Smart park multi-round intelligent question and answer system based on large language model

Abstract

The invention relates to the technical field of computers, and discloses a smart park multi-round intelligent question answering system based on a large language model. Comprising a display terminal and a platform server. The platform server comprises a session state management module, an authority projection module, a park data access module, an evidence construction module, a tool arrangement module, a large language model generation module and a credibility verification module. The session state management module generates or updates a session state structure. And the permission projection module forms a retrieval filtering condition and a tool calling white list. The evidence construction module retrieves the static knowledge data source of the park and queries the dynamic business data source of the park to generate an evidence packet. And the tool arrangement module executes the tool calling plan to obtain an execution result. A large language model generation module generates candidate replies carrying evidence references. And the credibility verification module outputs a clear question or a degraded reply when the verification is not passed, and outputs a final reply when the verification is passed.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically to a multi-turn intelligent question-answering system for smart parks based on a large language model. Background Technology

[0002] Smart parks typically integrate systems such as access control, parking, energy consumption, work orders, visitor and meeting resources. Daily operation and maintenance require providing services such as policy inquiry, event tracing, equipment maintenance and process handling to both resident companies and property staff. The interaction interface is gradually shifting from web pages and hotlines to mobile conversational services.

[0003] Existing intelligent customer service systems in industrial parks mostly use FAQ matching, keyword retrieval, or rule engines to achieve question and answer. Some solutions also introduce vector retrieval and large language models, using policy documents, equipment manuals, and announcements as context to generate responses, and return data such as access control records, parking space availability, or work order progress through interface queries. In multi-turn dialogues, the context is usually maintained by splicing historical messages or simple conversation memory.

[0004] In a multi-tenant and hierarchical permission environment, existing solutions struggle to apply consistent permission constraints to static knowledge retrieval, real-time business queries, and model generation outputs within the same session, and to form a traceable chain of evidence. This leads to easy deviations in object references and time ranges in multi-turn dialogues, and the generated content and business results lack verifiable references, making it difficult to simultaneously meet the integrated requirements of security compliance and result credibility in park scenarios. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a multi-turn intelligent question-answering system for smart parks based on a large language model, in order to solve the technical problems existing in the prior art.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution: A multi-turn intelligent question-answering system for smart parks based on a large language model includes: It includes a display terminal and a platform server, with the display terminal communicating with the platform server; the platform server includes a session state management module, a permission projection module, a park data access module, an evidence construction module, a tool orchestration module, a large language model generation module, and a credibility verification module; The session state management module is used to assign session identifiers to sessions and generate or update the dialogue state structure based on the current round of messages and the historical round of messages. The dialogue state structure includes tenant identifier, user role identifier, spatiotemporal anchor information, park object identifier set, task stage identifier, and slot parameter set. The permission projection module is used to generate a permission projection set based on the tenant identifier and user role identifier, and convert the permission projection set into search filtering conditions and tool call whitelist. The park data access module is used to access the park's static knowledge data source and the park's dynamic business data source. The park's dynamic business data source can be one or more of the following: access control system, parking system, energy consumption system, and work order system. The evidence construction module is used to perform keyword retrieval and vector retrieval on the static knowledge data source of the park and perform structured query on the dynamic business data source of the park under the constraints of retrieval filtering conditions, so as to obtain a set of candidate evidence and generate an evidence package. The evidence package includes multiple evidence items, and each evidence item includes evidence content, source identifier, timestamp and evidence citation identifier. The tool orchestration module is used to determine the intent type of the question-and-answer request based on the dialogue state structure, and to generate and execute a tool invocation plan under the constraint of the tool invocation whitelist to obtain the execution result; The large language model generation module is used to input the current round message, the summary information of the dialogue state structure, the evidence package and the execution result into the large language model to generate candidate responses with labeled evidence citations; The credibility verification module is used to evaluate the evidence coverage and verify the consistency of candidate responses. The consistency verification includes matching the key entities, key times and key values ​​in the candidate responses with the evidence package and execution results. If the coverage assessment and consistency check fail, a clarification question or a downgraded response based on the evidence package will be output. If they pass, the final response will be output to the display terminal.

[0007] Preferably, the session state management module includes a slot parsing unit, a digest generation unit, and a state storage unit; The slot parsing unit extracts slot parameters based on the current round message and writes them into the slot parameter set; The summary generation unit generates a session summary for the historical round messages; The state storage unit stores the dialogue state structure using the session identifier as an index, and persists the updated dialogue state structure after each round of question and answer. The session summary is written into the dialogue state structure or stored in association with the session identifier, and is called as the summary information of the dialogue state structure.

[0008] Preferably, the park data access module includes an object standardization unit and an object mapping table. The object standardization unit standardizes the park object names in the current round of messages and the historical round of messages into park object identifiers, and writes the park object identifiers into a park object identifier set. The spatiotemporal anchor information includes the park location identifier, start time, and end time, wherein the park location identifier is determined by the object mapping table.

[0009] Preferably, the permission projection module includes a permission rule table and a projection generation unit; The permission rule table records the correspondence between user role identifiers and data source identifiers, the set of accessible fields, and the set of accessible conditions; The projection generation unit generates a permission projection set based on the permission rule table, and forms a search filtering condition and a tool call whitelist accordingly. The search filtering condition includes tenant identifier filtering condition and field clipping rules.

[0010] Preferably, the park data access module includes a static knowledge connector and a dynamic business connector; The static knowledge connector is connected to the regulations and rules document library and the equipment operation and maintenance knowledge base; The dynamic business connector connects to one or more of the access control system, parking system, energy consumption system, and work order system, and converts the records returned by the dynamic business data source into unified event records containing park object identifiers, event type identifiers, event occurrence times, and record primary keys.

[0011] Preferably, the evidence construction module includes a keyword retrieval unit, a vector retrieval unit, a structured query unit, and an evidence merging unit; The keyword retrieval unit retrieves candidate evidence from the park's static knowledge data source based on an inverted index; The vector retrieval unit retrieves candidate evidence from the static knowledge data source of the park based on the vector index; The structured query unit performs structured queries on the dynamic business data source of the park based on the park object identifier set and spatiotemporal anchor information; The evidence merging unit deduplicates and sorts the candidate evidence to form a candidate evidence set.

[0012] Preferably, the evidence package consists of multiple evidence items, each of which includes, in addition to the evidence content, source identifier and timestamp, source type identifier and location identifier; The source type identifier is used to distinguish whether the evidence entry originates from the park's static knowledge data source or the park's dynamic business data source. The location identifier is either a document paragraph identifier or a record primary key; the evidence citation identifier corresponds one-to-one with the location identifier.

[0013] Preferably, the tool orchestration module includes an intent determination unit, a parameter completion unit, and an execution unit; The intent determination unit limits the intent type to either query intent or processing intent. When the intent type is a processing intent, the parameter completion unit completes the parameter set based on the slot parameter set in the dialogue state structure and the evidence package parameter set. The execution unit executes the tool calls in the order of the tool call plan and generates an execution result containing the identifier of the called tool, the parameter set, and the receipt identifier.

[0014] Preferably, the tool orchestration module further includes an idempotency identifier generation unit and a rollback unit; The idempotency flag generation unit generates an idempotency flag based on the session flag, the called tool flag, and the parameter set, and writes the idempotency flag into the execution result. The rollback unit stores a rollback operation table, which records the correspondence between the called tool identifier, the rollback tool identifier, and the rollback parameter rules. When the tool call plan fails to execute, the rollback unit executes the rolled-back tool call sequentially according to the reverse order of the executed steps in the rollback operation table, and writes the rollback result into the execution result.

[0015] Preferably, the platform server further includes an audit log module, which records the tenant identifier, user role identifier, search filter conditions, tool call plan, source identifier and timestamp in the evidence package, evidence citation identifier, and receipt identifier in the execution result corresponding to the session identifier, and writes the recorded content to the audit log storage area.

[0016] In summary, the present invention has the following main beneficial effects: This invention uses a session state management module to solidify tenant identifiers, user role identifiers, spatiotemporal anchor information, park object identifier sets, task stage identifiers, and slot parameter sets from multi-turn dialogues into a dialogue state structure. The summary information of the dialogue state structure is used to drive subsequent retrieval and parameter completion, thereby achieving the goal of consistent constraints on object reference, time range, and processing progress during multi-turn question and answer processes. This reduces the interaction costs caused by cross-turn semantic drift and repeated clarification, and enables query-type intents and processing-type intents to advance continuously in the same session state.

[0017] This invention generates a permission projection set based on tenant identifiers and user role identifiers through a permission projection module, and converts the permission projection set into search filtering conditions and a tool invocation whitelist. This ensures that keyword retrieval, vector retrieval, and structured querying in the evidence construction module are all executed under unified permission constraints. At the same time, it ensures that the tool orchestration module can only generate and execute tool invocation plans within the whitelist. This achieves the goal of simultaneously implementing tenant isolation and field pruning in the two links of evidence acquisition and tool execution, thereby reducing the risks of unauthorized retrieval, unauthorized citation, and unauthorized invocation, and improving the controllability and consistency of data access in multi-tenant scenarios in the park.

[0018] This invention generates an evidence package containing evidence content, source identifier, timestamp, and evidence citation identifier through an evidence construction module, and evaluates the evidence coverage and consistency of candidate responses through a credibility verification module. This achieves the purpose of matching and verifying key entities, key times, and key values ​​in the response with the evidence package and execution results. In case of insufficient or inconsistent evidence, it outputs clarifying questions or downgraded responses based on the evidence package, and outputs the final response when the verification is successful. This improves the traceability and consistency of the output results and reduces the probability of content that is not supported by evidence entering the final response. Attached Figure Description

[0019] Figure 1 This is a system flowchart of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1 refer to Figure 1 A multi-turn intelligent question-answering system for smart parks based on a large language model includes: It includes a display terminal and a platform server, with the display terminal communicating with the platform server; the platform server includes a session state management module, a permission projection module, a park data access module, an evidence construction module, a tool orchestration module, a large language model generation module, and a credibility verification module; The session state management module is used to assign session identifiers to sessions and generate or update the dialogue state structure based on the current round of messages and the historical round of messages. The dialogue state structure includes tenant identifier, user role identifier, spatiotemporal anchor information, park object identifier set, task stage identifier, and slot parameter set. The permission projection module is used to generate a permission projection set based on the tenant identifier and user role identifier, and convert the permission projection set into search filtering conditions and tool call whitelist. The park data access module is used to access the park's static knowledge data source and the park's dynamic business data source. The park's dynamic business data source can be one or more of the following: access control system, parking system, energy consumption system, and work order system. The evidence construction module is used to perform keyword retrieval and vector retrieval on the static knowledge data source of the park and perform structured query on the dynamic business data source of the park under the constraints of retrieval filtering conditions, so as to obtain a set of candidate evidence and generate an evidence package. The evidence package includes multiple evidence items, and each evidence item includes evidence content, source identifier, timestamp and evidence citation identifier. The tool orchestration module is used to determine the intent type of the question-and-answer request based on the dialogue state structure, and to generate and execute a tool invocation plan under the constraint of the tool invocation whitelist to obtain the execution result; The large language model generation module is used to input the current round message, the summary information of the dialogue state structure, the evidence package and the execution result into the large language model to generate candidate responses with labeled evidence citations; The credibility verification module is used to evaluate the evidence coverage and verify the consistency of candidate responses. The consistency verification includes matching the key entities, key times and key values ​​in the candidate responses with the evidence package and execution results. If the coverage assessment and consistency check fail, a clarification question or a downgraded response based on the evidence package will be output. If they pass, the final response will be output to the display terminal.

[0022] The terminal demonstrates a communication connection with the platform server, used to collect the user's current round of messages and send question-and-answer requests to the platform server, as well as receive and output the final response returned by the platform server.

[0023] The platform server includes a session state management module, a permission projection module, a campus data access module, an evidence construction module, a tool orchestration module, a large language model generation module, and a credibility verification module. The input and output interfaces of each module are associated under the same session identifier to ensure consistency, traceability, and executability of multi-turn question-and-answer sessions.

[0024] In this embodiment, the static knowledge data sources for the park include at least a regulations and rules document library and an equipment operation and maintenance knowledge base, while the dynamic business data sources for the park are one or more of the access control system, parking system, energy consumption system, and work order system. The platform server completes the connection to the above data sources and data read / write permission control through the park data access module, and conducts searches and tool calls under the constraints of the search filtering conditions and tool call whitelist generated by the permission projection module.

[0025] In this application, the dialogue state structure is denoted as... The permission projection set is denoted as The evidence package is recorded as follows: The tool invocation plan is denoted as Plan, and the execution result is denoted as Candidate responses are denoted as .

[0026] The session state management module is used to assign session identifiers to sessions and generate or update the dialogue state structure based on the current round of messages and previous rounds of messages. The dialogue state structure includes tenant identifier, user role identifier, spatiotemporal anchor information, a set of park object identifiers, task stage identifiers, and a set of slot parameters.

[0027] After receiving a question-and-answer request from the display terminal, the session state management module generates a new session identifier if the request does not carry a session identifier; if it does carry a session identifier, it indexes the corresponding historical round messages and dialogue state structure using that session identifier.

[0028] The session state management module performs slot parsing on the current round of messages. The slot parameter set is stored in key-value pairs and includes at least one or more of the following: business domain identifier, target object, time range, location range, action type, and constraints. For example, if a user enters "Check the energy consumption of the third floor of Building A this morning", the slot parameter set will at least include the object name "third floor of Building A", start time, end time, and business domain "energy consumption".

[0029] The spatiotemporal anchor information includes the park location identifier, start time, and end time. The park object identifier set is used to store park object identifiers that have been confirmed in the session, for subsequent structured queries and tool parameter completion.

[0030] The session state management module generates session summaries from historical rounds of messages. The session summaries are written to the dialogue state structure or stored in association with the session identifier, and are called as summary information of the dialogue state structure for subsequent input into the large language model generation module, avoiding the occurrence of contextual noise and referential bias caused by directly inputting all historical text.

[0031] For ease of formal description, the dialogue state structure in this embodiment is defined as follows: ; in, Indicates tenant identifier, Indicates user role identifier, Represents spatiotemporal anchor point information. Represents the set of object identifiers in the park. Indicates the task phase identifier. Represents the set of slot parameters. This indicates a summary of the session information.

[0032] The permission projection module is used to generate permission projection sets based on tenant identifiers and user role identifiers, and convert the permission projection sets into search filtering conditions and tool invocation whitelists.

[0033] The permission rules table records the mapping between user role identifiers and data source identifiers, the set of accessible fields, and the set of accessible conditions. The set of accessible conditions must include at least tenant isolation conditions and business domain conditions. The set of accessible fields must include at least the whitelist of fields required by the field pruning rules.

[0034] Permission projection set Generated by the permission projection module based on the permission rule table. The permission projection module will... This is converted into search filtering conditions, which include tenant identifier filtering conditions and field pruning rules. Tenant identifier filtering conditions limit the search scope to static knowledge documents and dynamic business records within the current tenant. Field pruning rules are used to prune fields not included in the accessible field set before returning evidence and execution results, ensuring the quality of subsequent evidence packages. With execution results It does not contain fields that exceed privileges.

[0035] The permission projection module will This is converted into a tool invocation whitelist, which includes a set of allowed tool identifiers and a set of allowed parameter fields corresponding to those tool identifiers. The tool orchestration module can only generate and execute tool invocation plans under the constraints of the tool invocation whitelist, thereby restricting cross-domain and unauthorized invocations at the tool level.

[0036] The above structure is used to apply permission constraints simultaneously during the retrieval and tool invocation phases, avoiding the risk of unauthorized citations and invocations caused by filtering only after the response is generated.

[0037] The park data access module is used to access the park's static knowledge data source and dynamic business data source, and includes object standardization units, object mapping tables, static knowledge connectors and dynamic business connectors.

[0038] The object standardization unit standardizes the park object names in the current round of messages and historical rounds into park object identifiers, and writes the park object identifiers into the park object identifier set. The object mapping table records the mapping relationship between park object names, park object identifiers, and park location identifiers. The park location identifier in the spatiotemporal anchor information is determined by the object mapping table.

[0039] The static knowledge connector links with the regulations and rules document library and the equipment operation and maintenance knowledge base, and is responsible for incremental synchronization of documents, version identification maintenance, and index construction. When a document is added to the library, a paragraph identifier is assigned to each document paragraph for subsequent location identification of evidence items.

[0040] The dynamic business connector connects to one or more of the access control system, parking system, energy consumption system, and work order system. It converts records returned by the dynamic business data source into unified event records. These unified event records include at least a park object identifier, event type identifier, event occurrence time, and record primary key, used for structured querying and evidence item location in the evidence construction module.

[0041] The evidence construction module is used to perform keyword and vector searches on static knowledge data sources within the park, and to perform structured queries on dynamic business data sources within the park, under the constraints of search and filtering conditions, to obtain a set of candidate evidence and generate an evidence package. .

[0042] The keyword retrieval unit searches for candidate evidence in the park's static knowledge data source based on the inverted index. During the retrieval, the tenant identifier filter condition and the field range corresponding to the field pruning rule are applied in the retrieval filter conditions.

[0043] The vector retrieval unit encodes the current round message and session summary information into a query vector, and calculates its similarity with the document paragraph vector to retrieve candidate evidence. The similarity is defined using cosine similarity. ; in, This indicates that the query text is processed by a vector encoder. The output vector, This indicates that the candidate paragraph text is processed by the vector encoder. The output vector, Represents the vector dot product. The L2 norm of a vector.

[0044] The structured query unit performs structured queries on the dynamic business data source of the park based on the park object identifier set and spatiotemporal anchor information. The structured query must include at least the following constraints: tenant identifier, park object identifier, start time, and end time. The results of the structured query, after being processed by field pruning rules, are input into the evidence merging unit as part of the candidate evidence.

[0045] The evidence merging unit deduplicates and sorts the candidate evidence from keyword retrieval, vector retrieval, and structured queries to form a candidate evidence set. In this embodiment, the sorting rule uses a weighted fusion score. ; in, This represents the fusion score of candidate evidence i. Represents vector similarity, This indicates the keyword search score. This indicates the score for time freshness. , , For preset weight parameters, and .

[0046] Evidence package It includes multiple evidence entries, each containing evidence content, a source identifier, a timestamp, and an evidence citation identifier. Furthermore, each evidence entry also includes a source type identifier and a location identifier. The source type identifier distinguishes whether the evidence entry originates from a static knowledge data source or a dynamic business data source within the park. The location identifier is either a document paragraph identifier or a record primary key. The evidence citation identifier corresponds one-to-one with the location identifier.

[0047] For ease of formal description, the evidence items are defined as follows: ; in, Indicate the content of the evidence, Indicates the source identifier. Represents a timestamp. Indicates evidence citation identifier, Indicates the source type identifier. This indicates a location identifier. and It is optional to output, but There must be supporting evidence citations in the candidate response.

[0048] The tool orchestration module is used to determine the intent type of the question-and-answer request based on the dialogue state structure, and to generate and execute the tool invocation plan under the constraints of the tool invocation whitelist to obtain the execution result.

[0049] The intent determination unit limits the intent type to either query intent or processing intent. Query intents correspond to scenarios where a response is generated primarily from evidence packages, while processing intents correspond to scenarios where the park's business system needs to be invoked to perform operations and return a receipt.

[0050] When the intent type is a processing intent, the parameter completion unit completes the parameter set based on the slot parameter set in the dialogue state structure and the evidence package parameter set. If the slot parameter set is missing the required parameters for processing, the tool orchestration module writes the missing item into the item to be completed corresponding to the task stage identifier, and then passes it to the credibility verification module to trigger the output of the clarification question.

[0051] The execution unit executes tool calls sequentially according to the tool call plan. The execution result includes at least the identifier of the called tool, the parameter set, and the receipt identifier. The execution result is written back to the dialog state structure, allowing subsequent rounds to continue following up or querying the processing progress based on the receipt identifier.

[0052] The tool orchestration module also includes an idempotent flag generation unit and a rollback unit to ensure the controlled execution of processing intentions in the event of network retries, repeated submissions, or intermediate failures.

[0053] The idempotency flag generation unit generates an idempotency flag based on the session flag, the called tool flag, and the parameter set, and writes the idempotency flag into the execution result. To avoid different idempotency flags caused by the order of fields in the parameter set, this embodiment first performs normalized encoding on the parameter set. Then perform hash calculation: ; in, Indicates idempotency. Represents a hash function. Indicates the session identifier. Indicates the identifier of the invoked tool. This represents a string concatenation operation. This represents the normalized string obtained by encoding the parameter set after sorting it by field name.

[0054] The rollback unit stores a rollback operation table, which records the correspondence between the called tool identifier, the rollback tool identifier, and the rollback parameter rules. When a tool call plan fails, the rollback unit executes the rollback tool calls in reverse order of the executed steps according to the rollback operation table, and writes the rollback results into the execution results. This ensures that even if a process fails after some steps have succeeded, the process can still be restored to a consistent state.

[0055] The large language model generation module is used to input the current round message, the summary information of the dialogue state structure, the evidence package and the execution result into the large language model to generate candidate responses with labeled evidence citations.

[0056] The input must include at least: the current round message text, session summary information, a list of evidence items and their citation identifiers, and the receipt identifier and key fields in the execution result. The correspondence between evidence items and citation identifiers must remain unchanged before being submitted to the large language model to ensure that the citation annotations in the candidate responses can be retrieved by the credibility verification module.

[0057] Candidate responses must include at least natural language content and evidence citation markers. Evidence citation markers are embedded in the corresponding segment of the response in the form of evidence citation icons, used to indicate the supporting evidence items for that segment.

[0058] The credibility verification module is used to evaluate the evidence coverage and verify the consistency of candidate responses. The consistency verification includes matching the key entities, key times and key values ​​in the candidate responses with the evidence package and execution results.

[0059] The credibility verification module extracts a set of key entities from candidate responses. Key Time Sets Key numerical sets The extraction rules employ a combination of a pre-defined dictionary and a sequence labeling model. The extraction results are checked for consistency with the set of park object identifiers in the dialogue state structure to prevent objects that have not been confirmed in the session from being mistakenly introduced.

[0060] Evidence coverage is used to determine whether key elements in a candidate response are covered by the evidence package or execution results. Entity coverage, time coverage, and numerical coverage are defined separately:

[0061]

[0062]

[0063] in, , , They represent the evidence packages respectively. With execution results The supported entity set, supported time set, and supported numerical set are obtained from the analysis. The overall coverage is defined as:

[0064] in, , , These are preset weight parameters.

[0065] Consistency verification is implemented using a location-based matching method. Each piece of evidence cited in the candidate responses is annotated. The credibility verification module locates the corresponding evidence item in the evidence package. Then verify whether the key entities, key times, and key values ​​in the candidate response fragment are all present. Evidence content or execution result The match was found in the fields specified. The matching rules include exact matching and equivalent mapping matching. Equivalent mapping matching is done through an object mapping table, such as the mapping between building aliases and park object identifiers.

[0066] The credibility verification module has a preset coverage threshold. With consistency threshold .like If the consistency check fails, the application is deemed unsuccessful, and a clarifying question or a downgraded response based on the evidence package is output. The clarifying question is used to fill in missing slots for processing intents or missing constraints for query intents. The downgraded response is generated only based on evidence entries that have passed the location matching and retains the evidence citation annotations. If the coverage assessment and consistency check pass, the final response is output to the display terminal.

[0067] The platform server also includes an audit logging module. The audit logging module records the tenant identifier corresponding to the session identifier, user role identifier, search filter conditions, tool call plan, source identifier and timestamp in the evidence package, evidence citation identifier, and receipt identifier in the execution result, and writes the recorded content to the audit log storage area.

[0068] The audit log storage area uses the session identifier as the primary key and the round number as the sorting key, persistently storing evidence package summaries, tool call plan summaries, and execution result summaries by round. The audit recording module and the permission projection module share field pruning rules, pruning fields not in the accessible field set before writing to the audit log to ensure that the audit log itself does not create unauthorized data copies.

[0069] The terminal displays the current round messages of the user in a multi-round dialogue and sends them to the platform server with the session identifier. The session state management module reads the historical round messages with the session identifier as the index and generates or updates the dialogue state structure. The tenant identifier, user role identifier, spatiotemporal anchor information, park object identifier set, task stage identifier and slot parameter set are uniformly solidified into a computable session state. The permission projection module generates a permission projection set based on tenant identifiers and user role identifiers, and converts it into search filtering conditions and tool invocation whitelists, so that evidence retrieval and tool invocation are executed under the same permission constraints. The park data access module connects to the park's static knowledge data source and the park's dynamic business data source respectively, and completes the standardized mapping from park object names to park object identifiers; The evidence construction module performs keyword and vector searches on static knowledge data sources under the constraints of search and filtering conditions, and performs structured queries on dynamic business data sources according to the park object identifier set and spatiotemporal anchor information to form an evidence package containing evidence content, source identifier, timestamp, and evidence citation identifier; the tool orchestration module determines the intent type based on the dialogue state structure and generates a tool invocation plan under the constraints of the tool invocation whitelist. After supplementing the parameters for the processing intent, it executes the tool invocation to obtain the execution result containing the receipt identifier; The large language model generation module inputs the current round message, the summary information of the dialogue state structure, the evidence package and the execution result into the large language model to generate candidate responses and marks the evidence citations in the responses; The credibility verification module evaluates the evidence coverage and consistency of candidate responses, checking whether the key entities, key times, and key values ​​in the candidate responses match the evidence package and execution results. If they fail, it outputs a clarifying question or a downgraded response based on the evidence package to fill in missing constraints or avoid outputting content that is not supported by evidence. If they pass, it outputs the final response to the display terminal. In this way, a secure question-and-answer and processing loop for smart parks is realized through a multi-round dialogue driven by session state, with permission constraints throughout, traceable evidence, executable tools, and verifiable output.

[0070] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart park multi-round intelligent question and answer system based on a large language model, comprising a display terminal and a platform server, the display terminal being in communication connection with the platform server; characterized in that, The platform server comprises a session state management module, an authority projection module, a park data access module, an evidence construction module, a tool arrangement module, a large language model generation module and a credibility verification module; The session state management module is configured to assign a session identification to a session and generate or update a dialogue state structure according to a current round message and a historical round message, the dialogue state structure comprising a tenant identification, a user role identification, space-time anchor information, a park object identification set, a task phase identification and a slot parameter set; The authority projection module is configured to generate an authority projection set based on the tenant identification and the user role identification, and convert the authority projection set into a retrieval filtering condition and a tool calling white list; The park data access module is configured to access a park static knowledge data source and a park dynamic business data source, the park dynamic business data source being one or more of an access control system, a parking system, an energy consumption system and a work order system; The evidence construction module is configured to perform keyword retrieval and vector retrieval on the park static knowledge data source and structured query on the park dynamic business data source under the constraint of the retrieval filtering condition, to obtain a candidate evidence set and generate an evidence package, the evidence package comprising a plurality of evidence entries, each evidence entry comprising evidence content, a source identification, a timestamp and an evidence reference identification; The tool arrangement module is configured to determine an intention type of a question and answer request based on the dialogue state structure, and generate and execute a tool calling plan under the constraint of the tool calling white list to obtain an execution result; The large language model generation module is configured to input the current round message, abstract information of the dialogue state structure, the evidence package and the execution result into a large language model to generate a candidate reply with a labeled evidence reference; The credibility verification module is configured to perform evidence coverage evaluation and consistency verification on the candidate reply, the consistency verification comprising matching key entities, key times and key values in the candidate reply with the evidence package and the execution result; When the coverage evaluation and the consistency verification fail, a clarification question or a degraded reply based on the evidence package is output, and when the coverage evaluation and the consistency verification pass, a final reply is output to a display terminal.

2. The smart park multi-round intelligent question and answer system based on a large language model according to claim 1, characterized in that, The session state management module comprises a slot analysis unit, an abstract generation unit and a state storage unit; The slot analysis unit extracts slot parameters from the current round message and writes the slot parameters into the slot parameter set; The abstract generation unit generates a session abstract from the historical round message; The state storage unit stores the dialogue state structure with the session identification as an index, and persistently saves the updated dialogue state structure after each round of question and answer is completed; The session abstract is written into the dialogue state structure or stored in association with the session identification, and is called as abstract information of the dialogue state structure.

3. The smart park multi-round intelligent question and answer system based on a large language model according to claim 2, characterized in that, The park data access module comprises an object standardization unit and an object mapping table, the object standardization unit standardizing park object names in the current round message and the historical round message into park object identifications, and writing the park object identifications into the park object identification set; The space-time anchor information comprises a park location identification, a start time and an end time, wherein the park location identification is determined by the object mapping table.

4. The smart park multi-round intelligent question and answer system based on a large language model according to claim 3, characterized in that, The authority projection module comprises an authority rule table and a projection generation unit; The permission rule table records the correspondence between the user role identifier, the data source identifier, the accessible field set, and the accessible condition set; The projection generation unit generates a permission projection set based on the permission rule table, and forms a retrieval filtering condition and a tool call whitelist based thereon, the retrieval filtering condition including a tenant identifier filtering condition and a field pruning rule.

5. The smart park multi-round intelligent question and answer system based on a large language model according to claim 4, characterized in that, The park data access module includes a static knowledge connector and a dynamic business connector; The static knowledge connector is connected with a regulation and system document library and a device operation and maintenance knowledge base; The dynamic business connector is connected with one or more of an access control system, a parking system, an energy consumption system, and a work order system, and converts records returned by a dynamic business data source into uniform event records containing a park object identifier, an event type identifier, an event occurrence time, and a record primary key.

6. The smart park multi-round intelligent question and answer system based on a large language model according to claim 5, characterized in that, The evidence construction module includes a keyword search unit, a vector search unit, a structured query unit, and an evidence merging unit; The keyword search unit searches for candidate evidence in park static knowledge data sources based on an inverted index; The vector search unit searches for candidate evidence in park static knowledge data sources based on a vector index; The structured query unit performs a structured query on park dynamic business data sources based on a set of park object identifiers and spatio-temporal anchor point information; The evidence merging unit forms a candidate evidence set after deduplication and sorting of the above candidate evidence.

7. The smart park multi-round intelligent question and answer system based on a large language model according to claim 6, characterized in that, The evidence package is composed of multiple evidence items, each of which includes, in addition to evidence content, a source identifier and a timestamp, a source type identifier and a positioning identifier; The source type identifier is used to distinguish whether the evidence item is derived from a park static knowledge data source or a park dynamic business data source; The positioning identifier is a document paragraph identifier or a record primary key; an evidence reference identifier and the positioning identifier correspond to each other.

8. The smart park multi-round intelligent question and answer system based on a large language model according to claim 7, characterized in that, The tool orchestration module includes an intent determination unit, a parameter completion unit, and an execution unit; The intent determination unit limits the intent type to a query intent or a handling intent; When the intent type is a handling intent, the parameter completion unit completes the parameter set based on the slot parameter set in the dialogue state structure body and the parameter set of the evidence package; The execution unit executes the tool call according to the step order of the tool call plan, and generates an execution result containing the called tool identifier, the parameter set, and the receipt identifier.

9. The smart park multi-round intelligent question and answer system based on a large language model according to claim 8, characterized in that, The tool orchestration module further includes an idempotent identifier generation unit and a rollback unit; The idempotent identifier generation unit generates an idempotent identifier based on the session identifier, the called tool identifier, and the parameter set, and writes the idempotent identifier into the execution result; The rollback unit stores a rollback action table, which records the correspondence between the called tool identifier and the rollback tool identifier and the rollback parameter rule; When the tool call plan fails to execute, the rollback unit executes the rollback tool call in the order opposite to the executed steps according to the rollback action table, and writes the rollback result into the execution result.

10. The smart park multi-round intelligent question and answer system based on a large language model of claim 9, wherein, The platform server further comprises an audit record module, which records the tenant identification corresponding to the session identification, the user role identification, the search filter condition, the tool call plan, the source identification and the time stamp in the evidence package, the evidence reference identification and the receipt identification in the execution result, and writes the record content into an audit log storage area.

Images