Purchase quotation generation method, system and device and storage medium

Abstract

The invention belongs to the technical field of purchase management, and provides a purchase quotation list generation method, system and device and a storage medium, and the method comprises the steps: carrying out the escape of a user demand, transmitting an escape result to a model, generating a purchase list, and carrying out the quotation; and then judging whether the quotation result meets the requirement, if so, outputting the quotation result, otherwise, transferring the meaning of the proposed new user requirement and outputting the result until the final quotation result meets the requirement. According to the method, the purchasing personnel can be helped to continuously and automatically generate new purchasing lists and quoted prices until the requirements of the purchasing personnel are met; in the process, information of a demand side can be automatically collected and sorted, required purchased commodities can be automatically sorted and classified, the commodities can be automatically selected, quoted prices of different suppliers can be compared, and a final quotation list can be automatically generated and sorted. In addition, results of retrieval, screening and list generation can be manually modified in the purchasing process, feedback can be continuously provided for the model, and the generation effect is continuously optimized.

Description

Technical Field

[0001] The present invention belongs to the technical field of procurement management, and in particular relates to a procurement quotation generation method, system, device and storage medium. Background Art

[0002] In the current procurement field, the process for suppliers to complete demand quotations is: collect and organize demander information, sales staff understand the demand and sort out the required purchase goods, sales staff select goods based on the purchase goods, and finalize the quotation based on the selection results. The following are the current situations:

[0003] 1. A lot of manual operations are required: manually collect and organize demander information, manually sort out and classify the required purchase goods, manually select products and compare quotations from different suppliers, and manually generate and organize the final quotation.

[0004] 2. Salesperson’s subjective experience and judgment: Salespersons understand demand and sort out products based on their personal experience. The product selection process relies on salesperson’s subjective judgment. Quotation compilation and price negotiation rely on salesperson’s experience and judgment.

[0005] 3. Lack of standardization and automation: Lack of standardization of information formats and document templates, insufficient utilization of automation tools and systems, and inconsistent processes and operating procedures.

[0006] In the process of generating purchase quotations, procurement companies currently rely on manual entry and search, which leads to low efficiency, a large impact of personal subjective factors, and difficulty in quickly responding to changes and fluctuations in customer demand. Existing quotation systems usually only support commodity screening functions, lack intelligent and automated support, and cannot fully utilize massive historical data and real-time market information to optimize procurement decisions. Even if some system software has the ability to generate, it can usually only be generated based on preset templates and rules, lacking flexibility and intelligence. Summary of the invention

[0007] In order to solve at least one problem in the background technology, the present invention provides a purchase quotation generation method, system, device and storage medium.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] A method for generating a purchase quotation includes the following steps:

[0010] Input the nth user demand, n≥1;

[0011] Escape the nth user demand and output the nth escape result;

[0012] If n=1, the following steps are performed in sequence based on the nth escape result:

[0013] Retrieval step: search the database through the retrieval model and output the product label information corresponding to the user's needs;

[0014] Screening step: obtain the product set corresponding to the product tag information through the screening model;

[0015] List step: Generate a purchase list through the product set;

[0016] If n>1, jump to the search step, screening step or list step based on the nth escape result, and continue to execute the next step until a purchase list is generated;

[0017] Make quotations based on purchase lists;

[0018] Determine whether the quotation result meets the requirements. If so, output the quotation result. Otherwise, escape the n+1th user requirement and output the escape result, and then continue to the next step until the final quotation result meets the requirements.

[0019] Preferably, the user demand is:

[0020] The user's initial needs, including the types and quantities of goods to be purchased;

[0021] Or, the user provides or modifies product label information and selects the product;

[0022] Or, the user gives or modifies the purchase list and needs to generate the corresponding optimal commodity quotation.

[0023] Preferably, searching in a database through a retrieval model and outputting product label information corresponding to user needs includes the following steps:

[0024] Input the nth escape result into the retrieval enhancement generation model;

[0025] Generate label classifications through retrieval-augmented generative models;

[0026] Retrieve and match corresponding product tag information based on tag classification;

[0027] Output the product label information, or manually modify the product label information before outputting it.

[0028] Preferably, manually modifying the product label information and then outputting it includes the following steps:

[0029] Calculate the scores of the output product label information, each product label information corresponds to a set score;

[0030] Input the score of the product label information into the reward model to obtain the reward parameter;

[0031] Optimize the product label information generation strategy based on reward parameters.

[0032] Preferably, the formula of the reward model is as follows:

[0033]

[0034] In the formula, x i is the label information of the i-th product; y i is the score of the i-th product label information; N is the number of samples of product label information; R θ (x i ) represents the reward model for product label information x i Ratings;

[0035] The formula for the optimized product label information generation strategy is as follows:

[0036]

[0037] Where πθ is the product label information generation strategy under the current model parameter θ; R θ (x) is the score of the reward model for the product label information x; Indicates the calculation of the gradient of the objective function J(θ) to the model parameter θ; E πθ represents the expected value of the product label information generation strategy; logπθ(x) represents the logarithmic probability of πθ on the product label information x.

[0038] Preferably, obtaining a product set corresponding to product tag information through a screening model includes the following steps:

[0039] Input product label information into the iFlytek Spark model;

[0040] Generate structured query language through iFlytek Spark model, and query based on structured query language to obtain the product set corresponding to product tag information;

[0041] Export the product set, or manually modify the product set before exporting it.

[0042] Preferably, manually modifying the product set and then outputting it includes the following steps:

[0043] Convert manually modified data into reward signals;

[0044] Input the reward signal into the satisfaction model;

[0045] Optimize product set generation strategy based on satisfaction model.

[0046] Preferably, the formula of the satisfaction model is as follows:

[0047]

[0048] In the formula, r i is the satisfaction after the i-th manual modification, and M is the total number of manual modification operations;

[0049] The formula for the optimized product set generation strategy is as follows:

[0050]

[0051] Where A(x') represents the advantage of the current product set generation strategy over the benchmark strategy; πδ is the product set generation strategy under the current model parameter δ; Indicates the calculation of the gradient of the objective function J(δ) to the model parameter δ; E πδ represents the expected value of the product set generation strategy πδ; logπδ(x') represents the logarithmic probability of πδ on the product set x'.

[0052] Preferably, generating a purchase list through a commodity set includes the following steps:

[0053] Combine past purchase records with this set of goods to generate a preliminary purchase list.

[0054] Preferably, if n>1, jumping to the search step, the screening step or the list step based on the nth escape result includes the following steps:

[0055] When n is greater than 1, if the escape result is the product and quantity that the user initially needs to purchase, jump to the search step, and execute the screening step and the list step in sequence;

[0056] If the result of the escape is that the user modifies the product label information and selects the product, jump to the screening step and execute the list step;

[0057] If the result of the escape is that the user modifies the purchase list and needs to generate the corresponding optimal product quotation, jump to the list step.

[0058] Preferably, making a quotation based on the purchase list includes the following steps:

[0059] Generate commodity datasets through purchase lists;

[0060] Set the quotation type, including customer acquisition type, high-end type and money-saving type;

[0061] Based on the product data set, quotes are obtained to obtain customer acquisition quotation sheets, high-end quotation sheets, and money-saving quotation sheets;

[0062] Select one or more outputs from customer acquisition quotation, high-end quotation, and money-saving quotation.

[0063] Preferably, it is determined whether the quotation result meets the requirement. If so, the quotation result is output. Otherwise, the n+1th user requirement is escaped and the result is output, and then the next step is continued until the final quotation result meets the requirement, including the following steps:

[0064] Compare the quotation result with the preset price range. If the quotation result exceeds the preset price range or is lower than the preset price range, escape the n+1th user demand and output the result, and then continue to the next step until the final quotation result meets the demand; otherwise, output the quotation result.

[0065] Preferably, before escaping the first user demand, the following steps are also included:

[0066] Record the information of past completed quotations;

[0067] For the existing product library data, the products are labeled based on their attributes and descriptions to obtain a collection of product label information.

[0068] A purchase quotation generation system, comprising:

[0069] Input unit, used to input the nth user demand, n≥1;

[0070] An escaping unit, used for escaping the nth user demand and outputting the nth escaping result;

[0071] A jump unit, for jumping to the search unit, the screening unit and the list unit in sequence to execute corresponding steps based on the nth escape result when n=1; and for jumping to the search unit, the screening unit or the list unit to execute corresponding steps based on the nth escape result when n>1;

[0072] A retrieval unit, used to search in the database through the retrieval model and output product label information corresponding to the user's needs;

[0073] A screening unit, used to obtain a product set corresponding to product tag information through a screening model;

[0074] List unit, used to generate a purchase list from a commodity set;

[0075] The quotation unit is used to make quotations based on the purchase list;

[0076] The judgment unit is used to judge whether the quotation result meets the requirements. If so, the quotation result is output. Otherwise, the n+1th user requirement is escaped and the result is output, and then the next step is executed until the final quotation result meets the requirements.

[0077] Preferably, the retrieval unit comprises:

[0078] A first retrieval module is used to input the nth escape result into the retrieval enhancement generation model;

[0079] The second retrieval module is used to generate label classification by retrieving the enhanced generation model;

[0080] The third search module is used to search and match the corresponding product tag information based on the tag classification;

[0081] The fourth search module is used to output product label information;

[0082] The first modification module is used to manually modify the product label information before outputting it.

[0083] Preferably, the screening unit comprises:

[0084] The first screening module is used to input product label information into the iFlytek Spark model;

[0085] The second screening module is used to generate a structured query language through the iFlytek Spark model, and perform a query based on the structured query language to obtain a product set corresponding to the product tag information;

[0086] The third screening module is used to output the product set;

[0087] The second modification module is used to manually modify the product set before output.

[0088] Preferably, the inventory unit includes:

[0089] The list module is used to generate a preliminary purchase list by combining past purchase records with the current set of goods.

[0090] A device comprising:

[0091] Memory, used to store computer programs;

[0092] The processor is used to implement the above-mentioned method for generating a purchase quotation when executing the program stored in the memory.

[0093] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned method for generating a purchase quotation.

[0094] Beneficial effects of the present invention:

[0095] 1. The present invention sets a purchasing agent to identify user needs, and then executes corresponding steps according to the user needs, generates a purchase list and makes a quotation. Then the purchasing staff can choose whether to re-enter new user needs according to the quotation results. The agent continues to perform corresponding operations according to the new user needs until no new user needs are input. This process can help the purchasing staff to continuously and automatically generate new purchase lists and quotations until the needs of the purchasing staff are met; this process can automatically collect and organize the demand side information, automatically sort out and classify the required purchase goods, automatically select and compare the quotations of different suppliers, and automatically generate and organize the final quotation sheet;

[0096] 2. The present invention can manually modify the results of retrieval, screening and list generation during the procurement process. This process introduces reinforcement learning technology based on human feedback, which can continuously provide feedback to the model, continuously improve the model recall rate, and continuously optimize the generation effect;

[0097] 3. The present invention uses a retrieval generation model to perform data retrieval and generation, thereby improving the accuracy of obtaining related products from the product library and solving the shortcomings of rule matching and screening products. In addition, by training historical user needs and product list data into model parameters as a reference for subsequent model quotations, the accuracy and practicality of quotation generation are further improved, solving the problem of subjective experience judgment of salesmen.

[0098] Other features and advantages of the present invention will be described in the following description, and partly become obvious from the description, or be understood by implementing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0100] Figure 1 A flow chart of a method for generating a purchase quotation of the present invention is shown;

[0101] Figure 2 A schematic diagram of the operation of a method for generating a purchase quotation of the present invention is shown;

[0102] Figure 3 A framework diagram of a purchase quotation generation system of the present invention is shown. DETAILED DESCRIPTION

[0103] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0104] A method for generating a purchase quotation, such as Figure 1 As shown, the following steps are included:

[0105] S1: input the nth user demand, escape the nth user demand, and output the nth escape result, n≥1; specifically, first input the nth user demand into the procurement agent, then escape the nth user demand through the procurement agent, and output the nth escape result;

[0106] If n=1, then S2-S4 are performed in sequence based on the result of the nth escape, followed by S5 and S6:

[0107] S2: Retrieval step: search the database through the retrieval model and output the product label information corresponding to the user's needs;

[0108] S3: Screening step: obtaining a product set corresponding to product tag information through the screening model;

[0109] S4: List step: Generate a purchase list through the commodity set;

[0110] If n>1, jump to the search step, screening step or list step based on the nth escape result, and continue to execute the next step until a purchase list is generated;

[0111] S5: Make a quotation based on the purchase list;

[0112] S6: Determine whether the quotation result meets the requirements. If so, output the quotation result. Otherwise, input the n+1th user requirement into the procurement agent for translation and output the result, and then continue to the next step until the final quotation result meets the requirements.

[0113] It should be noted that the purchased agent is a computing system that can autonomously perceive the environment, make decisions and perform actions. The core characteristics of agents include autonomy, adaptability, interactivity and goal orientation. They are often used in the fields of complex system control, data analysis, automation tasks and human-computer interaction.

[0114] Autonomy: Agents are able to operate independently without human intervention. They can make decisions and take actions based on current environmental information according to preset rules or learned knowledge.

[0115] Adaptability: The agent has the ability to learn and adapt. Through machine learning and artificial intelligence technology, the agent can obtain feedback from the environment, adjust its own behavior, and optimize the performance of tasks.

[0116] Interactivity: Agents can interact with other agents or humans. They are able to understand and process information from multiple sources and work together in multi-agent systems to complete complex tasks.

[0117] Goal-oriented: Agents are usually designed to achieve a specific goal or task. They are able to plan and execute a series of actions in a dynamic and uncertain environment to achieve their goals efficiently.

[0118] It should be further explained that the present invention uses an intelligent agent to identify user needs, and then executes corresponding steps according to the user needs, generates a purchase list and makes a quotation. The purchasing personnel can then choose whether to re-enter new user needs based on the quotation results. The intelligent agent continues to execute corresponding operations according to the new user needs until no new user needs are input. This process can help purchasing personnel to continuously and automatically generate new purchase lists and quotations until the purchasing personnel's needs are met.

[0119] In addition, the present invention proposes a method for automatically generating purchase quotations based on the fusion of large model retrieval and generation technology (RAG) and multi-agent system (MAS). The method aims to realize the automatic generation of quotations, improve generation efficiency and accuracy, reduce manual screening and intervention, and optimize customer demand analysis, commodity screening and quotation processes in procurement. In addition, the method introduces the human reinforcement feedback learning technology (RLHF) to integrate the scoring mechanism of the reward model into the process. The model output is scored by procurement experts to continuously guide the model to generate outputs that meet the expert experience preferences. Ultimately, the model can learn and accumulate the subjective knowledge and experience of procurement experts to avoid experience loss. At the same time, a procurement quotation agent is designed and constructed to further realize the intelligence and automation of the procurement process. In addition, the automated processing process reduces manual intervention, reduces the error rate, supports manual final decision-making assistance, and improves the work efficiency of the entire procurement process quotation. The system can not only automatically generate preliminary quotations, but also provide decision-making assistance functions to help purchasing personnel quickly review and confirm the final quotations.

[0120] Combine the following Figure 2 The process of S1-S6 is further explained.

[0121] Further, in Figure 2 The purchasing agent is mainly responsible for identifying the demand description information entered by the user, then translating it according to different demands, dispatching other agents to complete tasks, and judging and optimizing the results of the entire process. The agent's translation results for user demands include the following three situations:

[0122] The first case: the user describes the most original demand and describes the information. The demand is the type and quantity of the purchased goods, that is, describing the purchase demand.

[0123] The second situation: the user provides or modifies product label information and needs to select the product.

[0124] The third situation: The user has given or modified a specific purchase list and needs to generate the corresponding optimal commodity quotation.

[0125] Furthermore, the agent is implemented using a large model (QWen-72B), which deconstructs requirements based on user input, executes intent distribution, downstream function scheduling, and performs further scheduling based on function return results. The specific steps are described in detail below.

[0126] In this agent, the user demand is set as Q, that is, Q = the demand described by the user. The QWen-72B model completes the intention recognition based on the demand and determines the demand type I, which corresponds to the three escape results mentioned above.

[0127] Further, S2 comprises the following steps:

[0128] S201: Input the nth escape result into the Retrieval-Augmented Generation (RAG) model.

[0129] S202: Generate label classification through retrieval enhancement generation model.

[0130] S202: Retrieve corresponding product tag information based on tag classification.

[0131] S204: Output the product label information, or manually modify the product label information before outputting it.

[0132] It should be noted that, combined with Figure 2 It can be seen that if there is a document left for the product label information, output results and other files in S2, the document can be stored in the document library.

[0133] It should be further explained that the RAG model can be obtained by fine-tuning the QWen-72B model. In S2, the model can match the most relevant product label information for label classification. It can use the classification model in the QWen-72B model for detailed classification to obtain the classification label L, that is, L = f 分类模型 (Q), where f 分类模型 It is a function used for classification in the QWen-72B model, and outputs a classification label L, such as desktop office supplies, laptop peripherals, electronic products, souvenirs, employee New Year goods, etc. Then, the product label information T can be obtained according to L, which can be searched according to the search function in the QWen-72B model to obtain the product label information T corresponding to the classification label L.

[0134] Furthermore, in S204, the product label information may be manually modified before being output, which includes the following steps:

[0135] S2041: Calculate the score of the output product label information, each product label information corresponds to a set score; S2042: Input the score of the product label information into the reward model to obtain the reward parameter; S2043: Optimize the label information generation strategy based on the reward parameter.

[0136] It should be noted that in S204, a large model (such as RAG) can be used to retrieve the input information to obtain a preliminary set of label information. Then, each label is scored according to the set score of the label, for example, the score range can be 1 to 5 points. Then, a reward model is constructed and trained using the scoring data. The goal of the reward model is to predict the quality score of the label information.

[0137] Specifically, the formula of the reward model trained using supervised learning method is as follows:

[0138]

[0139] In the formula, x i is the label information of the i-th product, y i is the score of the i-th product label information, N is the number of samples of product label information; R θ (x i ) represents the reward model for product label information x i The score of the label information is L(θ), which is used to measure the difference between the model's prediction results and the actual results. In formula (1), the loss function is used to train the reward model so that the model can more accurately predict the quality score of the label information.

[0140] The formula for the product label information generation strategy optimized using the Proximal Policy Optimization (PPO) method in reinforcement learning is as follows:

[0141]

[0142] In the formula, πθ is the product label information generation strategy under the current model parameter θ; R θ (x) is the score of the reward model for the product label information x; Indicates the calculation of the gradient of the objective function J(θ) to the model parameter θ; E πθ represents the expected value of the product label information generation strategy; logπθ(x) represents the logarithmic probability of πθ on the product label information x.

[0143] It should be noted that in the task of generating product labels, πθ determines which product label information should be generated under each product state (such as the current attributes of the product, the user's behavior data, etc.). For example, for a product that a user browses on an e-commerce platform, πθ may determine which product label information should be recommended (such as "hot sale", "discount", "new product", etc.) to attract the user's attention.

[0144] Further, S3 includes the following steps:

[0145] S301: Input the product label information into the iFlytek SparkDesk model.

[0146] S302: Generate a structured query language through the SparkDesk model, and perform a query based on the structured query language to obtain a product set corresponding to the product tag information, wherein the structured query language is S = "SELECT * FROM products WHERE category = " + T.

[0147] S303: Output the product set, or manually modify the product set and then output it. Specifically, this step executes the above structured query language, and finally obtains the product list P of the product set, where P = execution (S). For example, filter out products under corresponding tags, such as notebooks, markers, blackboards, napkins, etc.

[0148] It should be noted that the iFlytek SparkDesk large model training and fine-tuning is used in S3 to achieve the ability to generate corresponding SQL (Structured Query Language) statements based on user demand description Q and product tag information T, and filter out the corresponding products under the tag. The tag classification here can be generated by the upstream RAG (Retrieval-Augmented Generation) model or organized by the user.

[0149] Furthermore, in S303, the product label information is manually modified and then output, including the following steps:

[0150] S3031: Convert manually modified data into reward signals; S3032: Input the reward signals into the satisfaction model; S3033: Optimize the product set generation strategy based on the satisfaction model.

[0151] It should be noted that in S303, a preliminary product set result is generated according to the tag information. The user is supported to manually intervene in the product set result, including adding and removing some products. Finally, the user's intervention operation data (adding and removing some products) is recorded as a satisfaction signal in reinforcement learning.

[0152] Furthermore, the user's intervention operation data is converted into a satisfaction signal. The satisfaction of adding a product operation is positive, and the satisfaction of removing a product operation is negative. The satisfaction signal r represents the user satisfaction and is defined as:

[0153]

[0154] In the formula, r i is the satisfaction after the i-th manual modification, and M is the total number of manual modification operations;

[0155] The formula for the product set generation strategy optimized using the reinforcement learning algorithm (A2C, Advantage Actor-Critic) is as follows:

[0156]

[0157] Where A(x') represents the advantage of the current product set generation strategy over the benchmark strategy; πδ is the product set generation strategy under the current model parameter δ; Indicates the calculation of the gradient of the objective function J(δ) to the model parameter δ; E πδ represents the expected value of the product set generation strategy πδ; logπδ(x') represents the logarithmic probability of πδ on the product set x'.

[0158] It should be noted that in S2 and S3, when manually modifying the large model generation part in the link, reinforcement learning technology based on human feedback (RLHF, Reinforcement Learning from Human Feedback) was introduced to continuously provide feedback to the model, continuously improve the model recall rate, and continuously optimize the generation effect.

[0159] Furthermore, S4 generates a purchase list through the commodity set, including the following steps:

[0160] S401: Generate a preliminary purchase list based on past purchase records and this commodity set.

[0161] It should be noted that when generating a purchase list, RAG technology is used to retrieve past records and generate Top-K related data. Among them, RAG retrieval will retrieve similar quotation records in the document library based on the user's demand description and product list to obtain Top-K (indicates the first k data sorted by relevance or matching) data. Then splice the Top-K data into Prompt (template information) to generate the initial screening results. Among them, the generation model will combine the user's demand description and the retrieved Top-K data to generate a preliminary purchase list.

[0162] Furthermore, if n>1, jumping to the search step, the screening step or the list step based on the nth escape result includes the following steps:

[0163] When n is greater than 1, if the escape result is the user's initial product and quantity to be purchased, jump to the search step and execute the screening step and list step in sequence; if the escape result is that the user modifies the product label information and selects the product, jump to the screening step and execute the list step; if the escape result is that the user modifies the purchase list and needs to generate the corresponding optimal product quotation, jump to the list step.

[0164] Further, S5 includes the following steps:

[0165] S501: Generate a commodity data set through a purchase list; S502: Set the quotation type, including customer acquisition type, high-end type and money-saving type; S503: Make a quotation based on the commodity data set to obtain a customer acquisition type quotation, a high-end type quotation and a money-saving type quotation; S504: Select one or more outputs from the customer acquisition type quotation, the high-end type quotation and the money-saving type quotation.

[0166] It should be noted that the goal of customer acquisition is to attract new customers, and price-sensitive products may be discounted. Premium configuration is to provide high-end configurations, highlighting high value and high quality. Economy configuration is to save costs and provide affordable options for price-sensitive customers.

[0167] Furthermore, when making a customer acquisition quotation, the formula is as follows:

[0168] discount(c i )=α i base_price(c i ); (5)

[0169] In the formula, discount means discount; c i represents the product set; α i Indicates the discount coefficient of the product, the value is between 0.1 and 0.3; base_price(c i ) represents the base price of the commodity;

[0170] When making a quotation for a high-end model, the formula is as follows:

[0171] Premium_price(c i )=base_price(c i )+β i ·value_added(c i ); (6)

[0172] Where Premium_price(c i ) represents the premium price; β i Indicates c i The added value coefficient; value_added(c i ) indicates the added price of a commodity;

[0173] When making a money-saving quote, the formula is as follows:

[0174] Economy_price(c i )=base_price(c i )-γ i base_price(c i ); (7)

[0175] Where, Economy_price(c i ) represents economic price; γ i Indicates c i The discount factor.

[0176] Further, S6 includes the following steps:

[0177] Compare the quotation result with the preset price range. If the quotation result exceeds the preset price range or is lower than the preset price range, the n+1th user demand is input into the procurement agent, and the next step is executed until the final quotation result meets the demand; otherwise, the quotation result is output.

[0178] It should be noted that the current and previous quotation results and business data can be stored in the storage database. In addition, during the judgment process, the user can reject one or all quotations, write down the reason after rejecting, and then let the model generate a new quotation, and continue to cycle until the desired quotation is obtained.

[0179] It should be noted that a salesperson feedback mechanism can be introduced for the quotation. Generally, the salesperson will give feedback on the generated quotation: too high, too low, or appropriate. Then adjust the weight value:

[0180] If the feedback is "too high", reduce the weight value w: ω←ω-τ;

[0181] If the feedback is "too low", increase the weight value w: ω←ω+τ;

[0182] If the feedback is "appropriate", the weight value remains unchanged, where τ is the adjustment step size, which will be dynamically set according to the specific quotation requirements.

[0183] It should be further explained that the weight value U can be updated through reinforcement learning, and the formula is as follows:

[0184]

[0185] Where s is the current state (current quotation configuration); a is the current action (adjusted weight value); r is the current satisfaction (salesman feedback); s′ is the next state (new quotation configuration); a′ is the next action (new adjusted weight value); α is the learning rate; γ is the discount factor.

[0186] Furthermore, before escaping the first user demand, the following steps are also included:

[0187] First, the information of the completed quotations in the past is recorded, which is recorded as T = {(Q1, R1), (Q2, R2), ..., (Q n ,R n )}, where Q1~Qn represents the customer demand description, R1~Rn represents the purchase commodity list and quotation information; then the existing commodity library data is labeled based on the attributes and descriptions of the commodities to obtain the set of commodity label information G i ={T i1 , T i2 , ..., T ik}, where G i represents the set of label information of the i-th product, T i1 ~T ik Indicates that in G i After processing the product and its label information, store it in Figure 2 In the graph database, the big model will filter and match the data based on the generated label results.

[0188] like Figure 3 As shown, a purchase quotation generation system includes an input unit, an escape unit, a jump unit, a search unit, a screening unit, a list unit, a quotation unit and a judgment unit. The input unit is used to input the nth user demand, n≥1; the escape unit is used to escape the nth user demand and output the nth escape result, n≥1.

[0189] The jump unit is used to jump to the search unit, the screening unit and the list unit in sequence based on the nth escape result when n=1; and to jump to the search step, the screening step or the list step based on the nth escape result when n>1.

[0190] The retrieval unit is used to search in the database through the retrieval model and output the product label information corresponding to the user's needs; the screening unit is used to obtain the product set corresponding to the product label information through the screening model; the list unit is used to generate a purchase list through the product set; the quotation unit is used to quote based on the purchase list; the judgment unit is used to judge whether the quotation result meets the needs, and if so, output the quotation result; otherwise, escape the n+1th user demand and output the result, and then continue to the next step until the final quotation result meets the needs.

[0191] The search unit includes:

[0192] A first retrieval module is used to input the nth escape result into the retrieval enhancement generation model;

[0193] The second retrieval module is used to generate label classification by adding a generation model through retrieval;

[0194] The third search module is used to search and match the corresponding product tag information based on the tag classification;

[0195] The fourth search module is used to output product label information;

[0196] The first modification module is used to manually modify the product label information before outputting it.

[0197] The screening unit includes:

[0198] The first screening module is used to input product label information into the SparkDesk model;

[0199] The second screening module is used to generate a structured query language through the SparkDesk model, and perform a query based on the structured query language to obtain a product set corresponding to the product tag information;

[0200] The third screening module is used to output the product set;

[0201] The second modification module is used to manually modify the product set before output.

[0202] Listing units include:

[0203] The list module is used to generate a preliminary purchase list by combining past purchase records with the current set of goods.

[0204] A device comprising:

[0205] Memory, used to store computer programs;

[0206] The processor is used to execute the program stored in the memory to achieve Figure 1 A method for generating a purchase quotation.

[0207] A computer-readable storage medium having a computer program stored thereon, which when executed by a processor implements Figure 1 A method for generating a purchase quotation.

[0208] Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent substitutions for some of the technical features therein; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating a purchase quotation, characterized in that: The following steps are involved: Input the nth user demand, n≥1; Escape the nth user demand and output the nth escape result; If n=1, the following steps are performed in sequence based on the nth escape result: Retrieval step: search the database through the retrieval model and output the product label information corresponding to the user's needs; Screening step: obtain the product set corresponding to the product tag information through the screening model; List step: Generate a purchase list through the product set; If n>1, jump to the search step, screening step or list step based on the nth escape result, and continue to execute the next step until a purchase list is generated; Make quotations based on purchase lists; Determine whether the quotation result meets the requirements. If so, output the quotation result. Otherwise, escape the n+1th user requirement and output the escape result, and then continue to the next step until the final quotation result meets the requirements.

2. A method for generating a purchase quotation according to claim 1, characterized in that: The user requirements are: The user's initial needs, including the types and quantities of goods to be purchased; Or, the user provides or modifies product label information and selects the product; Or, the user gives or modifies the purchase list and needs to generate the corresponding optimal commodity quotation.

3. A method for generating a purchase quotation according to claim 1, characterized in that: Searching in the database through the retrieval model and outputting the product label information corresponding to the user's needs includes the following steps: Input the nth escape result into the retrieval enhancement generation model; Generate label classifications through retrieval-augmented generative models; Retrieve and match corresponding product tag information based on tag classification; Output the product label information, or manually modify the product label information before outputting it.

4. A method for generating a purchase quotation according to claim 3, characterized in that: Manually modify the product label information and then output it, including the following steps: Calculate the scores of the output product label information, each product label information corresponds to a set score; Input the score of the product label information into the reward model to obtain the reward parameter; Optimize the product label information generation strategy based on reward parameters.

5. A method for generating a purchase quotation according to claim 4, characterized in that: The formula of the reward model is as follows: In the formula, x i is the label information of the i-th product; y i is the score of the i-th product label information; N is the number of samples of product label information; R θ (x i ) represents the reward model for product label information x i Ratings; The formula for the optimized product label information generation strategy is as follows: Where πθ is the product label information generation strategy under the current model parameter θ; R θ (x) is the score of the reward model for the product label information x; Indicates the calculation of the gradient of the objective function J(θ) to the model parameter θ; E πθ represents the expected value of the product label information generation strategy; logπθ(x) represents the logarithmic probability of πθ on the product label information x.

6. A method for generating a purchase quotation according to claim 1, characterized in that: Obtaining a product set corresponding to product tag information through a screening model includes the following steps: Input product label information into the iFlytek Spark model; Generate structured query language through iFlytek Spark model, and query based on structured query language to obtain the product set corresponding to product tag information; Export the product set, or manually modify the product set before exporting it.

7. A method for generating a purchase quotation according to claim 6, characterized in that: Manually modify the product set and then output it, including the following steps: Convert manually modified data into reward signals; Input the reward signal into the satisfaction model; Optimize product set generation strategy based on satisfaction model.

8. A method for generating a purchase quotation according to claim 7, characterized in that: The formula of the satisfaction model is as follows: In the formula, r i is the satisfaction after the i-th manual modification, and M is the total number of manual modification operations; The formula for the optimized product set generation strategy is as follows: Where A(x') represents the advantage of the current product set generation strategy over the benchmark strategy; πδ is the product set generation strategy under the current model parameter δ; Indicates the calculation of the gradient of the objective function J(δ) to the model parameter δ; E πδ represents the expected value of the product set generation strategy πδ; logπδ(x,) represents the logarithmic probability of πδ on the product set x,.

9. A method for generating a purchase quotation according to claim 1, characterized in that: Generating a purchase list from a commodity set includes the following steps: Combine past purchase records with this set of goods to generate a preliminary purchase list.

10. A method for generating a purchase quotation according to claim 9, characterized in that: If n>1, jump to the search step, screening step or list step based on the nth escape result, including the following steps: When n is greater than 1, if the result of the escape is the product and quantity that the user initially needs to purchase, then jump to the search step, and execute the screening step and the list step in sequence; If the result of the escape is that the user modifies the product label information and selects the product, jump to the screening step and execute the list step; If the result of the escape is that the user modifies the purchase list and needs to generate the corresponding optimal product quotation, jump to the list step.

11. A method for generating a purchase quotation according to claim 1, characterized in that: Producing a quotation based on a purchase list includes the following steps: Generate commodity datasets through purchase lists; Set the quotation type, including customer acquisition type, high-end type and money-saving type; Based on the product data set, quotes are obtained to obtain customer acquisition quotation sheets, high-end quotation sheets, and money-saving quotation sheets; Select one or more outputs from customer acquisition quotation, high-end quotation, and money-saving quotation.

12. A method for generating a purchase quotation according to claim 1, characterized in that: Determine whether the quotation result meets the requirements. If so, output the quotation result. Otherwise, escape the n+1th user requirement and output the result, and then continue to the next step until the final quotation result meets the requirements, including the following steps: Compare the quotation result with the preset price range. If the quotation result exceeds the preset price range or is lower than the preset price range, escape the n+1th user demand and output the result, and then continue to the next step until the final quotation result meets the demand; otherwise, output the quotation result.

13. A method for generating a purchase quotation according to any one of claims 1 to 12, characterized in that: Before escaping the first user demand, the following steps are also included: Record the information of past completed quotations; For the existing product library data, the products are labeled based on their attributes and descriptions to obtain a collection of product label information.

14. A purchase quotation generation system, characterized in that: include: Input unit, used to input the nth user demand, n≥1; An escaping unit, used for escaping the nth user demand and outputting the nth escaping result; A jump unit, used for jumping to the search unit, the screening unit and the list unit in sequence to execute corresponding steps based on the nth escape result when n=1; And, when n>1, jump to the search unit, the screening unit or the list unit to execute the corresponding steps based on the nth escape result; A retrieval unit, used to search in the database through the retrieval model and output product label information corresponding to the user's needs; A screening unit, used to obtain a product set corresponding to product tag information through a screening model; List unit, used to generate a purchase list from a commodity set; The quotation unit is used to make quotations based on the purchase list; The judgment unit is used to judge whether the quotation result meets the requirements. If so, the quotation result is output. Otherwise, the n+1th user requirement is escaped and the result is output, and then the next step is executed until the final quotation result meets the requirements.

15. A purchase quotation generation system according to claim 14, characterized in that: The retrieval unit comprises: A first retrieval module is used to input the nth escape result into the retrieval enhancement generation model; The second retrieval module is used to generate label classification by retrieving the enhanced generation model; The third search module is used to search and match the corresponding product tag information based on the tag classification; The fourth search module is used to output product label information; The first modification module is used to manually modify the product label information before outputting it.

16. A purchase quotation generation system according to claim 14, characterized in that: The screening unit comprises: The first screening module is used to input product label information into the iFlytek Spark model; The second screening module is used to generate a structured query language through the iFlytek Spark model, and perform a query based on the structured query language to obtain a product set corresponding to the product tag information; The third screening module is used to output the product set; The second modification module is used to manually modify the product set before output.

17. A purchase quotation generation system according to claim 14, characterized in that: The inventory unit includes: The list module is used to generate a preliminary purchase list by combining past purchase records with the current set of goods.

18. A device, characterized in that include: Memory, used to store computer programs; A processor, for implementing a method for generating a purchase quotation as described in any one of claims 1 to 13 when executing a program stored in a memory.

19. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, a method for generating a purchase quotation as described in any one of claims 1 to 13 is implemented.

Images