AI transaction question and answer multi-source data retrieval method and system for commodity market research

Abstract

The application provides an AI transaction question and answer multi-source data retrieval method and system for commodity market research and judgment. It belongs to the field of artificial intelligence and agricultural information retrieval technology. The method includes: performing semantic analysis and processing on the bulk commodity transaction question and answer problems transmitted by the user interaction layer to generate structured search instructions, and deploying the required model type and search range according to the semantic analysis result; performing multi-MCP parallel calling processing according to the search instruction; through the multi-MCP parallel calling mechanism combined with resource pool matching and weighted round robin scoring, different models can simultaneously process price prediction, supply and demand analysis and risk assessment tasks in their respective containerized environments, avoiding the problem of too long processing time caused by single model serial calculation, and reducing the running lag caused by resource contention between models.

Description

Technical Field

[0001] This invention proposes an AI-based transaction question-and-answer multi-source data retrieval method and system for commodity market analysis, belonging to the fields of artificial intelligence and agricultural and food industry information retrieval technology. Background Technology

[0002] The commodity trading sector involves multiple categories such as agriculture, food, energy, and metals. Price fluctuations are influenced by a complex interplay of factors, including supply and demand, macroeconomic conditions, policy changes, and unforeseen risk events. Currently, most trading question-and-answer systems on the market employ a single-model retrieval approach. This means that after a user asks a question, a single model retrieves information from a fixed data source and generates an answer. This approach often falls short when dealing with complex, multi-dimensional, and cross-category questions. For example, if a user inquires about the price trend of a certain grain commodity over the next three months, the system needs to simultaneously retrieve information from multiple sources, including historical prices, inventory data, import and export figures, and policy changes. A single model struggles to encompass all dimensions in a single retrieval, easily leading to information omissions or biased analysis. Furthermore, differences in format and terminology between different data sources, coupled with a lack of a unified scheduling strategy when calling models, can easily result in contradictory outputs due to data conflicts or unclear permissions. Moreover, existing systems generally underutilize user feedback. Users often find answers inaccurate, yet there is a lack of effective backtracking and optimization channels, hindering the system's ability to learn from errors and making it difficult to continuously improve search quality in the long run. Therefore, how to enable multiple models to work together, accurately extract and integrate analysis results from multi-source data, and at the same time have the ability to evolve on their own is a problem that urgently needs to be solved in the current question-and-answer scenario for commodity trading. Summary of the Invention

[0003] This invention provides an AI-powered trading question-and-answer multi-source data retrieval method and system for commodity market analysis, in order to solve the problems mentioned in the background section above: The present invention proposes an AI-based multi-source data retrieval method for commodity market analysis, comprising: S1. Perform semantic analysis on the bulk commodity transaction Q&A questions transmitted from the user interaction layer, generate structured search instructions, and allocate the required model type and search scope according to the semantic analysis results; S2. Perform parallel multi-MCP call processing according to the search command; S3. Perform data fusion processing on the calculation results of multiple MCP outputs to form a complete search report, which is presented in a combination of charts and text descriptions; S4. After the user obtains the search report, feedback is collected and processed. Based on the user feedback, machine learning algorithms are used to analyze the search process, identify semantic understanding biases or unreasonable model calls, adjust the semantic analysis model parameters and optimize the MCP calling strategy, and store the user feedback data in the database.

[0004] The present invention proposes an AI-powered multi-source data retrieval system for commodity market analysis, comprising: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors are made to implement the method described in any one of the above.

[0005] The beneficial effects of this invention are as follows: By combining a multi-MCP parallel invocation mechanism with resource pool matching and weighted round-robin scoring, different models can simultaneously process price prediction, supply and demand analysis, and risk assessment tasks in their own independent containerized environments. This avoids the problem of excessive processing time caused by serial computation of a single model and reduces runtime lag caused by resource contention between models. Through multiple rounds of iterative convergence and weighted voting conflict resolution of the MCP calculation results, abnormal intermediate data can be eliminated, and logical contradictions between the conclusions of different models can be resolved, preventing users from receiving contradictory judgments due to data conflicts. Furthermore, by performing cross-model correlation analysis and complementary information filling on conflict-free results, and then performing multi-source fusion according to logical hierarchy and time dimension, scattered price, supply and demand, and risk data can be integrated into a structured report containing trend judgments and risk warnings, reducing the workload for users to piece together multi-dimensional information themselves. With the support of feedback-driven root cause localization and incremental training mechanisms, the system can automatically identify semantic comprehension biases or inappropriate model calling strategies from user reviews and continuously iterate and optimize them, avoiding the recurrence of similar errors and thus improving the overall accuracy of commodity market analysis and search response efficiency. Attached Figure Description

[0006] Figure 1 This is a diagram illustrating the steps of the method described in this invention. Detailed Implementation

[0007] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0008] One embodiment of the present invention, such as Figure 1 As shown, an AI-based trading question-and-answer multi-source data retrieval method for commodity market analysis includes: S1. Perform semantic analysis on the bulk commodity transaction Q&A questions transmitted from the user interaction layer, use natural language processing technology to understand the complete intent of the question, identify the key information in the question, including agricultural and food categories, time range and analysis dimensions, generate structured search instructions, and allocate the required model type and search range according to the semantic analysis results. S2. Perform parallel call processing of multiple MCPs according to the search command, and start multiple MCPs to run in their own independent computing environments at the same time. The price prediction MCP calculates and analyzes the price trend, the supply and demand analysis MCP calculates and analyzes the market supply and demand situation, and the risk assessment MCP calculates and analyzes the potential risks. Each MCP outputs its own calculation results without interfering with each other. S3. Perform data fusion processing on the calculation results of multiple MCP outputs, remove duplicate information, correlate and supplement the results of different models, combine the analysis results of each model according to preset logic and rules to form a complete search report, and present it in a combination of charts and text descriptions. S4. After the user obtains the search report, feedback is collected and processed to guide the user to evaluate the satisfaction of the search results. Based on the user feedback, machine learning algorithms are used to analyze the search process to identify semantic understanding biases or unreasonable model calls. The semantic analysis model parameters are adjusted and the MCP calling strategy is optimized to address the issues. The user feedback data is stored in the database for continuous training of the subsequent model.

[0009] The working principle and effects of the above technical solution are as follows: By performing semantic analysis on user questions and generating structured search instructions, key elements such as category, time, and dimension in commodity trading questions can be accurately broken down, avoiding deviations in search direction due to misunderstandings, thereby significantly improving the retrieval hit rate for complex questions. Simultaneously, the multi-MCP parallel invocation mechanism allows multiple professional models such as price forecasting, supply and demand analysis, and risk assessment to operate simultaneously. This enables comprehensive analysis of the same question from different perspectives and reduces the response time for complex questions from seven or eight minutes to less than one minute, significantly lowering user waiting costs. Furthermore, the feedback-driven reflective optimization layer continuously iterates and adjusts the semantic analysis parameters and MCP invocation strategy, allowing the system to continuously converge to a better state during use, effectively preventing the problem of stagnant search quality and enhancing the system's adaptability to the ever-changing demands of the agricultural and food market. By integrating and deduplicating multi-model results, scattered analytical data can be integrated into a complete report that includes trend judgments, key factors, and risk warnings. This reduces the workload of users in secondary processing, enabling practitioners to quickly obtain the information needed for decision-making and reducing the risk of misjudgments due to missing information.

[0010] In one embodiment of the present invention, S1 includes: S11. Perform semantic recognition and text processing on the bulk commodity transaction Q&A questions transmitted from the user interaction layer, perform format verification and filtering on the text, and generate standardized Q&A text data; perform word segmentation and part-of-speech tagging on the standardized Q&A text data, extract agricultural and food category keywords, time range keywords and analysis dimension keywords from the text, and generate a multi-dimensional keyword feature set; S12. Perform semantic dependency analysis on the multi-dimensional keyword feature set, use the intent recognition algorithm in natural language processing to understand the complete intent of the question, map the question intent to the preset intent classification system, and generate intent classification label data; perform context association expansion on the intent classification label data to supplement the implicit association analysis needs in the question and generate extended intent data. S13. Perform structured transformation processing on the extended intent data, converting user questions in natural language form into machine-executable structured search instructions, and generating structured search instruction data; perform model matching analysis on the structured search instruction data, and select model types that match the search instructions from the preset MCP model library, generating model type matching data. S14. Perform search range definition processing on the model type matching data, and limit the search range of each matching model by combining agricultural and food categories, time range, and analysis dimensions to generate search range limit data for each MCP; prioritize the search range limit data of each MCP to generate MCP call priority sequence data. S15. Perform resource pre-allocation processing on the MCP call priority sequence data, allocate independent computing resources and running environment to each MCP, and generate MCP resource pre-allocation data; prepare for parallel calls using the MCP resource pre-allocation data, and generate a multi-MCP parallel call ready signal.

[0011] The working principle and effects of the above technical solution are as follows: By sequentially processing user questions through text processing, word segmentation and annotation, semantic dependency analysis, and intent mapping, the intent of the question can be accurately decomposed from multiple levels such as agricultural and food categories, time ranges, and analysis dimensions. This avoids deviations in search direction due to semantic understanding biases, thereby improving the accuracy of intent recognition for complex questions. Simultaneously, implicit requirements are supplemented through contextual association, and then structured transformation generates machine-executable search instructions. This allows the system to accurately identify the required MCP model type and efficiently convert natural language questions into structured instructions, reducing information loss during instruction generation. Furthermore, defining the search range and prioritizing the matching models, pre-allocating independent computing resources to each MCP, and generating parallel call ready signals ensure that multiple models do not compete for resources when starting simultaneously, avoiding computational blocking due to resource contention. This improves the execution efficiency of multi-MCP parallel calls and the overall system response speed.

[0012] In one embodiment of the present invention, S15 includes: S151. Perform resource feature parsing on the MCP call priority sequence data, extract the computational complexity level, data throughput requirements, and real-time response time requirements of each MCP, and generate MCP resource requirement feature data; perform resource pool matching analysis on the MCP resource requirement feature data, and select computing nodes and storage nodes that are compatible with the resource requirements of each MCP from the preset distributed computing resource pool, and generate resource node matching data. S152. Perform load balancing assessment on the resource node matching data. Use a weighted round-robin algorithm to comprehensively quantify and score the current load rate, memory usage rate, and network bandwidth utilization rate of each candidate computing node to generate node load score data. Perform optimal node screening on the node load score data and bind the computing node with the highest score to the corresponding MCP one by one to generate MCP and computing node binding data. S153. Perform runtime environment isolation configuration processing on the data bound to the MCP and the compute node, create an independent containerized runtime space for each MCP on its bound compute node, configure exclusive memory quotas and CPU core counts, and generate containerized runtime environment data; perform network path establishment processing on the containerized runtime environment data, build an independent internal communication channel between the runtime environments of each MCP, and at the same time establish a dedicated data transmission link between each MCP and the external data source, and generate network path configuration data. S154. Perform data permission allocation processing on network channel configuration data. Assign corresponding data access permission levels to each MCP according to the analysis task type. Perform hierarchical authorization on the access interfaces of multi-source market data database, upstream and downstream industry chain database and external risk event database, and generate data permission allocation data. Perform security sandbox deployment processing on data permission allocation data. Embed security sandbox modules in the running environment of each MCP to perform real-time auditing and anomaly interception on data flows entering and leaving the running environment, and generate security protection configuration data. S155. Perform parallel call readiness verification processing on the security protection configuration data, send heartbeat detection signals to the operating environment of each MCP, verify whether the computing engine, data interface and communication link of each MCP are in normal working condition, and generate operating status detection data for each MCP; perform readiness judgment processing on the operating status detection data of each MCP, mark the MCP with all normal operating status as ready, mark the MCP with abnormality as pending repair and trigger the automatic repair process, and generate MCP ready status marking data. S156. Perform parallel call preparation and integration processing on the MCP ready state marker data. Summarize and encapsulate the runtime environment parameters, data permission configurations, and network path information of all MCPs in the ready state to generate a multi-MCP parallel call configuration package. Perform call timing arrangement processing on the multi-MCP parallel call configuration package. Arrange the execution timing of each MCP according to the priority order in the MCP call priority sequence data. At the same time, mark the data dependencies and result transmission paths between each MCP to generate multi-MCP parallel call timing arrangement data. Perform ready signal broadcast processing on the multi-MCP parallel call timing arrangement data. Synchronously send multi-MCP parallel call ready signals to the search processing layer and the model call layer to trigger the next stage of the multi-MCP parallel call processing flow.

[0013] The working principle and effects of the above technical solution are as follows: By analyzing the resource characteristics of the MCP call priority sequence and combining it with a weighted round-robin algorithm to complete node load scoring and optimal node selection, models with high computational complexity and large data throughput can be matched with the most suitable computing nodes, avoiding problems such as model lag or response delays caused by resource mismatch. Simultaneously, through containerized runtime environment isolation configuration and the establishment of independent internal communication channels, it ensures that each MCP executes tasks without interference within its dedicated space, while also enabling efficient connection between the model and external data sources through dedicated data transmission links, reducing congestion and packet loss during data transmission. Furthermore, hierarchical authorization of data access permissions and the deployment of a security sandbox module can effectively intercept abnormal data access behavior, reducing the risk of information leakage during multi-source data interaction. By verifying the running status through heartbeat detection and triggering an automatic repair process, services can be quickly restored when an MCP malfunctions, avoiding single-point failures slowing down the entire parallel call process, thereby improving the stability and overall execution efficiency of multi-MCP collaborative calls.

[0014] In one embodiment of the present invention, step S152 includes: Load factor extraction processing is performed on the resource node matching data. Multi-dimensional load factors such as current CPU utilization, memory usage percentage, network input / output throughput, and disk I / O queue depth are extracted from the running logs of each candidate computing node to generate a real-time load factor group for candidate nodes. Time window aggregation processing is performed on the real-time load factor group of candidate nodes. The load factors are weighted and averaged according to a preset sliding time window to eliminate instantaneous fluctuation interference and generate steady-state load data for candidate nodes. The steady-state load data of candidate nodes is subjected to weighted round-robin scoring. Differentiated weight coefficients are configured for CPU utilization, memory usage percentage, and network bandwidth utilization. The weighted round-robin algorithm is used to perform comprehensive quantitative calculation on the steady-state load data of each candidate computing node to generate an initial comprehensive score value for each node. The initial comprehensive score value is then normalized and mapped to uniformly map the scores of each node to a preset percentage score range to generate node load score data. The node load score data is sorted and layered. Candidate computing nodes are sorted in descending order according to their score values. The sorted nodes are divided into three levels: the preferred layer, the candidate layer, and the elimination layer, generating node layer data. Cross-layer cross-validation is performed on the node layer data. Preferred layer nodes are paired with candidate layer nodes for comparison to verify the score stability of preferred layer nodes under sudden high load scenarios, generating node stability verification data. The node stability verification data is processed to select the optimal node. Nodes that fail the stability verification are removed from the preferred layer. The computing node with the highest score and that passes the stability verification among the remaining nodes is selected as the optimal node and bound to the corresponding MCP one by one to generate MCP and computing node binding data. The MCP and computing node binding data is processed to perform redundancy backup. For each bound MCP, a redundant computing node is specified in the candidate layer as a hot standby node to generate MCP hot standby binding data.

[0015] The working principle and effects of the above technical solution are as follows: By aggregating the multi-dimensional load factors of candidate computing nodes over a time window and using a weighted round-robin algorithm for comprehensive scoring, the interference of instantaneous fluctuations on node evaluation can be eliminated. This avoids the problem of high-quality nodes being misjudged as high-load due to load spikes at a single moment, thus missing the opportunity to call the node. Simultaneously, through hierarchical sorting and cross-layer cross-validation, nodes can be divided into a preferred layer, a candidate layer, and an elimination layer for rapid screening. This also verifies the scoring stability of preferred layer nodes under sudden high-load scenarios, reducing the risk of misjudgment during node selection. Furthermore, by eliminating unqualified nodes through stability verification and assigning a hot standby node to each MCP, a rapid switch to a standby node can be achieved when the primary node fails, avoiding the problem of single-point failures causing interruptions in the parallel call process. This enhances the fault tolerance and overall operational reliability of multi-MCP parallel calls.

[0016] In one embodiment of the present invention, S156 includes: The MCP ready status marker data is processed by aggregating runtime parameters. The containerized runtime environment parameters, data permission configuration information, and network path connection information of each MCP in the ready state are extracted to generate MCP runtime parameter aggregate data. The MCP runtime parameter aggregate data is then processed by cross-validation for compatibility. The matching relationship between the data permission level and the network path link of each MCP is compared one by one. Parameter items with permission conflicts or link redundancy are eliminated to generate a parameter set that passes the compatibility verification. The parameter set that passes compatibility verification is encapsulated into a configuration package. According to the priority in the MCP call priority sequence data, the runtime environment parameters, data permission configurations, and network path information of each MCP are embedded layer by layer into a unified configuration package structure. Each MCP is assigned an independent configuration package partition identifier, generating a multi-MCP parallel call configuration package. The multi-MCP parallel call configuration package is then processed for call timing orchestration. Based on the MCP call priority sequence data, the execution order of each MCP is arranged, and the data dependencies and intermediate result transmission paths between adjacent MCPs are marked, generating multi-MCP parallel call timing orchestration data. The timing conflict detection process is performed on the timing orchestration data of multi-MCP parallel calls. The execution timing interval of each MCP is scanned one by one to identify conflict nodes with overlapping time or resource contention. The timing offset of the detected conflict nodes is adjusted to eliminate the risk of resource contention in the parallel execution process and generate conflict-free timing orchestration data. The ready signal broadcast process is performed on the conflict-free timing orchestration data. The multi-MCP parallel call ready signal is sent synchronously to the search processing layer and the model calling layer. The signal carries the configuration package partition identifier and execution timing number of each MCP, triggering the next stage of the multi-MCP parallel call processing flow.

[0017] The working principle and effects of the above technical solution are as follows: By aggregating runtime parameters and performing compatibility cross-validation on the MCP ready status marker data, parameter items with permission conflicts and redundant links can be eliminated before the configuration package is encapsulated. This avoids data access conflicts or communication link deadlocks between MCPs during parallel calls due to configuration defects. Simultaneously, by embedding configuration packages layer by layer according to priority and marking data dependencies and result transmission paths, each MCP can clearly understand its execution sequence and upstream / downstream connections. Furthermore, timing conflict detection and offset adjustment can eliminate resource contention risks and reduce computational resource waste caused by time overlap during parallel execution. In addition, by broadcasting ready signals synchronously with configuration package partition identifiers and execution sequence numbers, the search processing layer and model calling layer can quickly identify the startup order and identity information of each MCP, avoiding instruction confusion or call misalignment when multiple models respond simultaneously. This effectively improves the startup efficiency of multi-MCP parallel calls and the accuracy of overall collaborative execution.

[0018] In one embodiment of the present invention, S2 includes: S21. The price prediction MCP is activated based on the priority sequence data of the MCP. The price prediction MCP retrieves historical price data, macroeconomic indicator data, and policy change data from a multi-source market database, and uses time series analysis algorithms to perform multi-period extrapolation and calculation of price trends to generate price trend prediction data. At the same time, the supply and demand analysis MCP is activated. The supply and demand analysis MCP retrieves capacity data, inventory data, and import and export data from upstream and downstream databases of the industrial chain, and uses a supply and demand balance model to perform quantitative analysis of market supply and demand conditions to generate supply and demand balance analysis data. S22. While the price forecasting MCP and supply and demand analysis MCP are running, the risk assessment MCP is started. The risk assessment MCP retrieves meteorological disaster data, geopolitical event data and emergency policy data from the external risk event database, uses the risk quantification model to conduct probability assessment of potential risks, and generates risk assessment quantification data. Each MCP runs in parallel in its own independent computing environment, and the calculation process does not interfere with each other. Each MCP writes the real-time calculation intermediate data into its own independent cache. S23. Perform phased snapshot processing on the real-time intermediate calculation data in the independent buffer of each MCP to generate phased calculation result data for each MCP; perform integrity verification on the phased calculation result data of each MCP, remove intermediate data with calculation abnormalities, and generate verified phased calculation result data. S24. Perform multiple rounds of iterative calculations on the verified interim calculation results data. Each MCP performs convergence judgment on its own analysis results. When the calculation results meet the preset convergence conditions, the iteration stops and the final calculation result data of each MCP is generated. The final calculation result data of each MCP is timestamped to generate a multi-MCP calculation result dataset with timestamps. S25. Perform consistency comparison processing on the multi-MCP calculation result dataset with timestamps, detect whether there are logical conflicts between the calculation results of each MCP, mark the conflicting results, and generate conflict-marked data; perform conflict resolution processing on the conflict-marked data, and use a weighted voting mechanism to generate conflict-free final calculation result data for each MCP.

[0019] The working principle and effects of the above technical solution are as follows: By allowing the three MCPs—price prediction, supply and demand analysis, and risk assessment—to run in parallel in their respective independent computing environments and write data to independent caches in real time, it ensures that different models do not compete for resources when retrieving multi-source market data, industry chain data, and external risk data. This avoids problems such as model lag or data read timeouts caused by competition for computing resources. Simultaneously, periodic snapshots and integrity checks are performed on the real-time intermediate data. This not only promptly removes abnormal intermediate data but also ensures the stability of the final result through multiple rounds of iterative convergence checks, reducing the risk of erroneous conclusions due to single-calculation deviations. Furthermore, the use of a weighted voting mechanism to resolve logical conflicts between the calculation results of each MCP effectively eliminates contradictions in conclusions caused by differences in data sources between different models. This avoids situations where conflicting results from different models prevent users from obtaining accurate analytical information, thereby improving the overall credibility and logical rigor of commodity market analysis results.

[0020] In one embodiment of the present invention, step S3 includes: S31. Perform duplicate information detection processing on the conflict-free final calculation result data, use a similarity matching algorithm to identify duplicate content in different MCP calculation results, and generate duplicate information marked data; perform deduplication processing on the duplicate information marked data, retain the calculation result with the highest information density, and generate deduplicated multi-MCP calculation result data; S32. Perform cross-model correlation analysis on the deduplicated multi-MCP calculation results data to explore the causal relationship between price forecast results and supply and demand analysis results, as well as the influence relationship between risk assessment results and price forecast results, and generate cross-model correlation data; perform complementary information filling processing on the cross-model correlation data to supplement the correlation data with the unique analysis dimension information of each MCP, and generate complementary enhanced data. S33. Perform data fusion processing on complementary and enhanced data, and use multi-source data fusion algorithms to combine and arrange the analysis results of each MCP according to logical hierarchy and time dimension to generate fused comprehensive analysis data; perform structured report generation processing on the comprehensive analysis data, and transform the analysis data into a structured search report containing trend judgment, key factor list and risk warning, and generate structured search report data. S34. Perform visualization transformation on the structured search report data, converting price trend data into line charts, supply and demand comparison data into bar charts, and risk distribution data into pie charts to generate multiple types of chart data; perform layout processing on the multiple types of chart data and the text descriptions in the structured search report to generate a search report presentation data that combines text and graphics. S35. Perform user-adaptation processing on the search report data that combines text and images. Based on the user preference information recorded in the user interaction layer, personalize the display format and information emphasis of the report to generate personalized search report data. Push the personalized search report data to the user interaction layer for display.

[0021] The working principle and effects of the above technical solution are as follows: By performing similarity matching to deduplicate the final calculation results without conflicts and retaining the results with the highest information density, redundant analysis content between different MCPs can be effectively eliminated, avoiding the problem of the same conclusion appearing repeatedly and causing the report to be bloated. At the same time, cross-model correlation analysis and complementary information filling are performed on the deduplicated data, which can not only explore the causal chain between price forecasting and supply and demand analysis, but also supplement the related data with the unique analysis dimensions of each model, enhancing the information coverage of comprehensive judgment. In addition, the analysis results are combined and arranged according to logical hierarchy and time dimension using multi-source data fusion algorithms, and transformed into a structured report containing trend judgment, key factors and risk warnings. This can integrate the scattered model outputs into a clear and logical judgment conclusion, reducing the workload of users piecing together information themselves. Through visualization transformation and layout, data is presented in a combination of text and graphics, and then personalized according to user preferences. This allows users with different roles to quickly obtain information that matches their own interests, and avoids the situation where key information is buried due to monotonous report display, thereby improving the readability of search reports and the efficiency of decision support.

[0022] In one embodiment of the present invention, S33 includes: The complementary and enhanced data are logically hierarchically divided, and the analysis results of each MCP are classified into price prediction layer, supply and demand analysis layer, risk assessment layer and policy impact layer to generate hierarchical analysis data. The hierarchical analysis data is then aligned with the time dimension, and the analysis results produced by different MCPs in different time windows are uniformly mapped to the same time axis to generate time-aligned hierarchical data. Cross-layer causal link construction is performed on time-aligned hierarchical data. Causal inference algorithms are used to mine the positive driving relationship between the price prediction layer and the supply and demand analysis layer, as well as the corrective influence path of the risk assessment layer on the price prediction layer, to generate cross-layer causal association data. Weight factor allocation is performed on the cross-layer causal association data. According to the information density and historical verification accuracy of the analysis results of each layer, fusion weights are assigned to the analysis data of different layers to generate weighted hierarchical data. The weighted hierarchical data is processed by multi-source fusion. A fusion algorithm combining weighted summation and logical constraints is used to stack and cross-validate the analysis data of each level according to preset logical rules to generate fused comprehensive analysis data. The comprehensive analysis data is then subjected to consistency verification to detect whether there are logical contradictions between the analysis results of different levels. Data nodes with contradictions are backtracked and corrected to generate comprehensive analysis data that has passed the verification. The validated comprehensive analysis data is processed to generate a structured report. The trend judgment information, key influencing factor list and risk warning information in the comprehensive analysis data are extracted and filled into the preset report template. A structured search report containing trend judgment paragraphs, key factor list paragraphs and risk warning paragraphs is generated, and structured search report data is generated.

[0023] The working principle and effects of the above technical solution are as follows: By logically dividing complementary and enhanced data into hierarchical levels and mapping the analysis results of different MCPs to a unified time axis, data from price forecasting, supply and demand analysis, risk assessment, and policy impact can be aligned on the same time base, avoiding deviations in cross-layer correlation analysis caused by inconsistent time windows. Simultaneously, by using a causal inference algorithm to construct cross-layer causal links and assign fusion weights, it can both uncover the positive driving relationship between price and supply and demand and identify the correction path of the risk layer to price forecasts, enhancing the logical depth of comprehensive judgment. Furthermore, by using a fusion algorithm combining weighted summation and logical constraints for layer-by-layer superposition and cross-validation, followed by consistency checks to eliminate logically contradictory nodes, it can effectively prevent conflicting analysis results from different levels, reducing the confusion users face when dealing with contradictory data. Finally, by filling in the trend judgment, key factors, and risk warnings into the report template, a structured search report is generated. This not only makes the analysis conclusions clear and highlight the key points, but also reduces the time cost for users to extract key information from scattered data, thereby improving the efficiency of the search report in supporting trading decisions and the accuracy of information transmission.

[0024] In one embodiment of the present invention, step S4 includes: S41. After the user obtains the search report, the feedback collection module is triggered to guide the user to rate the accuracy, completeness and timeliness of the search results in multiple dimensions. At the same time, binary feedback data on whether the user found the information they need is collected to generate user feedback evaluation data. Sentiment analysis is performed on the user feedback evaluation data to identify the user's positive or negative emotions towards the search results and generate sentiment label data. S42. Perform joint analysis and processing of sentiment tendency tag data and user feedback evaluation data. Use the anomaly detection model in machine learning algorithm to perform backtracking analysis on the entire search process, locate the specific links of semantic understanding deviation or unreasonable model calling, and generate problem location data. Perform root cause analysis processing on problem location data to distinguish whether it is a comprehension deviation of the semantic analysis layer or an inappropriate strategy of the model calling layer, and generate root cause classification data. S43. For issues in the root cause classification data that fall under semantic understanding bias, adjust the parameters of the semantic analysis model. Use gradient optimization algorithm to fine-tune the weight parameters of the natural language processing model and generate adjusted semantic analysis model parameter data. For issues in the root cause classification data that fall under unreasonable model calls, optimize the MCP calling strategy. Recalculate the calling priority and search range boundary of each MCP and generate optimized MCP calling strategy data. S44. Update the adjusted semantic analysis model parameter data and the optimized MCP calling strategy data, write the updated parameters and strategies into the model running configuration file, and generate model configuration update data; perform A / B testing on the model configuration update data, enable the new configuration in some user requests and compare the search results, and generate effect comparison verification data. S45. Perform continuous training data construction processing on the effect comparison verification data, associate and package the verified user feedback data with the corresponding problem location data and root cause classification data to generate a model continuous training dataset; perform incremental training processing on the model continuous training dataset, and synchronize the trained model parameters to the running environment of each MCP to complete the continuous iterative evolution of the model.

[0025] The working principle and effects of the above technical solution are as follows: By triggering feedback collection after the user obtains the report and combining it with sentiment analysis, it can capture the user's true feelings from multiple dimensions such as accuracy, completeness, and timeliness. This avoids missing implicit user dissatisfaction with search results by relying solely on a single rating dimension. Simultaneously, by using an anomaly detection model to perform backtracking analysis of the entire search process, it can accurately locate specific points where semantic understanding deviations or unreasonable model calls occur, preventing subsequent optimization directions from going astray due to unclear issues. Furthermore, by differentiating between semantic layer and model call layer problems through root cause analysis and performing parameter fine-tuning and strategy optimization separately, the semantic analysis model can continuously converge to a better state under gradient optimization. It can also recalculate the call priority and search boundaries of each MCP, reducing the probability of wasted search resources or result deviations due to improper model call strategies. Moreover, after configuration updates, A / B testing is performed first, and then the passed data is packaged into a training set for incremental training and synchronized to each MCP environment. This effectively avoids the risk of search quality fluctuations that may occur with full updates, reduces service instability during the switch between old and new configurations, and thus enhances the system's self-evolution capability and the continuous stability of search results during long-term use.

[0026] In one embodiment of the present invention, S42 includes: The system performs feature concatenation processing on sentiment tendency label data and user feedback evaluation data, and integrates user satisfaction scores, binary feedback results, and positive or negative sentiment labels with multi-dimensional features to generate joint feature vector data. The joint feature vector data is then processed for outlier pattern matching. The isolated forest model in machine learning algorithms is used to score the outlier of each stage of the search process, locate specific stages of semantic understanding bias or unreasonable model calls, and generate problem location data. The problem location data is labeled with the process of assigning the process to a specific stage. The specific problem stages are mapped and matched with the semantic analysis layer and the model calling layer to generate stage assignment data. The stage assignment data is then quantified to measure the degree of deviation. The deviation of the outlier score of each stage from the historical baseline is calculated to generate deviation quantification data. The deviation quantification data is processed for root cause determination. The decision tree classification model is used to identify the cause of the abnormal link, distinguishing whether it is the understanding deviation of the semantic analysis layer or the improper strategy of the model calling layer, and generating root cause classification data. The root cause classification data is then processed for confidence assessment. Each attribution result is assigned a confidence score, and high-confidence root cause data with confidence scores higher than the preset threshold are selected to generate a high-confidence root cause dataset. The high-confidence root cause dataset is subjected to impact range diffusion analysis to assess the impact range of a single root cause problem on the upstream and downstream search processes, generating impact diffusion data. The impact diffusion data is then prioritized and sorted in descending order according to the size of the impact range and the confidence level, generating root cause treatment priority sequence data.

[0027] The working principle and effects of the above technical solution are as follows: By concatenating user satisfaction scores, binary feedback results, and sentiment tags into multi-dimensional features, and then using an isolated forest model to score outliers at each stage of the entire search process, it can accurately capture abnormal stages with semantic comprehension biases or unreasonable model calls from massive user feedback, avoiding the situation where problem localization is unclear due to fragmented feedback data. Simultaneously, by using a decision tree classification model to attribute causes to abnormal stages and assign confidence scores, it can distinguish between comprehension biases at the semantic analysis layer and inappropriate strategies at the model call layer, and filter out low-confidence attribution results through confidence scores, enhancing the accuracy of root cause determination. Furthermore, by performing impact range diffusion analysis on high-confidence root causes and sorting them in descending order of impact range and confidence, the system can prioritize the issues that have the greatest impact on the search process, avoiding the situation where multiple issues coexist without priority, leading to scattered repair resources, thereby improving the targeting of feedback-driven optimization and the iterative efficiency of overall search quality.

[0028] One embodiment of the present invention provides an AI-powered multi-source data retrieval system for commodity market analysis, comprising: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors are made to implement the method described in any one of the above.

[0029] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A multi-source data retrieval method for AI-based trading question-and-answer for commodity market analysis, characterized in that: The method includes: S1. Perform semantic analysis on the bulk commodity transaction Q&A questions transmitted from the user interaction layer, generate structured search instructions, and allocate the required model type and search scope according to the semantic analysis results; S2. Perform parallel multi-MCP call processing according to the search command; S3. Perform data fusion processing on the calculation results of multiple MCP outputs to form a complete search report, which is presented in a combination of charts and text descriptions; S4. After the user obtains the search report, feedback is collected and processed. Based on the user feedback, machine learning algorithms are used to analyze the search process, identify semantic understanding biases or unreasonable model calls, adjust the semantic analysis model parameters and optimize the MCP calling strategy, and store the user feedback data in the database.

2. The AI-based multi-source data retrieval method for commodity market analysis based on claim 1, characterized in that, S1 includes: S11. Perform semantic recognition and text processing on the bulk commodity transaction Q&A questions transmitted from the user interaction layer to generate standardized Q&A text data; perform word segmentation and part-of-speech tagging on the standardized Q&A text data to generate a multi-dimensional keyword feature set; S12. Perform semantic dependency analysis on the multi-dimensional keyword feature set, use intent recognition algorithm to understand the complete intent of the question, map the question intent to the preset intent classification system, and generate intent classification label data; perform context association expansion on the intent classification label data to generate extended intent data; S13. Perform structured transformation on the extended intent data to generate structured search instruction data; perform model matching analysis on the structured search instruction data to generate model type matching data; S14. Perform search range definition processing on the model type matching data to generate search range limitation data for each MCP; sort the search range limitation data of each MCP by priority to generate MCP call priority sequence data. S15. Perform resource pre-allocation processing on the MCP call priority sequence data to generate MCP resource pre-allocation data; prepare for parallel calls using the MCP resource pre-allocation data to generate a multi-MCP parallel call ready signal.

3. The AI-based multi-source data retrieval method for commodity market analysis based on claim 2, characterized in that, S15 includes: S151. Perform resource feature parsing on the MCP call priority sequence data to generate MCP resource demand feature data; perform resource pool matching analysis on the MCP resource demand feature data to generate resource node matching data. S152. Perform load balancing assessment on the resource node matching data to generate node load score data; perform optimal node selection on the node load score data to generate MCP and compute node binding data. S153. Perform runtime environment isolation configuration processing on the MCP and compute node binding data to generate containerized runtime environment data; perform network path establishment processing on the containerized runtime environment data to generate network path configuration data. S154. Perform data permission allocation processing on the network path configuration data to generate data permission allocation data; perform security sandbox deployment processing on the data permission allocation data to generate security protection configuration data. S155. Perform parallel call readiness verification processing on the security protection configuration data to generate operational status detection data for each MCP; perform readiness judgment processing on the operational status detection data for each MCP to generate MCP readiness status marker data. S156. Perform parallel call preparation and integration processing on the MCP ready status marker data to generate a multi-MCP parallel call configuration package; perform call timing orchestration processing on the multi-MCP parallel call configuration package to generate multi-MCP parallel call timing orchestration data; perform ready signal broadcast processing on the multi-MCP parallel call timing orchestration data to synchronously send multi-MCP parallel call ready signals to the search processing layer and the model call layer, triggering the next stage of the multi-MCP parallel call processing flow.

4. The AI-based transaction question-and-answer multi-source data retrieval method for commodity market analysis according to claim 3, characterized in that, S152 includes: Load factor extraction is performed on the resource node matching data to generate a real-time load factor group for candidate nodes; time window aggregation is then performed on the real-time load factor group for candidate nodes to generate steady-state load data for candidate nodes. The steady-state load data of candidate nodes is subjected to weighted round-robin scoring to generate an initial comprehensive score value for each node; the initial comprehensive score value is then normalized and mapped to generate node load score data. The node load score data is sorted and layered to generate node layered data; cross-layered verification is performed on the node layered data to generate node stability verification data. The node stability verification data is processed to select the optimal node and generate MCP and compute node binding data; the MCP and compute node binding data is processed to perform redundant backup and generate MCP hot standby binding data.

5. The AI-based multi-source data retrieval method for commodity market analysis according to claim 3, characterized in that, S156 includes: The MCP ready status marker data is processed by aggregating operating parameters to generate MCP operating parameter aggregate data; the MCP operating parameter aggregate data is then processed by cross-validation for compatibility to generate a parameter set that passes compatibility validation. The parameter set that passes the compatibility check is encapsulated into a configuration package to generate a multi-MCP parallel call configuration package; the multi-MCP parallel call configuration package is then processed for call timing orchestration to generate multi-MCP parallel call timing orchestration data. Perform timing conflict detection and processing on the timing orchestration data of multiple MCP parallel calls, and generate conflict-free timing orchestration data; The ready signal broadcast process is performed on the conflict-free timing orchestration data. The multi-MCP parallel call ready signal is sent synchronously to the search processing layer and the model calling layer. The signal carries the configuration package partition identifier and execution timing number of each MCP, triggering the next stage of the multi-MCP parallel call processing flow.

6. The AI-based multi-source data retrieval method for commodity market analysis according to claim 1, characterized in that, S2 includes: S21. Based on the MCP priority sequence data, start the price prediction MCP, use time series analysis algorithm to perform multi-period extrapolation calculation on price trend, and generate price trend prediction data; at the same time, start the supply and demand analysis MCP, use supply and demand balance model to perform quantitative analysis on market supply and demand, and generate supply and demand balance analysis data. S22. Start the risk assessment MCP, use the risk quantification model to conduct a probability assessment of potential risks, and generate risk assessment quantification data; each MCP runs in parallel in its own independent computing environment, and each MCP writes the real-time calculation intermediate data into its own independent cache. S23. Perform phased snapshot processing on the real-time intermediate calculation data in the independent buffer of each MCP to generate phased calculation result data for each MCP; perform integrity verification on the phased calculation result data of each MCP to generate verified phased calculation result data. S24. Perform multiple rounds of iterative calculations on the verified interim calculation results data to generate the final calculation results data for each MCP; timestamp the final calculation results data for each MCP to generate a multi-MCP calculation result dataset with timestamps. S25. Perform consistency comparison processing on the multi-MCP calculation result dataset with timestamps, detect whether there are logical conflicts between the calculation results of each MCP, mark the conflicting results, and generate conflict-marked data; perform conflict resolution processing on the conflict-marked data, and use a weighted voting mechanism to generate conflict-free final calculation result data for each MCP.

7. The AI-based multi-source data retrieval method for commodity market analysis based on claim 1, characterized in that, The S3 includes: S31. Perform duplicate information detection processing on the conflict-free final calculation result data to generate duplicate information marked data; perform deduplication processing on the duplicate information marked data to generate deduplicated multi-MCP calculation result data; S32. Perform cross-model correlation analysis on the deduplicated multi-MCP calculation results data to generate cross-model correlation data; perform complementary information filling on the cross-model correlation data to generate complementary enhanced data. S33. Perform data fusion processing on the complementary and enhanced data to generate fused comprehensive analysis data; perform structured report generation processing on the comprehensive analysis data to transform the analysis data into a structured search report and generate structured search report data; S34. Perform visualization transformation on the structured search report data to generate various types of chart data; perform layout processing on the various types of chart data and the text descriptions in the structured search report to generate a search report presentation data that combines text and graphics; S35. Perform user-adaptation processing on the search report data presented by combining text and images to generate personalized search report data; push the personalized search report data to the user interaction layer for display.

8. The AI-based multi-source data retrieval method for commodity market analysis according to claim 7, characterized in that, S33 includes: The complementary and enhanced data are logically hierarchically divided, and the analysis results of each MCP are classified into price prediction layer, supply and demand analysis layer, risk assessment layer and policy impact layer to generate hierarchical analysis data; the hierarchical analysis data is then aligned with the time dimension to generate time-aligned hierarchical data. Cross-layer causal link construction is performed on time-aligned hierarchical data to generate cross-layer causal association data; weight factor allocation is performed on the cross-layer causal association data to generate weighted hierarchical data. Multi-source fusion processing is performed on the weighted hierarchical data to generate fused comprehensive analysis data; consistency verification is performed on the comprehensive analysis data to generate comprehensive analysis data that passes the verification. The validated comprehensive analysis data is processed to generate a structured report, which includes a trend judgment paragraph, a key factor list paragraph, and a risk warning paragraph. The structured search report data is generated.

9. The AI-based multi-source data retrieval method for commodity market analysis based on claim 1, characterized in that, The S4 includes: S41. After the user obtains the search report, the feedback collection module is triggered to generate user feedback evaluation data; the user feedback evaluation data is processed by sentiment analysis to generate sentiment tag data. S42. Perform joint analysis and processing of sentiment tendency tag data and user feedback evaluation data to generate problem location data; perform root cause analysis processing on the problem location data to generate root cause classification data; S43. Perform semantic analysis model parameter adjustment processing, fine-tune the weight parameters of the natural language processing model, and generate adjusted semantic analysis model parameter data; perform MCP invocation strategy optimization processing, and generate optimized MCP invocation strategy data. S44. Update the adjusted semantic analysis model parameter data and the optimized MCP calling strategy data to generate model configuration update data; perform A / B testing on the model configuration update data to generate effect comparison verification data. S45. Perform continuous training data construction processing on the effect comparison and verification data to generate a model continuous training dataset; perform incremental training processing on the model continuous training dataset, and synchronize the trained model parameters to the running environment of each MCP to complete the continuous iterative evolution of the model.

10. An AI-powered multi-source data retrieval system for trading question-and-answering in commodity market analysis, characterized in that: The system includes: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1 to 9.

Images