Intelligent agent task situation recognition and fact matching rate optimization method, system, device and medium based on AI twin information warehouse

Abstract

The application discloses an intelligent agent task situation recognition and fact matching rate optimization method, system, device and medium based on an AI twin information warehouse, and belongs to the field of artificial intelligence. The method comprises the following steps: receiving an intelligent agent task execution signal set; generating an intelligent agent task situation representation based on the signal set; deriving a fact matching constraint set from the task situation representation; obtaining a commodity or service fact unit from an AI twin information warehouse; calculating fact correlation characteristics and matching scores; generating a matching explanation path from a task situation or a fact matching constraint to a fact node, an evidence node and a matching conclusion node; and generating a machine-readable context package and updating a twin layer optimization strategy. Through the above scheme, the current task situation of the intelligent agent can be recognized when the original human instructions are invisible or incomplete, the secondary analysis cost of the intelligent agent is reduced, the fact reasoning path is shortened, and the fact coverage rate and evidence traceability are improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, and specifically relates to a method, system, device and medium for intelligent agent task situation recognition and fact matching rate optimization based on AI twin information warehouse. Background Technology

[0002] With the development of large language models and intelligent agent technology, software intelligent agents are gradually gaining the ability to represent users in reading web pages, retrieving products, comparing solutions, verifying policies, calling interfaces, and executing procurement tasks.

[0003] Unlike traditional searches or site-specific filtering, modern intelligent agents typically do not simply relay raw human instructions when performing tasks. Instead, they may have developed a phased plan based on the user's long-term memory, session history, organizational rules, previous search results, tool returns, and their own reasoning. This plan then continues to advance the task within the network environment through MCP tools, web searches, browser proxies, API calls, and form interactions. For example, after initially receiving the message "Find me suitable monitors for my office," the agent might generate a comparison table schema with fields including price, delivery time, warranty policy, invoice, inventory, and source of evidence. When the agent subsequently visits a merchant's website, the system may only observe MCP resource requests, search terms, form fields, or context packet requests, rather than the complete raw human instruction.

[0004] Current internet product and service information is primarily geared towards human visual reading and traditional search engine crawling. Merchants typically publish information through web pages, platform stores, product detail pages, PDF manuals, image posters, FAQs, customer service knowledge bases, product databases, pricing systems, inventory systems, and after-sales policy pages. Traditional search engine optimization (SEO) techniques mainly improve the crawling and display of web pages by optimizing titles, keywords, page structure, links, sitemaps, page speed, and structured markup. These techniques typically serve search ranking, summary display, and human clicks.

[0005] Structured data technology can embed fields such as product name, price, brand, and reviews into web pages using methods like schema.org, JSON-LD, and Microdata. E-commerce platforms can also synchronize some structured product information through product feeds, inventory feeds, or advertising feeds.

[0006] Retrieval Augmentation (RAG) systems typically segment documents into fragments, write them into a vector database, and retrieve fragments based on text similarity for a language model to generate responses. Knowledge graph systems can represent entities and relationships such as products, brands, categories, specifications, policies, qualifications, stores, and reviews.

[0007] AI twin information warehouse technology can transform raw information warehouses into agent-readable intermediate representations, relational networks, memory summaries, task intent indexes, and context packages. This technology solves the problems of general information warehouses for agent reading, memorization, retrieval, and action.

[0008] However, existing technologies still have the following shortcomings when intelligent agents autonomously perform tasks and participate in product discovery, solution comparison, or procurement decisions.

[0009] First, traditional SEO and search engine marketing mainly focus on text relevance, page weight, and display effect, which cannot identify the current task status of the agent, nor can they guarantee that the agent obtains a verifiable chain of facts.

[0010] Second, the field coverage of page-level product structured data is limited, and it cannot align with comparison schemas, inquiry schemas, evidence citation schemas, or tool parameter schemas generated by the agent itself.

[0011] Third, ordinary RAG and vector retrieval rely on text similarity, which cannot reliably determine whether a product or service meets the hard constraints, exclusion conditions, evidence requirements, and version expiration under the current task situation.

[0012] Fourth, although knowledge graphs can express entity relationships, they typically do not process execution signals such as agent MCP calls, web page search paths, browser navigation paths, API parameters, and task memories, nor do they directly generate machine-readable context packets for the current task stage.

[0013] Fifth, while general AI twin information warehouses can improve the efficiency of intelligent agents in reading information, they are not specifically optimized for task situation recognition, schema alignment, fact relevance matching rate, interpretation path, and feedback tuning in the autonomous execution behavior of intelligent agents.

[0014] Therefore, a technical solution is needed that, based on an existing or available AI twin information warehouse, first identifies the current task status from the set of task execution signals of the intelligent agent, then derives a set of fact matching constraints based on the task status, and generates matching scores, explanation paths, machine-readable context packets, and feedback optimization strategies for the fact units of goods or services. Summary of the Invention

[0015] To address the aforementioned issues, the present invention aims to provide a method, system, device, and medium for intelligent agent task situation recognition and fact matching rate optimization based on an AI twin information warehouse. This addresses the technical problems existing in the process of intelligent agents autonomously executing tasks, such as the invisibility of original human instructions, difficulty in aligning internal schemas of intelligent agents, difficulty in identifying task stages, difficulty in matching factual constraints, difficulty in tracing evidence chains, excessively long inference paths, lack of task adaptation in context packages, and inability to achieve closed-loop optimization of feedback.

[0016] To achieve the above objectives, the technical solution of the present invention is as follows:

[0017] This invention also provides a method for intelligent agent task situation recognition and fact matching rate optimization based on AI twin information warehouse, including the following steps:

[0018] S1: The system receives a set of task execution signals from the intelligent agent;

[0019] S2: The system generates an agent task situation representation based on the set of agent task execution signals;

[0020] S3: The system generates a set of fact-matching constraints based on the agent's task situation representation;

[0021] S4: Based on the fact matching constraint set, the system retrieves fact units related to goods or services from the AI ​​twin information warehouse containing semantic nodes, source anchors, relation edges and version records;

[0022] S5: The system calculates fact relevance features based on the fact matching constraint set and the fact unit;

[0023] S6: The system generates a fact relevance matching score based on the aforementioned fact relevance features;

[0024] S7: The system generates a matching interpretation path;

[0025] S8: The system generates or adjusts a machine-readable context packet for the agent based on the fact relevance matching score and the matching interpretation path;

[0026] S9: The system updates the twin layer tuning strategy based on the fact relevance matching score, the matching explanation path, or the agent feedback event.

[0027] Further, in step S1, the set of agent task execution signals includes at least one of the following: original user goal, task memory summary, agent generation schema, model context protocol resource request, tool call trajectory, web page search trajectory, browser navigation trajectory, application programming interface call parameters, context packet request, feedback event, and environmental observation.

[0028] Furthermore, the task memory summary includes at least one of long-term memory, conversational memory, task memory, historical comparison results, user preferences, organizational procurement rules, failed attempt records, or intermediate conclusions.

[0029] Furthermore, the intelligent agent generation schema includes at least one of the following: product comparison schema, supplier screening schema, inquiry parameter schema, compliance verification schema, evidence citation schema, output table field or tool call parameter schema.

[0030] Furthermore, the model context protocol resource request includes at least one of the following: resource URI, resource type, tool name, tool parameters, returned error, retry record, or resource list read record.

[0031] Furthermore, the webpage search trajectory and browser navigation trajectory include at least one of the following: search terms, search result click path, accessed page, document object model path, page fragment reading record, return to previous page event, repeated query event, or page inaccessibility event.

[0032] Furthermore, the environmental observation includes at least one of the following: insufficient inventory, price conflict, policy conflict, missing evidence, missing fields, inaccessible page, unreachable action entry point, or insufficient permissions.

[0033] Further, S2 includes:

[0034] S21: Deducing task stages and field requirements based on tool call trajectories;

[0035] S22: Generate schema-aligned fact fields for goods or services based on the agent;

[0036] S23: Identify task objectives and constraints based on webpage search trajectories and browser navigation trajectories;

[0037] S24: Intent disambiguation based on task memory summary;

[0038] S25: Generate multiple candidate agent task situation representations and their confidence levels when the signal is insufficient or ambiguous.

[0039] Furthermore, in steps S2 and S3, the agent task situation representation includes at least one of the following: candidate task target, task stage, activity sub-target, agent generation schema, schema source, required fact slots, evidence requirements, exclusion conditions, action intention, uncertainty, and confidence level.

[0040] Furthermore, the task phases include at least one of the following: information discovery, condition screening, scheme comparison, evidence verification, price inquiry, purchase redirection, appointment, after-sales judgment, or compliance check.

[0041] Furthermore, when the intelligent agent's task situation corresponds to a procurement task, the fact matching constraint set includes a procurement requirement constraint set, which includes at least one of the following: budget range, specifications, service area, delivery conditions, inventory status, qualification requirements, after-sales conditions, compatibility conditions, exclusion conditions, evidence requirements, or action requirements.

[0042] Furthermore, the fact matching constraint set includes at least one of the following: source task situation identifier, target category, task stage, schema mapping, hard constraints, preference constraints, exclusion constraints, evidence requirements, action requirements, context budget, and constraint confidence.

[0043] Further, step S3 includes:

[0044] S31: Map the fields in the agent generation schema to the fact fields in the AI ​​twin information warehouse;

[0045] S32: Generate synonym candidates, gap completion tasks, or uncertainties for schema fields that fail to be mapped;

[0046] S33: Map task phases to fact slot priorities and context package templates.

[0047] Furthermore, the AI ​​twin information warehouse includes semantic nodes, source anchors, relation edges, task intent indexes, memory summaries, permission information, and version records; S4 includes obtaining the fact units based on at least one of entity indexes, task intent indexes, source anchor indexes, full-text indexes, vector indexes, or graph indexes.

[0048] Furthermore, the fact unit includes at least one of the following: field value, source anchor, evidence reference, validity period, confidence level, conflict status, permission scope, and action entry point.

[0049] Furthermore, when the fact unit triggers the exclusion constraint, it marks the corresponding product or service as not recommended, downgraded to recommended, or pending manual confirmation.

[0050] Furthermore, in step S5, the fact relevance features include at least one of field coverage, constraint satisfaction, evidence completeness, version freshness, conflict status, permission matching, reasoning path length, and action entry accessibility.

[0051] Further, in step S6, the fact relevance matching score is calculated based on at least one of the following factors: hard constraint satisfaction rate, field coverage rate, preference constraint satisfaction, evidence completeness, version freshness, action entry accessibility, conflict penalty, reasoning path penalty, exclusion condition penalty, and task situation confidence.

[0052] Furthermore, in step S7, the matching interpretation path includes task status nodes or fact matching constraint nodes, product or service fact nodes, evidence nodes, relationship edges, and matching conclusion nodes.

[0053] Furthermore, the matching interpretation path is generated by searching for paths in the relational network of the AI ​​twin information warehouse. The path starts with a task status node or fact matching constraint node, ends with an evidence node or matching conclusion node, and passes through at least one product or service fact node.

[0054] Furthermore, the selection of the matching interpretation path is based on at least one of the following: path length, node confidence, source anchor coverage, evidence completeness, version freshness, conflict status, and task situation confidence.

[0055] Further, in step S8, the machine-readable context package includes at least one of the following: candidate goods or services ranking table, task status summary, schema field mapping, required constraint satisfaction status, difference fact matrix, constraints, evidence citation, source anchor, conflict indication, version time, action entry point, and mismatch reason.

[0056] Furthermore, when the context budget is limited, priority should be given to retaining content related to the current task phase, hard constraints, limitations, conflict warnings, evidence citations, and action entry points, while reducing or omitting repetitive explanatory text.

[0057] Furthermore, the twin layer optimization strategy includes at least one of the following: supplementing task tags, generating procurement scenario summaries, adding applicable scenario relationship edges, adding constraint condition fields, adjusting sorting weights, updating context package templates, generating field completion tasks, generating schema mapping rules, or generating conflict review tasks.

[0058] Further, step S9 includes: identifying the reasons for matching failure, wherein the reasons for matching failure include at least one of the following: uncertain task situation, schema field cannot be mapped, field is missing, insufficient evidence, ambiguous expression, unresolved conflict, version expired, inference path is too long, and action entry point is unreachable.

[0059] Furthermore, in step S9:

[0060] When the task situation is identified as uncertain, multiple candidate task situations are generated or a clarifying context is returned to the agent.

[0061] When a schema field is identified as unmappable, a schema mapping rule or field completion task is generated.

[0062] When the evidence is deemed insufficient, the original webpage fragment, document page number, table cell, image area, or database primary key is bound;

[0063] When the reasoning path is identified as too long, add direct relationship edges, task labels, or compress the explanation path.

[0064] Furthermore, step S9 also includes: gap filling and counterfactual inquiry:

[0065] Generate completion tasks based on failure reasons such as missing fields, insufficient evidence, inability to map schema fields, or irrelevant matching.

[0066] Generate a counterfactual query task based on the agent's task situation representation or fact matching constraint set;

[0067] The counterfactual query task is used to verify the matching status of goods or services under changing constraints, and the task tags, matching weights, or publication status are updated based on the verification results.

[0068] Furthermore, step S9 also includes: agent simulation evaluation step: using a preset agent task set to request machine-readable context packets before and after optimization, and generating evaluation results based on at least one of the following: task situation recognition accuracy, schema mapping success rate, matching task answerability rate, required field coverage, evidence backlink rate, average inference path length, irrelevant recall rate, correct citation rate, constraint recognition rate, and action entry recognition rate.

[0069] Furthermore, in step S9, the agent feedback event includes at least one of the following: task situation misjudgment event, schema mapping failure event, field missing event, evidence unreachable event, irrelevant matching event, parsing failure event, conflict report event, inquiry result event, or procurement result event.

[0070] Further, step S9 classifies the agent feedback events into situational correction, schema repair, field completion, evidence repair, correlation correction, parsing repair, conflict resolution, or action result categories, and updates the weights, task labels, field integrity requirements, summary templates, relation edges, schema mapping rules, or context package templates according to the classification results.

[0071] This application provides a system for intelligent agent task situation recognition and fact matching rate optimization based on an AI twin information warehouse, including:

[0072] The task signal receiving module is used to receive the set of task execution signals of the intelligent agent;

[0073] The task situation recognition module is used to generate a task situation representation of the agent based on the set of task execution signals of the agent;

[0074] The schema alignment module is used to map fields in the schema generated by the agent to fact fields in the AI ​​twin information warehouse;

[0075] The fact matching constraint generation module is used to generate a set of fact matching constraints based on the agent's task situation representation;

[0076] The fact unit acquisition module is used to acquire fact units related to goods or services from the AI ​​twin information warehouse based on the fact matching constraint set;

[0077] The matching and scoring module is used to calculate fact relevance features based on the fact matching constraint set and the fact unit, and generate a fact relevance matching score;

[0078] The interpretation path generation module is used to generate a matching interpretation path that includes task status nodes or fact matching constraint nodes, product or service fact nodes, evidence nodes, relationship edges, and matching conclusion nodes.

[0079] A context packet generation module is used to generate or adjust a machine-readable context packet for an agent based on the fact relevance matching score and the matching interpretation path.

[0080] The twin layer optimization module is used to update the twin layer optimization strategy based on the fact relevance matching score, the matching explanation path, or the agent feedback event.

[0081] Furthermore, the system also includes at least one of the following: a trajectory analysis module, a webpage behavior recognition module, a memory fusion module, an evaluation module, and a feedback optimization module;

[0082] The trajectory analysis module is used to infer the task stage and field requirements based on the tool call trajectory, MCP resource request, or API call parameters.

[0083] The webpage behavior recognition module is used to identify task objectives and constraints based on search terms, click paths, browser navigation paths, or page fragment reading records.

[0084] The memory fusion module is used to merge the task memory summary with the current execution signal;

[0085] The evaluation module is used to generate task situation recognition accuracy, schema mapping success rate, matching task answerability rate, required field coverage, evidence backlink rate, or average inference path length.

[0086] The feedback optimization module is used to receive agent feedback events and update weights, task labels, field integrity requirements, summary templates, relation edges, schema mapping rules, or context package templates.

[0087] This invention can be used for product discovery, condition filtering, solution comparison, evidence verification, price inquiry, reservation, purchase redirection, after-sales judgment, and compliance inspection for procurement agents, search agents, browser agents, personal assistant agents, enterprise procurement systems, or platform agents.

[0088] in:

[0089] The set of intelligent agent task execution signals refers to the signals carried, generated, or exposed by the intelligent agent during task execution, including but not limited to the original user goal, task memory summary, intelligent agent generation schema, MCP resource request, tool call trajectory, web page search trajectory, browser navigation trajectory, API call parameters, context packet request, feedback events, and environmental observations.

[0090] The agent task situation representation refers to the current task state inferred by the system based on the set of agent task execution signals. It may include candidate task objectives, task stages, activity sub-objectives, agent generation schema, schema source, required fact slots, evidence requirements, exclusion conditions, action intentions, uncertainties, and confidence levels.

[0091] The fact-matching constraint set refers to the set of computable constraints derived from the agent's task situation representation. In a procurement scenario, the fact-matching constraint set can be represented as the procurement requirement constraint set.

[0092] Fact-related matching score refers to the numerical value of the matching degree of candidate goods or services calculated based on fact-related features for a single intelligent agent task situation.

[0093] Fact matching rate refers to the optimization performance metric obtained by statistically analyzing indicators such as task situation identification accuracy, schema mapping success rate, matching score, task answerability rate, correct citation rate, irrelevant recall rate, and action entry point identification rate across multiple tasks, test tasks, or feedback events.

[0094] Furthermore, the present invention also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, implements the above-described method.

[0095] Furthermore, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0096] The beneficial effects of this invention are: compared with the prior art:

[0097] First, this invention enables the system to identify its current task status based on the agent's tool calls, MCP requests, web page searches, browser navigation, API parameters, task memory, and environmental observations, even when it cannot fully obtain the original human instructions.

[0098] Second, this invention generates schema alignment through intelligent agents, mapping the comparison tables, inquiry parameters, evidence citation formats, or tool parameters generated by the intelligent agents themselves to the product or service fact fields in the AI ​​twin information warehouse, thereby reducing field missing and mismatch caused by schema inconsistency.

[0099] Third, this invention obtains factual units of goods or services with source anchors, evidence citations, validity periods, confidence levels, and conflict states from an AI twin information warehouse, reducing the cost for intelligent agents to directly parse complex web pages, PDFs, images, and business systems.

[0100] Fourth, this invention generates a matching score based on field coverage, constraint satisfaction, evidence completeness, version freshness, conflict state, task situation confidence, inference path length, and action entry accessibility, so that agent matching no longer depends solely on text similarity.

[0101] Fifth, this invention generates a matching interpretation path, enabling the agent to trace the task status, fact nodes, and evidence nodes corresponding to a certain matching conclusion.

[0102] Sixth, when the context budget is limited, the present invention prioritizes the retention of the current task stage, hard constraints, limiting conditions, conflict prompts, evidence citations, and action entry points, thereby improving the task adaptability of the context package.

[0103] Seventh, this invention generates optimization strategies based on failure reasons such as uncertain task situation, unmapped schema fields, missing fields, insufficient evidence, ambiguous expression, unresolved conflicts, outdated versions, excessively long inference paths, and unreachable action entry points, enabling continuous optimization of the agent-oriented information layer. Attached Figure Description

[0104] Figure 1 This is the flowchart of the implementation method.

[0105] Figure 2 This is a system architecture diagram of this embodiment.

[0106] Figure 3 This is a schematic diagram of the data structure of the intelligent agent task execution signal set in this embodiment.

[0107] Figure 4This is a schematic diagram of the data structure for representing the task status of the intelligent agent in this embodiment.

[0108] Figure 5 This is a flowchart of the matching score and interpretation path generation process in this embodiment.

[0109] Figure 6 This is a flowchart of the twin layer optimization process in this embodiment.

[0110] Figure 7 This is a flowchart of the intelligent agent feedback closed-loop optimization process in this embodiment.

[0111] Figure 8 This is a diagram illustrating the code implementation principle of "initiating a resource request via MCP".

[0112] Figure 9 This is a schematic diagram illustrating the code implementation principle of "the system generating a task situation representation of the intelligent agent through signals".

[0113] Figure 10 This is a diagram illustrating the code implementation principle of "output schema requested by the agent".

[0114] Figure 11 This is a code diagram of "fact fields in an AI twin information warehouse".

[0115] Figure 12 This is a diagram illustrating the code implementation principle of "the system performs field name similarity, field description similarity, historical mapping rules, and unit type matching to generate a schema mapping".

[0116] Figure 13 This is a diagram illustrating the code implementation principle of "a set of fact-matching constraints".

[0117] Figure 14 This is a diagram illustrating the code implementation principle of "a unit of fact for a commodity".

[0118] Figure 15 This is a diagram illustrating the code implementation principle of the "matching scoring formula".

[0119] Figure 16 This is a diagram illustrating the code implementation principle of the "Example Context Package". Detailed Implementation

[0120] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0121] To achieve the above objectives, the technical solution of the present invention is as follows:

[0122] See Figure 1This embodiment also provides a method for intelligent agent task situation recognition and fact matching rate optimization based on AI twin information warehouse, including the following steps:

[0123] S1: The method first receives a set of agent task execution signals. The set of agent task execution signals may include at least one of the following: original user goal, task memory summary, agent generation schema, MCP resource request, tool call trajectory, web page search trajectory, browser navigation trajectory, API call parameters, context packet request, feedback event, and environmental observation.

[0124] S2: The system generates a task situation representation of the agent based on the set of agent task execution signals.

[0125] An agent's task situation representation can include candidate task objectives, task stages, activity sub-objectives, agent-generated schema, schema source, required fact slots, evidence requirements, exclusion conditions, action intentions, uncertainties, and confidence levels. The system can generate task situation representations in various ways. For example, the system can infer task stages and field requirements based on tool call trajectories; align product or service fact fields based on the agent-generated schema; identify task objectives and constraints based on web search trajectories and browser navigation trajectories; perform intent disambiguation based on long-term memory, session memory, or organizational procurement rules; and generate multiple candidate task situations when signals are insufficient or highly ambiguous.

[0126] S3: The system then generates a fact-matching constraint set based on the agent's task situation representation. This set may include the source task situation identifier, target category, task stage, schema mapping, hard constraints, preference constraints, exclusion constraints, evidence requirements, action requirements, contextual budget, and constraint confidence. In a procurement scenario, the fact-matching constraint set can be specifically represented as a procurement requirement constraint set.

[0127] S4: The system retrieves fact units related to goods or services from the AI ​​twin information warehouse based on the fact matching constraint set. The fact units of goods or services may include fact identifier, entity identifier, entity type, product name or service name, normalized field, field value, unit, applicable conditions, source anchor, evidence citation, validity period, timeliness status, confidence level, conflict status, permission scope, task tag, and action entry point.

[0128] S5: The system calculates fact relevance features based on the fact matching constraint set and fact units. Fact relevance features may include field coverage, hard constraint satisfaction rate, preference constraint satisfaction, exclusion constraint triggering status, evidence completeness, version freshness, conflict status, permission matching degree, inference path length, action entry accessibility, and task situation confidence.

[0129] S6: The system generates a fact relevance matching score based on fact relevance features.

[0130] S7: The system also generates a matching explanation path. The matching explanation path includes a task status node or fact matching constraint node, a product or service fact node, an evidence node, a relation edge, and a matching conclusion node. This path starts at the task status node or fact matching constraint node, ends at the evidence node or matching conclusion node, and passes through at least one product or service fact node. It is used to explain to the agent why a certain product or service satisfies, partially satisfies, or does not satisfy the current task status.

[0131] S8: The system generates or adjusts a machine-readable context package based on the factual relevance matching score and the matching interpretation path. The machine-readable context package may include a candidate goods or services ranking table, a task status summary, schema field mapping, required constraint satisfaction status, a difference fact matrix, constraints, evidence citations, source anchors, conflict warnings, version time, action entry points, and reasons for mismatch.

[0132] S9: The system updates the twin layer optimization strategy based on matching scores, explanation paths, or agent feedback events. The twin layer optimization strategy may include supplementing task tags, generating procurement scenario summaries, adding applicable scenario relationship edges, adding constraint fields, adjusting sorting weights, updating context package templates, generating field completion tasks, generating schema mapping rules, or generating conflict review tasks.

[0133] See Figure 2-7 As shown, this embodiment provides an intelligent agent task situation recognition and fact matching rate optimization system based on an AI twin information warehouse, including:

[0134] The task signal receiving module is used to receive the set of task execution signals of the intelligent agent;

[0135] The task situation recognition module is used to generate a task situation representation of the agent based on the set of task execution signals of the agent;

[0136] The schema alignment module is used to map fields in the schema generated by the agent to fact fields in the AI ​​twin information warehouse;

[0137] The fact matching constraint generation module is used to generate a set of fact matching constraints based on the agent's task situation representation;

[0138] The fact unit acquisition module is used to acquire fact units related to goods or services from the AI ​​twin information warehouse based on the fact matching constraint set;

[0139] The matching and scoring module is used to calculate fact relevance features based on the fact matching constraint set and the fact unit, and generate a fact relevance matching score;

[0140] The interpretation path generation module is used to generate a matching interpretation path that includes task status nodes or fact matching constraint nodes, product or service fact nodes, evidence nodes, relationship edges, and matching conclusion nodes.

[0141] A context packet generation module is used to generate or adjust a machine-readable context packet for an agent based on the fact relevance matching score and the matching interpretation path.

[0142] The twin layer optimization module is used to update the twin layer optimization strategy based on the fact relevance matching score, the matching explanation path, or the agent feedback event.

[0143] Furthermore, the system also includes at least one of the following: a trajectory analysis module, a webpage behavior recognition module, a memory fusion module, an evaluation module, and a feedback optimization module;

[0144] The trajectory analysis module is used to infer the task stage and field requirements based on the tool call trajectory, MCP resource request, or API call parameters.

[0145] The webpage behavior recognition module is used to identify task objectives and constraints based on search terms, click paths, browser navigation paths, or page fragment reading records.

[0146] The memory fusion module is used to merge the task memory summary with the current execution signal;

[0147] The evaluation module is used to generate task situation recognition accuracy, schema mapping success rate, matching task answerability rate, required field coverage, evidence backlink rate, or average inference path length.

[0148] The feedback optimization module is used to receive agent feedback events and update weights, task labels, field integrity requirements, summary templates, relation edges, schema mapping rules, or context package templates.

[0149] This invention can be used for product discovery, condition filtering, solution comparison, evidence verification, price inquiry, reservation, purchase redirection, after-sales judgment, and compliance inspection for procurement agents, search agents, browser agents, personal assistant agents, enterprise procurement systems, or platform agents.

[0150] in:

[0151] The set of intelligent agent task execution signals refers to the signals carried, generated, or exposed by the intelligent agent during task execution, including but not limited to the original user goal, task memory summary, intelligent agent generation schema, MCP resource request, tool call trajectory, web page search trajectory, browser navigation trajectory, API call parameters, context packet request, feedback events, and environmental observations.

[0152] The agent task situation representation refers to the current task state inferred by the system based on the set of agent task execution signals. It may include candidate task objectives, task stages, activity sub-objectives, agent generation schema, schema source, required fact slots, evidence requirements, exclusion conditions, action intentions, uncertainties, and confidence levels.

[0153] The fact-matching constraint set refers to the set of computable constraints derived from the agent's task situation representation. In a procurement scenario, the fact-matching constraint set can be represented as the procurement requirement constraint set.

[0154] Fact-related matching score refers to the numerical value of the matching degree of candidate goods or services calculated based on fact-related features for a single intelligent agent task situation.

[0155] Fact matching rate refers to the optimization performance metric obtained by statistically analyzing indicators such as task situation identification accuracy, schema mapping success rate, matching score, task answerability rate, correct citation rate, irrelevant recall rate, and action entry point identification rate across multiple tasks, test tasks, or feedback events.

[0156] This embodiment has the following application scenarios:

[0157] I. In the case of task situation recognition based on MCP invocation and task memory, this embodiment provides a method for intelligent agent task situation recognition based on MCP invocation and task memory:

[0158] Early in the conversation, the agent receives the user's objective: "Purchase a batch of commercial monitors suitable for use in the Shanghai office." The agent writes the budget, region, invoice requirements, and historical brand preferences into its task memory. Subsequently, when the agent accesses the merchant's AI twin information layer, it does not resubmit the complete user objective but instead initiates a resource request via MCP. See the code implementation details below. Figure 8 .

[0159] The system receives the MCP resource request, tool parameters, and task memory summary, forming a set of agent task execution signals. The system identifies the tool name. During the comparison phase, the fields price, delivery_time, warranty_policy, and invoice_supported correspond to the price, delivery, warranty, and invoice fact slots, region:Shanghai corresponds to the region constraint, and output_schema:comparison_table corresponds to the comparison table schema generated by the agent.

[0160] The system generates a task situation representation for the agent based on the above signals. See the code implementation below. Figure 9 .

[0161] The system further generates a set of fact matching constraints from the task situation representation and then proceeds to the fact unit retrieval and matching scoring process.

[0162] II. Task Situation Recognition Based on Web Search and Browser Navigation: This embodiment provides a method for task situation recognition when the agent does not directly expose structured requests through MCP.

[0163] The intelligent agent searches for "Shanghai commercial monitor with next-day delivery and three-year warranty invoice" through a web search engine, and then visits the product list page, product details page, after-sales policy page, and inquiry form in sequence.

[0164] The system can observe search terms, click paths, page types viewed, page fragment reading records, and form fields.

[0165] The system extracts the target product category "commercial display", the region "Shanghai", the delivery constraint "next-day delivery", the warranty constraint "three-year warranty", and the invoice requirement from the search terms.

[0166] The system identifies the task stage from the browsing path, transitioning from the discovery stage to the evidence verification stage, and identifies the action intent as a price inquiry after accessing the price inquiry form.

[0167] The system generates a task status representation for the intelligent agent, in which the task phase is "evidence verification / inquiry preparation". The required fact slots include price, inventory, delivery range, warranty period, invoice policy and inquiry entry.

[0168] If the system cannot determine whether the agent is purchasing for an individual or a company, it records "Purchasing entity uncertain" in the uncertainty section and generates multiple candidate task states. If the agent subsequently requests invoice or after-sales service fields, the system can increase the confidence level of the company purchasing state.

[0169] III. Alignment of Agent Generation Schema: This embodiment provides a method for aligning the agent generation schema with the fact fields of the AI ​​twin information warehouse.

[0170] See the code implementation of the output schema requested by the agent. Figure 10 :

[0171] See the fact fields in the AI ​​twin information warehouse. Figure 11 .

[0172] The system performs field name similarity, field description similarity, historical mapping rules, and unit type matching to generate schema mappings. See details below. Figure 12 :

[0173] If a field cannot be mapped, for example, if the agent requests installation_service_window but the repository lacks the installation service time field, the system generates a gap completion task and marks the missing field in the context package.

[0174] IV. Generation of Fact Matching Constraint Set: The system generates a fact matching constraint set based on the agent's task situation representation.

[0175] A set of fact-matching constraints; see code implementation below. Figure 13 :

[0176] This set of constraints is not generated directly from the original human instructions, but is derived from the agent's task situation representation. Therefore, even if the original human instructions are not visible, the system can generate computable constraints based on the agent's behavioral signals.

[0177] V. Acquisition of factual units of goods or services: The system accesses the AI ​​twin information warehouse based on the fact matching constraint set.

[0178] The AI ​​twin information warehouse can be the warehouse described in the previous patent, or it can be a compatible warehouse that can provide semantic nodes, source anchors, relation edges, task intent indexes, version records, and permission information.

[0179] The system first searches for candidate product entities in the entity index based on the target product category.

[0180] The system then looks up the fact units related to price, inventory, delivery, warranty, invoice, evidence, and inquiry entry based on the task stage, schema mapping, and fact slot.

[0181] See the code implementation for a single commodity fact unit. Figure 14 .

[0182] If the system finds that the same SKU has a one-year warranty period on the product page, but a three-year warranty period in the after-sales policy document, the system can mark the relevant fact unit as conflicting and add conflict penalty or output conflict description in subsequent matching scores.

[0183] VI. Fact Relevance Matching Scoring: The system calculates fact relevance features for each candidate product or service. Field coverage represents the ratio of the number of found fact fields to the number of fact matching constraint fields.

[0184] The hard constraint satisfaction rate can be expressed as the ratio of the number of satisfied hard constraints to the total number of hard constraints.

[0185] The completeness of evidence can be expressed as the ratio of the number of key facts with locatable source anchors and valid evidence citations to the total number of key facts.

[0186] Task situation confidence level represents the reliability of the system's inferences about the current task objective, task phase, and schema mapping. See also Figure 15 In one implementation, a matching scoring formula is used, where W1 to W10 are weights. These weights can be preset by the system or adjusted based on category, task stage, agent identity, schema source, or feedback results.

[0187] If a candidate product triggers an exclusion constraint, such as if the product is a refurbished product and the exclusion condition in the task state requires that refurbished products are not accepted, the system can mark it as not recommended.

[0188] VII. The matching interpretation path generation system generates matching interpretation paths in the relational network of the AI ​​twin information warehouse.

[0189] Taking "warranty period of not less than three years" as an example, the matching interpretation path may include:

[0190] 1. Task Status Node: Solution Comparison Phase, field: warranty_years.

[0191] 2. Schema mapping edge: warranty_years->warranty_period.

[0192] 3. Fact matching constraint node: warranty__period>=3 years.

[0193] 4. Product factual details: SKU-AX27 warranty__period=3 years.

[0194] 5. Evidence: Page 2, paragraph 4 of the after-sales policy PDF.

[0195] 6. Conclusion node: This hard constraint is satisfied.

[0196] The system can calculate path length, path confidence, evidence anchor coverage, version freshness, conflict status, and task situation confidence for each path.

[0197] When multiple paths exist, the system prioritizes the path that meets the confidence threshold, has complete evidence, and is the shortest.

[0198] If the path is long, such as requiring a transition from the agent's schema field to the merchant field, then to the product series, then to the brand policy, and finally to the PDF attachments to reach a conclusion, the system can generate direct relationship edges, schema mapping rules, or procurement scenario summaries during the optimization phase to shorten the subsequent agent reading path.

[0199] 8. Machine-readable context packet generation: The system generates machine-readable context packets based on task status, matching score, and interpretation path.

[0200] The context packet can be in JSON, YAML, Markdown, JSON-LD, MCP resource response, HTTP API response, or other machine-readable format.

[0201] See the example context package implementation. Figure 16 .

[0202] When the context budget is small, the system prioritizes retaining relevant facts, hard constraints, limitations, evidence citations, and action entry points for the current task phase.

[0203] 9. Multiple candidate matching for uncertain situations: In one implementation, the system cannot determine the current task situation of the agent.

[0204] For example, the agent searches for "monitor invoice warranty" but doesn't specify the region, budget, or whether it's a business purchase. The system can generate multiple candidate task states:

[0205] 1. Comparative trends in enterprise procurement.

[0206] 2. Verification status of individual purchases.

[0207] 3. Status of after-sales policy inquiry.

[0208] The system derives fact matching constraint sets and generates different context packages.

[0209] If a subsequent agent accesses the inquiry form and requests the invoice field, the system increases the confidence level of the enterprise's purchasing comparison and decreases the confidence level of the individual's purchasing.

[0210] 10. Gap filling and counterfactual query generation: In another preferred embodiment, the system generates a gap filling task based on the matching failure log.

[0211] The system identifies gaps that cause matching failures, such as no delivery cycle, unclear after-sales restrictions, lack of qualification evidence, inability to map the agent schema field, unreachable interface entry, or unclear price validity period.

[0212] The system generates completion requests for the merchant backend, product database, customer service knowledge base, or management and review module.

[0213] After the completed content is written into the AI ​​twin information repository, the system performs source binding, conflict detection, and version recording.

[0214] The system can also generate counterfactual queries.

[0215] For example, the system generates a counterfactual task: "If the region in the agent's task situation changes from Shanghai to Hangzhou, will product A still meet the next-day delivery requirement?"

[0216] The system updates task labels, schema mapping rules, matching weights, and publication status based on the counterfactual task results.

[0217] XI. Intelligent Agent Simulation Evaluation and Feedback Closed Loop In another preferred embodiment, the system uses a preset intelligent agent task set to evaluate the information layer before and after optimization.

[0218] The preset intelligent agent task set may include MCP resource request task, web page search task, browser navigation task, product discovery task, condition filtering task, comparison task, evidence verification task, and action entry point identification task.

[0219] The system records the accuracy of task situation identification, schema mapping success rate, whether the target product was recalled, whether facts were correctly cited, whether the limiting conditions were identified, whether the action entry point was found, and whether incorrect matching occurred.

[0220] The system categorizes the failure reasons as: misjudgment of task status, schema mapping failure, missing fields, ambiguous expression, insufficient evidence, unreachable entry point, unresolved conflict, or outdated version.

[0221] The system updates the weights, task labels, field integrity requirements, schema mapping rules, summary templates, relation edges, and context package templates based on the failure reason.

[0222] This embodiment also provides an electronic device, which includes a processor, a memory, a communication interface, and a bus.

[0223] The memory contains computer programs.

[0224] When the processor executes the computer program, it implements the method described in any of the above embodiments.

[0225] This embodiment also provides a computer-readable storage medium.

[0226] The computer-readable storage medium stores a computer program.

[0227] When the computer program is executed by a processor, it implements the method described in any of the above embodiments.

[0228] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent agent task situation recognition and fact matching rate optimization based on AI twin information warehouse, characterized in that, Includes the following steps: S1: The system receives a set of task execution signals from the intelligent agent; S2: The system generates an agent task situation representation based on the set of agent task execution signals; S3: The system generates a set of fact-matching constraints based on the agent's task situation representation; S4: Based on the fact matching constraint set, the system retrieves fact units related to goods or services from the AI ​​twin information warehouse containing semantic nodes, source anchors, relation edges and version records; S5: The system calculates fact relevance features based on the fact matching constraint set and the fact unit; S6: The system generates a fact relevance matching score based on the aforementioned fact relevance features; S7: The system generates a matching interpretation path; S8: The system generates or adjusts a machine-readable context packet for the agent based on the fact relevance matching score and the matching interpretation path; S9: The system updates the twin layer tuning strategy based on the fact relevance matching score, the matching explanation path, or the agent feedback event.

2. The method for intelligent agent task situation recognition and fact matching rate optimization based on AI twin information warehouse as described in claim 1, characterized in that, S2 includes: S21: Deducing task stages and field requirements based on tool call trajectories; S22: Generate schema-aligned fact fields for goods or services based on the agent; S23: Identify task objectives and constraints based on webpage search trajectories and browser navigation trajectories; S24: Intent disambiguation based on task memory summary; S25: Generate multiple candidate agent task situation representations and their confidence levels when the signal is insufficient or ambiguous.

3. The method for intelligent agent task situation recognition and fact matching rate optimization based on AI twin information warehouse as described in claim 1, characterized in that, Step S3 includes: S31: Map the fields in the agent generation schema to the fact fields in the AI ​​twin information warehouse; S32: Generate synonym candidates, gap completion tasks, or uncertainties for schema fields that fail to be mapped; S33: Map task phases to fact slot priorities and context package templates.

4. The method for intelligent agent task situation recognition and fact matching rate optimization based on AI twin information warehouse as described in claim 1, characterized in that, Step S9 includes: identifying the reasons for matching failure, wherein the reasons for matching failure include at least one of the following: uncertain task status, schema field cannot be mapped, field is missing, insufficient evidence, ambiguous expression, unresolved conflict, version expired, inference path is too long, and action entry point is unreachable.

5. The method for intelligent agent task situation recognition and fact matching rate optimization based on AI twin information warehouse as described in claim 4, characterized in that, In step S9: When the task situation is identified as uncertain, multiple candidate task situations are generated or a clarifying context is returned to the agent. When a schema field is identified as unmappable, a schema mapping rule or field completion task is generated. When the evidence is deemed insufficient, the original webpage fragment, document page number, table cell, image area, or database primary key is bound; When the reasoning path is identified as too long, add direct relationship edges, task labels, or compress the explanation path.

6. The method for intelligent agent task situation recognition and fact matching rate optimization based on AI twin information warehouse as described in claim 1, characterized in that, Step S9 also includes: gap filling and counterfactual inquiry: Generate completion tasks based on failure reasons such as missing fields, insufficient evidence, inability to map schema fields, or irrelevant matching. Generate a counterfactual query task based on the agent's task situation representation or fact matching constraint set; The counterfactual query task is used to verify the matching status of goods or services under changing constraints, and the task tags, matching weights, or publication status are updated based on the verification results.

7. The method for intelligent agent task situation recognition and fact matching rate optimization based on AI twin information warehouse as described in claim 6, characterized in that, Step S9 also includes: agent simulation evaluation step: using a preset agent task set to request machine-readable context packets before and after optimization, and generating evaluation results based on at least one of the following: task situation recognition accuracy, schema mapping success rate, matching task answerability rate, required field coverage, evidence backlink rate, average inference path length, irrelevant recall rate, correct citation rate, constraint recognition rate, and action entry recognition rate.

8. A system for intelligent agent task situation recognition and fact matching rate optimization based on an AI twin information warehouse, characterized in that, The system is used to perform the method of claim 1, comprising: The task signal receiving module is used to receive the set of task execution signals of the intelligent agent; The task situation recognition module is used to generate a task situation representation of the agent based on the set of task execution signals of the agent; The schema alignment module is used to map fields in the schema generated by the agent to fact fields in the AI ​​twin information warehouse; The fact matching constraint generation module is used to generate a set of fact matching constraints based on the agent's task situation representation; The fact unit acquisition module is used to acquire fact units related to goods or services from the AI ​​twin information warehouse based on the fact matching constraint set; The matching and scoring module is used to calculate fact relevance features based on the fact matching constraint set and the fact unit, and generate a fact relevance matching score; The interpretation path generation module is used to generate a matching interpretation path that includes task status nodes or fact matching constraint nodes, product or service fact nodes, evidence nodes, relationship edges, and matching conclusion nodes. A context packet generation module is used to generate or adjust a machine-readable context packet for an agent based on the fact relevance matching score and the matching interpretation path. The twin layer optimization module is used to update the twin layer optimization strategy based on the fact relevance matching score, the matching explanation path, or the agent feedback event.

9. An electronic device comprising a processor and a memory, characterized in that, The memory stores a computer program that, when executed by the processor, implements the method described in claim 1.

10. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the method described in claim 1.

Images