A supply-demand matching method and system based on a large model

Abstract

This application relates to the field of electronic digital data processing technology, and in particular to a supply and demand matching method and system based on a large model. The method includes: performing intent recognition on a user-inputted question based on a large model to obtain the user's intent; processing the user-inputted question according to the user's intent to obtain a refined question; obtaining a first prompt word based on the user-inputted question and inputting it into the large model to generate a solution corresponding to the user-inputted question; extracting tags from the refined question based on the large model to obtain a tag set for the refined question; performing multi-path recall and rearranging based on the solution corresponding to the user-inputted question and the refined question, returning the top preset number of documents in the rearranged result; and inputting a second prompt word into the large model so that the large model presents the top preset number of documents in the rearranged result to the user in a multimodal manner based on the second prompt word. This invention can improve the accuracy of supply and demand matching.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to a supply and demand matching method and system based on a large model. Background Technology

[0002] In various fields of modern society (such as the commercialization of scientific and technological achievements), accurate matching of supply and demand is a core element in improving resource allocation efficiency and optimizing service experience. The efficient connection between supply and demand directly impacts industry operational efficiency and user satisfaction. However, traditional supply and demand matching methods generally suffer from the following problems: traditional matching methods often rely on simple keyword searches or rule engines, lacking in-depth analysis of the demands or problems of the demand side, resulting in a large amount of invalid information with low matching accuracy. How to improve the accuracy of supply and demand matching is an urgent problem to be solved. Summary of the Invention

[0003] The purpose of this invention is to provide a supply and demand matching method and system based on a large model, so as to improve the accuracy of supply and demand matching.

[0004] According to a first aspect of the present invention, a supply and demand matching method based on a large model is provided, the method comprising the following steps:

[0005] S100 performs intent recognition on user-input questions based on a large model to obtain user intent.

[0006] S200 processes the user's input question based on the user's intent, removes irrelevant information, and extracts the refined question.

[0007] S300: Obtain the first prompt word based on the question input by the user, input the first prompt word into the large model, and generate a solution corresponding to the question input by the user.

[0008] S400, Based on the large model, extract labels from the refined questions to obtain a set of labels for the refined questions.

[0009] S500: Based on the solution corresponding to the user-input question and the extracted question, perform multi-way recall and rearrangement, and return the first preset number of documents of the rearranged result; the tag set of any returned document has an intersection with the tag set of the extracted question.

[0010] S600, input the second prompt word into the large model so that the large model presents the first preset number of documents of the rearranged results to the user in a multimodal manner based on the second prompt word.

[0011] According to a second aspect of the present invention, a supply and demand matching system based on a large model is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described supply and demand matching method based on a large model.

[0012] Compared with the prior art, the present invention has at least the following beneficial effects:

[0013] This invention first identifies the intent of the user's input question, accurately pinpointing the user's core purpose. Based on the user's intent, the input question is processed to remove redundant information, resulting in a refined question. A set of tags for the refined question is then obtained. A solution to the user's input question is obtained based on a large model. Based on the solution and the set of tags for the refined question, documents with semantically similar meanings and the same tags are searched from a pre-set database. This improves the accuracy of the output document and ensures a strong correlation between the final matched document and the core purpose of the need, thus improving the accuracy of supply and demand matching. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A flowchart illustrating the supply and demand matching method based on a large model provided in Embodiment 1 of the present invention;

[0016] Figure 2 A flowchart illustrating the process of obtaining user intent provided in Embodiment 1 of the present invention;

[0017] Figure 3 This is a flowchart of the process for obtaining and refining the product according to Embodiment 1 of the present invention;

[0018] Figure 4 A flowchart illustrating the process of obtaining a solution provided in Embodiment 1 of the present invention;

[0019] Figure 5 This is a flowchart illustrating the process of obtaining a tag set as provided in Embodiment 1 of the present invention;

[0020] Figure 6 This is a flowchart of the process for obtaining a preset number of documents with rearranged results, as provided in Embodiment 1 of the present invention.

[0021] Figure 7 The flowchart illustrates the process presented in Embodiment 1 of the present invention. Detailed Implementation

[0022] 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.

[0023] Example 1:

[0024] According to this embodiment, as Figure 1 As shown, a supply and demand matching method based on a large model is provided, the method including the following steps:

[0025] S100 performs intent recognition on user-input questions based on a large model to obtain user intent.

[0026] As one specific implementation, S100 includes, for example Figure 2 As shown:

[0027] S110, if the user does not input a question from the module with the specified intent, then based on the large model, a judgment is made according to the question input by the user to obtain which type of intent in the preset intent list the user's intent is. The preset intent list includes several types of intents and other intents.

[0028] In this embodiment, the preset intent list is a pre-constructed list. As a specific implementation, this embodiment is applied to the field of technology transfer, and the preset intent list includes: seeking results for technical problems, seeking results without technical problems, seeking technology transfer for existing results, seeking investment, seeking experts, industry research, writing business plans, and other intents.

[0029] As an optional specific implementation, based on the large model, the user's intention is determined by judging the question input by the user and which intention in the preset intention list it is. The intention with the highest probability obtained after inputting the user's question into the large model is determined as the user's intention.

[0030] As a preferred implementation, based on the large model, the user's input question is judged to determine which of the following is the user's intent from a preset intent list:

[0031] S111: Input the user's question into the large model and obtain the analysis text, several intents, and corresponding probabilities output by the large model.

[0032] As a specific implementation method, the user inputs a question into a large model, which is required to generate intents based on the user input questions, in the format of an intent list (including probabilities).

[0033] S112, if the probability of an intent output by the large model is greater than or equal to a preset probability threshold, then the intent is judged as a user intent; otherwise, keywords are extracted from the analysis text output by the large model, and several intents and their corresponding probabilities are obtained again based on the extracted keywords, and then proceed to S113.

[0034] Optionally, the preset probability threshold is an empirical value, for example, the preset probability threshold is 0.7.

[0035] Those skilled in the art will recognize that the process of extracting keywords is existing technology and will not be described in detail here.

[0036] As a specific implementation method, re-acquiring several intents and corresponding probabilities based on the extracted keywords includes: integrating the extracted keywords into prompt words, inputting them into a large model, and requiring it to regenerate more accurate intents based on the keywords, with the format still being: a list of intents (including probabilities).

[0037] S113, use the dialogue templates corresponding to the reacquired intents in descending order of probability to interact with the user.

[0038] In this embodiment, the dialogue templates corresponding to each intent are preset. Based on the dialogue template corresponding to any intent, it can be further determined whether the user's intent is that intent. For example, this embodiment is applied to the field of technology transfer. When the intent is to find results for a technical problem, the corresponding dialogue template includes: Are you looking for results?

[0039] S114. For any reacquired intent, obtain the probability of the intent based on the interaction results of each time the intent is acquired. If the probability of the intent shows an increasing trend, continue to interact with the user using the dialogue template corresponding to the intent until the probability of the intent is greater than a preset probability threshold, and then determine the intent as a user intent. If the probability of the intent does not show an increasing trend, then change to a dialogue template corresponding to another reacquired intent and interact with the user.

[0040] In this embodiment, the dialogue template corresponding to a certain intent and the user's response are input into the large model, and the probability of the intent can be obtained based on the large model.

[0041] In this embodiment, during the process of interacting with the user using the dialogue templates corresponding to the reacquired intents in descending order of probability, if an intent is determined to be a user intent, then the interaction with the user will cease; if an intent is determined not to be a user intent, then a different dialogue template corresponding to the reacquired intent will be used to interact with the user.

[0042] Based on S111-S114, more accurate user intent can be obtained from the analysis text output by the large model. The efficiency of obtaining user intent can be improved by using the dialogue templates corresponding to the re-obtained intents in descending order of probability.

[0043] S120, if the user inputs a question from the module with the specified intent, then the large model makes a judgment based on the question input by the user to obtain which intent in the preset intent list the user's intent is, and when the large model judges that the user's intent is another intent, it judges that the user's intent is the intent corresponding to the target module; the target module is the module with the specified intent of the question input by the user.

[0044] Based on S110-S120, user intent can be obtained.

[0045] S200 processes the user's input question based on the user's intent, removes irrelevant information, and extracts the refined question.

[0046] As a specific implementation method, such as Figure 3 As shown, S200 includes:

[0047] S210, Obtain key information corresponding to the user's intent; wherein, when the user's intent is a query-type intent, the key information corresponding to the user's intent is the query question itself.

[0048] In this embodiment, the key information corresponding to different user intentions is different, and the key information corresponding to any intention is preset. For example, when the user intention is a query-type intention, the key information corresponding to the user intention is the query question itself; when the user intention is a non-query-type intention, the key information corresponding to the user intention includes research direction, technical requirements, transformation needs, stage, and expected conditions, etc.

[0049] S220: Based on the large model, extract the content corresponding to the key information corresponding to the user's intent from the user's input question, and obtain the extracted question.

[0050] In this embodiment, the user's input question and corresponding key information are fed into a large model. Based on the large model, content corresponding to the user's intent can be extracted from the user's input question. For example, if the user's input question is to search for patents that use AI to design drugs, the query itself is about patents that use AI to design drugs, and the corresponding extracted question is about patents that use AI to design drugs.

[0051] Based on S210-S220, irrelevant information in the user's input question can be removed to obtain a refined question.

[0052] S300: Obtain the first prompt word based on the question input by the user, input the first prompt word into the large model, and generate a solution corresponding to the question input by the user.

[0053] As a specific implementation method, such as Figure 4 As shown, S300 includes:

[0054] S310, Obtain the first part of the first prompt word; the first part of the first prompt word includes: clearly defining the user's needs and problems, and extracting the user's problems; focusing on the technical issues and requirements themselves, and removing information irrelevant to the technology; retaining the complete user needs, and not losing key information; for query-type requirements, rewriting them into the question being queried itself.

[0055] S320, obtain the second part of the prompt words of the first prompt words; the second part of the prompt words of the first prompt words includes: detailed analysis of user needs and problems, and proposing the analysis process; for query-type needs, analyze the query problem itself; deeply analyze the application scenario, the sub-domains involved, the key problems to be solved, and the key technologies required.

[0056] S330, obtain the third part of the prompt words of the first prompt words; the third part of the prompt words of the first prompt words includes: based on the analysis of user needs and problems, propose solutions; for query-type needs, solve the query problem itself; if the number of solutions exceeds the preset target number, retain the most likely solutions from the preset target number.

[0057] S340, obtain the fourth part of the prompt word of the first prompt word; the fourth part of the prompt word of the first prompt word includes: for each idea, provide a technical solution to solve the problem; for user needs, provide a complete solution, including the key problems to be solved, specific key technologies, and specific technical paths.

[0058] S350 concatenates the first part of the prompt word, the second part of the prompt word, the third part of the prompt word, and the fourth part of the prompt word together to obtain the first prompt word, which is then passed to the large model interface to obtain the content generated by the large model.

[0059] S360 outputs the content generated from the large model according to the first preset format.

[0060] As a specific implementation method, the first preset format is JSON format, and the first preset format is as follows: {question: the question extracted, analysis: the analysis process, idea: [idea 1, idea 2], solution: [technical solution 1, technical solution 2]}.

[0061] Based on S310-S360, this embodiment applies Deep Thinking Tree (ToT) to the first prompt word, using the deep thinking ability of a large model to accurately and comprehensively answer user questions.

[0062] S400, Based on the large model, extract labels from the refined questions to obtain a set of labels for the refined questions.

[0063] As one specific implementation, S400 includes, for example Figure 5 As shown:

[0064] S410, based on the large model, extract the industry domain tags from the refined problem and output them according to the second preset format.

[0065] Optionally, extracting industry domain tags from the refined question includes: determining which industry tag in the preset industry domain tag list the most critical technical and application domain involved in the refined question belongs to.

[0066] As a specific implementation, the second preset format is JSON format, and the second preset format is as follows: {Technical Field: Tag, Application Field: Tag}.

[0067] As a specific implementation method, there are 21 industry domain labels, namely: next-generation information technology, big data and cloud computing, blockchain and advanced computing, software and information services, medical and health care, integrated circuits, new energy intelligent vehicles, high-end equipment manufacturing, green energy and energy conservation and environmental protection, new materials, synthetic biology, genetic engineering, carbon emission reduction and carbon neutrality, cutting-edge new materials, quantum information, optoelectronics, new memory, brain science and brain-computer interface, humanoid robots, low-altitude economy, and CGT (cell and gene therapy).

[0068] S420, based on the large model, extract the entity information and industry classification tags from the refined problem, and output them according to the third preset format.

[0069] In this embodiment, entity information includes entity nouns of the organization name and person name that the question seeks to search for.

[0070] As a specific implementation, the third preset format is JSON format, and the third preset format is as follows: {name:[person name1,person name2],comp:[organization1,organization2],chain_path:[industry chain1,industry chain2]}; if it does not contain a person name or cannot extract a person name, then name is an empty array: [], and if it does not contain or cannot extract an organization name, then comp is an empty array: [].

[0071] In this embodiment, the tags extracted based on S410 and S420 are merged to obtain the tag set of the refined problem.

[0072] S500: Based on the solution corresponding to the user-input question and the extracted question, perform multi-way recall and rearrangement, and return the first preset number of documents of the rearranged result; the tag set of any returned document has an intersection with the tag set of the extracted question.

[0073] As one specific implementation, S500 includes, for example Figure 6 As shown:

[0074] S510, based on the technical solution corresponding to the user-input question, search for similar documents in the specified documents of the preset database, and return the similar documents as a recall to the large model; the tag set of the specified documents and the tag set of the extracted question have an intersection; the similarity between any similar document and the technical solution is greater than or equal to a preset similarity threshold.

[0075] In this embodiment, a preset database stores a large number of documents. Each document in the preset database corresponds to a set of tags. If the tag set of a document intersects with the tag set of the extracted question, then the document is determined to be related to the extracted question and is identified as the specified document. By searching for specified documents in the preset database, the workload of the search can be reduced and the search efficiency improved while avoiding missing relevant documents.

[0076] In this embodiment, the process of obtaining the similarity between any specified document and the technical solution corresponding to the user-input question includes: using the word vector or sentence vector interface of the large model, converting the core text of the technical solution and the document into high-dimensional vectors respectively; calculating the similarity between the two high-dimensional vectors using cosine similarity or vector inner product. If the similarity is greater than a preset similarity threshold, the corresponding specified document is determined to be a similar document. Optionally, the preset similarity threshold is an empirical value.

[0077] S520: Search for documents in the specified documents of the preset database whose titles are semantically similar to the questions entered by the user. This includes: obtaining word vector a for the question entered by the user through the word vector interface, obtaining word vector b for the title of the first type of document through the word vector interface, obtaining the inner product c of the vectors of a and b, and returning documents with a value greater than a first preset value as one recall path to the large model.

[0078] As a specific implementation method, this embodiment is applied to the field of technology transfer, and the first type of document is the document corresponding to the achievement.

[0079] In this embodiment, the process of obtaining the similarity between any title and the technical solution corresponding to the user-input question includes: using the word vector or sentence vector interface of the large model to convert the technical solution and the title into high-dimensional vectors respectively; and calculating the similarity between the two high-dimensional vectors using the vector inner product. Optionally, the first preset value is an empirical value.

[0080] S530: Search for documents from the specified documents in the preset database whose summaries of the first type of documents are semantically similar to the question entered by the user. This includes: obtaining word vector a for the question entered by the user through the word vector interface, obtaining word vector d for the summary of the first type of document through the word vector interface, obtaining the inner product e of the vectors of a and d, and returning documents with e values ​​greater than a second preset value as one recall path to the large model.

[0081] In this embodiment, the process of obtaining the similarity between any summary and the technical solution corresponding to the user-input question includes: using the word vector or sentence vector interface of the large model to convert the technical solution and the summary into high-dimensional vectors respectively; calculating the similarity between the two high-dimensional vectors using the vector inner product. Optionally, the second preset value is an empirical value.

[0082] S540: Search for documents in the specified documents of the preset database whose summaries of the second type of documents are semantically similar to the question entered by the user. This includes: obtaining word vector a through the word vector interface for the input question, obtaining word vector f through the word vector interface for the summary of the second type of document, obtaining the inner product g of the vectors of a and f, and returning documents with a value greater than a third preset value as one recall path to the large model.

[0083] As a specific implementation method, this embodiment is applied to the field of technology transfer, and the second type of document is the document corresponding to the patent.

[0084] In this embodiment, the process of obtaining the similarity between any summary and the technical solution corresponding to the user-input question includes: using the word vector or sentence vector interface of the large model to convert the technical solution and the summary into high-dimensional vectors respectively; calculating the similarity between the two high-dimensional vectors using the vector inner product. Optionally, the third preset value is an empirical value.

[0085] S550 deduplicates the recalled documents and then inputs the deduplicated documents into the preset rearrangement interface, returning the first preset number of documents in the rearrangement result.

[0086] As a specific implementation method, for any deduplicated document, a large model is used to score the document from multiple dimensions, and a weighted summation formula is used to obtain the score. Optionally, the multiple dimensions include semantic similarity, tag matching degree, and intent matching degree, etc.

[0087] In this embodiment, the documents are sorted in descending order of score, and the top preset number of documents in the rearranged results are returned. Optionally, the preset number is an empirical value.

[0088] Based on S510-S550, documents with similar semantics and the same tags can be searched from a preset database, improving the accuracy of the output documents.

[0089] S600, input the second prompt word into the large model so that the large model presents the first preset number of documents of the rearranged results to the user in a multimodal manner based on the second prompt word.

[0090] As one specific implementation, S600 includes, for example Figure 7 As shown:

[0091] S610, obtain the first part of the second prompt word; the first part of the second prompt word includes: deeply analyze and understand the user's needs or problems, and search for relevant information within a preset geographical range.

[0092] In this embodiment, the preset geographical range is a pre-defined range, such as XX city.

[0093] S620, obtain the second part of the prompt words for the second prompt words; the second part of the prompt words for the second prompt words includes: first, briefly informing the list of titles of the recommendation results, then expanding each one for detailed analysis and giving reasons for the recommendation, and selecting at least one of the following dimensions for analysis: key technical points, difficulties, explanation of key terms, and technical feasibility.

[0094] S630 concatenates the first part of the second prompt word and the second part of the second prompt word together to obtain the second prompt word, which is then input into the large model interface to obtain the returned result.

[0095] This embodiment applies a deep thinking tree to the second prompt word, and returns accurate and comprehensive results based on the deep thinking capabilities of a large model.

[0096] S640 will stream the returned results.

[0097] As a specific implementation, the streaming output of the returned results includes: if the content of the returned results contains the text corresponding to a first type of document, then inserting a card corresponding to the first type of document; if the content of the returned results contains the text corresponding to a second type of document, then inserting a list corresponding to the second type of document. It should be understood that when the first type of document is a document corresponding to an achievement, the text corresponding to the first type of document is the achievement, and the card corresponding to the first type of document is the achievement card; when the second type of document is a document corresponding to a patent, the text corresponding to the second type of document is the patent, and the list corresponding to the second type of document is the patent list. As a specific implementation, if the content contains text such as the text corresponding to the first type of document and the text corresponding to the second type of document, then inserting a reference chart, and hovering the mouse over it will display the corresponding card content. As a specific implementation, S600 also includes: searching the network for achievements and patents similar to the user's input question, and supplementing them into the large model's answer. Thus, this embodiment presents achievements to the user in the form of cards, patents in the form of lists, and references in the form of floats, improving the user experience.

[0098] As a preferred embodiment, this example further includes: determining the number of questions input by the user; if the number of questions input by the user is greater than or equal to a preset question number threshold, then clustering the user-input questions based on the semantic correlation between them; for any given cluster, inputting all the questions in that cluster as a whole into the large model, and executing the method of this embodiment, i.e., S100-S600. Here, the preset question number threshold is an empirical value, the semantic correlation between the user-input questions is the semantic similarity between them, and the clustering uses density-based clustering. This avoids redundant calculations caused by processing scattered questions one by one; the overall analysis after clustering integrates related information, reduces the number of times the large model is called, and lowers processing costs; after clustering related questions, the large model can identify the inherent logic between questions, generating a more systematic solution and avoiding fragmentation.

[0099] This embodiment first performs intent recognition on the user's input question, accurately locating the user's core purpose. Based on the user's intent, the input question is processed to remove redundant information, resulting in a refined question. Furthermore, a set of tags for the refined question is obtained. A solution to the user's input question is then obtained based on a large model. Based on the solution and the set of tags for the refined question, documents with semantically similar meanings and the same tags are searched from a pre-set database. This improves the accuracy of the output document and ensures that the final matched document is strongly relevant to the core purpose of the need, thus improving the accuracy of supply and demand matching.

[0100] Example 2:

[0101] This embodiment provides a supply and demand matching system based on a large model, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it performs the following steps:

[0102] Based on a large model, the intent of the user's input question is identified to obtain the user's intent.

[0103] The system processes the user's input based on their intent, removing irrelevant information and extracting the key questions.

[0104] The system obtains a first prompt word based on the user's input question, inputs the first prompt word into the large model, and generates a solution corresponding to the user's input question.

[0105] Based on the large model, labels are extracted from the refined questions to obtain a set of labels for the refined questions.

[0106] Based on the solution corresponding to the user-input question and the extracted question, a multi-way recall is performed and rearranged, returning the first preset number of documents of the rearranged result; the tag set of any returned document intersects with the tag set of the extracted question.

[0107] Input the second prompt word into the large model so that the large model presents the first preset number of documents of the rearranged results to the user in a multimodal manner based on the second prompt word.

[0108] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0109] While specific embodiments of the invention have been described in detail by way of example, those skilled in the art should understand that the examples are for illustrative purposes only and not intended to limit the scope of the invention. It should also be understood that various modifications can be made to the embodiments without departing from the scope and spirit of the invention. The scope of the invention is defined by the appended claims.

Claims

1. A large model-based supply and demand matching method, characterized in that, The method comprises the following steps: S100, performing intent recognition on the user input question based on a large model to obtain a user intent; S200, processing the user input question according to the user intent, removing irrelevant information, and obtaining an extracted question; S300, obtaining a first prompt word according to the user input question, inputting the first prompt word into the large model, and generating a solution corresponding to the user input question; S400, performing label extraction on the extracted question based on the large model to obtain a label set of the extracted question; S500, performing multi-path recall on the solution corresponding to the user input question and the extracted question and rearranging to return the top pre-set number of documents of the rearranged result; the label set of any returned document and the label set of the extracted question have an intersection; S600, inputting a second prompt word into the large model to enable the large model to present the top pre-set number of documents of the rearranged result to the user in a multi-modal manner according to the second prompt word; S500 comprises: S510, searching for similar documents in a specified document of a pre-set database based on a technical solution corresponding to the user input question, and returning the similar documents to the large model as one path of recall; the label set of the specified document and the label set of the extracted question have an intersection; and the similarity of any similar document to the technical solution is greater than or equal to a pre-set similarity threshold; S520, searching for a document with a title of a first type of document and a semantic similarity to the user input question from the specified document of the pre-set database, comprising: obtaining a word vector a of the user input question through a word vector interface, obtaining a word vector b of the title of the first type of document through the word vector interface, obtaining a vector inner product c of a and b, and returning the documents with a c value greater than a first pre-set value to the large model as one path of recall; S530, searching for a document with an abstract of a first type of document and a semantic similarity to the user input question from the specified document of the pre-set database, comprising: obtaining a word vector a of the user input question through a word vector interface, obtaining a word vector d of the abstract of the first type of document through the word vector interface, obtaining a vector inner product e of a and d, and returning the documents with an e value greater than a second pre-set value to the large model as one path of recall; S540, searching for a document with an abstract of a second type of document and a semantic similarity to the user input question from the specified document of the pre-set database, comprising: obtaining a word vector a of the input question through a word vector interface, obtaining a word vector f of the abstract of the second type of document through the word vector interface, obtaining a vector inner product g of a and f, and returning the documents with a g value greater than a third pre-set value to the large model as one path of recall; S550, performing deduplication on the recalled documents, and finally inputting the deduplicated documents into a pre-set rearrangement interface to return a number of documents with a rearrangement score greater than a pre-set score threshold.

2. The large model-based supply and demand matching method according to claim 1, characterized in that, S100 comprises: S110, if the user does not input a question from the specified intent module, performing judgment based on a large model according to the user input question to obtain which one of a pre-set intent list is the user's intent, the pre-set intent list comprising a number of types of intents and other intents; S120, if the user inputs a question from the module of the specified intention, the large model is used to determine the user's intention based on the user's input question, and determine which intention in the preset intention list the user's intention is, and when the large model determines that the user's intention is other intention, determine that the user's intention is the intention corresponding to the target module; the target module is the specified intention module of the user's input question.

3. The large model-based supply and demand matching method according to claim 2, characterized in that, The large model is used to determine the user's intention based on the user's input question, and determine which intention in the preset intention list the user's intention is, and when the large model determines that the user's intention is other intention, determine that the user's intention is the intention corresponding to the target module; the target module is the specified intention module of the user's input question. S111, input the user's input question into the large model, and obtain the analysis text, a plurality of intentions and corresponding probabilities output by the large model; S112, if the probability of one intention output by the large model is greater than or equal to a preset probability threshold, the intention is determined as the user's intention; otherwise, extract keywords from the analysis text output by the large model, and reacquire a plurality of intentions and corresponding probabilities based on the extracted keywords, and enter S113; S113, in order from high to low, use the dialog templates corresponding to the reacquired intentions to interact with the user; S114, for any intention reacquired, obtain the probability of the intention according to the interaction result of the intention each time, if the probability of the intention presents a growth trend, continue to use the dialog template corresponding to the intention to interact with the user, until the probability of the intention is greater than the preset probability threshold, and the intention is determined as the user's intention; if the probability of the intention does not present a growth trend, replace the dialog template corresponding to other reacquired intention to interact with the user. 4.The large model-based supply-demand matching method according to claim 1, wherein, S200 includes: S210, obtain the key information corresponding to the user's intention; when the user's intention is a query intention, the key information corresponding to the user's intention is the query question itself; S220, based on the large model, extract the content corresponding to the key information corresponding to the user's intention from the user's input question, and obtain the extracted question. 5.The large model-based supply-demand matching method according to claim 1, wherein, S300 includes: S310, obtain the first part of the first prompt word; the first part of the first prompt word includes: clearly expressing the user's demand and question, extracting the user's question; paying attention to technical problems and requirements themselves, removing technical irrelevant information; retaining the user's complete demand, without losing key information; for query demand, rewrite into the queried question itself; S320, obtain the second part of the first prompt word; the second part of the first prompt word includes: detailed analysis of user demand and question, and the analysis process is proposed; for query demand, the queried question itself is analyzed; deeply analyze the application scenario, the involved sub-field, the key problem to be solved, and the key technology required; S330, obtain the third part of the first prompt word; the third part of the first prompt word includes: according to the analysis of the user's demand and question, the idea of solving the problem is proposed; for query demand, the queried question itself is solved; if the number of ideas exceeds a preset target number, the first preset target number of ideas most likely to solve the problem are retained; S340, a fourth part prompt word of the first prompt word is obtained; the fourth part prompt word of the first prompt word includes: giving a technical solution to solve the problem for each idea; and giving a complete scheme including key problems to be solved, specific key technologies and specific technical paths for the user demand; S350, the first part prompt word, the second part prompt word, the third part prompt word and the fourth part prompt word of the first prompt word are spliced together to obtain the first prompt word, which is input into the large model interface to obtain the content generated by the large model; S360, the content generated by the large model is output according to the first preset format. 6.The large model-based supply-demand matching method according to claim 1, wherein, S400 includes: S410, based on the large model, extracting the industry domain label in the refined problem, and outputting according to the second preset format; S420, based on the large model, extracting the entity information and industry classification label in the refined problem, and outputting according to the third preset format. 7.The large model-based supply-demand matching method according to claim 1, wherein, S600 includes: S610, a first part prompt word of a second prompt word is obtained; the first part prompt word of the second prompt word includes: deeply analyzing and understanding the demand or problem of the user, and searching for related information within a preset geographical range; S620, a second part prompt word of the second prompt word is obtained; the second part prompt word of the second prompt word includes: briefly informing the title list of the recommended result, expanding one by one to make detailed analysis and give the recommended reason, and selecting at least one of the following dimensions for analysis: technical key points, difficulties, key term explanation and technical feasibility; S630, the first part prompt word and the second part prompt word of the second prompt word are spliced together to obtain the second prompt word, which is input into the large model interface to obtain the returned result; S640, the content of the returned result is output in a streaming manner. 8.The large model-based supply-demand matching method according to claim 7, characterized in that, The content of the returned result is output in a streaming manner, including: if the content of the returned result contains the word of the first type document, a card corresponding to the first type document is inserted; and if the content of the returned result contains the word of the second type document, a list corresponding to the second type document is inserted. 9.A large model-based supply-demand matching system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein, The processor executes the computer program to realize the supply and demand matching method based on the large model in any one of claims 1 to 8.

Images