Enterprise innovation portrait construction and recommendation method and system based on multi-mode RAG

Abstract

The invention discloses a multi-mode RAG-based enterprise innovation portrait construction and recommendation method and system, and relates to the technical field of artificial intelligence, and the method comprises the steps: collecting and analyzing enterprise multi-source heterogeneous data, and converting the data into a structured semantic object set; and constructing an enterprise semantic tree and a comprehensive representation vector by executing recursive semantic aggregation on the set. Secondly, mapping the semantic tree into an enterprise innovation knowledge graph, and performing fact verification and semantic rewriting on user query based on the graph and an external credible knowledge source to generate an optimized query instruction; finally, executing double-path recommendation: performing multi-granularity matching on a vector retrieval path by utilizing an enterprise semantic tree and a unified representation library, and returning a related enterprise list; the generative reasoning path takes a semantic tree, a representation library and optimization query as a joint knowledge source, and drives a large language model to generate an interpretable recommendation report containing detailed reasons and comparative analysis. According to the invention, deep description and interpretable intelligent recommendation of enterprise innovation ability are realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method and system for constructing and recommending enterprise innovation profiles based on multimodal RAG. Background Technology

[0002] Currently, in-depth assessment and intelligent recommendation of enterprise innovation capabilities have become an urgent need in fields such as industry analysis and investment decision-making. However, existing enterprise information platforms and various industry mapping systems have significant limitations in their technical architecture when it comes to constructing enterprise innovation profiles, making it difficult to meet the requirements of in-depth, dynamic, and intelligent analysis in professional scenarios. The main shortcomings of existing technologies are concentrated in four aspects: First, the information dimensions are singular, relying excessively on structured business and financial data, lacking effective semantic parsing and fusion capabilities for unstructured multimodal data such as patents, technical reports, and product images, resulting in profiles that cannot deeply depict the core technology. Second, the profiles are static and superficial, mostly using predefined tags or keyword lists, unable to construct a dynamic knowledge system with clear relationships between entities that can support logical reasoning, and thus failing to reflect the dynamic process of enterprise innovation. Third, the recommendation logic is simple and mechanical, mainly based on matching industry classifications or explicit supply chain relationships, lacking deep semantic mining of technological similarity, innovation complementarity, etc., resulting in insufficient accuracy of recommendation results. Finally, the human-computer interaction is not intelligent. Users need to filter through fixed conditions and cannot flexibly express complex intentions using natural language. The system lacks the ability to understand natural language and interpret intentions.

[0003] While cutting-edge AI technologies such as Retrieval Augmentation (RAG) and Large Language Modeling (LLM) have demonstrated their potential in general domains, a systematic integration of these technologies in vertical enterprise innovation analysis scenarios, and the design of an end-to-end solution encompassing multimodal data processing, dynamic knowledge construction, and generative intelligent recommendation, remains a gap. Therefore, developing a method to overcome these shortcomings and achieve deep, dynamic, and interpretable enterprise innovation profiling and recommendation is of significant technological necessity and application value. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention proposes a method and system for constructing and recommending enterprise innovation profiles based on multimodal RAG, which enables in-depth and dynamic characterization of enterprise innovation capabilities and accurate and interpretable intelligent recommendations.

[0005] The first aspect of this invention provides a method for constructing and recommending enterprise innovation profiles based on multimodal RAG, comprising the following steps: Collect text, image, and relational data from enterprises, parse the collected multi-source heterogeneous data, and transform it into a unified set of structured semantic objects. Recursive semantic aggregation is performed on the structured semantic object set to build a top-level semantic vector representing the core innovation capabilities of the enterprise from bottom to top, and an interpretable enterprise semantic tree is generated. The top-level semantic vector is fused through a cross-modal attention mechanism to output a comprehensive enterprise representation vector and build a unified representation library. The entities and relationships in the enterprise semantic tree are mapped into a structured enterprise innovation knowledge graph. The system receives natural language queries from users and performs fact verification and semantic rewriting based on the large language model in the enterprise innovation knowledge graph and external trusted knowledge sources to generate optimized structured query instructions. Based on the optimized structured query instructions, a dual-path recommendation, including a vector retrieval path and a generative reasoning path, is executed. The vector retrieval path uses the optimized structured query instructions to perform multi-granular semantic matching based on the index constructed by the enterprise semantic tree and the unified representation library, retrieves and returns a list of relevant enterprises; the generative reasoning path uses the enterprise semantic tree, the unified representation library and the optimized structured query instructions as a joint enhanced knowledge source to drive the large language model to generate an interpretable recommendation report containing reasons and comparative analysis.

[0006] This solution collects text, image, and relational data from enterprises, parses the collected multi-source heterogeneous data, and transforms it into a unified set of structured semantic objects, including: Collect multi-source heterogeneous data from target companies through public APIs, web crawlers, and collaborative data interfaces. The multi-source heterogeneous data includes text data, image data, and relational data. The text data is processed using a multi-level semantic parsing model to extract basic semantic units containing technical entities, performance indicators, method steps, and application scenarios. The semantic associations of the basic semantic units are obtained. Based on the semantic associations, the basic semantic units with associations are dynamically combined into composite semantic blocks as structured semantic objects. The types of composite semantic blocks include technical solution semantic blocks and application result semantic blocks, and the logical relationships between internal units are recorded. The image data is parsed using a visual language model to generate a structured description tuple containing fields for visual subject, functional description, and associated technical items, which serves as a visual semantic object. The relational data is mapped to a primary triple object in the format of (head entity, relation type, tail entity); All generated structured semantic objects, visual semantic objects, and primary triplet objects are converted into a unified set of structured semantic objects.

[0007] In this solution, recursive semantic aggregation is performed on the structured semantic object set to construct a top-level semantic vector representing the core innovation capabilities of the enterprise from bottom to top, and an interpretable enterprise semantic tree is generated, including: The finest-grained semantic objects in the set of structured semantic objects of different modalities are vectorized, and bottom-up iterative semantic abstraction is performed. In each iteration, the semantic object vectors of the current level are clustered based on semantic similarity. For the generated clusters, the abstraction model is invoked to generate high-order semantic abstract vectors representing the overall semantics of the clusters and corresponding natural language abstract description texts. By using a cross-level attention mechanism, the generated higher-order semantic abstract vectors are corrected in a context-aware manner by utilizing the semantic object vectors within the current cluster, the semantic object vectors of the lower layer, and other related higher-order semantic abstract vectors at the same level. Repeated iterative loops until a preset abstraction level is reached, outputting hierarchical enterprise semantic trees of different modalities. The leaf nodes of the enterprise semantic tree are the initial semantic objects, and the top-level nodes are the top-level semantic vector set representing the core innovation capabilities of the enterprise.

[0008] In this solution, the top-level semantic vector is fused using a cross-modal attention mechanism to output a comprehensive enterprise representation vector and construct a unified representation library, including: Obtain the top-level semantic vector sets of different modalities, and input the top-level semantic vector sets of text, images, and relationships into the cross-modal attention fusion layer; In the cross-modal attention fusion layer, the top-level semantic vectors of different modalities are interacted, aligned and weighted in importance in the semantic space through the attention mechanism. The multimodal semantic information after interaction and weighting is integrated to output the fused comprehensive enterprise representation vector. All comprehensive enterprise representation vectors are stored to form a unified enterprise representation library.

[0009] In this solution, the entities and relationships in the enterprise semantic tree are mapped into a structured enterprise innovation knowledge graph, including: The semantic nodes at each level in the enterprise semantic tree are mapped to corresponding entities in the knowledge graph according to the node type, level position and abstract description text, and each entity is assigned a dynamic entity type determined by the enterprise semantic tree and a core attribute traceable to the original data source. Based on the inherent parent-child node hierarchy in the enterprise semantic tree, a vertical graph relationship is generated. The association weights between different semantic nodes learned through a cross-level attention mechanism during the recursive semantic aggregation process are obtained. Based on the association weights, a horizontal graph relationship is automatically identified and generated. The vertical graph relationship includes "contains", "realizes", or "decomposes into". The horizontal graph relationship includes "collaborates", "supports", "applies to", or "associates". A structured enterprise innovation knowledge graph is constructed based on the entities, vertical graph relationships, and horizontal graph relationships, and the enterprise innovation knowledge graph is used as a profile of enterprise innovation capabilities.

[0010] In this solution, the system receives natural language queries from users, performs fact-checking and semantic rewriting based on the enterprise innovation knowledge graph and external trusted knowledge sources using a large language model, and generates optimized structured query instructions, including: After receiving a user's natural language query, the large language model is invoked to parse and decompose the natural language query into several atomic fact subqueries. For each atomic fact subquery, dynamic evidence chain verification is performed, including internal graph tracing and external knowledge completion. In the internal graph tracing process, the atomic fact subqueries are interacted with the enterprise innovation knowledge graph in real time. The minimum semantic evidence path that can support or refute the corresponding atomic fact subqueries is constructed in the enterprise innovation knowledge graph. The completeness, matching degree and attribute consistency of the minimum semantic evidence path are evaluated, and the internal evidence evaluation result is generated. In the external knowledge completion process, elements involving timeliness and consensus knowledge are extracted from the atomic fact sub-queries and sent to external trusted knowledge sources for retrieval and comparison to obtain external timeliness evidence results. The internal evidence evaluation results and external timeliness evidence results are weighted, fused, and conflict resolved to generate a multidimensional confidence vector corresponding to each atomic fact subquery. The multidimensional confidence vector includes factual confidence, timeliness confidence, and evidence conflict identifier. Based on the multidimensional confidence vector, the large language model is invoked again to drive the context-aware semantic rewriting strategy engine to reconstruct the original natural language query, and output the optimized structured query instruction as well as natural language explanations of the correction items and the basis for the decision.

[0011] In this solution, the vector retrieval path includes: Based on the enterprise semantic tree and the enterprise unified representation library, a multi-granularity semantic index is constructed. The multi-granularity semantic index stores the comprehensive representation vector of the enterprise and associates it with the intermediate semantic vectors generated in the recursive semantic aggregation process. All vectors are bound to the source nodes in the enterprise semantic tree to record the hierarchical relationship. The optimized structured query command is parsed and vectorized synchronously at multiple granularities to generate a global query vector that represents the overall intent of the query, and a local query vector that represents the specific constraints in the query. The global query vector is used to calculate the similarity with the comprehensive enterprise representation vector of all enterprises in the multi-granularity semantic index to complete the top-level coarse screening and obtain a preliminary candidate enterprise set. In the preliminary candidate enterprise set, the local query vector is compared with the intermediate semantic vector of the preliminary candidate enterprise stored in the index using fine-grained similarity calculation. By combining the matching scores calculated from the top-level coarse screening and the fine-grained similarity, the preliminary candidate companies are weighted and sorted to generate the final ranking list of relevant companies, and preliminary structured matching reasons are generated based on the matching process.

[0012] In this scheme, the generative reasoning path includes: The system receives a ranking list of relevant enterprises as a candidate enterprise list, and integrates and encapsulates the enterprise semantic tree, the unified enterprise representation library, and the optimized structured query instructions to form a joint enhanced knowledge source input into the large language model. The enterprise semantic tree provides the logical concept hierarchy and entity relationship network required for analysis, the unified representation library provides the comprehensive enterprise representation vectors and their relative positional relationships for quantitative comparison, and the optimized query provides a precise analysis task framework and constraints. Based on the joint enhanced knowledge source, prompts are generated to guide the large language model to perform inference chain analysis. In the inference chain analysis, the matching of candidate enterprises with the query intent at different abstraction levels is identified and interpreted according to the enterprise semantic tree and the optimized structured query instructions. Based on the relationship between the comprehensive enterprise representation vector in the unified representation library and the enterprise semantic tree, the candidate enterprises are compared in multiple dimensions, including technological similarity, capability structure differences, and evidence characteristics. In addition, an explanation is generated for each recommendation conclusion based on the knowledge base. The large language model integrates all the outputs of the inference chain analysis based on the preset report template, and generates an interpretable recommendation report that includes a list of recommended companies, detailed reasons for the recommendations, multi-dimensional comparative analysis of companies, and risk warnings based on the analysis.

[0013] The second aspect of this invention provides an enterprise innovation profile construction and recommendation system based on multimodal RAG. The system includes: a multi-source heterogeneous data acquisition and access module, a semantic parsing and objectification module, a semantic aggregation and representation learning module, a knowledge graph construction and profile solidification module, a query intent understanding and purification enhancement module, a multi-granularity intelligent retrieval and matching module, and a generative reasoning and interpretable report generation module. The multi-source heterogeneous data acquisition and access module is responsible for acquiring multi-source heterogeneous data from various internal and external data sources in a standardized manner, including: text, images, and relational data; The semantic parsing and objectification module parses the collected multi-source heterogeneous data and transforms it into a unified set of structured semantic objects, including: structured semantic objects, visual semantic objects, and primary triplet objects; The semantic aggregation and representation learning module processes structured semantic objects, visual semantic objects, and primary triplet objects in parallel, performs recursive semantic aggregation, constructs a top-level semantic vector representing the core innovation capabilities of the enterprise from bottom to top, and generates an interpretable enterprise semantic tree. The top-level semantic vector is then fused through a cross-modal attention mechanism to output a comprehensive enterprise representation vector and construct a unified representation library. The knowledge graph construction and profile solidification module maps entities and relationships in the enterprise semantic tree into a structured enterprise innovation knowledge graph, which is defined as the enterprise's structured innovation profile. The query intent understanding and purification enhancement module receives the user's natural language query, decomposes the natural language query into atomic fact subqueries based on the large language model, performs fact verification in the enterprise innovation knowledge graph and external trusted knowledge sources, and, based on the verification results, calls the large language model again to perform semantic rewriting to generate an optimized structured query instruction. The multi-granularity intelligent retrieval and matching module performs multi-granularity semantic matching on the optimized structured query command based on the index constructed by the enterprise semantic tree and the unified representation library, retrieves and returns a list of relevant enterprises; The generative reasoning and interpretable report generation module uses the enterprise semantic tree, unified representation library, and optimized structured query instructions as joint enhanced knowledge sources to drive the large language model to perform semantic alignment, multi-dimensional comparison, and quantitative analysis based on the joint knowledge sources, thereby generating an interpretable recommendation report.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention utilizes recursive semantic aggregation to dynamically construct a rich hierarchical semantic tree and its corresponding enterprise innovation knowledge graph from multi-source heterogeneous data, building it from the bottom up. This enables enterprise profiles to clearly display a structured semantic network from core technologies to macro-level innovation capabilities, with traceable inter-entity relationships. The profiles dynamically evolve as enterprise data is updated, accurately reflecting the growth and changes in their innovation capabilities. Within each modality, refined semantic parsing and objectification are first performed, followed by deep fusion using a cross-modal attention mechanism in the high-level semantic space after recursive abstraction. This ensures that the final generated comprehensive enterprise representation vector deeply integrates technical details from patents, morphological features from images, and ecological information from collaborative relationships. A query purification and enhancement mechanism based on a dual-track evidence chain is introduced, comparing and verifying user queries against internal authoritative knowledge graphs and external trusted sources in real time. This effectively identifies and corrects noise and errors in queries, and transforms ambiguous intents into precise retrieval instructions through intelligent semantic rewriting.

[0015] This invention reduces the time and cost of traditional enterprise analysis, which heavily relies on manual research, information processing, and report writing. It enables intelligent enterprise innovation assessment and recommendation decision-making processes, providing an efficient tool for business scenarios such as industry analysis and technological cooperation. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating the method for constructing and recommending enterprise innovation profiles based on multimodal RAG is shown. Figure 2 A flowchart illustrating the process of performing recursive semantic aggregation is shown. Figure 3 A flowchart illustrating the process of generating optimized query instructions through fact-checking and semantic rewriting is shown. Figure 4 A block diagram of an enterprise innovation profile construction and recommendation system based on multimodal RAG is shown. Detailed Implementation

[0018] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0020] like Figure 1 As shown, this embodiment provides a method for constructing and recommending enterprise innovation profiles based on multimodal RAG, including: Collect text, image, and relational data from enterprises, parse the collected multi-source heterogeneous data, and transform it into a unified set of structured semantic objects. Recursive semantic aggregation is performed on the structured semantic object set to build a top-level semantic vector representing the core innovation capabilities of the enterprise from bottom to top, and an interpretable enterprise semantic tree is generated. The top-level semantic vector is fused through a cross-modal attention mechanism to output a comprehensive enterprise representation vector and build a unified representation library. The entities and relationships in the enterprise semantic tree are mapped into a structured enterprise innovation knowledge graph. The system receives natural language queries from users and performs fact verification and semantic rewriting based on the large language model in the enterprise innovation knowledge graph and external trusted knowledge sources to generate optimized structured query instructions. Based on the optimized structured query instructions, a dual-path recommendation, including a vector retrieval path and a generative reasoning path, is executed. The vector retrieval path uses the optimized structured query instructions to perform multi-granular semantic matching based on the index constructed by the enterprise semantic tree and the unified representation library, retrieves and returns a list of relevant enterprises; the generative reasoning path uses the enterprise semantic tree, the unified representation library and the optimized structured query instructions as a joint enhanced knowledge source to drive the large language model to generate an interpretable recommendation report containing reasons and comparative analysis.

[0021] It should be noted that multi-source heterogeneous data of the target company is collected through public APIs, web crawlers, and cooperative data interfaces. The multi-source heterogeneous data includes text data, image data, and relational data. The text data includes: patent texts, academic papers, project proposals, technical white papers, industry research reports, press releases, etc. The image data includes: product images, technical principle diagrams, laboratory environment photos, promotional posters, etc. The relational data includes supplier and customer lists, cooperative research institutions, investor information, etc.

[0022] The text data is processed using a multi-level semantic parsing model to extract basic semantic units corresponding to technical entities, performance indicators, methodological steps, and application scenarios. These basic semantic units are then encapsulated into objects containing content, type, context, and their position in the original text. For example, for battery companies, this involves identifying technical entities such as solid-state electrolytes, performance indicators such as energy density greater than 300Wh / kg, methodological steps such as co-precipitation synthesis, and automotive batteries. Semantic relationships among these basic semantic units are obtained. Based on these relationships, related basic semantic units are dynamically combined into composite semantic blocks, which serve as structured semantic objects. These composite semantic blocks include technical solution semantic blocks and application result semantic blocks, and the logical relationships between their internal units are recorded. For example, technical entities, performance indicators, and methodological steps are combined into a technical solution semantic block, and technical entities, application scenarios, and performance indicators are combined into an application result semantic block. Each semantic block object defines the logical relationships between its internal units. The image data is parsed using a visual language model to generate structured descriptive tuples containing fields for visual subjects, functional descriptions, and related technical items, serving as visual semantic objects. For example, for a precision instrument product image, the visual language model can not only identify the instrument but also generate structured descriptive tuples for functions and key technologies. For technical schematic diagrams such as chip architecture diagrams, the visual language model can parse components, connections, and purposes, converting image information into semantic symbols. The relational data is mapped to primary triple objects in the format of (head entity, relation type, tail entity), such as (Company A, Investment, Company B). These primary triple objects provide seed relationships for subsequent knowledge graph construction. All generated structured semantic objects, visual semantic objects, and primary triple objects are then converted into a unified set of structured semantic objects.

[0023] It should be noted that recursive semantic aggregation is performed on the structured semantic object set, such as... Figure 2As shown, the finest-grained semantic objects in the structured semantic object set of different modalities are vectorized. These finest-grained semantic objects include, for example, a single technical entity, a basic performance index, or a simple relation triple. A bottom-up iterative semantic abstraction is performed, calculating the semantic similarity between vectors in each iteration. Objects describing the same micro-topic are clustered together; for example, different technical solutions, performance parameters, and preparation methods of solid electrolytes are clustered into clusters. For the generated clusters, an abstraction model (a lightweight large language model) is invoked to generate high-order semantic abstract vectors representing the overall semantics of the clusters and corresponding natural language abstract description text. The high-order semantic abstract vectors generated in the first iteration are used as new input to repeat the clustering process at a higher level of abstraction. This round of clustering occurs at a higher, more abstract level. For example, the abstract vectors in the first round regarding sulfide electrolytes and polymer electrolytes both point to electrolyte material development; the abstraction model is invoked to generate higher-order abstract vectors and descriptions demonstrating solid electrolyte material development capabilities.

[0024] Introducing a cross-level attention mechanism in the iterative clustering loop, when generating new abstract vectors, it not only focuses on information within the current cluster (the vectors of objects within the current cluster refer to the vector representations of all member objects in the cluster currently generating the current higher-order abstract vector), but also pays attention to the vectors of lower-level sub-concepts and other related concepts at the same level. Lower-level sub-concepts refer to the vectors of the original components from the next lower level that constitute objects within the current cluster. Vectors of other related concepts at the same level refer to all higher-order abstract vectors that have been generated or are being generated at the Nth level, excluding the currently being processed cluster. Through context-aware vector correction, it ensures that the generated new concepts are consistent in the global semantic network, avoiding abstraction bias. The iterative loop is repeated until a preset abstraction level is reached, such as the core innovation capability dimension of an enterprise, outputting a hierarchical enterprise semantic tree with different modalities. The leaf nodes of the enterprise semantic tree are the initial semantic objects, and the top-level nodes are the set of top-level semantic vectors representing the core innovation capabilities of the enterprise.

[0025] Using structured semantic objects, visual semantic objects, and primary triplet objects as inputs, text semantic trees, image semantic trees, and relational semantic trees are constructed through iterative vectorization, clustering, abstract generation, and context correction. The outputs are top-level semantic vector sets representing the core innovative capabilities of the text, image, and relational dimensions, respectively. These top-level semantic vector sets are then input into a cross-modal attention fusion layer. An attention mechanism enables vectors from different sources to interact and dynamically allocate weights at a high-level semantic level, achieving deep semantic alignment. Within this cross-modal attention fusion layer, the top-level semantic vectors from different modalities interact, align, and are weighted in importance within the semantic space using an attention mechanism. The resulting multimodal semantic information is integrated to output a fused comprehensive enterprise representation vector. All these comprehensive enterprise representation vectors are stored to form a unified enterprise representation library. Furthermore, the semantic trees from each modality collectively constitute a complete multimodal semantic forest for the enterprise, serving as an interpretable knowledge foundation.

[0026] It should be noted that a hierarchical and semantically defined enterprise semantic tree, generated through recursive aggregation, is used as the foundation. The entities and relationships within this semantic tree are then mapped to a structured enterprise innovation knowledge graph. In the semantic tree, each node represents an abstractly defined semantic concept and is accompanied by abstract descriptive text. Each level of semantic node in the enterprise semantic tree is mapped to a corresponding entity in the knowledge graph based on its node type, hierarchical position, and abstract descriptive text. The entity type is dynamically determined by the hierarchy and abstract descriptive text in the enterprise semantic tree; for example, a middle-level node might map to a R&D capability entity, while a leaf node might map to a technology entity. Each entity uses the abstract descriptive text from the semantic tree as its core attribute and is also associated with its most relevant original data fragment, forming traceable evidence.

[0027] Vertical graph relationships are generated based on the inherent parent-child node hierarchy in the enterprise semantic tree. The association weights between different semantic nodes learned through a cross-level attention mechanism during the recursive semantic aggregation process are obtained. Horizontal graph relationships are automatically identified and generated based on these association weights. The vertical graph relationships include "included," "implemented," or "decomposed into," for example, (core technical capability: solid-state battery technology, includes, sub-technology: electrolyte development), (sub-technology: electrolyte development, implemented, specific solution: sulfide synthesis process). The horizontal graph relationships include "collaborative," "supportive," "applied to," or "associated." For example, the abstract vector of thermal management simulation capability and the abstract vector of cell design capability have attentional interaction, generating a horizontal relationship of (thermal management simulation capability, collaborative optimization, cell design capability) in the graph. A structured enterprise innovation knowledge graph is constructed based on the entities, vertical graph relationships, and horizontal graph relationships. The top-level entities of the graph represent the enterprise's core innovation positioning, the middle-level entities describe key capability domains, and the bottom-level entities describe specific technical points and achievements. This enterprise innovation knowledge graph serves as a portrait of the enterprise's innovation capabilities.

[0028] It should be noted that, as Figure 3 As shown, after receiving a user's natural language query, the large language model is invoked to parse and decompose the natural language query into several atomic fact subqueries. For each atomic fact subquery, a dynamic evidence chain verification is performed, including internal graph tracing and external knowledge completion. In the internal graph tracing process, the atomic fact subqueries interact with the enterprise innovation knowledge graph in real time. A minimal semantic evidence path capable of supporting or refuting the corresponding atomic fact subqueries is constructed within the enterprise innovation knowledge graph. The completeness, matching degree, and attribute consistency of the minimal semantic evidence path are evaluated to generate internal evidence evaluation results. In the internal evidence evaluation results, completeness is calculated as follows: based on the semantics of the atomic fact subqueries, a predefined evidence path template is used. If the found path completely matches the template, the completeness score is high; the more missing nodes or attributes, the lower the completeness score. Matching degree is calculated using a weighted calculation based on semantic similarity, with key dimensions being entity matching degree, relationship matching degree, and attribute value matching degree. Attribute consistency is calculated by checking for contradictions within the found evidence path, performing logical conflict detection, and comparing the time attributes in the evidence path with the implicit or explicit time requirements in the atomic fact subqueries to perform a failure consistency assessment. In the external knowledge completion process, elements related to timeliness and consensus knowledge (such as whether certain technologies are recognized as cutting-edge technologies by the industry) are extracted from atomic fact subqueries and sent to external credible knowledge sources such as industry databases and academic knowledge bases for retrieval and comparison to obtain external timeliness evidence results. The internal evidence evaluation results and external timeliness evidence results are weighted, fused, and conflict resolved to generate a multi-dimensional confidence vector corresponding to each atomic fact subquery. The multi-dimensional confidence vector includes factual confidence, timeliness confidence, and evidence conflict identifier. For example, the output may be (factual: high, timeliness: medium, relevance: high, conflict: external source suggests the existence of more recent cooperation stage information).

[0029] Based on the multidimensional confidence vector, the large language model is invoked again to drive the context-aware semantic rewriting strategy engine to reconstruct the original natural language query, outputting optimized structured query instructions and natural language explanations of the correction items and the basis for the decision. Preferably, for subqueries with high factual basis, they are mapped to technical classification nodes or standard scientific and technological terms in the graph during rewriting and set as rigid constraints for retrieval; for subqueries with conflicting evidence or questionable timeliness, multi-contextual expressions are performed, transforming them into logical groups containing alternative or range conditions; for subqueries with low relevance, they are downgraded to optional filtering parameters or removed from the core search conditions.

[0030] It should be noted that, based on the optimized structured query command, a dual-path recommendation including a vector retrieval path and a generative inference path is executed. The vector retrieval path, based on the enterprise semantic tree and the enterprise unified representation library, constructs a multi-granularity semantic index. The index is essentially a vector forest with explicit semantic labels and structural relationships. This multi-granularity semantic index stores the comprehensive enterprise representation vector and associates it with intermediate semantic vectors generated during recursive semantic aggregation. All vectors are bound to the source nodes in the enterprise semantic tree, recording hierarchical relationships. The optimized structured query command is synchronously parsed and vectorized at multiple granularities to generate a comprehensive representation query. The graph generates a global query vector and local query vectors representing specific constraints in the query. The global query vector is then used to perform a similarity calculation with the comprehensive enterprise representation vectors of all enterprises in the multi-granularity semantic index to complete the top-level coarse screening and obtain a preliminary candidate enterprise set. Within this preliminary candidate enterprise set, the local query vectors are respectively compared with the intermediate semantic vectors of the preliminary candidate enterprises stored in the index to perform fine-grained similarity calculations. The matching scores from the top-level coarse screening and the fine-grained similarity calculations are then fused to weightedly rank the preliminary candidate enterprises, generating a final ranking list of relevant enterprises. Preliminary structured matching reasons are also generated based on the matching process.

[0031] The generative inference path receives a ranking list of relevant companies as a candidate company list. It then integrates and structurally encapsulates the company semantic tree, the unified company representation library, and the optimized structured query instructions to form a joint enhanced knowledge source input to the large language model. The company semantic tree provides the logical concept hierarchy and entity relationship network required for analysis; the unified representation library provides comprehensive company representation vectors and their relative positional relationships for quantitative comparison; and the optimized query provides a precise analytical task framework and constraints. Based on the joint enhanced knowledge source, prompts are generated to guide the large language model in performing inference chain analysis. In this analysis, the company semantic tree and representation vectors are used to identify a shortlist of candidate companies that best match the query at each semantic level. Based on the company semantic tree and the optimized structured query instructions, the candidate companies are identified and interpreted. The system analyzes the matching results with the query intent at different abstraction levels. Based on the relationship between the comprehensive enterprise representation vector in the unified representation library and the enterprise semantic tree, it performs multi-dimensional comparisons of candidate enterprises based on factors such as technological similarity, capability structure differences (Company B and Company A have similar scores, but Company B has a stronger electrolyte R&D capability node), and evidence characteristics (Company A's cooperation evidence comes from its internal project graph, while Company B's relevant technical evidence is related to recently published patents). Combined with knowledge bases, including but not limited to node descriptions in the semantic tree, relationship paths in the knowledge graph, and quantified vector similarity values, it generates explanatory reasons for each recommendation conclusion. The large language model, based on a preset report template, integrates all outputs from the reasoning chain analysis to generate an interpretable recommendation report containing a list of recommended enterprises, detailed recommendation reasons, multi-dimensional comparative analysis of enterprises, and risk warnings based on the analysis.

[0032] like Figure 4 As shown, the second embodiment of the present invention provides an enterprise innovation profile construction and recommendation system based on multimodal RAG. The system includes: a multi-source heterogeneous data acquisition and access module, a semantic parsing and objectification module, a semantic aggregation and representation learning module, a knowledge graph construction and profile solidification module, a query intent understanding and purification enhancement module, a multi-granularity intelligent retrieval and matching module, and a generative reasoning and interpretable report generation module. The multi-source heterogeneous data acquisition and access module is responsible for acquiring multi-source heterogeneous data from various internal and external data sources in a standardized manner, including: text, images, and relational data; The semantic parsing and objectification module parses the collected multi-source heterogeneous data and transforms it into a unified set of structured semantic objects, including: structured semantic objects, visual semantic objects, and primary triplet objects; The semantic aggregation and representation learning module processes structured semantic objects, visual semantic objects, and primary triplet objects in parallel, performs recursive semantic aggregation, constructs a top-level semantic vector representing the core innovation capabilities of the enterprise from bottom to top, and generates an interpretable enterprise semantic tree. The top-level semantic vector is then fused through a cross-modal attention mechanism to output a comprehensive enterprise representation vector and construct a unified representation library. The knowledge graph construction and profile solidification module maps entities and relationships in the enterprise semantic tree into a structured enterprise innovation knowledge graph, which is defined as the enterprise's structured innovation profile. The query intent understanding and purification enhancement module receives the user's natural language query, decomposes the natural language query into atomic fact subqueries based on the large language model, performs fact verification in the enterprise innovation knowledge graph and external trusted knowledge sources, and, based on the verification results, calls the large language model again to perform semantic rewriting to generate an optimized structured query instruction. The multi-granularity intelligent retrieval and matching module performs multi-granularity semantic matching on the optimized structured query command based on the index constructed by the enterprise semantic tree and the unified representation library, retrieves and returns a list of relevant enterprises; The generative reasoning and interpretable report generation module uses the enterprise semantic tree, unified representation library, and optimized structured query instructions as joint enhanced knowledge sources to drive the large language model to perform semantic alignment, multi-dimensional comparison, and quantitative analysis based on the joint knowledge sources, thereby generating an interpretable recommendation report.

[0033] The third embodiment of the present invention provides a computer-readable storage medium, which includes a program for constructing and recommending enterprise innovation profiles based on multimodal RAG. When the program for constructing and recommending enterprise innovation profiles based on multimodal RAG is executed by a processor, it implements the steps of the method for constructing and recommending enterprise innovation profiles based on multimodal RAG.

[0034] In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, and can be electrical, mechanical, or other forms. Furthermore, in the various embodiments of the present invention, all functional units can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0035] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0036] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing and recommending enterprise innovation profiles based on multimodal RAG, characterized in that, Includes the following steps: Collect text, image, and relational data from enterprises, parse the collected multi-source heterogeneous data, and transform it into a unified set of structured semantic objects. Recursive semantic aggregation is performed on the structured semantic object set to build a top-level semantic vector representing the core innovation capabilities of the enterprise from bottom to top, and an interpretable enterprise semantic tree is generated. The top-level semantic vector is fused through a cross-modal attention mechanism to output a comprehensive enterprise representation vector and build a unified representation library. The entities and relationships in the enterprise semantic tree are mapped into a structured enterprise innovation knowledge graph. The system receives natural language queries from users and performs fact verification and semantic rewriting based on the large language model in the enterprise innovation knowledge graph and external trusted knowledge sources to generate optimized structured query instructions. Based on the optimized structured query instructions, a dual-path recommendation, including a vector retrieval path and a generative reasoning path, is executed. The vector retrieval path uses the optimized structured query command to perform multi-granular semantic matching based on the index constructed by the enterprise semantic tree and the unified representation library, and retrieves and returns a list of relevant enterprises. The generative reasoning path uses the enterprise semantic tree, unified representation library, and optimized structured query instructions as joint enhanced knowledge sources to drive the large language model to generate an interpretable recommendation report containing reasons and comparative analysis.

2. The method for constructing and recommending enterprise innovation profiles based on multimodal RAG according to claim 1, characterized in that, Collect text, image, and relational data from enterprises; parse and transform the collected multi-source heterogeneous data into a unified format of structured semantic object collections, including: Collect multi-source heterogeneous data from target companies through public APIs, web crawlers, and collaborative data interfaces. The multi-source heterogeneous data includes text data, image data, and relational data. The text data is processed using a multi-level semantic parsing model to extract basic semantic units containing technical entities, performance indicators, method steps, and application scenarios. The semantic associations of the basic semantic units are obtained. Based on the semantic associations, the basic semantic units with associations are dynamically combined into composite semantic blocks as structured semantic objects. The types of composite semantic blocks include technical solution semantic blocks and application result semantic blocks, and the logical relationships between internal units are recorded. The image data is parsed using a visual language model to generate a structured description tuple containing fields for visual subject, functional description, and associated technical items, which serves as a visual semantic object. The relational data is mapped to a primary triple object in the format of (head entity, relation type, tail entity); All generated structured semantic objects, visual semantic objects, and primary triplet objects are converted into a unified set of structured semantic objects.

3. The method for constructing and recommending enterprise innovation profiles based on multimodal RAG according to claim 1, characterized in that, Recursive semantic aggregation is performed on the structured semantic object set to construct a top-level semantic vector representing the enterprise's core innovation capabilities from bottom to top, and an interpretable enterprise semantic tree is generated, including: The finest-grained semantic objects in the structured semantic object set of different modalities are vectorized, and bottom-up iterative semantic abstraction is performed. In each iteration, the semantic object vectors of the current level are clustered based on semantic similarity. For the generated clusters, the abstraction model is invoked to generate high-order semantic abstract vectors representing the overall semantics of the clusters and corresponding natural language abstract description texts. By using a cross-level attention mechanism, the generated higher-order semantic abstract vectors are corrected in a context-aware manner by utilizing the semantic object vectors within the current cluster, the semantic object vectors of the lower layer, and other related higher-order semantic abstract vectors at the same level. Repeated iterative loops until a preset abstraction level is reached, outputting hierarchical enterprise semantic trees of different modalities. The leaf nodes of the enterprise semantic tree are the initial semantic objects, and the top-level nodes are the top-level semantic vector set representing the core innovation capabilities of the enterprise.

4. The method for constructing and recommending enterprise innovation profiles based on multimodal RAG according to claim 3, characterized in that, The top-level semantic vectors are fused using a cross-modal attention mechanism to output a comprehensive enterprise representation vector and construct a unified representation library, including: Obtain the top-level semantic vector sets of different modalities, and input the top-level semantic vector sets of text, images, and relationships into the cross-modal attention fusion layer; In the cross-modal attention fusion layer, the top-level semantic vectors of different modalities are interacted, aligned and weighted in importance in the semantic space through the attention mechanism. The multimodal semantic information after interaction and weighting is integrated to output the fused comprehensive enterprise representation vector. All comprehensive enterprise representation vectors are stored to form a unified enterprise representation library.

5. The method for constructing and recommending enterprise innovation profiles based on multimodal RAG according to claim 1, characterized in that, Mapping the entities and relationships in the enterprise semantic tree into a structured enterprise innovation knowledge graph includes: The semantic nodes at each level in the enterprise semantic tree are mapped to corresponding entities in the knowledge graph according to the node type, level position and abstract description text, and each entity is assigned a dynamic entity type determined by the enterprise semantic tree and a core attribute traceable to the original data source. A vertical graph relationship is generated based on the inherent parent-child node hierarchy in the enterprise semantic tree. The association weights between different semantic nodes learned through a cross-level attention mechanism during the recursive semantic aggregation process are obtained. A horizontal graph relationship is automatically identified and generated based on the association weights. The vertical graph relationship includes "contains", "implements", or "decomposes into". The horizontal graph relationship includes "collaborates", "supports", "applies to", or "associates". A structured enterprise innovation knowledge graph is constructed based on the entities, vertical graph relationships, and horizontal graph relationships, and the enterprise innovation knowledge graph is used as a profile of enterprise innovation capabilities.

6. The method for constructing and recommending enterprise innovation profiles based on multimodal RAG according to claim 1, characterized in that, The system receives natural language queries from users, performs fact-checking and semantic rewriting based on the enterprise innovation knowledge graph and external trusted knowledge sources using a large language model, and generates optimized structured query instructions, including: After receiving a user's natural language query, the large language model is invoked to parse and decompose the natural language query into several atomic fact subqueries. For each atomic fact subquery, dynamic evidence chain verification is performed, including internal graph tracing and external knowledge completion. In the internal graph tracing process, the atomic fact subqueries are interacted with the enterprise innovation knowledge graph in real time. The minimum semantic evidence path that can support or refute the corresponding atomic fact subqueries is constructed in the enterprise innovation knowledge graph. The completeness, matching degree and attribute consistency of the minimum semantic evidence path are evaluated, and the internal evidence evaluation result is generated. In the external knowledge completion process, elements involving timeliness and consensus knowledge are extracted from the atomic fact sub-queries and sent to external trusted knowledge sources for retrieval and comparison to obtain external timeliness evidence results. The internal evidence evaluation results and external timeliness evidence results are weighted, fused, and conflict resolved to generate a multidimensional confidence vector corresponding to each atomic fact subquery. The multidimensional confidence vector includes factual confidence, timeliness confidence, and evidence conflict identifier. Based on the multidimensional confidence vector, the large language model is invoked again to drive the context-aware semantic rewriting strategy engine to reconstruct the original natural language query, and output the optimized structured query instruction as well as natural language explanations of the correction items and the basis for the decision.

7. The method for constructing and recommending enterprise innovation profiles based on multimodal RAG according to claim 1, characterized in that, The vector retrieval path includes: Based on the enterprise semantic tree and the enterprise unified representation library, a multi-granularity semantic index is constructed. The multi-granularity semantic index stores the comprehensive representation vector of the enterprise and associates it with the intermediate semantic vectors generated in the recursive semantic aggregation process. All vectors are bound to the source nodes in the enterprise semantic tree to record the hierarchical relationship. The optimized structured query command is parsed and vectorized synchronously at multiple granularities to generate a global query vector that represents the overall intent of the query, and a local query vector that represents the specific constraints in the query. The global query vector is used to calculate the similarity with the comprehensive enterprise representation vector of all enterprises in the multi-granularity semantic index to complete the top-level coarse screening and obtain a preliminary candidate enterprise set. In the preliminary candidate enterprise set, the local query vector is compared with the intermediate semantic vector of the preliminary candidate enterprise stored in the index using fine-grained similarity calculation. By combining the matching scores calculated from the top-level coarse screening and the fine-grained similarity, the preliminary candidate companies are weighted and sorted to generate the final ranking list of relevant companies, and preliminary structured matching reasons are generated based on the matching process.

8. The method for constructing and recommending enterprise innovation profiles based on multimodal RAG according to claim 1, characterized in that, The generative reasoning path includes: The system receives a ranking list of relevant enterprises as a candidate enterprise list, and integrates and encapsulates the enterprise semantic tree, the unified enterprise representation library, and the optimized structured query instructions to form a joint enhanced knowledge source input into the large language model. The enterprise semantic tree provides the logical concept hierarchy and entity relationship network required for analysis, the unified representation library provides the comprehensive enterprise representation vectors and their relative positional relationships for quantitative comparison, and the optimized query provides a precise analysis task framework and constraints. Based on the joint enhanced knowledge source, prompts are generated to guide the large language model to perform inference chain analysis. In the inference chain analysis, the matching of candidate enterprises with the query intent at different abstraction levels is identified and interpreted according to the enterprise semantic tree and the optimized structured query instructions. Based on the relationship between the comprehensive enterprise representation vector in the unified representation library and the enterprise semantic tree, the candidate enterprises are compared in multiple dimensions, including technological similarity, capability structure differences, and evidence characteristics. In addition, an explanation is generated for each recommendation conclusion based on the knowledge base. The large language model integrates all the outputs of the inference chain analysis based on the preset report template, and generates an interpretable recommendation report that includes a list of recommended companies, detailed reasons for the recommendations, multi-dimensional comparative analysis of companies, and risk warnings based on the analysis.

9. A system for constructing and recommending enterprise innovation profiles based on multimodal RAG, characterized in that, To implement the enterprise innovation profile construction and recommendation method based on multimodal RAG as described in any one of claims 1-8, the system includes: a multi-source heterogeneous data acquisition and access module, a semantic parsing and objectification module, a semantic aggregation and representation learning module, a knowledge graph construction and profile solidification module, a query intent understanding and purification enhancement module, a multi-granularity intelligent retrieval and matching module, and a generative reasoning and interpretable report generation module. The multi-source heterogeneous data acquisition and access module is responsible for acquiring multi-source heterogeneous data from various internal and external data sources in a standardized manner, including: text, images, and relational data; The semantic parsing and objectification module parses the collected multi-source heterogeneous data and transforms it into a unified set of structured semantic objects, including: structured semantic objects, visual semantic objects, and primary triplet objects; The semantic aggregation and representation learning module processes structured semantic objects, visual semantic objects, and primary triplet objects in parallel, performs recursive semantic aggregation, constructs a top-level semantic vector representing the core innovation capabilities of the enterprise from bottom to top, and generates an interpretable enterprise semantic tree. The top-level semantic vector is then fused through a cross-modal attention mechanism to output a comprehensive enterprise representation vector and construct a unified representation library. The knowledge graph construction and profile solidification module maps entities and relationships in the enterprise semantic tree into a structured enterprise innovation knowledge graph, which is defined as the enterprise's structured innovation profile. The query intent understanding and purification enhancement module receives the user's natural language query, decomposes the natural language query into atomic fact subqueries based on the large language model, performs fact verification in the enterprise innovation knowledge graph and external trusted knowledge sources, and, based on the verification results, calls the large language model again to perform semantic rewriting to generate an optimized structured query instruction. The multi-granularity intelligent retrieval and matching module performs multi-granularity semantic matching on the optimized structured query command based on the index constructed by the enterprise semantic tree and the unified representation library, retrieves and returns a list of relevant enterprises; The generative reasoning and interpretable report generation module uses the enterprise semantic tree, unified representation library, and optimized structured query instructions as joint enhanced knowledge sources to drive the large language model to perform semantic alignment, multi-dimensional comparison, and quantitative analysis based on the joint knowledge sources, thereby generating an interpretable recommendation report.

Images