A RAG question-answering system that combines knowledge tracking and evidence annotation

Abstract

This invention relates to the field of intelligent question answering technology, specifically to a RAG question answering system combining knowledge tracking and evidence annotation. The system includes a semantic parsing module, a sentence localization module, a knowledge path annotation module, an evidence indexing module, and an enhanced question answering module. This invention extracts explicit verb semantics, indicator entities, and time-normalized expressions through structured parsing of natural language input, ensuring structural uniformity and processing consistency of the input sentences. A tag-based text filtering strategy effectively limits the text scanning range, improving the accuracy and efficiency of indicator localization. It clearly extracts statistically significant causal links from the text, and enhances the cross-coverage of indicator fragments by imposing indicator combination constraints and labeling key sentences in path nodes, thereby improving the density and logical order consistency of indicator integration in the answer.

Description

Technical Field

[0001] This invention relates to the field of intelligent question answering technology, and in particular to a RAG question answering system that combines knowledge tracking and evidence annotation. Background Technology

[0002] Intelligent question answering technology is an important branch of natural language processing, aiming to enable computer systems to understand user-posed questions and automatically generate targeted answers based on existing information resources. This field integrates multiple technologies such as information retrieval, machine reading comprehension, semantic parsing, and language generation to construct end-to-end or retrieval-enhanced generative question-answering models. Core challenges in this field include complex semantic understanding, multi-turn context modeling, accurate evidence localization and integration, answer consistency control, and dynamic knowledge updating capabilities. Application scenarios cover intelligent customer service, online education, legal consultation, medical diagnosis, and search engine enhancement. Systems typically utilize knowledge resources such as structured or unstructured documents, knowledge graphs, and large-scale language models to implement question-answering functions.

[0003] Among them, the RAG question-answering system, which combines knowledge tracking and evidence annotation, is an enhanced intelligent question-answering solution based on the RAG architecture. It aims to introduce user knowledge state modeling and explicit labeling of source information to improve knowledge adaptability and answer credibility during the question-answering process. By dynamically maintaining the user's knowledge mastery level, the system can accurately locate the knowledge points that need to be supplemented and generate answers by combining explicitly annotated evidence fragments in the document. This ensures that the answers not only meet the user's needs but are also verifiable and explainable. The system is suitable for scenarios requiring continuous learning or personalized responses, such as online education question-answering systems and professional domain-assisted question-answering platforms.

[0004] Traditional question-answering systems rely heavily on semantic representation and relevance scoring strategies to construct answers using question-answering models. When faced with complex questions containing time elements and indicators, they lack a mechanism for unified structural parsing of input statements. This results in some non-standard or missing fields failing to effectively match standard question templates, leading to insufficient intent recognition accuracy. During document retrieval, the lack of coverage constraints and the need to traverse the entire text for keyword extraction reduces localization efficiency and increases invalid interference. Furthermore, the failure to effectively model causal structures and semantic order during answer generation results in disorganized indicator organization and a lack of clear reasoning chains in the answers. This is particularly problematic in scenarios with highly coupled indicators, such as the power economy, where the question-answering results fail to reflect the logical connections between indicators, affecting their interpretability and practical value. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a RAG question-answering system that combines knowledge tracking and evidence annotation.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a RAG question-answering system combining knowledge tracking and evidence annotation, the system comprising:

[0007] The semantic parsing module obtains the natural language statement input by the user, extracts three types of structural content from the statement: predicate phrases, indicator noun groups, and time expression items, determines whether the three types of fields jointly meet the key trigger field coverage conditions set by the template, filters the matching template items, and generates a set of question structure tags.

[0008] The statement location module, based on the problem structure tag set, obtains three types of fields from the operating database: electricity purchase volume, load demand, and electricity price execution records, as well as the report page number location value. It then filters the paragraph set according to the set keyword coverage lower limit and generates a tag-driven text segment set.

[0009] The knowledge path identification module, based on the tag-driven text segment set, retrieves the node corresponding to each keyword in the power economic causal knowledge graph, filters the paths whose node frequency is above the target number in the TF-IDF node importance ranking as effective jump paths, and generates a knowledge jump path cluster;

[0010] The evidence indexing module calls the knowledge jump path cluster to number and label the sentence-level units in the text, extracts the combination of three indicators appearing in the sentence: electricity offset rate, electricity price settlement difference, and planned electricity error rate, identifies them according to the path number, and generates a list of combined feature evidence.

[0011] The present invention improves upon this invention by including the following improvements: the problem structure tag set includes semantic constraint items, indicator binding items, time boundary items, retrieval trigger items, and structure mapping items; the tag-driven text segment set includes similarity reference fragments, coverage verification values, indicator display words, position annotation values, and paragraph sequence groups; the knowledge jump path cluster includes graph jump numbers, semantic connection nodes, jump number path chains, mapping report identifiers, and path paragraph anchors; and the combined feature evidence list includes offset indicator statements, structure comparison numbers, indicator combination tags, context sentence boundaries, and segment-level index positions.

[0012] The present invention is improved in that the semantic parsing module includes:

[0013] The structure extraction submodule acquires the natural language sentences input by the user, identifies three types of language components in the sentences: predicate phrases, indicator noun groups, and time expression items. It performs syntactic role labeling on predicate phrases, extracts semantically significant items from indicator noun groups using keyword vector comparison, and identifies the start and end ranges of time expression items and standardizes their representation using regular expressions. It establishes the correspondence between multiple sets of language component structures and generates structure mapping relationship information.

[0014] The field validation submodule calls the structure mapping relationship information, compares the power generation, line loss rate, average monthly electricity purchase price, and transaction volume fields included in the business analysis indicator dictionary with the content of the indicator noun group, calculates the coverage ratio between the number of hit fields and the total number of mapped fields, judges based on the set key trigger field coverage rate threshold, filters the matching results and outputs the mapping set that passes the judgment, calculates the field coverage offset value, compares it with the coverage ratio indicator, and obtains the field matching filter set.

[0015] The tag generation submodule, based on the field matching filter set, calls the structured question template entries according to the marked indicator fields, standardized time expressions and verb phrase sets, fills the language components into the structure position field, constructs the question vector expression structure, and transcribes it into vector index prompt text format to generate a question structure tag set.

[0016] The present invention is improved in that the statement locating module includes:

[0017] The field acquisition submodule acquires the problem structure tag set, and based on the indicator fields, calls the three fields of electricity purchase, load demand and electricity price execution record in the operation database and the corresponding page number positioning index to establish a structured reference mapping group and obtain the field index structure;

[0018] The keyword statistics submodule, based on the field index structure, calls the corresponding original report text paragraph content according to the page number information, counts the number of tag keywords appearing in each paragraph, extracts the total number of keywords in each paragraph and normalizes it with the total number of words in the corresponding paragraph, calculates and obtains the keyword coverage strength value, and obtains the keyword coverage strength sequence.

[0019] The paragraph filtering submodule filters out paragraph index items that do not reach the lower limit of keyword coverage based on the keyword coverage intensity sequence, extracts paragraph numbers that meet the threshold, and combines page numbers, fields and keyword content to reorganize them into a vector index structure input body to generate a tag-driven text segment set.

[0020] The present invention is improved in that the knowledge path identification module includes:

[0021] The node extraction submodule obtains the tag-driven text segment set, and according to the keyword group of each segment, retrieves the graph node set corresponding to the keyword in the power economic causal knowledge graph, extracts the node identifier and associated upstream and downstream node number corresponding to each group of keywords, and obtains the graph node matching set.

[0022] The path filtering submodule, based on the graph node matching set, constructs causal connection edges and co-occurrence edge combinations between nodes according to the node number, extracts candidate paths with a path length in the range of 2 to 3 hops, and calculates the average TF-IDF importance, edge density, structural span, and average distance of each path. It weighs and compares each path feature, calculates and obtains path filtering reference values, sorts the paths according to the filtering reference values ​​from largest to smallest, and filters out paths below the threshold to obtain the causal association path filtering results.

[0023] Based on the causal path filtering results, the path mapping submodule maps each path, including the original report number corresponding to the node, the text segment index, and the path number information, according to the retained valid path numbers. It organizes the corresponding structural entries between the path and the text, establishes a multi-path cross-segment reference set, and generates a knowledge jump path cluster.

[0024] The present invention is improved in that the evidence index module includes:

[0025] The node location submodule calls the knowledge jump path cluster, and locates the paragraph position of each node in the corpus item by item according to the original report number and paragraph index value of each path, including the node, obtains the original text of the paragraph, and performs numbering and labeling of sentence-level units, establishes node paragraph sequence index, and generates sentence-level number index set.

[0026] The indicator extraction submodule scans the phrases and matches the three indicator items of electricity offset rate, electricity price settlement difference and planned electricity error rate for each sentence-level numbered statement according to the sentence-level numbered index set. It records the number of matched indicators in the statement and establishes a sentence-indicator comparison structure. It counts the number of matched indicators included in the sentence segment and obtains the indicator occurrence information.

[0027] The sentence segment filtering submodule filters sentence segments with more than two matches based on the frequency of occurrence of the indicator and the number of matches corresponding to the sentence segment. It then binds and identifies the filtered sentence segments with their corresponding path numbers, integrates them into a structure mapping table, and outputs the results to generate a list of combined feature evidence.

[0028] The present invention has an improvement, wherein the system further includes:

[0029] The enhanced question answering module, based on the combined feature evidence list, constructs an enhanced question answering context by embedding the structural semantics of the original user question according to the labeled sentence information. Using an AI model with enhanced retrieval generation capabilities, it combines structural tags, indicator causal order, and the semantics of the original sentence to generate response text with causal chains, structural guidance, and indicator integration sentences, thus obtaining knowledge-enhanced backtracking results.

[0030] The knowledge-enhanced backtracking results specifically include structural response sequences, guided question-and-answer segments, indicator interspersed phrases, evidence citation anchors, and output format templates.

[0031] The present invention is improved in that the enhanced question-answering module includes:

[0032] Based on the combined feature evidence list, the corpus splicing submodule splices sentence blocks according to the path number order based on the labeled sentence segment information, calls the corpus retrieval unit of the RAG system to perform vectorized embedding processing on each sentence segment, aligns the embedded vector with the homologous index field in the structural label, reconstructs the semantic position of the original corpus, establishes a corpus input group compatible with the RAG retrieval component, and generates a structured corpus fragment sequence.

[0033] The context construction submodule, based on the structured corpus fragment sequence, according to the indicator items, subject phrases and time constraints included in the question structure tags, combined with the relevance weights generated in the RAG retrieval stage, matches the trigger factor phrases in the embedded chunks, adjusts the corpus layout according to the reasoning order between indicators, assembles the semantic context to be generated, and obtains the embedded semantic context sequence.

[0034] The question-answering generation submodule calls the embedded semantic context sequence, and inputs it into the AI ​​model with retrieval enhancement generation according to the sorted sentence groups and embedding vectors. Under the control of the structure word template, it generates response content with an integrated structure of indicator reasoning chain, label mapping relationship and question background, splices and formats the output, and generates knowledge-enhanced backtracking results.

[0035] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0036] In this invention, structured parsing of natural language input extracts explicit verb semantics, indicator entities, and time-normalized expressions, establishing a structural label set with fixed coverage. This ensures structural uniformity and processing consistency of the input statements. A label-based text filtering strategy effectively limits the text scanning range, improving the accuracy and efficiency of indicator positioning. By combining causal and co-occurrence edges to construct a path set with limited hop counts, statistically significant causal node links in the text can be clearly extracted, ensuring the integrity of the causal structure between indicators in the answer generation. It also provides a traceable sentence index. By implementing indicator combination constraints and labeling key sentences in path nodes, the indicator cross-coverage of candidate evidence fragments is enhanced, improving the density and logical order consistency of indicator integration in the answer. The generation process integrates structural semantics, indicator order, and original sentence semantic information, resulting in output text with closed-loop indicator chains, stable semantic control, and verifiable content. Attached Figure Description

[0037] Figure 1 This is a system flowchart of the present invention;

[0038] Figure 2This is a flowchart of the semantic parsing module of the present invention;

[0039] Figure 3 This is a flowchart of the statement locating module of the present invention;

[0040] Figure 4 This is a flowchart of the knowledge path identification module of the present invention;

[0041] Figure 5 This is a flowchart of the evidence index module of the present invention;

[0042] Figure 6 This is a flowchart of the enhanced question-and-answer module of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0044] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0045] Please see Figure 1 The present invention provides a technical solution: a RAG question answering system that combines knowledge tracking and evidence annotation, the system including a semantic parsing module, a statement location module, a knowledge path annotation module, an evidence indexing module and an enhanced question answering module;

[0046] The semantic parsing module acquires the natural language statement input by the user, extracts three types of structural content from the statement: predicate phrases, indicator noun groups, and time expression items. It calls the preset structure in the standard question template library of natural language processing, performs verb semantic mapping on the predicate phrases, performs full matching on the indicator noun groups in the business analysis indicator dictionary, which includes power generation, line loss rate, average monthly electricity purchase price, and transaction volume, and uniformly normalizes the time expression items using the standard time regular expression pattern. It determines whether the three types of fields jointly meet the key trigger field coverage conditions set by the template, filters the matching template items, establishes a structured tag set, and generates a question structure tag set.

[0047] The statement location module is based on the problem structure tag set. According to the indicator fields, it obtains three types of fields from the operation database: electricity purchase, load demand, and electricity price execution records, as well as the report page number location value. It scans the text paragraphs page by page in the report text, counts the number of tag keywords appearing in the paragraphs, and filters the paragraph set according to the set keyword coverage lower limit to generate a tag-driven text segment set.

[0048] The knowledge path identification module is based on a tag-driven text segment set. According to the keyword group of each segment, it retrieves the node corresponding to each keyword in the causal knowledge graph of power economy, extracts the causal connection edges and co-occurrence edges between nodes, limits the path length to 2 to 3 hops, and selects the path with the node frequency above the target number of nodes in the TF-IDF node importance ranking as the effective jump path. For the report number and page segment number of the nodes included in the path, a path-text mapping is established to generate a knowledge jump path cluster.

[0049] The evidence index module calls the knowledge jump path cluster. Based on the paragraph number and text content corresponding to each path node, it numbers and marks the sentence-level units in the text. It extracts the combination of three indicators appearing in the sentence: electricity offset rate, electricity price settlement difference, and planned electricity error rate. Sentences with any two or more of the three indicators are screened out and marked according to the path number to generate a list of combined feature evidence.

[0050] The enhanced question answering module is based on a combined feature evidence list. According to the annotated sentence information, it serves as the input corpus for generation in the RAG system. It is embedded in the structural semantics of the original user question to construct an enhanced question answering context. Through the generator in the RAG generation process, it combines structural tags, indicator causal order and original sentence semantics to generate response text with causal chain, structural guidance and indicator integration sentence, and obtain knowledge-enhanced backtracking results.

[0051] The problem structure tag set includes semantic constraints, indicator bindings, time boundary items, retrieval triggers, and structure mapping items. The tag-driven text segment set includes similarity reference fragments, coverage verification values, indicator display words, position annotation values, and paragraph sequence groups. The knowledge jump path cluster includes graph jump numbers, semantic connection nodes, jump number path chains, mapping report identifiers, and path paragraph anchors. The combined feature evidence list includes offset indicator statements, structure comparison numbers, indicator combination labels, context sentence boundaries, and segment-level index positions. The knowledge-enhanced backtracking results specifically include structure response sequences, guided question-and-answer fragments, indicator interspersed blocks, evidence reference anchors, and output format templates.

[0052] Please see Figure 2 The semantic parsing module includes:

[0053] The structure extraction submodule acquires the natural language sentences input by the user, identifies three types of language components in the sentences: predicate phrases, indicator noun groups, and time expression items. It performs syntactic role labeling on predicate phrases, extracts semantically significant items from indicator noun groups using keyword vector comparison, and identifies the start and end ranges of time expression items and standardizes their representation using regular expressions. It establishes the correspondence between multiple sets of language component structures and generates structure mapping relationship information.

[0054] In the structure extraction submodule, the user-input natural language statement is obtained, such as "the total electricity purchase volume and corresponding electricity price in a certain region in July 2023". The sentence structure is decomposed, and "total electricity purchase volume" and "corresponding electricity price" are identified as indicator noun groups, "July 2023" is a time expression item, and "situation" is the syntactic focus in the predicate phrase. During the syntactic role labeling process, "situation" is marked as the central verb. For noun groups, keyword embedding vectors from the business indicator database are matched. Cosine similarity is used to identify whether it matches fields such as "average monthly electricity purchase price" and "transaction volume" in the indicator database. If the similarity is greater than the set standard of 0.85, it is considered a valid mapping item. The phrase "2023 July" is matched with the ISO-8601 regular expression time template. "2023" is extracted as the year and "07" as the month, and standardized into the timestamp "2023-07". Then, the order of the structural combination between the two phrases "total electricity purchase" and "corresponding electricity price" is determined. "Electricity purchase" is determined as the main structural unit and "electricity price" as the subordinate indicator. The structure mapper generates the following combination of time, indicator, and predicate structure: {Indicator = [total electricity purchase, electricity price], Time = 2023-07, Predicate = situation}. The obtained structural mapping relationship information will provide a position index for subsequent field comparison and template generation. The final generated result is the structural mapping relationship information.

[0055] The field validation submodule calls the structure mapping relationship information. Based on the content of the indicator noun group, it compares the power generation, line loss rate, average monthly electricity purchase price, and transaction volume fields included in the business analysis indicator dictionary. It calculates the coverage ratio between the number of matched fields and the total number of mapped fields. Based on the set key trigger field coverage threshold, it makes a judgment, filters the matching results, and outputs the set of mappings that pass the judgment, using the formula:

[0056] ;

[0057] The calculation obtains the field coverage offset value, compares it with the coverage ratio indicator, and obtains the field matching filter set;

[0058] in, The field represents the offset value. The normalized value representing the number of index fields hit. The normalized value representing the number of metrics extracted from the statement. The normalized value of the identified item in the indicator dictionary. Represents the field hierarchy weight value. This represents the number of noun phrases extracted from the original sentence;

[0059] In the field validation submodule, the indicator noun group {total electricity purchase, electricity price} extracted from the structural mapping relationship information is called and compared with the standard indicator items {power generation, line loss rate, average monthly electricity purchase price, transaction volume} already registered in the business analysis indicator dictionary. Among them, "total electricity purchase" matches "transaction volume", and "electricity price" matches "average monthly electricity purchase price", with a hit count of 2. The total number of extracted items is 2, and the corresponding identifier item in the indicator dictionary is 3. The field level weight is set to 1.5, and the total number of noun groups is 2. Substituting into the formula:

[0060] ;

[0061] The input values ​​are:

[0062] ;

[0063] The formula calculation steps are as follows:

[0064] ; ; ;

[0065] ;

[0066] The formula consists of three parts, among which This reflects the completeness of the indicator structure, among which The structural offset reflects the difference between the extracted statement quantity and the indicator library identifier. If this value is 0, it means that the number of indicators is completely consistent with the identified standards; intermediate items and The product of these terms represents the overall impact strength of the match under weight adjustment. The higher the value, the more critical the indicator; finally, divide by... This operation structure controls the smooth compression of the structural complexity of the original statement on the result value. It can reduce the interference of long sentences, while amplifying the structural offset of key fields.

[0067] The advantage of this formula lies in its ability to control the stability of field extraction by introducing a square root term for indicator differences, and to effectively improve the sensitivity of statement mapping to structural judgments by assigning weights to the degree of structural contribution. The final output field covers the offset value. This value is used to compare with the set key field coverage threshold to determine whether the field hit is valid.

[0068] Table 1 lists the variable names, meanings, and numerical settings of key parameters in this structure mapping process:

[0069] Table 1 Example table of structure mapping parameters

[0070]

[0071] Weight The weighting is based on the hierarchical classification of indicators within the operational system. Power generation and traded electricity are defined as core-level indicators with a weight of 1.5, while line loss rate and voltage fluctuation are mid-level indicators with a weight of 1.0. Auxiliary items such as contract number and power plant number have a weight of 0.5. This classification is based on the frequency of the indicator appearing in the main analysis field of the report. If an indicator appears in the top 20% of the frequency, it is classified as 1.5; if it appears in the 50% median, it is classified as 1.0; and if it appears in less than 80%, it is classified as 0.5. This ensures that the weighting has structural support and practical differentiation. The threshold of 0.8 is a preset field matching and identification benchmark value. It is set based on the average offset of 0.73 for the hit fields in 600 training statements, which is then shifted upward by 1.5 times the standard deviation to calculate a result of approximately 0.79, rounded to 0.8 for the judgment standard.

[0072] The results show that there is an effective comparison relationship between field structure hit and word count and hierarchical structure. The current value of 0.916 is significantly higher than the threshold of 0.8, so the field structure can be identified as meeting the matching criteria. Finally, the field matching filter set is output as the result for downstream modules to use.

[0073] The tag generation submodule uses a field matching filter set. Based on the tagged indicator fields, standardized time expressions, and verb phrase sets, it calls structured question template entries, fills the linguistic components into the structure position field, constructs the question vector expression structure, and transcribes it into a vector index hint text format to generate a question structure tag set.

[0074] In the tag generation submodule, based on the index fields "total electricity purchase" and "electricity price" included in the field matching filter set, the time expression "2023-07" and the predicate "situation", the structured question template entry is selected: "Please tell me the [predicate] of [indicator 1] and [indicator 2] of [time]". "2023-07" is replaced in the time field position, "total electricity purchase" and "electricity price" are replaced in the index position, and "situation" is filled in the predicate position. The structure is mapped to "Please tell me the situation of total electricity purchase and electricity price in 2023-07". Then, the keywords in the sentence are vectorized and a BERT embedded representation is generated. This is used as the input prompt vector structure in the RAG system retrieval stage. After embedding, a vectorized input set is formed, and the fields are marked with their tag positions in the database. The input expression structure required by the RAG system generation stage is output. The final result is the question structure tag set.

[0075] Please see Figure 3 The statement locating module includes:

[0076] The field acquisition submodule retrieves the problem structure tag set, and based on the indicator fields, calls the three fields of electricity purchase, load demand and electricity price execution record in the operation database and the corresponding page number positioning index to establish a structured reference mapping group and obtain the field index structure;

[0077] In the field acquisition submodule, the set of indicator fields specified in the problem structure tag set is obtained, such as {purchased electricity volume, load demand, executed electricity price}. Among them, "purchased electricity volume" is aligned with the main item "monthly electricity purchase summary" in the indicator database, "load demand" corresponds to "regional maximum load value", and "executed electricity price" corresponds to "average settlement electricity price". The structure field in the operation database is called to extract the report page number index position associated with each field. For example, the page number associated with "monthly electricity purchase summary" is page 12, "average settlement electricity price" is page 17, and "regional maximum load value" is page 15. At the same time, a structured correspondence between indicator items and index page numbers is established, the field names are numbered and added to the reference structure list, such as: field name: purchased electricity volume, page number index: 12, index type: whole table range, text range: paragraph index #45 to #53. The above content is integrated and organized into a list dictionary form, and a reverse pointer table between field and text index is constructed. The output content is the field index structure.

[0078] The keyword statistics submodule, based on the field index structure, retrieves the corresponding original report text paragraphs according to page number information, counts the number of tag keywords appearing in each paragraph, extracts the total number of keywords for each paragraph, and normalizes it with the total number of words in the corresponding paragraph using the following formula:

[0079] ;

[0080] The keyword coverage strength value is obtained through calculation, resulting in a keyword coverage strength sequence;

[0081] in, This represents the keyword coverage strength value, indicating the degree of fluctuation in the keyword distribution across the paragraph set. Indicates the first The number of tag keywords appearing in the paragraph Indicates the first Total word count of the paragraph Indicates the first Normalized values ​​of the importance of tags and keywords in a paragraph. Indicates the first The normalized value of the page number containing the paragraph. This represents the normalized mean of the tag numbers corresponding to all tag keywords. This indicates the total number of paragraphs in the paragraph set;

[0082] In the keyword statistics submodule, based on multiple indicator fields and corresponding paragraph numbers included in the field index structure, the number of keywords, total number of words, and paragraph page number position in each paragraph are extracted from the original report text. A weight query operation is performed on each keyword item in each paragraph to obtain the weight value. The proportion of keywords, position deviation, and weight response degree in each paragraph are calculated. The overall calculation uses the following formula:

[0083] ;

[0084] The calculation logic of this formula is as follows: First, calculate the keyword density (i.e., the number of keywords) for each segment. Total number of words The keywords are weighted according to their importance in the indicator system. This weight is compressed and amplified using the natural logarithm, and then normalized using the page number of each segment. Normalized mean of the tag numbers corresponding to all tag keywords The difference between the three factors measures the distance of the paragraph from the core topic position in the document structure. Taking the square root of the absolute value reflects the degree of deviation. The greater the deviation, the lower the weight, thus preventing paragraphs far from the core structure from being misled by high-frequency keywords. The product of the three factors is the paragraph's "keyword structure score." Then, the scores of all paragraphs are summed and averaged to obtain the final score. This represents the keyword coverage strength value, which reflects the degree of fluctuation in the keyword distribution across the paragraph set.

[0085] The keyword coverage intensity value The meaning is: it quantifies whether the keyword distribution in the current paragraph set is concentrated in paragraphs with high keyword intensity and located in the core structure. The larger the value, the more frequently the keywords appear in important paragraphs and the more concentrated their positions are. It can be used as an important basis for filtering text paragraphs for RAG context construction.

[0086] The following are the actual calculation data:

[0087] Table 2 Example of Keyword Coverage Strength Calculation

[0088]

[0089] Taking paragraph 1 as an example, the calculation process is as follows:

[0090] ;

[0091] The average of the three paragraph values ​​is calculated as follows:

[0092] ;

[0093] The ultimate purpose of this value is to serve as an input for determining paragraph importance. If the coverage strength value is higher than the lower threshold, it indicates an imbalance. If the keyword is selected, the paragraph is considered to have the ability to express the key points and is suitable for constructing RAG prompts; otherwise, it will be removed.

[0094] In addition, keyword weight The settings are based on the weight hierarchy of the fields in the database structure. First-level fields such as "transactional electricity volume" and "average monthly electricity purchase price" are set to a range of 1.0 to 1.2, while second-level nested fields such as "electricity purchase category" are set to 0.6 to 0.8. This setting is obtained by multiplying the field importance participation by the structure hierarchy factor, and the normalization range must meet the following requirements. ; Average of label positions The average of the normalized page numbers of all paragraphs containing the matched keywords is calculated. A lower average value indicates that the core keyword is located at the beginning of the document. The further the paragraph's page number deviates from this average, the lower its score. This design ensures that the formula can comprehensively judge the three dimensions of keyword quality, density, and structural position.

[0095] The paragraph filtering submodule filters out paragraph index items that do not reach the lower limit of keyword coverage based on the keyword coverage intensity sequence, extracts paragraph numbers that meet the threshold, and combines page numbers, fields and keyword content to reorganize them into a vector index structure input body to generate a tag-driven text segment set.

[0096] In the paragraph filtering submodule, each paragraph is selected based on its keyword coverage strength sequence. The value is compared to the lower limit of keyword coverage, and this lower limit is set as follows: This value is set based on the keyword distribution coverage statistics of 50 report documents, taking the sample point value at the 25th percentile, and selecting at least the top 75% of paragraphs from the sample for RAG generation. Based on the aforementioned calculation results for each paragraph, paragraph 1 scored 0.0171, paragraph 2 scored 0.0056, and paragraph 3 scored 0.0145; therefore, paragraph 2 was eliminated. The remaining paragraph numbers, page numbers, relevant fields, and keyword content are combined and reorganized into structured cue data according to vector retrieval format, ultimately generating a tag-driven text segment set.

[0097] Please see Figure 4 The knowledge path identification module includes:

[0098] The node extraction submodule obtains the tag-driven text segment set, and based on the keyword group of each segment, retrieves the graph node set corresponding to the keyword in the power economic causal knowledge graph, extracts the node identifier and associated upstream and downstream node number corresponding to each keyword group, and obtains the graph node matching set.

[0099] The keyword group content of each segment in the tag-driven text segment set is obtained. The keyword group is used as a search term and compared with the node entries in the power economic causal knowledge graph. According to the complete matching relationship between the terms in the keyword group and the node tags in the graph, the node number and its category index are extracted. The node identifiers matched by each keyword group are recorded and the direct predecessor and successor node number sets of the node are counted. The nodes are classified and labeled according to the cluster number of the path to which the node belongs. The node set dictionary is stored in the order of the keyword group to form the graph node matching set. For example, if a keyword group is "purchased electricity volume, marginal electricity price", the node numbers matched in the graph are N103 and N205 respectively. The upstream node N098 and the downstream nodes N209 and N210 are extracted, and their mapping path numbers are recorded as P12 and P15, forming the matching data item {"purchased electricity volume": [N103, N098, N209], "marginal electricity price": [N205, N210]}. After further processing, the graph node matching set is established.

[0100] The path selection submodule, based on the graph node matching set, constructs combinations of causal connections and co-occurrence edges between nodes according to node numbers. Candidate paths are extracted with a path length ranging from 2 to 3 hops. The TF-IDF average importance, edge density, structural span, and average distance of each path are statistically analyzed, and each path feature is compared and weighed using the following formula:

[0101] ;

[0102] The path filtering reference value is obtained through calculation. The paths are sorted from largest to smallest according to the filtering reference value, and paths below the threshold are filtered out to obtain the causal relationship path filtering results.

[0103] in, This represents the filtering reference value in the set of causal relationship path scores. Indicates the first The mean TF-IDF importance of all nodes in the path. Indicates the first The average number of hops in the structure of the path, Indicates the first Edge density of the path, Indicates the first The number of causal nodes in the path Indicates the index number of the path. This represents the total number of candidate paths;

[0104] Based on all node IDs recorded in the graph node matching set, a combined structure of causal connections and co-occurrence edges between nodes is constructed. Candidate path sets are extracted under the condition that the number of hops is limited to 2 to 3. The mean TF-IDF importance, average hop count, edge density, and number of causal nodes for all nodes in each path are calculated sequentially, denoted as follows: , , , Substitute the above four parameters into the following formula:

[0105] ;

[0106] Among them, the molecule of the first term This term is used to measure the importance of semantic keywords in a path. A higher TF-IDF indicates a more significant semantic meaning of the keyword within that path. This value is obtained by taking the natural logarithm of the arithmetic mean of the TF-IDF values ​​of all path nodes; the denominator of this term is... The first item is used to suppress path redundancy caused by hop count; the higher the hop count, the greater the impact. This is used to represent the degree to which structural complexity in a path weakens its semantic value, where Edge density represents the actual density of edges along the path; the denser the edges, the more likely they are to generate noise. The formula is the square root of the number of causal nodes, used to adjust the balance effect of causal strength in complex paths; the overall formula reflects the trade-off between semantic salience and structural simplicity; the formula calculation results... The score represents the causal representativeness of the path; a higher score indicates greater causal value. The system will evaluate all candidate paths... The values ​​are sorted from largest to smallest and a lower limit threshold for path scores is set for filtering. Paths with scores below the threshold will be eliminated.

[0107] Taking the feature data of three paths as an example, the complete calculation is explained by substituting them into the formula. The parameter values ​​are shown in Table 3.

[0108] Table 3 Example of Feature Data

[0109]

[0110] Substituting the parameters of path P12, the calculation is as follows:

[0111] The first item is:

[0112] ;

[0113] The second item is:

[0114] ;

[0115] The score for path P12 is:

[0116] ;

[0117] Similarly, calculate path P15:

[0118] First item:

[0119] ;

[0120] Second item:

[0121] ;

[0122] Score:

[0123] ;

[0124] Path P21:

[0125] First item:

[0126] ;

[0127] Second item:

[0128] ;

[0129] Score:

[0130] ;

[0131] The above score sequence , , After sorting, the average value is calculated as follows:

[0132] ;

[0133] Set the path score filtering threshold to Then only keep The path, i.e., P15, is selected, and the final causal relationship path screening results are generated.

[0134] The advantage of this formula is that by jointly modeling the number of hops in the path structure, keyword importance, and causal density, it avoids the problem of mis-screening caused by filtering based solely on path length or keyword quantity. While ensuring structural compactness, it also takes into account the structural role of semantic weight and causal nodes, thus more stably filtering out high-quality causal paths.

[0135] The path mapping submodule, based on the causal path filtering results, maps each path according to the retained valid path numbers, including the original report number corresponding to the node, the text segment index, and the path number information. It organizes the corresponding structural entries between the path and the text, establishes a multi-path cross-segment reference set, and generates a knowledge jump path cluster.

[0136] Based on the path number list retained by the path filtering submodule, the node numbers contained in each path are read from the graph node matching set. The page number index and paragraph number information in the original report file structure are matched to construct the correspondence between the node numbers and the report document structure. The combination of path numbers, node numbers, and original node text sentences is converted into a unified format structure entry to form a multi-paragraph text index set corresponding to the path number. Each record in this set contains path ID, node sequence number, document page number, and paragraph number index information. The path numbers are aggregated and categorized. Finally, a multi-path cross-paragraph reference set is established, and the output is a knowledge jump path cluster.

[0137] Please see Figure 5 The evidence index module includes:

[0138] The node location submodule calls the knowledge jump path cluster. Based on the original report number and paragraph index value of each path, it locates the paragraph position of each node in the corpus, obtains the original text of the paragraph, and performs numbering and labeling of sentence-level units. It then establishes a node paragraph sequence index and generates a sentence-level numbered index set.

[0139] After obtaining the node information corresponding to each path in the knowledge jump path cluster, the original report number and paragraph index value associated with each node in each path are read sequentially. The above number combinations are then searched sequentially, and the original report text structure in the corpus is invoked to locate the paragraph position indicated by the number, extracting the complete text content from the paragraph. For each text segment, punctuation marks are used as sentence-level boundaries to divide the entire segment into several clauses, and each clause is labeled with a unique number, such as "Sentence 1" and "Sentence 2". Subsequently, the corresponding sentence-level number is embedded in the paragraph text structure to establish a node-paragraph-sentence triple mapping structure. For example, if path P12 contains node N5, which is associated with paragraph 28 of report R3, then paragraph 28 of report R3 is located, and after sentence division, three clauses are obtained, labeled as Sentence 1, Sentence 2, and Sentence 3. Node N5 is mapped to R3-28-sentence 1, R3-28-sentence 2, and R3-28-sentence 3. The correspondence between its path number, node number, and sentence-level number is organized into entries, generating a five-tuple sequence of structure fields: "path number – node number – paragraph number – sentence-level number – sentence-level text," thus establishing a sentence-level number index set. This process not only clarifies the text location pointed to by the node but also provides a well-granular text location basis for subsequent indicator extraction. The generated result is the sentence-level number index set.

[0140] The indicator extraction submodule scans the phrases and matches the three indicator items of electricity offset rate, electricity price settlement difference and planned electricity error rate for each sentence-level numbered statement according to the sentence-level numbered index set. It records the number of matched indicators in the statement and establishes a sentence-indicator comparison structure. It counts the number of matched indicators included in the sentence segment and obtains the indicator occurrence information.

[0141] Based on the constructed sentence-level index set, the text content of the sentences corresponding to each sentence-level number is traversed and read. Semantic units are segmented according to phrase delimiters within the sentence, and the segmentation results are compared using indicator keywords. The matching targets are set as three commonly used indicators in power economic analysis: "electricity offset rate," "electricity price settlement difference," and "planned electricity error rate." For each matched indicator, a record is added to the sentence number structure corresponding to that sentence, representing the indicator hit status. The matching logic is based on complete keyword matching and semantic consistency with the context, excluding semantically ambiguous and vague terms; for example, "offset rate" appearing alone is not counted as "electricity offset rate." After this traversal and comparison process is completed, the types and number of indicators matched for each sentence are counted. For example, if sentence number R3-28-sentence 2 contains the terms "electricity price settlement difference" and "electricity offset rate," the match count is 2, and the sentence text, path number, and matching details are recorded in a table. The results are shown below:

[0142] Table 4 Indicator Matching Status

[0143]

[0144] As shown in Table 4, sentence segment R3-28-sentence 3 matched all three indicators, indicating a high indicator content within this path; while sentence segment R3-28-sentence 2 only matched one indicator, failing to meet the screening requirements. This statistical result represents the frequency of indicator occurrences and will serve as input data for subsequent screening decisions. The indicator matching behavior places demands on semantic recognition accuracy. The system performs structural tag cleaning on the input sentences and semantic consistency verification on the matching results to ensure the accuracy of the statistical data. Sentence segment length, keyword density, and contextual logic are used as auxiliary factors in the judgment. The number of matched sentence segments is recorded through structural organization, constructing a sentence number – indicator matching item list.

[0145] The sentence segment filtering submodule filters sentence segments with more than two matches based on the frequency of occurrence of the indicator and the number of matches corresponding to the sentence segment. It then binds and identifies the filtered sentence segments with their corresponding path numbers, integrates them into a structure mapping table, and outputs the results to generate a list of combined feature evidence.

[0146] Based on the frequency of occurrence of the constructed indicators, the number of indicators contained in the sentence segments is further judged and processed. The judgment threshold is set as "at least two indicators match", and this threshold is used as the basis for sentence segment selection. During the selection process, the "match count" field corresponding to each sentence segment is read. When the value is ≥2, the sentence segment is included in the selection result set, and its sentence number, path number, original text content, and set of matched indicators are extracted and integrated to form a structure mapping table. For example, if sentence segment R3-28-sentence 1 matches 2 indicators, it meets the threshold requirement; if R3-28-sentence 2 matches only 1 indicator, it is filtered out. If only some sentence segments in a path meet the selection requirements, only the matching items are retained and the corresponding path number is recorded. If there are no matching sentence segments, the path is not used in subsequent steps. Based on the structure of the selected sentence segments, its corresponding information is reorganized in vector format to generate a path-paragraph-sentence-indicator quadruple input body, which serves as the input corpus for the downstream generation module of the RAG system, and finally, a combined feature evidence list is constructed.

[0147] Please see Figure 6 The enhanced question-and-answer module includes:

[0148] The corpus splicing submodule is based on the combined feature evidence list. According to the labeled sentence segment information, it splices sentence blocks in the order of path number. It calls the corpus retrieval unit of the RAG system to perform vectorized embedding processing on each sentence segment, aligns the embedding vector with the homologous index field in the structure label, reconstructs the semantic position of the original corpus, establishes a corpus input group compatible with the RAG retrieval component, and generates a sequence of structured corpus fragments.

[0149] The text content of each path number in the combined feature evidence list is obtained. The text of each sentence segment is concatenated in the order of the numbers. The concatenation process must preserve the consistency between the sentence order and the original structure of the path. After obtaining the concatenated corpus, it is input into the vectorization encoding interface of the RAG system. The encoding model is called for vector conversion for each sentence segment. After obtaining the vector representation, the vector is compared with the index fields in the structure label set. The semantic matching process is performed. During the process, the cosine similarity between each embedded vector and its homologous index field is calculated and the corresponding alignment mapping index is generated. The matching is recorded according to the pairing with a cosine distance of less than 0.4 to form the semantic distribution matrix of the corpus. The embedded structure is divided into semantic regions according to the index fields and labeled with numbers according to the original sentence segment position. Finally, a structured corpus fragment sequence is constructed.

[0150] Table 5. Examples of semantic alignment between structural tags and corpus embedding vectors.

[0151]

[0152] As shown in Table 5, the system establishes a mapping alignment structure between the sentence segment embedding vector and the indicator fields in the structural label based on the similarity threshold, which is used for subsequent context construction tasks; this mapping relationship provides an indicator reference basis for subsequent embedding semantic localization.

[0153] The question-answering generation submodule calls the embedded semantic context sequence, and inputs the sorted sentence groups and embedding vectors into the AI ​​model with retrieval enhancement generation. Under the control of the structure word template, it generates response content with an integrated structure of indicator reasoning chain, label mapping relationship and question background. Finally, it splices and formats the output to generate knowledge-enhanced backtracking results.

[0154] Based on the semantic representation of the embedded vectors of each sentence segment in the aforementioned structured corpus fragment sequence, the label fields, subject phrases, and time item information are called to construct a semantic context construction benchmark based on the input structure of the RAG system. When performing similarity matching on each set of embedded vectors, the logical order of multiple indicators involved in the sentence segment is obtained according to the causal reasoning order in the structural labels. The system records the order of appearance of indicators in these reasoning orders as sequence vectors. Then, the original sentence segment order is adjusted through structural guidance rules. During the adjustment process, logical conditional phrases in the structural template, such as "due to," "therefore," and "leading to," are introduced and inserted into the segment connection position to complete the causal connection between the blocks. For sentence groups containing the same label fields but different time expression items in the corpus blocks, the time expression order is used to re-sort them. The trigger factor phrases involved in the structural blocks are explicitly marked and re-embedded into the semantic vector space to obtain an embedded semantic context sequence. This context has semantic continuity and causal order structure.

[0155] The context construction submodule, based on the structured corpus fragment sequence, the indicators, subject phrases and time constraints included in the question structure labels, and the relevance weights generated in the RAG retrieval stage, matches the triggering factor phrases in the embedded chunks, adjusts the corpus layout according to the reasoning order between indicators, assembles the semantic context to be generated, and obtains the embedded semantic context sequence.

[0156] The system calls the statement groups and their embedding vectors from the aforementioned embedded semantic context sequence and inputs them into a large language model generation component with RAG retrieval enhancement capabilities. During the generation phase, based on the preset template format in the structure word template, the system fills each set of context statements into the generated structure field in structural order. For example, in the format "[Indicator Inference Chain] → [Label Content] → [Background Statement]", the embedded content is injected sequentially into the corresponding structural positions. During the model generation phase, the system controls the output order and format. When generating text, it follows the order of the indicator causal chain expression, guiding the generation to output factors first, then result indicators, and finally semantic mapping phrases for label items. The entire process forms question-and-answer content with causal chains, indicator explanation labels, and nested contextual logic. The system merges these text blocks, formats them uniformly, adds sentence tags and field content identifiers according to field marking rules, and outputs structurally solvable response results. Finally, it concatenates these to generate knowledge-enhanced backtracking results. This result not only preserves the question context structure but also embeds indicators and semantic association paths of combined feature evidence, ensuring that the question-and-answer response has structural and semantic traceability.

[0157] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A RAG question answering system that combines knowledge tracking and evidence annotation, characterized in that, The system comprises: The semantic analysis module obtains the natural language sentence input by the user, extracts three types of structure contents of the predicate phrase, the index noun group and the time expression item in the sentence, judges whether the three types of fields jointly satisfy the key trigger field coverage condition set by the template, screens the hit template item, and generates a question structure label set; The sentence positioning module obtains the three types of fields of the electricity purchase capacity, load demand and electricity price execution record in the operation database and the report page code positioning value based on the question structure label set, screens the paragraph set according to the set lower limit of the keyword coverage, and generates a label-driven text segment set; The knowledge path calibration module retrieves the node corresponding to each keyword in the electric power economic causal knowledge graph based on the label-driven text segment set, screens the paths with the node frequency located in the front target number of the TF-IDF node importance ranking as the effective jump paths, and generates a knowledge jump path cluster; The evidence index module calls the knowledge jump path cluster, performs numbering annotation on the sentence-level unit in the text, extracts the three index combinations of the electricity capacity deviation rate, the electricity price settlement difference and the planned electricity capacity error rate appearing in the sentence, performs identification processing according to the path number, generates a combination feature evidence list, and the combination feature evidence list includes the deviation index sentence, the structure comparison number, the index combination label, the context sentence boundary and the paragraph-level index position; The enhanced question and answer module embeds the original user question structure semantics according to the annotated sentence and paragraph information based on the combination feature evidence list, constructs an enhanced question and answer context, uses an AI model with retrieval enhancement, combines the structural label, the index causal sequence and the original sentence semantics, generates a response text with a causal chain, a structure guide sentence and an integrated index sentence, and obtains a knowledge enhancement backtracking result. The knowledge enhancement backtracking result specifically includes a structure response sequence, a guided question and answer segment, an index intercalation language block, an evidence reference anchor point and an output format template.

2. The RAG Q&A system that combines knowledge tracing and evidence labeling of claim 1, wherein, The question structure label set includes a semantic constraint item, an index binding item, a time boundary item, a retrieval trigger item and a structure mapping item, the label-driven text segment set includes a similarity reference segment, a coverage verification value, an index appearance word, a position annotation value and a paragraph sequence group, and the knowledge jump path cluster includes a graph jump number, a semantic connection node, a jump number path chain, a mapping report identifier and a path paragraph anchor point.

3. The RAG Q&A system that combines knowledge tracing and evidence labeling of claim 2, wherein, The semantic analysis module comprises: The structure extraction submodule obtains the natural language sentence input by the user, identifies three types of language components including the predicate phrase, the index noun group and the time expression item in the sentence, performs syntactic role labeling on the predicate phrase, extracts the semantic significant item by using the keyword vector comparison method on the index noun group, identifies the start and end range by using the regular expression method on the time expression item and standardizes the representation, establishes the corresponding relationship between multiple groups of language component structures, and generates structure mapping relationship information; The field checking submodule calls the structure mapping relationship information, compares the power generation, line loss rate, monthly average power purchase price and transaction power field recorded in the business analysis index dictionary according to the index noun group content, calculates the coverage ratio between the number of hit fields and the total number of mapping fields, judges according to the set key trigger field coverage threshold, filters the matching result and outputs the mapping set passing the judgment, calculates the field coverage offset value, compares it with the coverage ratio index, and obtains the field matching filtering set; The label generation submodule generates a question structure label set based on the field matching filtering set, according to the marked index field, standardized time expression and verb phrase set, calls the structured question template item, fills the language component into the structure position field, constructs the question vector expression structure, and transcribes it into a vector index prompt text format, and generates a question structure label set.

4. The RAG Q&A system that combines knowledge tracing and evidence labeling of claim 3, wherein, The sentence positioning module comprises: The field acquisition submodule acquires the question structure label set, calls the power purchase amount, load demand and power price execution record three fields and corresponding page code positioning indexes in the business database according to the index field, establishes a structured reference mapping group, and acquires a field index structure body; The keyword statistics submodule calculates the number of label keywords appearing in each text according to the page code information, extracts the total number of keywords in each paragraph and normalizes it with the total number of words in the corresponding paragraph, calculates the keyword coverage intensity value, and obtains the keyword coverage intensity sequence based on the field index structure body; The paragraph filtering submodule filters out the paragraph index items that do not reach the lower limit of the keyword coverage rate according to the keyword coverage intensity sequence, extracts the paragraph numbers that meet the threshold, combines the page code, field and keyword content, reorganizes them into a vector index structure input body, and generates a label-driven text segment set.

5. The RAG Q&A system tracking knowledge pursuit with evidence annotation of claim 4, wherein, The knowledge path calibration module comprises: The node extraction submodule acquires the label-driven text segment set, retrieves the graph node set corresponding to the keywords in the power economic causal knowledge graph according to the keyword group of each paragraph, extracts the node identifier corresponding to each group of keywords and the associated upstream and downstream node numbers, and acquires a graph node matching set; The path filtering submodule constructs the combination of causal connection edges and co-occurrence edges between nodes according to the node number based on the graph node matching set, extracts candidate paths with a path length in the range of 2 to 3 hops, calculates the TF-IDF average importance, edge density, structure span and average distance of each path, compares each path feature, calculates the path filtering reference value, sorts the paths according to the filtering reference value from large to small, and filters out the paths below the threshold, to obtain the causal correlation path filtering result; The path mapping submodule maps the original report number, text page segment index and path number information corresponding to each node included in each path according to the effective path numbers retained based on the causal correlation path filtering result, arranges the corresponding structure items between the path and the text, establishes a multi-path cross-segment reference set, and generates a knowledge jump path cluster.

6. The RAG Q&A system of claim 5, wherein, The evidence index module comprises: The node positioning sub-module calls the knowledge jump path cluster, locates the paragraph position of each node in the corpus according to the original report number and paragraph index value of each path including the node, obtains the paragraph original text and performs numbered annotation processing on the sentence level unit, establishes a node paragraph sequence index, and generates a sentence level numbered index set; The index extraction sub-module scans the word groups of each sentence level numbered sentence according to the sentence level numbered index set, matches the power offset rate, electricity price settlement difference and planned power error rate three index items, records the number of matched index items in the sentence and establishes a sentence-index contrast structure, counts the number of matched index items included in the sentence paragraph, and obtains index occurrence frequency information; The sentence paragraph screening sub-module screens the sentence paragraph with a matching number greater than two according to the index occurrence frequency information and the matching number of the corresponding sentence paragraph, binds the screened sentence paragraph and the corresponding path number, integrates them into a structure mapping table, and outputs the integrated structure mapping table, to generate a combined feature evidence list.

7. The RAG Q&A system of claim 1, wherein, The enhanced question and answer module includes: The corpus splicing sub-module splices the sentence block according to the path number order based on the combined feature evidence list and the annotated sentence paragraph information, calls the corpus retrieval unit of the RAG system to perform vector embedding processing on each sentence paragraph, aligns the embedded vector with the homologous index field in the structure label, reconstructs the original corpus semantic position, establishes a corpus input group compatible with the RAG retrieval component, and generates a structured corpus segment sequence; The context construction sub-module matches the trigger factor phrase in the embedded sentence block according to the index item, subject word group and time constraint range included in the question structure label, combines the relevance weight generated in the RAG retrieval stage, adjusts the corpus layout according to the reasoning order between indexes, assembles the to-be-generated semantic context, and obtains an embedded semantic context sequence; The question and answer generation sub-module calls the embedded semantic context sequence, inputs the sorted sentence group and embedded vector into the AI model with retrieval enhancement generation, generates response content with index reasoning chain, label mapping relationship and question background integration structure under the control condition of the structure word template, splices and formats the output, and generates knowledge enhancement backtracking results.

Images