Autonomous agent system based on retrieval enhancement generation technology and generative large model

By introducing search-enhanced generation technology and autonomous proxy systems into the generative large model, the business knowledge base and memory module are built, and the problem of insufficient generation of illusions and timeliness is solved, more accurate and faster text generation is achieved, and more accurate and faster text generation is improved. Interpretability of the model.

CN119003592AActive Publication Date: 2024-11-22INSTITUTE OF COMPUTING INNOVATION ZHEJIANG UNIVERSITY

Patent Information

Application Number
CN202411076898.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-07
Publication Date
2024-11-22
Estimated Expiration
2044-08-07

AI Technical Summary

Technical Problem

Generative big models have problems such as generative illusions and insufficient timeliness, resulting in insufficient timeliness and accuracy of text generation in specific business scenarios.

Method used

Adopt an autonomous agent system based on search enhancement generation technology and generative big models. By building a business knowledge base and memory module, combining search enhancement generation (RAG) technology, assist in supervising model generation, formulate a closed-loop executable plan list, and reflect on it. Mechanisms adjust the plan to improve the accuracy of the generated results.

Benefits of technology

It effectively alleviates the illusion of generation, improves the timeliness, accuracy and relevance of generated content, enhances the generation speed, and improves the interpretability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119003592A_ABST
    Figure CN119003592A_ABST
Patent Text Reader

Abstract

The invention discloses an autonomous agent system based on a retrieval enhancement generation technology and a generative large model. A knowledge base module of the system is used for storing a knowledge base established after conversion and separation of a service field data set, and a memory module stores questions and dialogue contents of a user and associated knowledge recalled from the user and the knowledge base after RAG retrieval; the planning module formulates a plan list according to the questioning and dialogue content and the associated knowledge; and the execution module performs execution according to the plan list to obtain a plan execution result, judges whether the plan meets the expectation through reflection, and decides whether to perform re-planning until the plan meets the expectation. According to the system, the generation illusion problem of the large model can be effectively relieved, the generation speed of the large model is increased, and meanwhile the relevance, timeliness and accuracy of the generated content are improved; in addition, by designing a plan list of the plan module, step-by-step execution of the execution module and other process mechanisms, the user is helped to understand the execution process of the large model, and the interpretability of a final result is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to an autonomous agent system, and relates to the fields of autonomous agent, retrieval enhancement generation technology, and generative large models, and in particular to an autonomous agent system based on retrieval enhancement generation technology and generative large models. Background Art

[0002] In recent years, with the development of deep learning and large models, the shortcomings of large models such as generation hallucination and lack of timeliness have become increasingly prominent. Retrieval enhancement generation technology and autonomous agents based on large language models (LLM) have become hot research directions. These technologies can help large models overcome generation hallucinations, obtain more timely information, and have stronger language understanding and generation capabilities, and can complete more complex tasks.

[0003] Autonomous Agent refers to a software system or entity that can independently complete goal setting, decision-making and action to a certain extent. Autonomous agents are usually able to perform tasks independently in the environment, and can interact with the environment and other agents to obtain information or achieve collaboration, and further perform logical reasoning in order to make reasonable decisions.

[0004] Retrieval-Augmented Generation (RAG) refers to a method that assists the generation process by retrieving external knowledge sources in generation tasks. It combines the advantages of retrieval mechanisms and generation models to improve the quality, diversity, and coverage of generation results. RAG has achieved significant improvements in tasks such as dialogue systems, text summarization, and machine translation. The main challenges of retrieval-augmented generation technology include how to build an effective retrieval index and how to design the interaction between the model and the retrieval mechanism.

[0005] Generative Large Language Model is a type of natural language processing model trained with large-scale text data. This type of model is usually based on the transformer architecture and self-attention mechanism, and learns text representation from massive text through unsupervised or weakly supervised learning; the number of parameters of large models is huge, generally more than 6 billion parameters. Large models can generate coherent and grammatically correct text, and are widely used in tasks such as chatbots, text summarization, machine translation, and text generation. Although large models have significantly refreshed records in language understanding and text generation, they generally have problems such as generation hallucinations, security and privacy, and lack of timeliness. Summary of the invention

[0006] In order to solve the problems existing in the background technology, the present invention provides an autonomous agent system based on retrieval enhancement generation technology and generative big models. The method can improve the timeliness and accuracy of text generation of big models in specific business scenarios, overcome its generation illusion problem, and build an autonomous agent Agent based on RAG technology to assist in supervising model generation by building a business knowledge base, better completing the adaptive transformation of big models in vertical business fields, and providing important innovative ideas, technological breakthroughs and application value for related big model application solutions.

[0007] The technical solution adopted by the present invention is:

[0008] 1. An autonomous agent system based on retrieval-enhanced generation technology and generative large models, comprising:

[0009] The knowledge base module is used to transform and separate the collected data sets in specific business fields and then input them into the knowledge base for storage.

[0010] The memory module is used to store the query and conversation content input by the user in multiple rounds of dialogue, as well as the associated knowledge recalled from the memory module itself and the knowledge base of the knowledge base module after generating RAG retrieval through retrieval enhancement.

[0011] The planning module is used to formulate a closed-loop executable plan list based on the query and dialogue content and related knowledge input by the user from the memory module.

[0012] The execution module is used to obtain the plan execution results after executing the plan list obtained from the planning module, so as to judge whether the plan execution results meet expectations through reflection to obtain the reflection results, and store the plan execution results and the reflection results in the memory module; directly output the plan list for plan execution results that meet expectations for execution, and push the plan execution results that do not meet expectations to the planning module for re-planning until they meet expectations.

[0013] The data set of the specific business field includes long text, tables and databases in the specific business field; the data set is uniformly converted into text format, and then divided into blocks according to delimiters and the summary information of the blocks is generated, and each summary and its block text are used together to establish a knowledge base. The knowledge base data content is established in the index form of {summary: block text}. Specific business fields include finance, industrial manufacturing, medical and other fields.

[0014] The memory module includes a long-term memory submodule and a short-term memory submodule. The long-term memory submodule is used to store the associated knowledge recalled from the memory module itself and the knowledge base of the knowledge base module after the retrieval enhancement generation RAG retrieval, and the short-term memory submodule is used to store the query and dialogue content input by the user in the multi-round dialogue and the reflection results of the execution module. The two submodules can alleviate the limitation problem that the Agent's token length exceeds the maximum length of the large model during the multi-round dialogue process.

[0015] The planning module identifies the user's question intention through the prompt word Prompt according to the query input by the user in multiple rounds of dialogue and the dialogue content, that is, obtains the user's request content, and combines the user's request content with the associated knowledge obtained from the memory module, and finally formulates a closed-loop executable plan list; the user's request content is stored in the short-term memory submodule of the memory module, and the associated knowledge is stored in the long-term memory submodule of the memory module. The planning module is the core module of the entire autonomous agent system, usually including functions such as intention recognition, retrieval enhancement, reflection, and re-planning.

[0016] The execution module is composed of a large model and a tool function. The execution module calls the large model to obtain the plan execution text according to the plan list fed back by the plan module, calls the tool function to obtain the plan execution result, adds the plan execution result to the prompt word Prompt in the execution module and then feeds it back to the plan module. Large models such as GLM, Llama, etc., tool functions such as search, python code execution, drawing, etc.

[0017] 2. A processing method of an autonomous agent system, comprising:

[0018] Step 1: Based on the specific application scenario, collect data sets in specific business fields, and establish an industry knowledge base in specific business fields through data conversion, segmentation, summary generation and indexing, so as to facilitate subsequent RAG retrieval and then store it in the knowledge base module.

[0019] Step 2: The long-term memory submodule in the memory module stores the associated knowledge recalled from the memory module itself and the knowledge base of the knowledge base module after retrieval through retrieval enhancement generation RAG, and the short-term memory submodule stores the query and dialogue content input by the user in multiple rounds of dialogue.

[0020] Step 3: The planning module identifies the user's request content through the prompt word Prompt according to the query input by the user in multiple rounds of dialogue and the dialogue content, and combines the user's request content with the related knowledge obtained from the memory module to enhance the user's request content, and finally formulates a closed-loop executable plan list; the user's request content is stored in the short-term memory submodule of the memory module, and the related knowledge is stored in the long-term memory submodule of the memory module.

[0021] Step 4: The execution module accepts the specific plan in the plan list, executes the plan list according to the specific plan steps, calls the large model to obtain the plan execution text, calls the tool function to obtain the plan execution result, adds the plan execution result to the prompt word Prompt in the execution module and then feeds it back to the planning module.

[0022] Step 5: The planning module receives the plan execution results from the execution module, determines whether the plan execution results meet expectations through reflection, and obtains the reflection results. The plan list of plan execution results that meet expectations is directly output for execution, and the plan execution results that do not meet expectations are pushed to the planning module for re-planning until they meet expectations; and the plan execution results and reflection results are stored in the memory module; the reflection result is the judgment result of whether the plan execution results meet expectations.

[0023] Step 6: Loop through steps 4 and 6 until the final result is returned.

[0024] The beneficial effects of the present invention are:

[0025] 1. Alleviate the generation hallucination of large models: Generative large models have the problem of generation hallucination. The present invention can retrieve relevant facts, fragments or texts from a large-scale knowledge base by introducing external knowledge retrieval and introduce them into the generation process. This can make up for the lack of knowledge in the generation model and generate more accurate and rich content.

[0026] 2. Improve the timeliness of generated content: Since the corpus used in the pre-training stage of the large model has the problem of lagging timeliness, the present invention uses retrieval enhancement generation technology to retrieve related knowledge from the real-time updated knowledge base, thereby compensating for the problem of insufficient timeliness of the large model.

[0027] 3. Improve generation relevance: Interactive agents can interact with the environment and tools to obtain information or achieve collaboration. This enhances the flexibility and effectiveness of the agent. Through reflection and replanning, the accuracy of plan execution can be effectively improved. The relevant fragments retrieved by RAG contain a wider range of language patterns, which can improve the relevance of model results.

[0028] 4. Improve generation speed: Since the information stored in the knowledge base is generated by recall and does not rely on additional generation of large models, the generation process of the model can be effectively improved and accelerated through retrieval enhancement technology. Compared with pure generation models, retrieval enhancement generation can achieve better performance in many tasks.

[0029] 5. Help users understand the execution process of large models: Since the autonomous agent will display each step of the model plan in the form of a plan list, users can intuitively see the model's intention to decompose the user's questions and the execution process of each step of the plan, providing a basis for the final generation results and improving the interpretability of the generative large model.

[0030] In summary, the present invention realizes an autonomous agent system using retrieval enhanced generation technology, which effectively alleviates the generation hallucination problem of large models, improves the generation speed of large models, and at the same time improves the relevance, timeliness and accuracy of the generated content; in addition, the present invention helps users understand the execution process of large models and improves the interpretability of the final results by designing process mechanisms such as the plan list of the planning module and the step-by-step execution of the execution module. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is the overall flow chart of the system of the present invention;

[0032] Figure 2 It is a schematic diagram of the process of building a knowledge base in the present invention;

[0033] Figure 3 It is a schematic diagram of the framework of the interaction of the internal modules of the system of the present invention. DETAILED DESCRIPTION

[0034] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] The autonomous agent system based on retrieval enhancement generation technology and generative large models of the present invention includes a knowledge base module, a memory module, a planning module and an execution module. The knowledge base module transforms and separates the collected data sets in a specific business field and inputs them into the knowledge base for storage. The memory module stores the questions Query and conversation content input by users in multiple rounds of dialogues, as well as the related knowledge recalled from the memory module itself and the knowledge base of the knowledge base module after retrieval through retrieval enhancement generation RAG retrieval; the planning module formulates a closed-loop executable plan list based on the questions Query and conversation content and related knowledge input by the user obtained from the memory module; the execution module obtains the plan execution result after executing the plan list obtained from the plan module, thereby judging whether the plan execution result meets expectations in the form of reflection to obtain the reflection result, and storing the plan execution result and the reflection result in the memory module; the plan list of the plan execution result that meets expectations is directly output for execution, and the plan execution result that does not meet expectations is pushed to the plan module for re-planning until it meets expectations.

[0036] Data sets for specific business fields include long texts, tables, and databases in specific business fields. The data sets are uniformly converted into text format, and then divided into blocks according to delimiters and summary information of the blocks is generated. The knowledge base is established by combining each summary and its block text. The knowledge base data content is established in the index form of {summary: block text}. Specific business fields include finance, industrial manufacturing, medical care, etc.

[0037] The memory module includes a long-term memory submodule and a short-term memory submodule. The long-term memory submodule is used to store the associated knowledge recalled from the memory module itself and the knowledge base of the knowledge base module after the retrieval enhancement generation RAG retrieval, and the short-term memory submodule is used to store the query and dialogue content entered by the user in the multi-round dialogue and the reflection results of the execution module. The two submodules can alleviate the limitation problem that the Agent's token length exceeds the maximum length of the large model during the multi-round dialogue process.

[0038] The planning module uses the prompt word Prompt to identify the user's question intention based on the query input by the user in multiple rounds of dialogue and the dialogue content, that is, to obtain the user's request content, and combine the user's request content with the associated knowledge obtained from the memory module, and finally formulate a closed-loop executable plan list; the user's request content is stored in the short-term memory submodule of the memory module, and the associated knowledge is stored in the long-term memory submodule of the memory module. The planning module is the core module of the entire autonomous agent system, usually including intent recognition, retrieval enhancement, reflection, re-planning and other functions.

[0039] The execution module consists of a large model and tool functions. The execution module calls the large model to obtain the plan execution text according to the plan list fed back by the planning module, calls the tool function to obtain the plan execution result, adds the plan execution result to the prompt word Prompt in the execution module and then feeds it back to the planning module. Large models include GLM, Llama, etc., and tool functions include search, python code execution, drawing, etc.

[0040] The processing method of the autonomous agent system of the present invention is specifically as follows:

[0041] Step 1: Based on the specific application scenario, collect data sets in specific business fields, and establish an industry knowledge base in specific business fields through data conversion, segmentation, summary generation and indexing, so as to facilitate subsequent RAG retrieval and then store it in the knowledge base module.

[0042] Step 2: The long-term memory submodule in the memory module stores the associated knowledge recalled from the memory module itself and the knowledge base of the knowledge base module after retrieval through retrieval enhancement generation RAG, and the short-term memory submodule stores the query and dialogue content input by the user in multiple rounds of dialogue.

[0043] Step 3: The planning module identifies the user's request content through the prompt word Prompt according to the query input by the user in multiple rounds of dialogue and the dialogue content, and combines the user's request content with the related knowledge obtained from the memory module to enhance the user's request content, and finally formulates a closed-loop executable plan list; the user's request content is stored in the short-term memory submodule of the memory module, and the related knowledge is stored in the long-term memory submodule of the memory module.

[0044] Step 4: The execution module accepts the specific plan in the plan list, executes the plan list according to the specific plan steps, calls the large model to obtain the plan execution text, calls the tool function to obtain the plan execution result, adds the plan execution result to the prompt word Prompt in the execution module and then feeds it back to the planning module.

[0045] Step 5: The planning module receives the plan execution results from the execution module, determines whether the plan execution results meet expectations through reflection, and obtains the reflection results. The plan list of plan execution results that meet expectations is directly output for execution, and the plan execution results that do not meet expectations are pushed to the planning module for re-planning until they meet expectations; and the plan execution results and reflection results are stored in the memory module; the reflection result is the judgment result of whether the plan execution results meet expectations.

[0046] Step 6: Loop through steps 4 and 6 until the final result is returned.

[0047] The specific implementation methods of the present invention are as follows:

[0048] It is agreed that user query is Query, the plan of the autonomous agent is Plan, the memory is Mem, the execution is Act, the knowledge base block information is Chunk, the large model is LLM, the representation model is enc, and the representation vector is vec.

[0049] Step 1: Build a knowledge base:

[0050] According to the specific business, it is necessary to build a detailed business knowledge base before building an autonomous agent system. The data quality, data timeliness and relevance of the knowledge base directly determine the upper limit of the quality of subsequent agent model generation. Therefore, it is necessary to improve the information collection, data cleaning and data preprocessing of the knowledge base.

[0051] like Figure 2 As shown, the present invention lists several common business data storage forms: long text (such as Word documents, PDF documents, PPT manuscripts, etc.), tables (such as Excel reports, CSV reports, etc.), databases (such as MySQL databases, MongoDB databases, etc.).

[0052] First, the data file is converted into text format through the preprocessing script. The images, videos and other formats that cannot be directly converted in the data set are ignored to obtain the data text.

[0053] After that, use regular expressions to process the data text by delimiters. For example, in document type files, use paragraph format line breaks or title format rich list as delimiters for block division, in table type files, use subtable names as delimiters, and in database environments, use library names and table names as delimiters.

[0054] Afterwards, a language model is used to summarize each block of text one by one and generate a summary. At this step, you can choose a pre-trained language model, such as BERT (Bidirectional Encoder Representation from Transformers), RoBERTa, RoFormer, etc. Summary generation plays a role in data cleaning to a certain extent, while further condensing the core knowledge of the block and improving the accuracy of subsequent RAG retrieval.

[0055] Finally, the business knowledge base is constructed using the chunk information, and the knowledge information is stored in the index form of {Abstract: Chunk} in the present invention. At the same time, a vector database is constructed, and the representation vectors of all abstracts are calculated using one of the above language models to improve the speed of subsequent RAG retrieval.

[0056] Step 2: Make a plan:

[0057] like Figure 3 As shown, in the initial state, the memory module Mem of the autonomous agent system init The autonomous agent system receives the user input information Query, and the autonomous agent system encodes the user input information Query through one of the above language models to obtain a representation vector.

[0058] vec Query =enc(Query)

[0059] At the same time, through the vector similarity algorithm VectorSimilarity, relevant information can be matched from the knowledge base for this query vector. Here, it is assumed that only the blocks with the highest vector similarity are matched, that is, the recalled associated knowledge.

[0060] Chunk=max(VectorSimilary(vec Query ,vec Z ))

[0061] Among them, vec Z A representation vector representing the summary.

[0062] Add chunk information Chunk and user input information Query to PlanPrompt, and generate an initial plan through the planning module;

[0063] Plan init =LLM(PlanPrompt(Query, Chunk,...))

[0064] At the same time, the information in the long and short-term memory submodules of the memory module needs to be updated synchronously.

[0065] Mem = {ShortTerm: Plan init , LongTerm: (Query, Chunk)}

[0066] Step 3: Execution plan:

[0067] In a specific task scenario, there will be multiple execution tool functions Act. Here we assume that Plan init It is a single-step plan, only one tool is called, and the tool execution result is;

[0068] Resp=Act(Plan init )

[0069] At the same time, update this part of the execution results to the short-term memory.

[0070] Mem = {ShortTerm: Resp...}

[0071] Step 4: Reflect and re-plan:

[0072] In order to improve generation accuracy and suppress the illusion of large models, the autonomous agent system needs to do a reflection Prompt before confirming the final result. The present invention uses the prompt word Prompt to allow the planning module to reflect by comparing the plan and the execution result, to determine whether the execution result meets expectations, or to re-plan.

[0073] Reflection=LLM(ReflectionPrompt(Resp,Plan init ))

[0074] Step 5: Process the reflection results:

[0075] First, when Reflection shows that the execution result is as expected, the autonomous agent system will directly return the execution result of the tool function Act.

[0076] In addition, when Reflection shows that the execution result does not meet expectations, the autonomous agent system will update the reflection to the short-term memory.

[0077] Mem={ShortTerm: Reflection...}

[0078] At the same time, enter step 2 of the planning process and call the current status of the planning module to execute the new plan.

[0079] Plan new =LLM(PlanPrompt(Reflection,Query,Chunk,...))

[0080] Afterwards, execute step 3 and step 4 in sequence to obtain the execution result of the update plan.

[0081] The present invention also includes an electronic device, including a memory and a processor coupled to each other, wherein the memory stores program data, and the processor calls the program data to execute the above method.

[0082] The present invention also includes a computer-readable storage medium on which program data is stored, and the program data implements the above method when executed by a processor.

[0083] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware.

[0084] Moreover, the present application may be in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The solutions in the embodiments of the present application may be implemented in various computer languages.

[0085] Although the present invention has been disclosed as above by way of embodiments, it is not intended to limit the present invention. Appropriate modifications or equivalent substitutions of the technical solutions of the present invention made by ordinary technicians in the field should all be included in the protection scope of the present invention. The protection scope of the present invention shall be based on what is defined in the claims.

Claims

1. An autonomous agent system based on retrieval-enhanced generation technology and generative large models, characterized in that: include: The knowledge base module is used to transform and separate the collected data sets in specific business fields and then input them into the knowledge base for storage; The memory module is used to store the questions and conversation contents input by users in multiple rounds of dialogue, as well as the associated knowledge recalled from the memory module itself and the knowledge base of the knowledge base module after retrieval through retrieval enhancement generation RAG; A planning module, used for formulating a plan list according to the questions and dialogue contents and associated knowledge input by the user obtained from the memory module; An execution module is used to obtain a plan execution result after executing the plan list obtained from the plan module, thereby judging whether the plan execution result meets expectations through reflection to obtain a reflection result, and storing the plan execution result and the reflection result in the memory module; The plan list with the expected plan execution results is directly output for execution, and the plan execution results that do not meet expectations are pushed to the planning module for re-planning until they meet expectations.

2. The autonomous agent system based on retrieval-enhanced generation technology and generative large model according to claim 1, characterized in that: The data set of the specific business field includes long text, tables and databases in the specific business field; the data set is uniformly converted into a text format, then divided into blocks according to delimiters and summary information of the blocks is generated, and each summary and its block text are used together to establish a knowledge base.

3. The autonomous agent system based on retrieval-enhanced generation technology and generative large model according to claim 1, characterized in that: The memory module includes a long-term memory submodule and a short-term memory submodule. The long-term memory submodule is used to store associated knowledge recalled from the memory module itself and the knowledge base of the knowledge base module after RAG retrieval is generated through retrieval enhancement. The short-term memory submodule is used to store questions and dialogue content input by users in multiple rounds of dialogue and the reflection results of the execution module.

4. The autonomous agent system based on retrieval-enhanced generation technology and generative large model according to claim 3 is characterized by: The planning module identifies the user's question intention through the prompt word Prompt according to the questions input by the user in multiple rounds of dialogue and the content of the dialogue, that is, obtains the user's request content, and combines the user's request content with the associated knowledge obtained from the memory module, and finally formulates a plan list; the user's request content is stored in the short-term memory submodule of the memory module, and the associated knowledge is stored in the long-term memory submodule of the memory module.

5. The autonomous agent system based on retrieval-enhanced generation technology and generative large model according to claim 1, characterized in that: The execution module is composed of a large model and a tool function. The execution module calls the large model to obtain the plan execution text according to the plan list fed back by the plan module, calls the tool function to obtain the plan execution result, adds the plan execution result to the prompt word Prompt in the execution module and then feeds it back to the plan module.

6. The processing method of the autonomous agent system according to any one of claims 1 to 5, characterized in that: include: Step 1: Based on the specific application scenario, collect data sets in specific business fields, and establish an industry knowledge base in the specific business field through data conversion, segmentation, summary generation and indexing, and then store it in the knowledge base module; Step 2: The long-term memory submodule in the memory module stores the associated knowledge recalled from the memory module itself and the knowledge base of the knowledge base module after retrieval through retrieval enhancement generation RAG, and the short-term memory submodule stores the questions and dialogue content input by the user in multiple rounds of dialogue; Step 3: The planning module uses the prompt word Prompt to identify the user's request content based on the questions and conversation content input by the user in multiple rounds of dialogue, and combines the user's request content with the associated knowledge obtained from the memory module to enhance the user's request content and finally formulate a plan list; the user's request content is stored in the short-term memory submodule of the memory module, and the associated knowledge is stored in the long-term memory submodule of the memory module; Step 4: The execution module accepts the specific plan in the plan list, executes the plan list according to the specific plan steps, calls the large model to obtain the plan execution text, calls the tool function to obtain the plan execution result, adds the plan execution result to the prompt word Prompt in the execution module and then feeds it back to the planning module; Step 5: The planning module receives the plan execution results from the execution module, determines whether the plan execution results meet expectations through reflection, and obtains reflection results. The plan list of the plan execution results that meet expectations is directly output for execution, and the plan execution results that do not meet expectations are pushed to the planning module for re-planning until they meet expectations; and the plan execution results and reflection results are stored in the memory module; Step 6: Loop through steps 4 and 6 until the final result is returned.

Citation Information

Patent Citations

  • RPA intelligent agent generated based on large model retrieval enhancement and interaction method thereof

    CN118313411A

  • Systems and methods for synchronizing and protecting data

    GB202406625D0

Cited By

  • User intention alignment robot task planning method based on large language model

    CN119658692A

  • Method for enhancing RPA digital employee planning ability based on retrieval enhancement generation mode

    CN119831017A

  • Web 0Day vulnerability detection and mining method and platform based on security agent

    CN119892422A

  • Data cleaning scheme generation method and device and readable storage medium

    CN121561262A