AI-driven multi-modal data stream scheduling method and system

Abstract

The invention relates to the technical field of artificial intelligence and databases, in particular to an AI-driven multi-modal data stream scheduling method and system. The invention aims to realize automatic selection and query scheduling of the multi-source heterogeneous database through a real-time intelligent decision-making mechanism, and improve the enterprise data query efficiency and intelligence. According to the method, an intelligent closed-loop process is constructed, and the method comprises the following steps: firstly, ensuring the accuracy and integrity of query intention recognition by utilizing a semantic analysis technology combining a large language model and an ontology map; then, the AI Agent participates in database selection and query planning tasks, and an optimal database selection scheme is provided through an advanced semantic matching algorithm and intelligent decision-making ability; then, based on a selection result, the AI Agent automatically generates a cross-library query plan which covers query decomposition, rewriting, field mapping and execution sequence planning; performing parallel query execution and semantic fusion; and finally, a query execution result is deeply analyzed through a feedback optimization unit, and an AI Agent decision model is updated based on query performance and user satisfaction, so that the accuracy and adaptability of system query scheduling are continuously improved. The system can be widely applied to multi-database heterogeneous scenes such as supply chain management, scientific research data center, medical information integration and the like.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and database technology, specifically to an AI-driven multimodal data flow scheduling method and system. Background Technology

[0002] In modern enterprise data management, efficiently executing complex queries across various heterogeneous database environments is a key challenge. Traditional methods primarily rely on static routing rules and manual configuration; however, these methods often lack the ability to understand query semantics and adapt dynamically.

[0003] With the diversification and heterogeneity of enterprise data, a single database can no longer meet complex query needs. Currently, multi-database integration systems suffer from the following problems: First, static mapping: the routing rules between databases need to be maintained manually and cannot dynamically adapt to changes in query semantics, which severely restricts the flexibility and scalability of the system; Second, there is a lack of semantic understanding: the system cannot determine the most suitable data source based on the query intent, resulting in low query efficiency and wasted resources; Third, multimodal fragmentation: structured data and unstructured data cannot be integrated in the same semantic space, resulting in information silos and loss of data value; Fourth, it has weak optimization capabilities: traditional query scheduling lacks a feedback-based learning and evolution mechanism, and cannot optimize and improve itself based on actual query results.

[0004] Semantic understanding, as an important topic in the field of artificial intelligence, has been widely applied in various fields such as natural language processing, knowledge graphs, and intelligent question answering. In particular, when dealing with complex query intents, a two-layer semantic parsing method combining Large Language Model (LLM) and ontology graph can more accurately extract query entities and constraints, which provides a theoretical foundation for building intelligent database query systems.

[0005] AI Agent intelligent decision-making technology is an important method for autonomous decision-making in dynamic environments. Unlike traditional rule engines, AI Agents can automatically select the optimal database combination and generate cross-database query plans based on semantic relevance calculations and historical experience learning. In academia and industry, AI Agents have been widely used in fields such as intelligent recommendation, automated operation and maintenance, and resource scheduling, providing decision-makers with effective decision support tools. Their application potential in the field of database query scheduling is enormous.

[0006] Combining semantic parsing, AI agent decision-making, and multi-source data fusion can overcome the shortcomings of traditional multi-database query methods, improving the intelligence and accuracy of query scheduling. Semantic parsing helps identify the true intent of the query, while AI agents can make optimal decisions at the database selection and query planning levels. Semantic fusion technology enables a unified representation of heterogeneous data. This combination not only helps enterprises optimize data query efficiency but also provides users with more intelligent and accurate query results, thereby maximizing the value of data.

[0007] By comprehensively leveraging the advantages of semantic parsing, AI Agent intelligent decision-making, and multi-source data fusion, we can provide enterprises with more intelligent and data-driven multi-database query management solutions, effectively addressing the challenges of data heterogeneity and business complexity, and promoting enterprise digital transformation and intelligent upgrading. Summary of the Invention

[0008] To address the lack of semantic understanding and intelligent decision-making mechanisms in current multi-database query systems, this invention proposes an AI-driven multimodal data stream scheduling method and system.

[0009] The objective of this invention is achieved through the following technical solution: This invention provides an AI-driven multimodal data stream scheduling method, comprising the following steps: Step 1: Receive user query requests and generate semantic vector representations of the queries using a large language model and ontology graph; Step 2: Based on the query semantic vector, calculate the relevance score with the metadata of each database through the AI ​​Agent, and select the optimal set of databases; Step 3: Based on the characteristics of the selected database, automatically generate a cross-database query plan, including query rewriting, field mapping, and execution order; Step 4: Execute multiple database queries in parallel and fuse the results at the semantic level to generate a unified output; Step 5: Optimize the decision strategy parameters of the AI ​​Agent based on execution latency, result accuracy, and user feedback.

[0010] Further, step 1 includes: The user query is vectorized using a pre-trained language model to generate an LLM semantic vector; entities and relations are extracted from the query and mapped to an ontology graph to generate ontology augmentation vectors; the LLM semantic vectors and ontology augmentation vectors are weighted and fused to generate the final query semantic vector; and the optimal fusion weights are learned through training data.

[0011] Furthermore, the pre-trained language model includes BERT, RoBERTa, or a domain-fine-tuned model; the ontology graph includes a domain ontology, a data ontology, and a query ontology; the LLM semantic vector has a dimension of 512-1024; and the fusion weight parameters α and β are obtained through contrastive learning training.

[0012] Further, step 2 includes: Text descriptions and vector encodings are generated for the metadata of each database; the cosine similarity between the query semantics and the database metadata is calculated; a comprehensive score is calculated by combining semantic relevance, performance indicators, and historical success rate; and the Top-K databases are selected based on the comprehensive score.

[0013] Furthermore, step 3 includes: The query is decomposed into subqueries based on query semantics and database characteristics; the subqueries are rewritten into the query language of the target database; cross-database field mapping is performed based on the ontology graph; and the execution order is planned based on query dependencies and performance prediction.

[0014] Furthermore, step 4 includes: The query is executed in parallel by calling various database interfaces according to the query plan, ensuring strict control of query parameters; the query results are collected using an asynchronous parallel execution framework to ensure the completeness and accuracy of the query results; the heterogeneous results are mapped to a unified semantic space based on the ontology graph; the relevance between the calculation results and the query semantics is fused and sorted; and query results in a unified format are generated.

[0015] Furthermore, step 5 includes: Record query latency, database selection, query plan, and result quality; collect user feedback on result satisfaction; update decision strategy parameters using reinforcement learning algorithms; and continuously adjust model parameters and structure through feedback mechanisms.

[0016] The present invention also provides an AI-driven multimodal data stream scheduling system that implements the above method, comprising: The semantic parsing unit is used to receive user query requests and generate semantic vector representations of the queries using a large language model and ontology graph. The intelligent scheduling unit is used to calculate the relevance score with the metadata of each database based on the query semantic vector, select the optimal set of databases, and automatically generate a cross-database query plan. The query execution and fusion unit is used to execute multiple database queries in parallel and perform semantic alignment and fusion sorting on heterogeneous query results. The feedback optimization unit is used to optimize the decision model parameters of the AI ​​Agent based on query execution performance data and user feedback. If the query results are unsatisfactory, the feedback data is transformed into training samples for further training and optimization of the model. If the query results are good, the successful experiences are extracted to update the decision rule base.

[0017] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-described AI-driven multimodal data stream scheduling method.

[0018] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the above-described AI-driven multimodal data stream scheduling method. Attached Figure Description

[0019] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0020] Figure 1 This is a flowchart of the AI-driven multimodal data stream scheduling method proposed in this invention;

[0021] Figure 2 A schematic diagram of the structure of the AI-driven multimodal data stream scheduling system provided by the present invention;

[0022] Figure 3 This is a schematic diagram of an electronic device provided by the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0024] like Figure 1 As shown, the main steps of the AI-driven multimodal data stream scheduling method proposed in this invention include:

[0025] Step 1: Generate query semantic vectors using large language models and ontology graphs.

[0026] Specifically, it receives natural language queries or structured query statements input by the user, for example: Natural Language Query: "Find suppliers with the lowest costs for using medical-grade silicone within the last three months." Structured queries: SQL queries, Cypher graph queries, vector retrieval statements, etc. The query Q is encoded into a vector V_llm using a pre-trained language model (such as BERT, RoBERTa, or a domain-fine-tuned model), where V_llm = LLM_encode(Q). Entities and relations are extracted from the query Q and mapped to the ontology graph O, generating an ontology-enhanced vector V_onto = Ontology_embed(Q, O). The LLM vector is then fused with the ontology vector to generate the final semantic representation, Q' = α · V_llm + β · V_onto, where α and β are weight parameters learned through training.

[0027] For example, in a supply chain management scenario, the ontology graph O includes: Domain ontology O_domain: Concepts such as supplier, materials, cost, and time, and their relationships. Data ontology O_data: Describes the semantic structure of entities, relations, and attributes in each database. Query ontology O_query: Defines the classification system of query intent (such as aggregation, filtering, association). The entity extraction results include: "Medical grade silicone" (material type), "Supplier" (entity type), "Last three months" (time constraint), and "Lowest cost" (sorting criteria).

[0028] Step 2: The AI ​​Agent calculates database relevance and selects the optimal set of databases.

[0029] Specifically, the metadata of each database DB_i is vectorized, M_i = Encode(Meta-Profile_i), where the metadata includes database type, data schema, performance metrics, and semantic description; the similarity between the query semantics Q' and the database metadata M_i is calculated, Score_i = Similarity(Q', M_i) = cosine(Q', M_i); combining semantic relevance, performance metrics, and historical success rate, Final_Score_i = w1 · Score_i + w2 · Perf_i + w3 · History_i; the Top-K databases are selected based on the comprehensive score, D* = {DB_i | Final_Score_i ∈ Top-K}.

[0030] For example, in the supply chain query example: SQL database (MySQL): Meta-Profile contains "Supplier Cost Table, Purchase Record Table", Score_SQL = 0.85 The Graph database (Neo4j) contains a Meta-Profile that includes "supplier relationship network and collaboration history," with a Score_Graph value of 0.78. Vector database (Milvus): Meta-Profile includes "material description vectors and semantic similarity retrieval", Score_Vector = 0.72 Based on the overall score, the selection is: D* = {MySQL, Neo4j, Milvus}

[0031] Step 3: Automatically generate cross-database query plan.

[0032] Specifically, based on query semantics and database characteristics, query Q is decomposed into subqueries, SubQuery_i = Decompose(Q, DB_i); the subqueries are rewritten into the query language of the target database, such as SQL_i = Rewrite(SubQuery_i, SQL), Cypher_j = Rewrite(SubQuery_j, Cypher); cross-database field mapping is performed based on the ontology graph, FieldMap = {(field_A, DB_1) → (field_B, DB_2)}; and the execution order is planned based on query dependencies and performance prediction, ExecOrder = [Step_1, Step_2, ..., Step_n].

[0033] For example, in the supply chain query example: MySQL subquery: SELECT supplier_id, material_type, cost FROM procurement WHERE material_type LIKE '%medical grade silicone%' AND date >= DATE_SUB(NOW(), INTERVAL3 MONTH) ORDER BY cost ASC LIMIT 10 Neo4j subquery: MATCH (s:Supplier)-[r:PROVIDES]->(m:Material {type: 'Medical grade silicone'}) WHERE r.date >= date() - duration('P3M') RETURN s.id,r.reliability_score Milvus subquery: search(collection='materials', query_vector=embed('medical grade silicone'), top_k=10, filter='category == medical_grade') Field mapping: MySQL.supplier_id → Neo4j.s.id → Output.supplier_id Execution order: Parallel execution of MySQL, Neo4j, and Milvus → Result fusion → Final output

[0034] Step 4: Execute queries in parallel and perform semantic fusion.

[0035] Specifically, according to the execution plan, the database interfaces are called in parallel: Result_i = Execute(Query_i,DB_i); the heterogeneous results are mapped to a unified semantic space based on the ontology graph: Aligned_i = Align(Result_i,Ontology); the relevance between the results and the query semantics is calculated and merged and sorted: FusedResult = Merge([Aligned_1, ..., Aligned_n], Q'); and query results in a unified format are generated: Output = Format(FusedResult).

[0036] For example, the fusion strategy includes: Semantic alignment: The `cost` field in MySQL, the `price` field in Neo4j, and the `cost_estimate` field in Milvus are all uniformly mapped to "cost". Relevance calculation: Score_relevance = α1 · Score_semantic + α2 · Score_reliability + α3 · Score_cost Merged Ranking: Rank suppliers according to their overall score and output a list of Top-10 suppliers.

[0037] Step 5: Feedback and optimization.

[0038] Specifically, the query latency, database selection accuracy, and query success rate are recorded, with Metrics = {Latency, Accuracy, Success_Rate}; user satisfaction ratings for query results are collected, with Satisfaction ∈ [1,5]; the Agent policy parameters θ are updated using a reinforcement learning algorithm, with Reward = f(Latency, Accuracy, Satisfaction) and θ_new = θ_old + α · ∇Reward; and the database selection weights and query planning strategy are updated, with Policy_new = Update(Policy_old, Metrics).

[0039] For example, if the user is not satisfied with the result (Satisfaction < 3), the system will: Analysis of failure causes: inappropriate database selection, query rewrite errors, and poor fusion strategy. Generate negative samples: Mark failed query-database select pairs as negative samples. Model retraining: Using negative samples for comparative learning enhances the model's ability to identify "failure cases". Strategy Update: Adjust database scoring weights, prioritizing databases with high historical success rates. Example

[0040] like Figure 2 As shown, the present invention provides an AI-driven multimodal data stream scheduling system, the system comprising: Semantic parsing unit: Responsible for extracting semantic information from user queries and generating semantic vector representations of the queries. This unit uses a pre-trained language model to vectorize the queries and combines it with an ontology graph for semantic enhancement, ultimately generating a fused semantic vector.

[0041] The intelligent scheduling unit is responsible for calculating the relevance score between the query semantic vector and the metadata of each database, selecting the optimal set of databases, and automatically generating a cross-database query plan. This unit includes a metadata vectorization module, a semantic matching calculation module, and a query plan generation module.

[0042] Query Execution and Fusion Unit: Responsible for parallel execution of multiple database queries, performing semantic alignment and fusion sorting on heterogeneous query results. This unit supports multiple database query languages, provides an ontology graph-based semantic alignment engine, and calculates the relevance of results to query semantics in real time for fusion and sorting.

[0043] Feedback Optimization Unit: Responsible for optimizing the AI ​​Agent's decision model parameters based on query execution performance data and user feedback. This unit records query latency, database selection accuracy, query success rate, collects user satisfaction scores, and updates decision strategy parameters using reinforcement learning algorithms.

[0044] Furthermore, the intelligent scheduling unit invokes the model training and optimization unit during the decision-making process. The model training and optimization unit continuously receives inference requests from the intelligent scheduling unit during system operation. The model training and optimization unit is responsible for the training, tuning, and deployment of the semantic model and the decision model. It cleans, labels, and vectorizes the received data to generate a standardized training set. It adopts a pre-training and fine-tuning strategy, uses a large-scale dataset to train the semantic model, and improves the model's performance in specific business domains through contrastive learning and reinforcement learning techniques. It optimizes inference speed and resource consumption and outputs the optimized model parameters to the intelligent scheduling unit.

[0045] Furthermore, the semantic parsing unit and the model training optimization unit send computing power requests to the computing power management unit, which provides heterogeneous computing power support and optimizes resource allocation. The specific operation process is as follows: The computing cluster is built based on GPU / NPU to meet the computing power requirements of large model inference. It dynamically schedules GPU / NPU resources to perform multi-task parallel processing and low-latency response, thereby improving the overall system performance. It also dynamically adjusts resource allocation according to the load, supports large model inference and real-time query scheduling, and provides computing power support for model training optimization unit and intelligent scheduling unit.

[0046] Furthermore, the query execution and fusion unit is communicatively connected to the query performance monitoring unit. The query performance monitoring unit monitors the operation of the query execution and fusion unit, analyzes its operational status in real time, and determines whether to generate a performance anomaly alarm signal. If a performance anomaly alarm signal is generated, an alarm mechanism is triggered. The specific status analysis process is as follows:

[0047] The real-time query latency of the query execution and fusion unit is obtained and marked as a latency measurement value. The growth rate of the real-time query latency is marked as a latency growth value. The latency measurement value and the latency growth value are compared with the preset latency measurement threshold and the preset latency growth threshold respectively. If the latency measurement value or the latency growth value exceeds the corresponding preset threshold, a performance abnormality alarm signal is generated.

[0048] If neither the delay measurement nor the delay increase value exceeds the corresponding preset threshold, the resource consumption of the query execution and fusion unit is collected. The difference between the resource consumption and the set standard resource consumption is calculated and the absolute value is taken to obtain the resource deviation value. The resource deviation value is compared with the preset resource deviation threshold. If the resource deviation value exceeds the preset resource deviation threshold, a performance abnormality alarm signal is generated.

[0049] Furthermore, the query performance monitoring unit is also used to set a detection period of duration L1. When the duration reaches L1, the operational optimization status of the query execution and fusion unit during the detection period is comprehensively evaluated to determine whether to generate a system optimization alarm signal. When a system optimization alarm signal is generated, the alarm mechanism is triggered. The specific evaluation and analysis process is as follows:

[0050] The system acquires the generation time of the corresponding performance abnormality alarm signal and marks it as the first characteristic time, and acquires the time when the corresponding performance abnormality alarm is cleared after adjustment and marks it as the second characteristic time. The system calculates the characteristic duration value by performing a time difference calculation between the first characteristic time and the second characteristic time. The system compares the characteristic duration value with a preset characteristic duration threshold. If the characteristic duration value exceeds the corresponding preset characteristic duration threshold, the system marks the corresponding characteristic duration value as a characteristic abnormal value.

[0051] The system obtains the number of feature anomalies during the detection period and marks them as anomaly frequency values. It then compares these anomaly frequency values ​​with a preset anomaly frequency threshold. If the anomaly frequency value exceeds the preset anomaly frequency threshold, a system optimization alarm signal is generated. If the anomaly frequency value does not exceed the preset anomaly frequency threshold, the system obtains the number of database selection errors and query failures during the detection period and marks them as selection error rate and query failure rate, respectively. It then compares these rates with preset selection error rate thresholds and preset query failure rate thresholds, respectively. If either the selection error rate or the query failure rate exceeds the corresponding preset threshold, a system optimization alarm signal is generated.

[0052] If the error rate and query failure rate do not exceed the corresponding preset thresholds, then when a user performs a query operation in the query execution and fusion unit, the user's query buffer time is collected, and the average of all query buffer times within the detection period is calculated to obtain the average latency deviation; and when a user performs a query operation in the query execution and fusion unit, the user's actual query accuracy and target query accuracy are obtained, the difference between the actual query accuracy and the target query accuracy is calculated and the absolute value is taken to obtain the result accuracy deviation, and the average of all result accuracy deviations within the detection period is calculated to obtain the average accuracy deviation;

[0053] The system optimization urgency coefficient is calculated by weighting and summing the abnormal frequency value, selection error rate, query failure rate, average delay deviation, and average accuracy deviation. The system optimization urgency coefficient is then compared with a preset system optimization urgency coefficient threshold. If the system optimization urgency coefficient exceeds the preset system optimization urgency coefficient threshold, a system optimization alarm signal is generated. Example

[0054] Corresponding to the aforementioned embodiments of the AI-driven multimodal data stream scheduling method, this application also provides an electronic device, including: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the AI-driven multimodal data stream scheduling method as described above. Figure 3 The diagram shown illustrates a hardware structure of any data processing-capable device in which the AI-driven multimodal data stream scheduling method provided in this application embodiment is implemented. (Except for...) Figure 3 In addition to the processor, memory, network interface, and storage shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0055] Corresponding to the aforementioned embodiments of the AI-driven multimodal data stream scheduling method, this embodiment of the invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the AI-driven multimodal data stream scheduling method described in the above embodiments.

[0056] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., mounted on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0057] The beneficial effects of this invention are: This invention achieves accurate identification and understanding of user query intent by combining a large language model and an ontology graph as a two-layer semantic parsing mechanism, solving the problem of traditional systems lacking semantic understanding capabilities and significantly improving query accuracy.

[0058] This invention utilizes AI Agent intelligent decision-making technology to achieve automatic database selection and automatic generation of cross-database query plans, reducing manual configuration and maintenance costs and improving the system's automation level and operational efficiency.

[0059] This invention achieves unified representation and intelligent sorting of heterogeneous data through semantic fusion algorithm, solves the problem of fragmented multimodal data, and improves the quality of query results and user satisfaction.

[0060] This invention achieves continuous model optimization based on real query results through a feedback optimization mechanism, enabling the system to have self-learning and evolution capabilities and adapt to changes in data patterns and the evolution of business needs.

[0061] The intelligent method provided by this invention not only improves the efficiency and accuracy of multi-source database queries, but also promotes the intelligent upgrade of enterprise data management through continuous feedback and optimization mechanisms. Its significant scientific and commercial value lies in its ability to accelerate query response speed, reduce operational costs, improve query accuracy, and provide more reliable data analysis results. This invention offers an innovative solution for data query management in enterprise multi-database environments, with broad application prospects and practical value.

[0062] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0063] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. An AI-driven multimodal data stream scheduling system, characterized in that, It includes a semantic parsing unit, an intelligent scheduling unit, a multi-source database unit, a query execution and fusion unit, and a feedback optimization unit; The semantic parsing unit is responsible for receiving user query requests, generating semantic vector representations of the queries using large language models and ontology graph technology, and outputting the semantic vectors to the intelligent scheduling unit. The intelligent scheduling unit is responsible for calculating the relevance score and selecting the optimal database set based on the semantic matching of the query semantic vector and the metadata of each database, automatically generating a cross-database query plan, and sending the query plan and database selection results to the query execution and fusion unit. The query execution and fusion unit receives the query plan from the intelligent scheduling unit, executes multiple database queries in parallel, performs semantic alignment and fusion sorting on heterogeneous query results, and generates output results in a unified format. The feedback optimization unit is responsible for receiving performance data and user feedback from query execution, and transmitting the feedback data to the intelligent scheduling unit for updating decision model parameters and optimizing strategies.

2. The AI-driven multimodal data stream scheduling system according to claim 1, characterized in that, The semantic parsing unit uses a pre-trained language model to vectorize user queries, generating LLM semantic vectors. It also extracts entities and relations from the queries and maps them to an ontology graph to generate ontology-enhanced vectors. Finally, it performs weighted fusion of the LLM semantic vectors and ontology-enhanced vectors to generate the final query semantic vector representation.

3. The AI-driven multimodal data stream scheduling system according to claim 1, characterized in that, The intelligent scheduling unit includes a metadata vectorization module, a semantic matching calculation module, and a query plan generation module. The metadata vectorization module performs vector encoding on the metadata of each database. The semantic matching calculation module calculates the cosine similarity between the query semantics and the database metadata and combines it with performance indicators and historical success rates to make a comprehensive score. The query plan generation module performs query decomposition, query rewriting, field mapping, and execution order planning based on the characteristics of the selected database.

4. The AI-driven multimodal data stream scheduling system according to claim 1, characterized in that, The query execution and fusion unit supports multiple database query languages, including relational database SQL, graph database Cypher, vector database semantic retrieval, and document database NoSQL queries, enabling seamless communication between the system and heterogeneous databases. It also provides a semantic alignment engine based on ontology graphs and performs real-time calculations and fusion sorting based on the relevance of query semantics.

5. The AI-driven multimodal data stream scheduling system according to claim 1, characterized in that, The feedback optimization unit has performance monitoring, user feedback collection, and reinforcement learning optimization functions. Performance monitoring is used to record query latency, database selection accuracy, and query success rate. User feedback collection is used to obtain user satisfaction ratings for query results. Reinforcement learning optimization is used to update the decision strategy parameters of the intelligent scheduling unit based on the reward function.

6. An AI-driven multimodal data stream scheduling method, characterized in that, Includes the following steps: Step 1: Receive user query requests and generate semantic vector representations of the queries using a large language model and ontology graph; Step 2: Based on the query semantic vector, calculate the relevance score with the metadata of each database through the AI ​​Agent, and select the optimal set of databases; Step 3: Based on the characteristics of the selected database, automatically generate a cross-database query plan, including query rewriting, field mapping, and execution order; Step 4: Execute multiple database queries in parallel and fuse the results at the semantic level to generate a unified output; Step 5: Optimize the decision strategy parameters of the AI ​​Agent based on execution latency, result accuracy, and user feedback.

7. The AI-driven multimodal data stream scheduling method as described in claim 6, characterized in that, Step 1 includes: The user query is vectorized using a pre-trained language model to generate an LLM semantic vector; entities and relations are extracted from the query and mapped to an ontology graph to generate ontology augmentation vectors; the LLM semantic vectors and ontology augmentation vectors are weighted and fused to generate the final query semantic vector.

8. The AI-driven multimodal data stream scheduling method as described in claim 6, characterized in that, Step 2 includes: Text descriptions and vector encodings are generated for the metadata of each database; the cosine similarity between the query semantics and the database metadata is calculated; a comprehensive score is calculated by combining semantic relevance, performance indicators, and historical success rate; and the Top-K databases are selected based on the comprehensive score.

9. The AI-driven multimodal data stream scheduling method as described in claim 6, characterized in that, Step 3 includes: The query is decomposed into subqueries based on query semantics and database characteristics; the subqueries are rewritten into the query language of the target database; cross-database field mapping is performed based on the ontology graph; and the execution order is planned based on query dependencies and performance prediction.

10. An electronic device comprising a memory and a processor, characterized in that, The memory is coupled to the processor; The memory is used to store program data, and the processor is used to execute the program data to implement the AI-driven multimodal data stream scheduling method as described in any one of claims 6-9.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the AI-driven multimodal data stream scheduling method as described in any one of claims 6-9.

Images