Cloud computing-based self-developed AI research and development platform collaboration system

By introducing modules such as requirement data management, AI context construction, delivery scenario analysis, and version management into the cloud-based R&D platform, the problems of inconsistent generated results and obstructed collaborative processes in cloud-based R&D collaboration have been solved, achieving consistency of generated results and improved collaborative efficiency.

CN123067916APending Publication Date: 2026-11-10SHENZHEN DINGFENG TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611211838.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-11
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In cloud-based collaborative R&D scenarios, existing technologies suffer from issues such as repeated generation of multiple types of deliverables, unstable output parsing, and a disconnect between version updates and the current effective status. Furthermore, they lack a unified record of the invocation process, runtime status, and delivery chain, leading to obstructed collaborative processes and high system maintenance complexity.

Method used

The system uses a demand data management module to verify demand objects and access permissions, an AI context construction module to assemble the call context, an AI delivery scenario analysis module to normalize and deduplicate, a multi-output orchestration and parsing module to merge requests, a runtime record and persistence module to manage version numbers, a version and status closed-loop module to maintain the effective status, an AI call audit module to record the call chain, and a delivery center module to track the pipeline status, ensuring the consistency of generated results and collaborative efficiency.

Benefits of technology

It achieves consistency in generated results under the same requirements, avoids issues of repeated generation and unstable parsing, improves collaboration efficiency, and forms a closed loop between permission verification, exception rollback and delivery tracking, simplifying system maintenance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123067916A_ABST
    Figure CN123067916A_ABST
Patent Text Reader

Abstract

The application relates to the technical field of cloud computing and artificial intelligence research and development management, in particular to a self-research AI research and development platform collaboration system based on cloud computing, which comprises demand data management, demand AI context construction, AI delivery scene analysis, multi-output arrangement and analysis, large model interaction, operation record and persistence, version and state closed loop, AI calling audit and delivery center modules; is used for checking a demand object, normalizing a delivery object type, arranging and combining and requesting, response distribution analysis, version saving and state updating, and auditing and tracing a calling process and delivery tracking; so as to reduce the risk of analysis failure and error content storage in a database.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of cloud computing and artificial intelligence R&D management technology, specifically a collaborative system for a self-developed AI R&D platform based on cloud computing. Background Technology

[0002] In cloud-based collaborative R&D scenarios, requirement information, design documents, test cases, and delivery status usually need to be managed in conjunction with each other on the same platform. Currently, such platforms often rely on manual processes to break down the requirements and initiate generation and organization separately. Furthermore, the output format, result structure, and version saving methods of different deliverables lack unified constraints, making it difficult to maintain consistency in the generation results under the same requirement. However, when processing in the above way, problems such as duplicate generation of multiple types of deliverables, unstable output parsing, and disconnection between version updates and the current effective status are likely to occur, which in turn cause downstream content to fail to be synchronized in a timely manner after requirement changes, and make it difficult to trace the delivery results. Furthermore, the existing solution lacks a unified record of the invocation process, running status, and delivery chain, making it difficult to form a closed loop between permission verification, exception rollback, and delivery tracking. This results in problems such as obstructed collaborative processes and high system maintenance complexity. Summary of the Invention

[0003] The purpose of this invention is to provide a collaborative system for a self-developed AI R&D platform based on cloud computing, which avoids the repeated generation of multiple types of deliverables, unstable output parsing, and the disconnect between version updates and the current effective status. It also makes it easier to form a closed loop between permission verification, exception rollback, and delivery tracking, maintain the consistency of generated results under the same requirements, and improve collaborative efficiency.

[0004] The objective of this invention can be achieved through the following technical solutions: The requirements data management module manages cloud-based project data and verifies whether the requirements object exists and whether the current customer has access rights before delivery begins. The Demand AI Context Construction Module obtains the target demand identifier, reads basic information to form the demand AI context, and assembles the demand fields, target deliverable type and expected format into a call context. The AI ​​delivery scenario analysis module obtains the target deliverable type, parses it into the corresponding AI delivery scenario, and removes duplicates and normalizes it. The large model interaction module receives AI requests and sends them to the large model in the cloud to obtain AI responses. The multi-output orchestration and parsing module divides AI delivery scenarios into at least one orchestration group based on the calling context and according to the context length verification rules. Scenarios in the same group are merged into one AI request. The module obtains the AI ​​response returned by the large model interaction module and distributes and parses the AI ​​response according to the output identifier or default text. Run the logging and persistence module to obtain the parsing results. When the AI ​​generates the delivery successfully, it creates the sequential version number and semantic version number of the deliverable. When there is an error, it only records the result and does not generate a version. The version and status closed-loop module maintains the sequential version number, semantic version number, and effective status pointer, and updates the effective version after regeneration or revision. The AI ​​call audit module records the request identifier, running status, and result summary of the AI ​​call to form an audit link; The delivery center module verifies the readiness status of pipelines and branches, aggregates pipeline operation and continuous integration status, and displays delivery tracking information.

[0005] Optionally, AI delivery scenario analysis includes deliverable type enumeration normalization and scenario mapping. The target deliverable types include at least the final requirements document, front-end technical document, back-end technical document, and test case design. The deliverable types corresponding to the technical design document further include front-end technical document and back-end technical document.

[0006] Optionally, multi-output orchestration and parsing includes orchestration group aggregation and output identifier distribution; when the AI ​​response contains a named output set, the corresponding content is extracted according to the output identifier. When the AI ​​response contains only a single text or default output, the text is first assigned to the first-ranked delivery scenario in the orchestration group as the main scenario. Then, the remaining scenarios are split into single-scenario requests and called one by one to improve the efficiency of multi-deliverable generation and parsing stability.

[0007] Optionally, multi-output orchestration and parsing also includes format validation, which is used to verify the validity of the marked text format and structured data format; The marked text format is saved as normalized text after removing code fences; when the target deliverable type is test case design, the structured data format is parsed and the root node is required to be an object and contain a preset suite array and test case fields. When the verification fails, a parsing failure status is generated to avoid erroneous content from entering the subsequent process.

[0008] Optionally, when AI generation is successful, the execution record and persistence first create a deliverable version and then mark the execution record as successful; if an exception occurs in any parsing phase, persistence phase or state update phase, the execution record is marked as failed or parsing failed, and transaction rollback is used to prevent erroneous deliverables from entering the current effective version, thereby ensuring that the version data is consistent with the execution result.

[0009] Optionally, the semantic version number in the version and state closed loop includes a major version number and a minor version number; when the AI ​​generates a deliverable version, the semantic version number is generated according to the major version number increment rule, and the sequential version number is generated according to the increment sequence number of the same requirement and the same deliverable type. In this process, AI regenerates the corresponding major version number in increments, while manual storage preserves the corresponding minor version number in increments, thus forming a clear version evolution relationship.

[0010] Optionally, after the final requirements document is regenerated or manually saved, the current final requirements document status is set to pending review, and the existing front-end technical documents, back-end technical documents, and test case designs are set to expired.

[0011] Optionally, AI call auditing records the call status by updating the request identifier in place, recording the status as pending execution, running, successful or failed in sequence, and saving the model configuration snapshot version, target deliverable type, expected output identifier and result deliverable summary in the record.

[0012] Optionally, the delivery center performs structured validation and sensitive variable filtering before the work item is delivered. The structured validation includes at least whether the template is published and enabled, whether the template node is empty, whether the delivery track exists, whether the version control system project identifier is complete, and whether the branch mode is legal. Sensitive variable filtering removes version control system tokens, private tokens, and network hook keys to improve delivery security.

[0013] Optionally, the delivery center aggregates delivery operation status, track status, and branch context based on pipeline operation, node operation, and pipeline summary. When it receives a manual decision or takeover instruction, it can pause or modify the automatic execution status of the current pipeline node. It also supports canceling, retrying, and skipping operations on the operation or node to enhance the controllability and recoverability of the delivery process.

[0014] The beneficial effects of this invention are: 1. This invention verifies the existence of the requirement object and the current customer's access permissions before delivery is initiated through the requirement data management module. After verification, the requirement AI context construction module assembles the requirement fields, target deliverable type, and expected format into a calling context. This prevents unauthorized requirements from entering the generation process and ensures that the model input is consistent with the target requirement, thereby improving the security and generation accuracy in cloud-based collaborative R&D scenarios. The AI ​​delivery scenario parsing module enumerates, normalizes, deduplicates, and maps scenarios to deliverable types. The multi-output orchestration and parsing module merges scenarios in the same orchestration group into a single AI request, which reduces redundant generation and lowers the resource consumption caused by calling multiple types of deliverables separately, thereby improving collaborative processing efficiency. 2. This invention extracts content based on output identifiers when the AI ​​response contains a named output set through a multi-output orchestration and parsing module. When it contains only single text or default output, it prioritizes the text and assigns it to the first-ranked delivery scenario, while splitting the remaining scenarios into single-scenario requests and calling them one by one. This approach is compatible with different model output formats and improves the stability of multi-deliverable parsing. The format validation unit performs legality checks on the format of the marked text and the structured data format, and constrains the root node, suite array, and test case fields under the test case design scenario. This approach can intercept results with incomplete structure or invalid format before writing to the official version, thereby reducing the risk of parsing failure and erroneous content being stored in the database. Attached Figure Description

[0015] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a structural diagram of the self-developed AI R&D platform collaborative system based on cloud computing provided in the embodiments of this application. Detailed Implementation

[0016] 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.

[0017] Please see Figure 1 A self-developed AI R&D platform collaborative system based on cloud computing, including: The requirements data management module manages cloud-based project data and verifies whether the requirements object exists and whether the current customer has access rights before delivery begins. The Demand AI Context Construction Module obtains the target demand identifier, reads basic information to form the demand AI context, and assembles the demand fields, target deliverable type and expected format into a call context. The AI ​​delivery scenario analysis module obtains the target deliverable type, parses it into the corresponding AI delivery scenario, and removes duplicates and normalizes it. The large model interaction module receives AI requests and sends them to the large model in the cloud to obtain AI responses. The multi-output orchestration and parsing module divides AI delivery scenarios into at least one orchestration group based on the calling context and according to the context length verification rules. Scenarios in the same group are merged into one AI request. The module obtains the AI ​​response returned by the large model interaction module and distributes and parses the AI ​​response according to the output identifier or default text. Run the logging and persistence module to obtain the parsing results. When the AI ​​generates the delivery successfully, it creates the sequential version number and semantic version number of the deliverable. When there is an error, it only records the result and does not generate a version. The version and status closed-loop module maintains the sequential version number, semantic version number, and effective status pointer, and updates the effective version after regeneration or revision. The AI ​​call audit module records the request identifier, running status, and result summary of the AI ​​call to form an audit link; The delivery center module verifies the readiness status of the pipeline and its branches, aggregates pipeline operation and continuous integration status, and displays delivery tracking information. In a preferred embodiment of the present invention, the AI ​​delivery scenario analysis module includes a deliverable type enumeration and normalization unit and a scenario mapping unit. The target deliverable types include at least final requirement documents, technical design documents, and test case designs. The deliverable types corresponding to the technical design documents include front-end technical documents and back-end technical documents. The multi-output orchestration and parsing module includes an orchestration group aggregation unit and an output identifier distribution unit; when an AI response contains a named output set based on an output identifier, the corresponding content is extracted according to the output identifier. When the AI ​​response contains only a single text or a default text, the text is first assigned to the first-ranked delivery scenario in the orchestration group as the main scenario, and then the remaining scenarios are split into single-scenario requests and invoked one by one.

[0018] In this implementation, the cloud server is responsible for performing the tasks of demand data management, context construction, scenario resolution, request orchestration, model invocation, result parsing, version saving, and audit logging. The system processes individual requirement objects as units. The requirement data management module first reads the requirement object from the cloud project data and verifies whether the requirement object exists before delivery starts. At the same time, it verifies whether the current customer has permission to access the requirement. If the verification fails, the system directly terminates the subsequent delivery request to prevent unauthorized data from entering the generation process. Once the requirement object passes the validation, the requirement AI context construction module reads the basic information of the requirement according to the requirement identifier, and organizes the requirement title, description, status, priority, module, tag, planned version, iteration, responsible person, proposer, developers, testers, product owner, acceptance notes and comment summary into the requirement AI context; The demand AI context refers to the structured set of input data formed around the target demand. Its role is to bring the demand fields and the current demand status into the request payload before calling the large model in the cloud, so that the subsequent generated content corresponds to the demand object. During the call context assembly, the system concatenates the requirement fields, target deliverable type, and expected format into a call context, and uses this call context as the unified input for subsequent requests. The expected format is used to limit the content format that the AI ​​output should adopt, such as tagged text format or structured data format, thereby providing a basis for subsequent parser selection. After receiving the target deliverable type, the AI ​​delivery scenario analysis module first performs enumeration and normalization on the deliverable type, and then converts it into the corresponding AI delivery scenario according to the scenario mapping relationship. If there are duplicate deliverable types in the input, the module will remove duplicates according to the insertion order to avoid the same scenario from being repeatedly entered into the subsequent arrangement. Define the input request as a dynamic dimension vector. When the dimension checking function When the dimension checking function terminates parsing and returns an exception; Scalar analysis is performed when; Expand the tensor list and apply the filtering function. After removing empty nodes, a valid set is formed, in which For dynamic dimension vectors The elements in Indicates an empty node; For the final requirements document, front-end technical document, back-end technical document and test case design, the system establishes corresponding scenario codes and output identifiers. In this implementation, the technical design document is further divided into front-end technical document and back-end technical document. During normalization, the system will unify the final requirement documents into requirement document scenarios, and classify the front-end technical documents, back-end technical documents and test case designs into front-end design scenarios, back-end design scenarios and test scenarios respectively, and assign a fixed scenario code and output identifier to each scenario. If the input contains aliases, Chinese names, and internal codes, the internal code will be matched first. If the match fails, the name will be matched. Unrecognized items will be directly discarded. To facilitate implementation, the system uses a fixed mapping table to normalize the AI ​​delivery scenarios, which maps the internal code, Chinese name and alias to the scenario code. When deduplicating, only the first occurrence of the scenario is retained. The division of orchestration groups is based on whether the output identifiers share the same prompt template. If multiple scenarios correspond to the same output identifier, they are merged into the same orchestration group. If there is only a single scenario, a single request is formed directly. After receiving multiple AI delivery scenarios, the multi-output orchestration and parsing module first merges scenarios in the same orchestration group into one AI request according to preset rules; Before merging requests, the system performs a context physical dimension check: calculating the length of the current requirement context. The sum of the expected maximum output length for each scenario If the inequality is satisfied: in, This represents the expected maximum output length for each scenario. The upper limit of the absolute physical context window for the large model. If the safety threshold coefficient is met, then the orchestration group restriction is broken, and the task is forcibly split into independent micro-task requests in descending order of target scenario weight; if the inequality is satisfied: If so, the current orchestration group restrictions will be maintained when making the merge request; For scenarios where front-end technical documentation, back-end technical documentation, and test case designs share the same orchestration group, the system will write them together in a single request and include their respective output identifiers in the call context; this allows the large cloud model to return multiple types of results in a single call, avoiding duplicate requests. Once the cloud-based large model returns an AI response, the large model interaction module sends the AI ​​response back to the multi-output orchestration and parsing module; If the response contains a named set of outputs, such as output identifiers corresponding to front-end technical documents, back-end technical documents, and test case designs, the system extracts the corresponding content according to the output identifiers and distributes the different content to their respective scenario processing branches. The default output is the unstructured global default response text returned by the large cloud model when no specific output identifier is recognized; Even if a certain AI channel only returns a single text result, the system executes the supplementary generation logic for the remaining deliverables; When the parsing results are available, the runtime logging and persistence module first records the request identifier, runtime status, and result summary of this run; when the AI ​​is successfully generated, the system then creates the sequential version number and semantic version number of the deliverable. Sequential version numbers are used to indicate the incremental saving order under the same requirement and the same deliverable type, while semantic version numbers are used to distinguish between AI regeneration and manual revision. If an anomaly occurs during the parsing or persistence phase, the system only retains the running results and does not create a formal deliverable version to prevent erroneous content from entering the currently effective data. The version and status closed-loop module maintains the sequential version number, semantic version number, and pointer to the currently effective version; when a deliverable under the same requirement is regenerated, the system updates the pointer to the currently effective version, making the new version the currently used version; After manual revision, the system updates the version number and switches to the effective version according to the revision rule of incrementing the corresponding minor version number; the AI ​​call audit module writes the request identifier, running status and result summary into the audit record, so that a call can be traced back to the corresponding model call, output content and version result according to the request identifier; Before a work item is delivered, the delivery center module verifies the readiness status of the pipeline and its branches, and aggregates the pipeline running status and continuous integration status to display delivery tracking information, ensuring consistency between the requirements, generated results, and delivery execution.

[0019] In a preferred embodiment of the present invention, the multi-output orchestration and parsing module further includes a format verification unit, wherein the expected format includes a marked text format and a structured data format, and the format verification unit performs format validity verification on the marked text format and the structured data format; Among them, the marked text format is saved as normalized text after removing code fences; when the target deliverable type is test case design, the structured data format is parsed and the root node is required to be an object and contain a preset suite array and test case fields. When the verification fails, a parsing failure status is generated. When AI generation is successful, the runtime log and persistence module first creates a deliverable version and then marks the runtime log as successful. If an exception occurs in any of the parsing, persistence, or state update phases, the run record will be marked as failed or parsing failed, and transaction rollback will be used to prevent erroneous deliverables from entering the current effective version. The semantic version number in the version and status closed-loop module includes a major version number and a minor version number. When the AI ​​generates a deliverable version, the semantic version number is generated according to the major version number increment rule, and the sequential version number is generated according to the increment sequence number of the same requirement and the same deliverable type. Among them, the AI ​​regenerates the corresponding major version number increment, and the corresponding minor version number increment is manually saved. After the final requirements document is regenerated or manually saved, the version and status closed-loop module sets the current final requirements document status to pending review and sets the existing front-end technical documents, back-end technical documents, and test case designs to expired.

[0020] In this embodiment, the cloud server continues to act as the execution entity, first performing format verification on the AI ​​response, and then deciding whether to proceed with the version saving process. After receiving the AI ​​response content distributed by the multi-output orchestration and parsing module, the format verification unit first determines whether the content belongs to the tagged text format or the structured data format. For the tagged text format, the system first checks whether the response text is empty. If the text has code fences on the outside, the first and last fences are removed and the text is saved as normalized text. If it is still empty after removal, the parsing is determined to be failed. For structured data formats, the system first extracts the first structured data object from the response; among them, the suite array is used to define the business module classification and hierarchical structure of test cases, and the test case fields include at least the test steps, expected results and preconditions; Structured data validation in test case design is a necessary format legality check; the system does not directly write the AI-returned content into the official deliverable, but first completes the field existence and structural integrity judgment; only the content that passes the validation will enter the subsequent persistence stage, thereby avoiding incomplete data from polluting the official version; When AI generation is successful, the runtime log and persistence module first creates a deliverable version and then marks the runtime log as successful. Creating the deliverable version first means writing the deliverable entity, version number, and content within the transaction before updating the runtime log status. If an exception occurs during any of the parsing, persistence, or state update phases, the system will mark the execution record as failed or parsing failed, and will prevent erroneous deliverables from entering the current effective version through transaction rollback; that is, the system adopts an asynchronous compensation timing control model: network large-scale requests are strictly separated from local database operations. Version writing, state updates, and effective pointer movement are completed only after the callback response arrives, within independent microsecond-level local transaction boundaries; For large model call time In this case, distributed locks and state machines are used to ensure eventual consistency. A preset timeout threshold is used to avoid prolonged periods of inactivity. The blockage led to the depletion of physical resources; In the version and status closed-loop module, the semantic version number is composed of a major version number and a minor version number; for the same requirement and the same deliverable type, the sequential version number is accumulated by incrementing the serial number. Semantic version numbers are distinguished based on the version source. When AI regenerates, the major version number increments, and when manually saved, the minor version number increments. The system uses the version number to determine whether the deliverable is regenerated by artificial intelligence or revised manually on an existing manuscript. When the final requirements document is regenerated or manually saved, the version and status closed-loop module will set the current final requirements document status to pending review and set the existing front-end technical documents, back-end technical documents and test case designs to expired. The system marks existing downstream deliverables as expired, indicating that their correspondence with the latest final requirements document has become invalid. This is used to determine which downstream deliverables need to be regenerated or confirmed after requirements change. For format validity checks, the system executes the following sequence: null value judgment - outer fence stripping - structure field verification - result storage in the database. If the marked text is still empty after stripping, or the structured data cannot be parsed into a single object, the system will directly return a parsing failure and will not proceed to version writing.

[0021] In a preferred embodiment of the present invention, the AI ​​call audit module records the call status by updating the request identifier in place, and records the status of pending execution, running, success or failure in sequence, and saves the model configuration snapshot version, target deliverable type, expected output identifier and result deliverable summary in the record; Before a work item is delivered, the delivery center module performs structured validation and sensitive variable filtering. Structured validation includes at least checking whether the template is published and enabled, whether the template node is empty, whether the delivery track exists, whether the version control system project identifier is complete, and whether the branch mode is valid. Sensitive variable filtering removes version control system tokens, private tokens, and network hook keys. The delivery center module aggregates delivery operation status, track status, and branch context based on pipeline operation, node operation, and pipeline summary. When it receives manual decision or takeover instructions, it can pause or modify the automatic execution status of the current pipeline node, and supports cancel, retry, and skip operations on the operation or node.

[0022] In this implementation, the AI ​​call auditing module updates the call status in place by the cloud server according to the request identifier; the system writes the pending status before the request is sent, updates it to running after the model call starts, updates it to successful after the model returns and completes parsing and saving, and updates it to failed if the call or parsing is abnormal. The audit log saves the model configuration snapshot version, target deliverable type, expected output identifier, and result deliverable summary, so that the same AI call can reproduce its input, model configuration, and output results according to the run identifier; The model configuration snapshot version is used to represent the model configuration state used in this call. Even if the configuration changes later, the call conditions at that time can be traced back based on the snapshot version. Before a work item is delivered, the delivery center module first performs structured validation and sensitive variable filtering on the cloud server. The structured validation includes whether the template has been published and enabled, whether the template node is empty, whether the delivery track exists, whether the version control system project identifier is complete, and whether the branch mode is legal. If any of these conditions are not met, the system will return the verification result and prevent startup; Sensitive variable filtering is used to remove version control system tokens, private tokens and network hook keys from client-input variables to prevent these fields from being written to runtime variables or returned to the front end; After the verification is passed, the delivery center module aggregates the delivery operation status, track status and branch context based on pipeline operation, node operation and pipeline summary; Track status refers to the operational status associated with a specific delivery track, including information such as repository, branch, and continuous integration status; branch context is used to describe the source branch, target branch, and related commit and merge request information associated with the current operation; When the system receives a manual decision or takeover instruction, it will pause or modify the automatic execution status of the current pipeline node, and support cancellation, retry, and skip operations on the run or node; in this way, the delivery center can display the requirement object, code repository, continuous integration results, and manual operation in the same delivery view, which is convenient for subsequent processing and tracking. For structured validation, the system returns Boolean results item by item in the following order: template release, activation, node not empty, track exists, project identifier complete, and branch mode valid. If any item is negative, the system will terminate the process. The four statuses—pending execution, running, success, and failure—are updated once overwritten according to the request identifier. The failure status and the parsing failure status share the same request identifier.

[0023] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A collaborative system for a self-developed AI R&D platform based on cloud computing, characterized in that: include: The requirements data management module manages cloud-based project data and verifies whether the requirements object exists and whether the current customer has access rights before delivery begins. The Demand AI Context Construction Module obtains the target demand identifier, reads basic information to form the demand AI context, and assembles the demand fields, target deliverable type, and expected format into a call context. The AI ​​delivery scenario analysis module obtains the target deliverable type, parses it into the corresponding AI delivery scenario, and removes duplicates and normalizes it. The large model interaction module receives the AI ​​request and sends it to the large model in the cloud to obtain the AI ​​response; The multi-output orchestration and parsing module, based on the calling context, divides the AI ​​delivery scenario into at least one orchestration group according to the context length verification rule, and merges scenarios in the same group into one AI request; it obtains the AI ​​response returned by the large model interaction module, and distributes and parses the AI ​​response according to the output identifier or default text; Run the logging and persistence module to obtain the parsing results. When the AI ​​generates the delivery successfully, it creates the sequential version number and semantic version number of the deliverable. When there is an error, it only records the result and does not generate a version. The version and status closed-loop module maintains the sequential version number, semantic version number, and effective status pointer, and regenerates or revises the effective version. The AI ​​call audit module records the request identifier, running status, and result summary of the AI ​​call to form an audit link; The delivery center module verifies the readiness status of pipelines and branches, aggregates pipeline operation and continuous integration status, and displays delivery tracking information.

2. The collaborative system of the self-developed AI R&D platform based on cloud computing as described in claim 1, characterized in that, The AI ​​delivery scenario analysis module includes a deliverable type enumeration and normalization unit and a scenario mapping unit. The target deliverable types include at least the final requirement document, technical design document and test case design. The deliverables corresponding to the technical design documents include front-end technical documents and back-end technical documents.

3. The collaborative system of a self-developed AI R&D platform based on cloud computing as described in claim 1, characterized in that, The multi-output orchestration and parsing module includes an orchestration group aggregation unit and an output identifier distribution unit; When the AI ​​response contains a named output set based on the output identifier, the corresponding content is extracted according to the output identifier; when the AI ​​response contains only a single text or a default text, the text is first assigned to the first-ranked delivery scenario in the orchestration group as the main scenario, and then the remaining scenarios are split into single-scenario requests and called one by one.

4. The collaborative system of the self-developed AI R&D platform based on cloud computing as described in claim 1, characterized in that, The multi-output orchestration and parsing module further includes a format verification unit. The expected format includes a marked text format and a structured data format. The format verification unit performs format validity verification on the marked text format and the structured data format. Among them, the marked text format is saved as normalized text after removing code fences; When the target deliverable type is test case design, the structured data format, after parsing, requires the root node to be an object and to contain a preset suite array and test case fields. If the verification fails, a parsing failure status is generated.

5. The collaborative system of a self-developed AI R&D platform based on cloud computing as described in claim 1, characterized in that, When the AI ​​is successfully generated, the runtime record and persistence module first creates a version of the deliverable, and then marks the runtime record as successful. If an exception occurs during any of the resolution, persistence, or state update phases, the run record will be marked as failed or resolution failed, and a transaction rollback will be used to prevent erroneous deliverables from entering the current effective version.

6. The collaborative system of a self-developed AI R&D platform based on cloud computing as described in claim 1, characterized in that, The semantic version number in the version and state closed-loop module includes the major version number and the minor version number. When AI generates deliverable versions, semantic version numbers are generated according to the major version number increment rule, and sequential version numbers are generated according to the increment sequence number of the same requirement and the same deliverable type. In this process, AI regenerates the corresponding major version number in increments, while manual storage preserves the corresponding minor version number in increments.

7. The collaborative system of a self-developed AI R&D platform based on cloud computing as described in claim 2, characterized in that, After the final requirement document is regenerated or manually saved, the version and status closed-loop module sets the current final requirement document status to pending review and sets the existing front-end technical documents, back-end technical documents, and test case designs to expired.

8. The collaborative system of a self-developed AI R&D platform based on cloud computing as described in claim 1, characterized in that, The AI ​​call auditing module records the call status by updating the request identifier in place, and records the status of pending execution, running, success or failure in sequence. The record saves the model configuration snapshot version, target deliverable type, expected output identifier and result deliverable summary.

9. The collaborative system of a self-developed AI R&D platform based on cloud computing as described in claim 1, characterized in that, Before a work item is delivered, the delivery center module performs structured validation and sensitive variable filtering. The structured validation includes at least whether the template has been published and enabled, whether the template node is empty, whether the delivery track exists, whether the version control system project identifier is complete, and whether the branch mode is legal. The sensitive variable filter removes version control system tokens, private tokens, and network hook keys.

10. The collaborative system of a self-developed AI R&D platform based on cloud computing as described in claim 1, characterized in that, The delivery center module aggregates delivery operation status, track status, and branch context based on pipeline operation, node operation, and pipeline summary. When it receives manual decision-making or takeover instructions, it can pause or modify the automatic execution status of the current pipeline node, and support cancellation, retry, and skip operations on the operation or node.