero when all candidate documents in the set have the same score, which would lead to calculation errors. Normalized sparse retrieval score. The formula is completely symmetrical to the one described above.
[0100] Based on the fusion weights, the normalized scores are linearly weighted and fused to obtain the coarse ranking score for each document. The calculation formula is as follows:
[0101]
[0102] In the formula, and These are the normalized dense retrieval score and normalized sparse retrieval score for document d, respectively; the system sorts all candidate documents by... Sort from highest to lowest to form a coarse-ranked candidate set. .
[0103] After obtaining the coarsely ranked candidate set, the system scores based on intent. With preset precise threshold Based on the comparison results, a threshold truncation strategy based on intent confidence is implemented to determine the number M of candidate documents to be sent for fine ranking. An adaptive principle is established that "if the intent is clear, the fine ranking scale is reduced; if the intent is ambiguous, the fine ranking scale is maintained." The specific truncation logic is as follows:
[0104] The system first determines the intent score. Is it greater than the preset precision threshold? . In this embodiment, the system presets an intent confidence threshold value. The preferred value is 0.5.
[0105] Scenario 1: If If the query is determined to be a high-confidence exact match, the number of candidates for fine-ranking is dynamically reduced according to the following formula:
[0106]
[0107] In the formula, The default number of candidates for fine ranking is 30; To set the lower limit of the number of candidates for fine ranking, the optimal setting is... This is used to ensure that even for the most confident exact queries, the fine-ranking window still contains at least [a certain number of elements]. 10 candidate documents to avoid outputting empty results in extreme cases; This is the floor function;
[0108] Scenario 2: If Keep the default number of candidates for fine ranking unchanged, that is .
[0109] Finally, from the coarse-ranked candidate set The first M documents are extracted to form a candidate set for fine-grained ranking. .
[0110] As described above, the effect of this mechanism is as follows: For high-confidence exact-match queries (such as queries containing explicit filenames and file extensions), the top-ranked documents in the coarse-rank fusion sort already have extremely high relevance confidence, eliminating the need for expensive cross-encoder inference on a large number of candidate documents. The M-value is automatically reduced through a formula to decrease computational overhead. For semantically understanding queries, the complete fine-ranking window is maintained to fully explore semantic relationships. Thus, an adaptive matching is achieved between query intent confidence and fine-ranking computational overhead.
[0111] Rearrange the candidate set Each result is fed into the cross-encoder model for deep inference scoring, generating a refined ranking score. The formula for calculating the ranking score is:
[0112] ,
[0113] In the formula: This means concatenating the user query q and candidate documents d using a special delimiter to form the model input sequence; This refers to the hidden vector of the [CLS] token output by the cross-encoder model after processing the input sequence. In this embodiment, if the BGE-Reranker-base model is used, W is a 768-dimensional floating-point vector; W and B are the weight and bias parameters of the fully connected output layer of the cross-encoder model. These parameters are fixed parameters learned during the model pre-training and fine-tuning stages and remain unchanged during the inference stage.
[0114] Finally, the scores are determined by ranking. The candidate documents are sorted in descending order, and the final Top-K documents are extracted as the search results and output to the downstream large language model generation module. In this embodiment, the final output quantity K is set to 5, for example.
[0115] Example 2:
[0116] When a user enters the query "What are the main causes of ocean acidification?" into the RAG system, the system first performs feature extraction on the query:
[0117] (Query length ratio): L=12, r=30, u=8, The query should be short but contain a complete sentence structure. Preferring to the precision end.
[0118] (Keyword density): After word segmentation, a total of 7 words were obtained: ["ocean", "acidification", "of", "main", "reason", "is", "what"]. Among them, the function words are "of", "is", and "what", and the content words are "ocean", "acidification", "main", and "reason". .
[0119] (File search signal strength): The search did not trigger any signals for items containing book titles, file extensions, quotation marks, or the verb "search". =0.00.
[0120] (Interrogative signal): Contains the interrogative word "what" (+0.40), ends with a question mark "?" (+0.30). =min(0.40+0.30,1.0)=0.70.
[0121] (Embedded semantic similarity difference): Query vector With semantic understanding feature vectors cosine similarity , and precisely match the feature vector cosine similarity 0.35. By normalizing the function . (Semantic bias).
[0122] (Percentage of English technical terms): No English terms were found in the search. .
[0123] Calculate Intent Score (The system determines that this query is a purely semantic understanding query.)
[0124] Based on the above intention scoring The system calculates the weights for intensive retrieval. Dense retrieval accounts for 77% of the dominant weight, while sparse retrieval accounts for 23%. The system heavily relies on vector semantic matching to find documents related to the concept of "cause of ocean acidification." The system performs FAISS dense retrieval and BM25 sparse retrieval in parallel, each recalling the top-30 documents. After merging and deduplication, the scores from both approaches are normalized using Min-Max. Linear weighted fusion is used to calculate a coarse ranking score for each candidate document: .
[0125] Arrange them in descending order to obtain the coarsely ranked candidate set 𝐶𝑐𝑜𝑎𝑟𝑠𝑒.
[0126] System intent confidence level: due to Less than the preset precision threshold This query is classified as a non-high-confidence exact match and the default number of candidates for fine-grained ranking is maintained at 𝑀= The system performs cross-encoder inference depth scoring on each of the coarsely ranked TOP-30 candidates. The documents are sorted in descending order of their ranking scores, and the top-5 documents are output as the search results. During the fusion phase, the system heavily relies on the FAISS vector search results, successfully recalling and prioritizing conceptual documents related to "causes of ocean acidification," "carbon dioxide dissolution to form carbonic acid," and "pH decrease mechanisms."
[0127] Example 3:
[0128] When a user enters a query into the RAG system, such as "View 'Research on Ocean Acidification.pdf'", the system first extracts features from the query:
[0129] (Query length ratio): L=12, r=30, u=8, It is the same length as Example 2, but with a completely different intent; subsequent features will reveal this key difference.
[0130] (Keyword density): After word segmentation, we get ["Check", "《", "Ocean", "Acidification", "Research", ".pdf", "》"], where the function words are only the punctuation marks "《" "》", and the content words are "Check", "Ocean", "Acidification", "Research", ".pdf", a total of 5 words. .
[0131] (File search signal strength): Includes book title marks 《 (+0.35), includes file extension .pdf (+0.35), includes the verb "search" (+0.20). .
[0132] (Interrogative Sentence Signal): The query does not contain interrogative words, question marks, interrogative modifiers, or semantic verbs; all signals are not triggered. .
[0133] (Embedded semantic similarity difference): Query vector With precise query feature vector The cosine similarity is high (because the words "look up" and document-related words are highly consistent with the exact prototype). , . . . (Preferring to the precise end).
[0134] (Proportion of English technical terms): . = .
[0135] Calculate Intent Score .
[0136] Based on the above intention scoring The system determined that the query was a high-confidence exact match query. (This is in contrast to Example 2.) In comparison, both queries had the same query length (L=12), but their intent scores differed by a factor of 13, which fully demonstrates the intent discrimination capability of the 6-dimensional continuous feature system.
[0137] Based on the extremely high accuracy of the intent matching score, the system calculates the dynamic fusion weights.
[0138] .
[0139] The weight of dense retrieval is reduced to about 40%, while the weight of sparse retrieval is increased to about 60%. The system will mainly rely on BM25 keyword matching to find documents with specific filenames, while still retaining a certain proportion of semantic vector signals to prevent missed recalls caused by changes in filename spelling.
[0140] The system performs FAISS intensive search and BM25 sparse search in parallel, each recalling the top-30 documents. After merging and deduplication, it calculates the results for each candidate document. Arrange the candidates in descending order to obtain the coarsely ranked candidate set.
[0141] The target document "Research on Ocean Acidification.pdf" achieved a very high score in the BM25 sparse search due to its exact match of the keyword "Research on Ocean Acidification.pdf", ranking among the top in the fusion ranking.
[0142] Because the query intent is extremely clear ( The system triggers a truncation strategy, dynamically reducing the fine-sorting window. The number of fine-sorting operations, M, is calculated according to the formula:
[0143]
[0144] The system performs cross-encoder depth scoring only on the top 20 documents in the coarse-ranked fusion sort results:
[0145] Reducing M from the default 30 to 20 saves 33.3% of the computational cost of fine-ranking. This is based on a high-confidence judgment: the top 20 documents in the coarse-ranked fusion sort already contain the target file with a very high probability, and fine-ranking inference for the subsequent 10 documents is an unnecessary waste of computation.
[0146] The fine-ranking engine performs cross-encoder deep scoring only on the top 20 documents from the coarse-ranked results. The target document "Ocean Acidification Research.pdf" scored extremely high in precise keyword matching during the sparse retrieval stage, and "ocean acidification" and ".pdf" received significant weights in the cross-attention mechanism. This document ranked first with an overwhelming advantage in the fine-ranking stage. The top-5 documents are output in descending order of fine-ranking score.
[0147] The embodiments illustrate the complete workflow of this invention in an enterprise-level RAG system user retrieval scenario. To quantitatively verify the technical advantages of this invention compared to traditional vector retrieval schemes, a test environment was built, and multiple sets of controlled experiments were conducted for verification and analysis. Two sets of controlled experiments were set up: Experiment 1 verified the effectiveness of intent-aware weight adjustment based on six-dimensional continuous features; Experiment 2 verified the effect of the threshold truncation mechanism based on intent confidence in saving computational overhead while ensuring retrieval accuracy.
[0148] The experimental environment configuration for the two experiments is as follows:
[0149] CPU: Intel Core i9 / AMD Ryzen9;
[0150] GPU: NVIDIA GeForce RTX 4090 (24GB);
[0151] Memory: 64GB DDR5;
[0152] Operating system: Windows 10 Pro;
[0153] Python: 3.10;
[0154] Embedding model: BAAI / bge-m3 (1024-dimensional);
[0155] Vector index: FAISS IndexFlatIP;
[0156] Dual recall: FAISS + BM25
[0157] Dataset: 100 manually annotated queries (35 semantic queries, 30 mixed queries, and 35 exact match queries)
[0158] Experiment 1: Validation of the effectiveness of dynamic weight adjustment for intent perception based on 6-dimensional continuous features: The control group used three fixed weights. The experimental group used intent-aware dynamic weight adjustment to process 100 manually labeled queries (35 semantic, 30 mixed, and 35 exact match) according to specified criteria. Perform dual-path recall (FAISS+BM25), and press after Min-Max normalization. Weighted fusion ranking was performed, and the Top-30 results were fed into a cross-encoder for fine ranking. Pre-Rerank NDCG@30 and Recall@30 were used to evaluate the quality of the fusion ranking, and the relative performance loss of each scheme relative to the optimal scheme was calculated. Paired t-tests were used to verify the significance of differences (*p<0.05, **p<0.01).
[0159] Table 1 Comparison of Intent-Aware Dynamic Weight Adjustment and Fixed Weight Data
[0160]
[0161] (Note: ★ indicates the optimal value of the solution for this query type, * indicates p<0.05 in paired t-test, and ** indicates p<0.01; Relative loss = optimal NDCG@30 for this type - current solution NDCG@30, and the global maximum value is the maximum relative loss.)
[0162] As shown in Table 1, the retrieval quality difference is most significant in semantic query scenarios: Fixed (α=0.3) has an NDCG@30 of only 0.3243, significantly lower than the 0.8346 of the proposed solution, resulting in a relative performance loss of 0.5103. This means that if the production system chooses a fixed weight biased towards sparse retrieval, the ranking quality of semantic queries will suffer a loss of approximately 61%. The proposed solution eliminates the risk of missing key documents due to improper weight settings by dynamically increasing the weight of dense retrieval. In terms of recall assurance, the Recall@30 of the proposed solution remains stable at 100%, while Fixed (α=0.3) is only 75%, a difference of 25 percentage points.
[0163] Minimax analysis based on maximum relative loss provides the strongest support for the robustness of the patented solution: the maximum relative loss of Fixed (α=0.3) is as high as 0.5103, posing an extremely high risk of unavailability; while the maximum relative loss of the proposed solution is only 0.0303, the lowest among all solutions. This proves that the proposed solution is Minimax optimal: that is, it minimizes the "loss in the worst case". Although the proposed solution sacrifices a small amount of precision (loss of 0.0303) for precise queries, it gains significant benefits for semantic queries. This strategy of "trading a small local loss for maximum global security" ensures that the system's retrieval quality is always stably controlled at an optimal level, regardless of the query type.
[0164] Experiment 2: Verification of the effectiveness of the threshold truncation mechanism based on intent confidence in saving computational overhead while ensuring retrieval accuracy: The control group adopted four fixed ranking truncation strategies: fixed number of ranked candidates. The experimental group adopted a method based on Threshold truncation mechanism, , Based on the coarse-ranking fusion of the dynamic α scheme in Experiment 1, the RPR, NDCG@5, and M of each scheme on semantic, mixed, and precise queries are recorded as the core evaluation indicators.
[0165] Table 2: Comparison of RPR, NDCG@5, and M by query type for each solution
[0166]
[0167] Table 3: Overall Summary of Each Plan
[0168]
[0169] Tables 2 and 3 show that adaptive truncation results in near-zero information loss: the overall RPR of this invention is 0.9903, with a median of 1.0, meaning that for more than half of the queries, no relevant documents found in the coarse-ranked search were discarded. For precise queries, with M reduced from 30 to 24.5 (-18.3%), the RPR is 0.9724, meaning 97.24% of the coarse-ranked relevant documents are still retained in the precise-ranked search. In contrast, Fixed-M5 has an RPR of only 0.7143 (0.7882 for precise queries), meaning that 28.6% of the coarse-ranked recall information is irreversibly discarded. This significant difference in RPR (0.9903 vs 0.7143) directly distinguishes intelligent truncation from blind truncation at an information theory level.
[0170] Self-statistical tests prove that adaptive truncation is equivalent to full-scale fine-grained ranking: Paired t-tests show that the NDCG@5 difference between the proposed solution and Fixed-M30 is 0.0002 (t=1.00, p>0.05). The NDCG@5 of the proposed solution is completely consistent with Fixed-M30 across the three query types (semantic 0.6658 vs 0.6658, mixed 0.3313 vs 0.3313, precise 0.3470 vs 0.3463), with differences falling within the statistical noise range. In contrast, Fixed-M5's NDCG@5 plummeted from 0.3463 to 0.2532 (-26.9%, p<0.01) in the precise query type, primarily due to the RPR dropping to 0.7882—discarded relevant documents cannot be recovered during the fine-grained ranking stage. This comparison demonstrates the optimal position of adaptive truncation on the precision-efficiency Pareto front.
[0171] The core of adaptive truncation lies not in 'how much' is reduced, but in 'knowing which queries should not be reduced': As shown in Table 2, the proposed solution maintains M=30 for semantic and mixed queries (accounting for 65% of the total) without any reduction, and only reduces high-resolution queries. The exact queries (35%) are reduced by confidence gradient. The rating distribution shows that semantic types are concentrated in [0.00, 0.25), mixed types in [0.25, 0.50), and precise types expand upwards from 0.50. The decision boundaries are highly consistent. This differentiated strategy of "not reducing semantic content and moderately reducing precision" eliminates the precision risk caused by a one-size-fits-all approach in fixed truncation schemes from the mechanism level.
[0172] The cost-saving scenario is scalable—the current overall savings of 6.5% stem from the fact that only 35% of queries are exact queries. In a separate analysis of exact queries, Reducing M from 30 to 24.5 saves 18.3% of the computational cost for precise queries alone. In real-world enterprise knowledge base scenarios (such as technical document retrieval and regulatory review systems), queries containing precise features such as filenames, numbers, and clauses typically account for 50%-70%, resulting in an expected overall savings rate of 9.2%-11.9%. The cost-benefit ratio of this mechanism increases monotonically with the increase in the proportion of precise queries, demonstrating clear value for scenario expansion.
[0173] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.
Claims
1. An adaptive hybrid retrieval method for multidimensional spatial data, characterized in that, Includes the following steps: S1. Obtain the query text input by the user, extract features from the query text, and obtain a query feature set containing at least one feature; S2. Calculate the intent score based on the set of query features, whereby the intent score is used to characterize the probability that the user query tends to be an exact match; S3. Calculate the fusion weight dynamically based on the intent score. The fusion weight includes dense retrieval weight and sparse retrieval weight. S4. Perform dense retrieval based on vector similarity and sparse retrieval based on keyword matching in parallel to obtain the dense retrieval score set and the sparse retrieval score set respectively; S5. Normalize the dense retrieval score set and the sparse retrieval score set respectively to obtain the normalized dense score set and the normalized sparse score set. S6. Based on the fusion weight, the normalized dense retrieval score and the normalized sparse retrieval score are weighted and fused to obtain the fusion ranking score of each document, and the documents are sorted from high to low according to the fusion ranking score to determine the coarse ranking candidate set. S7. Based on the comparison result between the intent score and the preset precision threshold, perform dynamic truncation based on intent to determine the number of candidate documents to be sent for fine ranking. Based on the number of candidate documents in the fine ranking stage, truncate the coarse ranking candidate set to obtain the fine ranking candidate set. S8. Based on the refined candidate set, the candidate documents are sent to the cross-encoder model for refined ranking and reasoning one by one, and the final retrieval results are output in descending order of the refined ranking score.
2. The method according to claim 1, characterized in that, In S1: The query feature set includes one or more of the following features: feature (Query length ratio): A smooth mapping function is used to map the number of characters in the query text to the range [0,1]. The input of the mapping function includes the number of characters in the query text and a preset center length parameter. feature (Keyword density): Calculates the proportion of content words in the total number of words in the query. The content words are words that exclude the preset functional word set. feature (File search signal strength): Calculated based on a combination of specific symbols, file extensions, or specific verbs contained in the query, with values mapped to [0,1]. feature (Interrogative signal): Calculated based on the combination of interrogative words, interrogative symbols, or interrogative modifiers contained in the query, with a value range mapped to [0,1]. feature (Embedded semantic similarity difference): Calculate the first similarity between the query vector and the preset precise prototype vector, and the second similarity between the query vector and the preset semantic prototype vector, based on the difference between the first similarity and the second similarity. feature (Proportion of English technical terms): Calculated based on the ratio of the number of English words to the number of Chinese characters in the query.
3. The method according to claim 1, characterized in that, In S2, intent scoring The calculation uses a linear weighted clipping formula: In the formula, For the first Preset weight coefficients for each feature; It is a bias term used to adjust the baseline offset of the overall intention score. This is used to limit the results to a preset range.
4. The method according to claim 1, characterized in that, The S3 dynamically calculates the intensive retrieval weight. The calculation formula is: In the formula, The base weight represents the initial trust ratio assigned to the two search paths when the intent is unclear. To adjust the amplitude factor, the value range is greater than 0, which is used to control the intensity of the influence of the intent score on the final weight; For the boundary truncation function, ensure The value range is always between [0, 1].
5. The method according to claim 1, characterized in that, The S5 normalization process specifically includes: Min-Max normalization maps the two scores to the (0,1) interval to eliminate dimensional differences and normalizes dense scores. The calculation formula is as follows: In the formula, This represents the specific document for which the normalized score is to be calculated. Represents any one document in the candidate set C; This is the collection of candidate documents for the current rough ranking stage; This is a very small positive number, used to prevent the denominator from being zero and causing calculation errors when all candidate documents in the set have the same score. Normalized sparse search score. The formula is completely symmetrical to the above formula.
6. The method according to claim 1, characterized in that, Specifically, S6 includes: The weighted fusion ranking score is calculated using the following formula: In the formula, and Documents respectively Normalized dense retrieval score and normalized sparse retrieval score; For the dense retrieval weights dynamically calculated in S3, ( This corresponds to the sparse retrieval weight. The system performs sparse retrieval on all candidates. Document by Sort the documents in descending order from highest to lowest quality, and select the top N documents as a coarse candidate set. N is the preset upper limit of the capacity of the coarse-ranked candidate set.
7. The method according to claim 1, characterized in that, In step S7, the dynamic truncation based on intent scoring specifically includes the following steps: S7a, Determine the intent score Is it greater than the preset precision threshold? ,in The preset confidence threshold value for the system's intent; S7b, when When a query is determined to be a high-confidence exact match, the number of candidates for fine-grained ranking is determined according to the following formula: In the formula, The default number of candidates for fine-grained ranking, representing the standard fine-grained ranking window size for semantically or ambiguous queries; This is a lower bound on the number of candidates for fine-grained ranking, used to ensure that even for the most confident exact query, the fine-grained ranking window still contains at least [number missing]. 10 candidate documents to avoid outputting empty results in extreme cases; This is the floor function. S7c, when When (i.e., semantically understandable or ambiguous intent queries), the default number of fine-ranking candidates remains unchanged: .
8. The method according to claim 1, characterized in that, In step S8, the cross encoder fine sorting score is generated according to the following formula: , In the formula: This means concatenating the user query q with the candidate documents d into a unified input sequence. W represents the hidden layer vector of the [CLS] token output by the cross-encoder model, and W and B are the weight and bias parameters of the fully connected output layer of the cross-encoder model. The candidate documents are sorted in descending order, and the final Top-K documents are extracted as the search results.
9. An adaptive hybrid retrieval system for multidimensional spatial data, characterized in that, include: The system comprises a unified retrieval interface layer, a dynamic routing and scheduling layer, a multi-path recall execution layer, and a coarse-ranking and fine-ranking processing layer, among which: The unified retrieval interface layer is used to receive user queries, encapsulate the query text into a retrieval context object, and send the retrieval context object to the dynamic routing scheduling layer. The dynamic routing scheduling layer includes an intent parsing module and a dynamic weight calculation module: the intent parsing module is configured to perform the steps described in S1 and S2 of claim 1, extracting query feature vectors and calculating intent scores. The dynamic weight calculation module is configured to perform the steps described in S3 of claim 1, according to... Preset basic weights And dynamically calculate the dense retrieval weight α using the adjustment amplitude factor λ. ; The multi-path recall execution layer includes a dense retrieval engine and a sparse retrieval engine; the dense retrieval engine is configured to encode the query text into a high-dimensional vector and perform an approximate nearest neighbor search; the sparse retrieval engine is configured to perform a keyword matching algorithm based on an inverted index; the dense retrieval engine and the sparse retrieval engine are configured to be invoked in parallel by the dynamic routing scheduling layer through a concurrency mechanism. The coarse and fine ranking processing layer includes a coarse ranking fusion module, a threshold truncation module, and a fine ranking engine; the coarse ranking fusion module is configured to execute the steps described in S5 and S6 of claim 1; the threshold truncation module is configured to execute the intent-based truncation step described in claim 7; the fine ranking engine is configured to load a cross-encoder model, perform deep reasoning scoring on the fine ranking candidate set, and output the final retrieval result.
10. The system according to claim 9, characterized in that, It also includes a configuration center module; The configuration center module is configured to store and manage the set of parameters required for the operation of the intent parsing module, the dynamic weight calculation module, and the threshold truncation module. The set of parameters includes feature weight coefficients, bias terms, basic weights, adjustment amplitude factors, precise thresholds, and truncation quantity parameters. The intent parsing module, the dynamic weight calculation module, and the threshold truncation module are all configured to read the latest parameter values from the parameter set in real time from the configuration center module.
11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.
12. The method according to claim 2, characterized in that, The features The calculation of the embedding semantic similarity difference specifically includes: 12a. Pre-construct a precise prototype vector set and a semantic prototype vector set; the precise prototype vector set contains several texts used to represent precise retrieval intent, and the semantic prototype vector set contains several texts used to represent semantic understanding intent; 12b. Using an embedding model, encode the exact prototype vector set and the semantic prototype vector set into vectors respectively, and calculate the center vector for each to obtain the exact matching feature vector. and semantic understanding feature vectors ; 12c. Encode the user query text q into a query vector. Calculate the query vector respectively The exact matching feature vector The semantic understanding feature vector The cosine similarity is used to obtain the first and second similarities; 12d. Calculate the original difference based on the difference between the first similarity and the second similarity. ; 12e. The original difference is normalized using a normalization function. Mapped to the [0,1] interval, 0 corresponds to the exact matching intent, and 1 corresponds to the semantic understanding intent. This normalized value is used as a feature. Participating in the intent scoring The calculation.
