Auxiliary understanding and intelligent question and answer for electronic design document

Abstract

The invention discloses an electronic design document-oriented auxiliary understanding and intelligent question-answering system, which comprises a data processing module for uniformly filing electronic design documents in different formats; the document analysis module is used for analyzing multi-modal contents such as texts, tables and images in the document and representing the multi-modal contents as indexable semantic blocks; the knowledge graph construction module is used for integrating the loose information dispersed in different documents into a structured knowledge graph; the vector database construction module is used for vectorizing all semantic blocks and storing the semantic blocks into a vector database; and the RAG reasoning module is used for combining the multi-modal retrieval result with large language model reasoning to generate an intelligent question and answer result oriented to an engineering scene. By means of the system, scattered, heterogeneous and multi-mode electronic design data are converted into a knowledge base capable of being queried, reasoned and explained, the time cost for research and development personnel to search and understand technical documents in the chip design process is remarkably reduced, and efficient knowledge assistance for the design process is achieved.

Description

Technical Field

[0001] This invention relates to intelligent question-answering technology, and more particularly to an intelligent question-answering system for assisting understanding electronic design documents. Background Technology

[0002] In chip design projects, researchers need to frequently write, review, and reference a large number of electronic design documents, which cover key information at each stage of chip development.

[0003] Typical document types include, but are not limited to: (1) Design Specification: Describes the module functions, interfaces, timing logic, etc., mainly in natural language, supplemented by structural diagrams; (2) Product Datasheet: Provides electrical parameters, pinout, etc., and is usually presented in tabular form; (3) Register Manual: Defines register addresses, bit widths, default values, etc., and has strong structured semantics; (4) Circuit schematic diagram: It shows electrical connections in the form of an image, and it is difficult to extract semantics directly; (5) Package and Layout Diagram: Primarily illustrated, supplemented by text descriptions; (6) Simulation and Test Reports: including waveforms, parameter tables, etc., in various formats; (7) Design Change Records: The structure is loose but the key is high.

[0004] However, the aforementioned documents come from diverse sources, are modally heterogeneous, and have inconsistent structures. Furthermore, a large amount of content is embedded in PDFs or images, making information extraction difficult and semantically fragmented. Developers often face the dilemma of not being able to find, understand, or match the information, which seriously affects development efficiency and delivery quality. Summary of the Invention

[0005] The purpose of this invention is to provide an auxiliary understanding and intelligent question-answering system for electronic design documents, which transforms scattered, heterogeneous, and multimodal electronic design data into a queryable, reasonable, and interpretable knowledge base, thereby significantly reducing the time cost for R&D personnel to find and understand technical documents.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A system for assisting in understanding and intelligently answering questions related to electronic design documents, comprising: The data processing module is used to archive documents of different formats in a unified manner. The document parsing module is used to parse documents and represent the different modalities of the document into indexable semantic blocks. The knowledge graph construction module is used to transform loose information distributed across different documents into a structured knowledge graph; The vector database building module is used to vectorize all semantic blocks and store them in the vector database. The RAG reasoning module is used to combine retrieval results with large language model reasoning to form an intelligent question-answering function for engineering.

[0007] Furthermore, the data processing module supports batch reading of different design document formats and performs the following steps: using rule templates and a pre-trained named entity recognition model, entity recognition is performed on the document content, and one or more core tags including project number, module name, register definition, and circuit category are extracted; based on the extracted entity information, the system generates a standardized path tree and archives the documents uniformly.

[0008] Furthermore, the document parsing module performs the following steps: Based on semantic segmentation and title structure splitting, the text is transformed into logically complete paragraphs while retaining chapter hierarchy information; The table content is extracted into a two-dimensional array structure, and the semantics of the column names are identified through a deep learning model, while also associating them with their context headings and paragraph information; The OCR technology is used to extract text information from image content, and the CLIP model is used to generate visual embeddings, supporting image segmentation and semantic retrieval. The parsing results of the above modalities are uniformly encapsulated into a standardized semantic block structure.

[0009] Furthermore, the knowledge graph construction module performs the following steps: Based on the HTMLDOM structure and the hierarchical relationship of document headings, a complete chapter tree is formed, and paragraphs, tables and images are mounted to the corresponding hierarchical paths; The tables are paired with their adjacent descriptive paragraphs, and the images are paired with their corresponding explanatory texts based on positional proximity, title path consistency, and semantic similarity to form logically related modal pairs. Identify key entities in documents and extract relation edges to construct a preliminary knowledge graph; Based on the preliminary knowledge graph, a lightweight graph neural network is introduced to optimize edge weights and sparse redundant connections.

[0010] Furthermore, the formation of the mode pairs includes: Based on the document parsing results, the physical location of each table or image element in the original document is obtained, and it is combined with several preceding and following paragraphs to form a candidate association set. Extract the chapter title paths of the modal element and its candidate paragraphs, and perform consistency checks. If they share at least one common parent node, they are considered consistent; otherwise, they are considered inconsistent. The candidate paragraph text and table title are vectorized separately, and the cosine similarity is calculated using a Chinese semantic embedding model. When the similarity exceeds a preset threshold, it is determined that the semantic similarity requirement is met. Only when semantic similarity meets the criteria and at least one of positional proximity or title path consistency is satisfied, a modal anchor edge is created in the knowledge graph to explicitly bind the table or image to its corresponding paragraph, and the metadata on which the binding is based is recorded, including relative position offset, path matching score and semantic similarity value.

[0011] Furthermore, the lightweight graph neural network adopts a two-layer GraphSAGE architecture, which only dynamically aggregates and updates the weights of subgraphs that are within two hops of the current query context or core design entity.

[0012] Furthermore, in the lightweight graph neural network, the node embeddings in the preliminary knowledge graph are used as input, the semantic block vectors generated in the document parsing stage are initialized and the initial confidence of the edges are set, a dynamic threshold is set, and redundant edges that are unrelated to the core entity path are removed.

[0013] Furthermore, the vector database construction module performs the following steps: For text, a semantic transformation tool is used to convert semantics into vectors; for tables, the table structure is broken down into text and converted into vectors using a text processing model; for images, a visual feature recognition tool is used to convert image features into vectors; all vectors are L2 normalized, and the normalized vectors are subjected to dimensionality reduction processing. Meta-information is added to each vector, including one or more of modality, source path, chapter level, and chunk_id. Relevant candidate results are first searched in the text, table, and image libraries respectively; weighted scores are assigned to the candidate results based on structural information; finally, the results are returned based on the comprehensive score.

[0014] Furthermore, the weighting strategy includes one or more of the following: keyword enhancement, version decay, and modal filtering: if the title of a semantic block or its chapter path contains key engineering keywords, it is determined to have high reference value, and its score is multiplied by an enhancement factor greater than 1; if the semantic block comes from a document that is not the latest version, its score is multiplied by a decay factor less than 1 according to a preset decay strategy; users can specify to recall only specific modal results through the interactive interface, and the system filters out non-target modal results after weighting.

[0015] Furthermore, the RAG inference module performs the following steps: based on vector similarity matching, it retrieves text, tables, and images that are semantically closest to the user's question; it prioritizes recalling content consistent with the target module or title path and expands related entities based on the knowledge graph; it combines the recall results into a structured format, uniformly presenting text paragraphs, table content, and image links to ensure that the context input to the large language model is complete and has structural information; it calls the large language model to generate a natural language answer in combination with the context, and includes the source of the citation and its original path in the answer.

[0016] By adopting the above solution, the present invention has the following advantages: 1. Researchers no longer need to search for electronic design documents. Instead, they can obtain the information they need directly in the form of questions and answers through the assisted understanding and intelligent question-and-answer system. This transforms the loose information scattered in register manuals, schematics, and datasheets into a well-structured and reliable cross-modal knowledge graph, which significantly reduces the time cost for researchers to search for and understand technical documents. 2. This invention achieves unified encapsulation of information from different modalities by defining semantic blocks that include modality type, hierarchical path, and content index; 3. In the knowledge graph construction module, this invention adapts to the structured layout habits of electronic design documents, achieves fine-grained cross-modal alignment, improves the integrity of the knowledge graph, and provides a reliable contextual basis for subsequent RGA inference; 4. This invention introduces a lightweight graph neural network, which dynamically aggregates and updates weights only for subgraphs within a two-hop range of the current query context or core design entity, reducing training overhead; at the same time, it performs sparsity pruning on the graph to improve the accuracy of information. 5. In the vector database construction module, this invention improves the accuracy of the answer information by dynamically adjusting the scores of candidate results through weighted measurement configuration. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the topology of the system of the present invention; Figure 2 This is a schematic diagram illustrating the generation of semantic blocks in this invention; Figure 3 This is a schematic diagram of the knowledge graph of the present invention; Figure 4 This is a schematic diagram illustrating the database creation logic of the multimodal vector database of the present invention; Figure 5 This is a flowchart illustrating the RAG question-and-answer process of the present invention. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0019] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the invention to those skilled in the art. Example

[0020] Please refer to Figure 1 As shown, this invention proposes an auxiliary understanding and intelligent question-answering system for electronic design documents, including a data processing module, a document parsing module, a knowledge graph construction module, a vector database construction module, and a RAG reasoning module.

[0021] The data processing module archives documents of different formats in a unified manner; the document parsing module parses documents and represents the different modalities of the document as indexable semantic blocks; the knowledge graph construction module transforms the loose information distributed in different documents into a structured knowledge graph; the vector database construction module vectorizes all semantic blocks and stores them in the vector database; the RAG reasoning module combines the retrieval results with large model reasoning to form an intelligent question-answering function for engineering.

[0022] Specifically, the data processing module supports batch reading of common design document formats such as doc, docx, pdf, txt, ppt, and schdoc, to ensure that files from different sources can enter a unified processing flow.

[0023] The execution steps are as follows: entity recognition is performed on the document content using rule templates and a pre-trained NER model, and one or more core tags, including project number, module name, register definition, and circuit category, are extracted. These tags provide semantic indexes for document archiving and subsequent retrieval. A standardized path tree is generated based on the extracted entity information, and the documents are archived uniformly.

[0024] In this way, the data processing module unifies and archives the previously chaotic document set, while providing a clear contextual environment for subsequent text parsing.

[0025] Here, "rule template" refers to a set of text matching rules built based on document format features, such as regular expressions or pattern matching rules based on document structure, used to quickly locate entity information with fixed formats in electronic design documents. "NER model" refers to a named entity recognition model trained with labeled data, such as a BERT-based sequence labeling model, used to identify key entities in unstructured text within documents. "Core tags" refer to a set of entity categories strongly related to the electronic design field, including standardized named entities such as project numbers, register addresses, and circuit module names, used to establish a semantic index of document content. "Path tree" refers to a tree-like storage structure built based on entity hierarchy relationships. For example, by using the project number as the root node, the module name as the child node, and the register definition as the leaf node, all register definition documents under the power management module of project A are stored in the path " / project A / power management / register definition". In this way, electronic design documents of different formats are unified in storage structure.

[0026] The document parsing module implements text parsing, table parsing, and image parsing respectively, specifically: First, it identifies paragraph boundaries and chapter heading hierarchy through a pre-trained language model, thereby transforming text into logically complete paragraphs based on semantic segmentation and heading structure splitting. The document is then divided into semantic paragraphs with hierarchical relationships.

[0027] Secondly, the table content is extracted into a two-dimensional array structure, and the semantics of the column names are identified through a deep model. The parsing results are used to save the table content and associate it with context headings and paragraph information. The two-dimensional array structure refers to converting the row and column data of the table into a matrix form for storage. This can be achieved by parsing the coordinate information of the table or HTML table tags, in order to preserve the original data relationship of the table. Taking the register manual as an example, its address mapping table is converted into a two-dimensional array corresponding to rows and columns. At the same time, the context is automatically associated by analyzing the register definition headings of adjacent paragraphs.

[0028] Simultaneously, image analysis is performed on the circuit schematic diagram. The image analysis uses OCR to extract text information from the image content, extracts information such as pin numbers in the diagram annotation, and then uses the CLIP model to extract image feature vectors and semantic vectors, establishes the association between image content and text description, and generates visual embeddings. Finally, by defining a JSON format that includes modality type, hierarchical path, and content index, the parsing results of different modalities are uniformly encapsulated into standardized, indexable semantic blocks.

[0029] Thus, this invention effectively solves the technical problem of the difficulty in uniformly parsing and associating multimodal content in electronic design documents. For example, in the scenario of register manual parsing, table data and corresponding circuit schematics can be automatically associated, so that relevant illustrations can be retrieved simultaneously when querying register functions; for example, when processing Design Specifications containing structural diagrams, image block embedding technology can support the retrieval of illustration areas of specific functional modules, avoiding the defect of traditional methods that can only return the entire image.

[0030] The knowledge graph construction module sequentially executes the steps of hierarchical tree construction, modal anchor binding, entity relation extraction, and lightweight graph neural network optimization.

[0031] First, based on the HTMLDOM and the hierarchical relationship of headings, a complete chapter tree structure is formed, and paragraphs, tables, and images are mounted to their corresponding hierarchical paths to complete the hierarchical tree construction, thus solving the problem of chaotic search paths caused by document content being scattered across different chapters.

[0032] Secondly, tables and descriptive paragraphs, as well as images and explanatory text, are paired based on positional relationships, title path consistency, and semantic similarity to form logically related modal pairs, thus eliminating logical gaps caused by the isolation of different modal information.

[0033] Specifically, the system first obtains the physical location of each table or image element in the original document (such as PDF page offset or HTML DOM node index) based on the document parsing results, and then combines it with several preceding and following paragraphs (the default window is ±3 paragraphs) to form a candidate association set.

[0034] Subsequently, the system extracts the chapter title paths to which the modal element and its candidate paragraphs belong, and performs consistency checks. If the two share at least one common parent node, the title path consistency condition is satisfied. For example: if the table is located in "4.3.1 Register Definition Table" and the paragraph above it belongs to "4.3 Interface Registers", they share the parent node "4.3", and are therefore considered consistent; if the image is located in "5.1 Circuit Schematic" and the explanatory text below it belongs to "5.1.1 Pin Distribution", they share the parent node "5.1", and are also considered consistent; if the paths are "3.2.1 Power Control" and "4.1 Clock Management" respectively, and they do not share a common parent node, they are considered inconsistent.

[0035] At the same time, the system vectorizes the candidate paragraph text and table title respectively, and uses Chinese semantic embedding models such as BGE-large-zh or Qwen-Embedding to calculate cosine similarity. When the similarity exceeds the preset threshold, it is considered to meet the semantic similarity requirements.

[0036] Ultimately, only when semantic similarity meets the standard and at least one of positional proximity or title path consistency is satisfied will the system create a "modal anchor" edge in the knowledge graph, explicitly binding the table or image to its corresponding paragraph and recording the metadata on which the binding is based, including relative position offset, path matching score and semantic similarity value.

[0037] This binding mechanism, combined with the structured layout conventions and hierarchical heading system unique to electronic design documents (description first, then tables, centered diagrams, and followed by annotations), achieves fine-grained cross-modal alignment.

[0038] Thus, taking a register manual as an example, the system can accurately identify the "REG_VDD_CTRL" table under the "4.3 Interface Registers" section and bind it to the paragraph above, "This register is used to enable core voltage output and is disabled by default after reset," even if the paragraph does not explicitly mention the table name. This kind of association not only improves the completeness of the knowledge graph but also provides a reliable multimodal context for the subsequent RAG inference stage. When a user queries "How to enable VDD_CTRL?", the system can simultaneously recall the function description, register bit definition table, and associated power supply block diagram, achieving truly intelligent question answering for engineering scenarios.

[0039] Next, entities are identified and relation edges are extracted to form a preliminary knowledge graph. A lightweight graph neural network is introduced into the knowledge graph construction module to optimize the edge weights and sparsify the structure of the preliminary knowledge graph, thereby preserving high-value associations, suppressing noise interference, and improving the engineering practicality and reasoning efficiency of the graph.

[0040] The lightweight graph neural network adopts a two-layer GraphSAGE architecture, which dynamically aggregates and updates weights only on subgraphs that are within two hops of the current query context or core design entities (such as registers, modules, and pins), thus avoiding high-overhead training of the entire graph.

[0041] Based on a lightweight graph neural network, this system takes node embeddings from a preliminary knowledge graph as input. The initial confidence of each edge is calculated from the semantic block vectors generated during document parsing, derived from modal anchor binding scores or entity co-occurrence frequencies. The optimized weights of each edge are output through neighbor node feature aggregation and a learnable weight matrix. The training process employs weak supervision. Positive samples are derived from manually labeled high-frequency effective relation pairs (such as "register-address" and "module-power supply voltage"), while negative samples are constructed by randomly disconnecting low-frequency edges or introducing cross-module unrelated entity pairs. The loss function includes both edge existence binary classification loss and weight regression loss to balance structure preservation and numerical calibration.

[0042] In the preliminary knowledge graph, the initial confidence C_init of each relation edge is determined by one of the following two methods: (1) if the edge originates from "modal anchor binding", then C_init = S_binding (i.e., the aforementioned binding score); (2) if the edge originates from "entity co-occurrence", then C_init = N_co-occur / N_total. Where: N_co-occur represents the number of times two entities co-occur in the same document or related documents; N_total represents the total number of occurrences of all different entity pairs in all documents processed by the system.

[0043] In the initial knowledge graph, the initial confidence of each relation edge is determined by its source. If it originates from "modal anchor binding", the binding score is used. If it originates from "entity co-occurrence", it is calculated as the ratio of the number of documents in which the two entities co-occur to the total number of occurrences of all entity pairs. For example, if "temporal margin" and "temperature condition" co-occur in 5 documents, and there are a total of 50 entity pair occurrence records in the system, then its initial confidence is 0.1.

[0044] During the inference phase, the system performs sparsity pruning on the graph based on edge weights optimized by a lightweight graph neural network: setting a dynamic threshold (e.g., retaining the top 80% of edges by weight, or connections with an absolute weight greater than 0.6) removes redundant edges unrelated to the core entity paths. In this way, high-frequency and semantically close relationships (such as "timing margin - temperature condition") are strengthened, while weak relationships without direct engineering relevance (such as "package diagram number" from the package diagram and "register default value" from the register manual) are pruned due to their low co-occurrence frequency and loose semantics, resulting in weights below the threshold.

[0045] It should be noted that the modal anchor binding described in this invention is only applicable to table-paragraph or image-description text pairs that are physically adjacent within the same document (within the same page or ±3 paragraphs) and have contextual dependencies; entity associations across pages or documents do not constitute modal pairs and may only be included in the preliminary graph through global co-occurrence, but such weak associations are usually suppressed during the graph neural network optimization stage.

[0046] Thus, this invention can transform loose information scattered in register manuals, schematics, and datasheets into a well-structured and reliably related cross-modal knowledge graph, effectively solving the problem that designers need to manually link multiple documents and significantly improving information retrieval efficiency and accuracy.

[0047] The vector database construction module performs the following steps: Modality-independent embedding is employed. For text content, a semantic transformation tool is used to convert its semantic information into vector representations. For table content, its structured data (including table headers, row and column content, and contextual descriptions) is first converted into natural language descriptive text, and then vectors are generated using a text processing model. For image content, OCR is first used to extract text information from the image, and then visual feature recognition tools such as CLIP are combined to generate overall or segmented visual embedding vectors. Specifically, the semantic transformation tool uses Chinese embedding models such as Qwen3-Embedding; the text processing model uses a sequence encoder based on the BERT architecture; and the visual feature recognition tool uses the CLIP multimodal pre-trained model to adapt to the data characteristics of different modalities and ensure the semantic fidelity of the vector representations.

[0048] Unified vector processing involves uniformly L2 normalizing the vectors generated from each modality to eliminate inconsistencies in vector scale caused by model differences or modal characteristics. Subsequently, structured meta-information is appended to each vector. This meta-information includes one or more of the following: modality type, original document source path, chapter level, unique semantic block identifier (chunk_id), and document version number. This information is used for structure-aware sorting in the subsequent retrieval stage.

[0049] This method fully integrates the unique engineering semantic structure and version management requirements of electronic design documents, and dynamically weights and intelligently sorts the multimodal retrieval results.

[0050] Specifically, when a user initiates a query, the system first independently searches the text, table, and image libraries based on vector similarity to obtain candidate semantic blocks and their initial similarity scores for each modality. Subsequently, the system dynamically adjusts the initial scores based on the structured metadata carried by each candidate semantic block. This structured metadata includes: the semantic block title, the path to its corresponding chapter, the document version number, and the modality type.

[0051] Weighting strategies include one or more of the following: keyword enhancement, version decay, and modal filtering. Specifically: If the title of a semantic block or its chapter path contains key engineering keywords such as "definition", "description", "configuration", or "register mapping", it is considered to have high reference value, and an improvement factor (e.g., multiplied by 1.2) is applied to its score. If the semantic block comes from a document that is not the latest version (such as a version number lower than the current trunk version), its weight will be appropriately reduced according to the preset attenuation strategy (e.g., multiplied by 0.8) in order to prioritize the timeliness of the information. Users can specify to recall only specific modal results (such as "tables only" or "images only") through the interactive interface, and the system will filter out non-target modal results after weighting.

[0052] Finally, all candidate results are sorted according to the weighted composite score, and the Top-N results are returned as context input for the RAG inference module.

[0053] In this way, the engineering semantic conventions of electronic design documents (such as "prioritizing definition-related content" and "prioritizing the latest version") are transformed into quantifiable retrieval and ranking rules. This allows the system to not only find semantically similar content but also the most authoritative and relevant content in the engineering field. For example, when a user queries "default value of the VDD_CTRL register," the system will not only recall multiple paragraphs mentioning "VDD_CTRL" but will also prioritize the semantic score of the table block located under the "4.3 Register Definition" section with the title explicitly marked "default value," while suppressing weakly relevant results from older manuals or the "Package Description" section.

[0054] Thus, through the above-mentioned vector construction and fusion, the present invention effectively solves practical problems such as scattered multimodal data retrieval results, broken cross-modal associations, and interference from historical versions in electronic design documents, enabling the register definition table to be recalled in a single retrieval along with its corresponding functional description paragraphs and circuit schematics, significantly improving the accuracy and engineering practicality of the question-and-answer system.

[0055] The reasoning steps of the RAG reasoning module include: The multi-layered recall process begins with the following steps: First, based on vector similarity matching, candidate semantic blocks that are semantically closest to the user's question are retrieved from text, table, and image databases to obtain initial recall results. Second, core engineering entities in the user's query are identified as "target modules" (e.g., register name "VDD_CTRL", functional module "Power Management Unit", or signal name "CLK_EN"), and semantic blocks whose chapter paths are consistent with or highly related to the target module are prioritized (e.g., when querying "VDD_CTRL", the content under the path "4.3 Interface Register" is prioritized for recall). Finally, related entities are expanded based on the constructed knowledge graph: starting from the target module, its directly connected neighboring entities are searched in the knowledge graph (e.g., "Power Supply Voltage", "Reset Value", "Enable Signal", etc. associated with "VDD_CTRL"), and the semantic blocks corresponding to these neighboring entities are included in the recall scope, thereby expanding context coverage and improving the completeness and accuracy of the answer.

[0056] The Prompt build combines the recall results into a Markdown structure, uniformly displaying text paragraphs, table content, and image links, ensuring that the context input to the large language model is complete and has structural information; The system generates answers by calling a large language model and combining it with the structured context to produce natural language answers. The answers automatically include the source of the citation and the path to the original document, enabling traceable engineering question answering.

[0057] Thus, this invention achieves collaborative reasoning of cross-modal information in electronic design documents through multi-layered recall and knowledge graph-driven context expansion. This allows users to accurately obtain composite answers that integrate text descriptions, register tables, and circuit diagrams through a single question, effectively reducing the risk of misreading due to information fragmentation. Furthermore, the reference tracing mechanism improves the reliability and efficiency of design verification.

[0058] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An auxiliary understanding and intelligent question-answering system for electronic design documents, characterized in that, Comprise: a data processing module for archiving documents of different formats uniformly; a document parsing module for parsing documents and representing different modal contents of the documents as indexable semantic blocks uniformly; a knowledge graph construction module for converting loose information distributed in different documents into a structured knowledge graph; a vector database construction module for vectorizing all semantic blocks and storing them in a vector database; a RAG inference module for combining search results with large language model inference to form an engineering-oriented intelligent question and answer function.

2. The electronic design document oriented aided understanding and intelligent question answering system of claim 1, wherein, The data processing module supports batch reading of different design document formats and performs the following steps: entity recognition of document content is performed through rule templates and pre-trained named entity recognition models, and one or more core tags including project number, module name, register definition, and circuit category are extracted; based on the extracted entity information, the system generates a standardized path tree to uniformly archive the documents.

3. The electronic design document oriented aided understanding and intelligent question answering system as claimed in claim 1, wherein, The document parsing module performs the following steps: Based on semantic segmentation and title structure splitting, the text is converted into logically complete paragraphs while retaining chapter hierarchy information; Table content is extracted as a two-dimensional array structure, and a deep learning model is used to identify the semantics of column names while associating them with their context titles and paragraph information; OCR technology is used to extract text information from image content, and a CLIP model is used to generate visual embeddings to support image segmentation and semantic retrieval; The parsing results of the above modalities are encapsulated into a standardized semantic block structure.

4. The electronic design document oriented aided understanding and intelligent question answering system of claim 3, wherein, The knowledge graph construction module performs the following steps: Based on HTML DOM structure and document title hierarchy, a complete chapter tree is formed, and paragraphs, tables, and images are mounted under the corresponding hierarchical path; Tables and their adjacent descriptive paragraphs, images and their corresponding explanatory text are paired based on location proximity, title path consistency and semantic similarity to form logically associated modal pairs; Identify key entities in the document and extract relationship edges to construct a preliminary knowledge graph; Based on the preliminary knowledge graph, a lightweight graph neural network is introduced to optimize edge weights and sparsify redundant connections.

5. The electronic design document oriented aided understanding and intelligent question answering system as claimed in claim 4, wherein, The formation of the modal pair includes: Based on the document parsing results, the physical location of each table or image element in the original document is obtained, and it is combined with the front and rear several natural paragraphs to form a candidate association set; Extract the chapter title path to which each modal element and its candidate paragraph belongs, and perform consistency checking. If they share at least one common parent node, they are determined to be consistent, otherwise, they are determined to be inconsistent; The candidate paragraph text and table title are respectively vectorized, and the cosine similarity is calculated using a Chinese semantic embedding model. When the similarity exceeds a predetermined threshold, it is determined to meet the semantic similarity requirement; only when the semantic similarity meets the requirement and at least one of the location proximity or title path consistency is met, a modal anchor edge is created in the knowledge graph to explicitly bind the table or image to its corresponding paragraph, and the metadata of the binding is recorded, including the relative position offset, path matching score and semantic similarity value.

6. The electronic design document oriented aided understanding and intelligent question answering system as claimed in claim 4, characterized in that: The lightweight graph neural network adopts a two-layer GraphSAGE architecture and only performs dynamic aggregation and weight update on a subgraph within a two-hop range of the current query context or core design entity.

7. The electronic design document oriented aided understanding and intelligent question answering system as claimed in claim 4, characterized in that: In the lightweight graph neural network, the node embedding in the preliminary knowledge graph is taken as input, the semantic block vector generated by the document parsing stage is initialized and the edge initial confidence is set, a dynamic threshold is set, and redundant edges unrelated to the core entity path are removed.

8. The electronic design document oriented aided understanding and intelligent question answering system as claimed in claim 4, wherein, The vector database construction module performs the following steps: The text is converted into a vector using a semantic conversion tool, the table structure is disassembled into text, the text processing model is used to convert it into a vector, and the image features are converted into a vector using a visual feature recognition tool; All vectors are L2 normalized, the normalized vectors are dimensionally reduced, and meta information is added to each vector, including one or more of modalities, source paths, chapter levels, and chunk_ids; First, relevant candidate results are searched in the text, table, and image libraries respectively, the candidate results are weighted according to the structural information, and finally the results are returned according to the comprehensive score.

9. The electronic design document oriented aided understanding and intelligent question answering system as claimed in claim 8, wherein, The weighting strategy includes one or more of keyword enhancement, version decay, and modality filtering: if the semantic block title or its chapter path contains engineering key keywords, it is determined to have high reference value, and its score is multiplied by an enhancement factor greater than 1; if the semantic block is derived from a non-latest version of the document, according to the preset decay strategy, its score is multiplied by a decay factor less than 1; the user can specify to recall only specific modality results through the interactive interface, and the system filters non-target modality results after weighting.

10. The electronic design document oriented aided understanding and intelligent question answering system as claimed in claim 4, wherein, The RAG reasoning module performs the following steps: Based on vector similarity matching, retrieve the text, table, and image closest to the user's question semantics; Prioritize recalling content consistent with the target module or title path, and expand related entities based on the knowledge graph; Combine the recall results into a structured format, and present the text paragraphs, table contents, and image links uniformly, ensuring that the context input into the large language model is complete and has structural information; Call the large language model to generate a natural language answer based on the context, and include the reference source and its original path in the answer.

Images