Method for managing a temporary data storage, and corresponding entity and computer program
By employing a feedback-based cache management system with a confidence score for Large Language Models, the method optimizes storage and response relevance, addressing inefficiencies in existing cache systems and ensuring timely, accurate responses.
Patent Information
- Application Number
- PCT/EP2026/062453
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-05
- Filing Date
- 2026-04-30
- Publication Date
- 2026-11-12
AI Technical Summary
Existing cache management systems for Large Language Models (LLMs) face challenges such as excessive storage consumption, confidentiality concerns, limited lifespan, and the risk of outdated responses due to inadequate invalidation strategies, as well as inefficiencies in handling subtly evolving requests.
A method that utilizes feedback information to process responses stored in the cache, including a confidence score system to determine the relevance and validity of cached responses, allowing for their modification and removal based on user feedback, thereby optimizing storage and ensuring timely updates.
This approach enhances cache management by reducing storage usage, improving response relevance, and minimizing resource consumption while ensuring timely and accurate responses to user queries.
Smart Images

Figure EP2026062453_12112026_PF_FP_ABST
Abstract
Description
Method for managing temporary data storage, entity and corresponding computer program.
[0001] The invention belongs to the general field of communications, such as, for example, the field of telecommunications.
[0002] It falls more specifically within the context of managing temporary data storage mechanisms or "cache".
[0003] The invention more specifically aims at mechanisms to improve the processing of requests issued by user equipment and the data obtained in response to these requests.
[0004] Large Language Models (LLMs) are language models designed to intelligently understand and generate text. These models are often based on deep neural networks and can learn to predict and produce text based on the data they have been trained to process. They are used in various fields such as machine translation, text generation, virtual assistants, communication network management, and many more. Their ability to understand context and generate coherent responses makes them particularly useful in applications where natural language understanding is essential, such as chatbots that can answer questions, guide users through processes, or provide customer support.
[0005] LLMs respond to each user query submitted to them independently. In other words, they don't know if the user who submitted the query or other users have already asked or are simultaneously asking the same or a similar question, and they have no overall view of past or parallel interactions. However, the infrastructures that host or interact with LLMs can implement optimization mechanisms. One of the most common is caching the generated responses.
[0006] An application cache is a local storage space—that is, specific to the entity hosting the application or shared between applications—that allows for the temporary storage of data or resources used or generated by the entity during request processing. This storage in the entity's cache or in a cache associated with the entity, or caching, aims to reduce latency in providing a response to a request, limit bandwidth consumption within the communication network used to generate a response, and improve data processing time for subsequent requests identical or similar to a previously processed request. It thus enables faster and more efficient request processing, while reducing the software and energy load on the servers hosting the LLM (Local Management Interface) as well as the load on the communication networks used to generate a response to a request.
[0007] Thus, when the same or a similar query to a previous query is sent to an LLM by one or more users, and it does not depend on specific contextual data, a previously generated response can be reused instead of requesting the model used to generate a response to a query again.
[0008] However, while caching mechanisms offer significant advantages in terms of performance and reduced resource consumption, they also have several limitations, including excessive local storage consumption, confidentiality concerns regarding sensitive data used to generate a response, a limited lifespan for cached information, and the limitation of a single cache per application or for a set of applications, including web browsers. These issues highlight the need for optimized and adaptable cache management capable of meeting the diverse needs of users and the constraints of dynamic environments.
[0009] Despite their many advantages, caches introduce a risk of outdated responses: if information changes over time, such as news, a price, or dynamic data, the system may continue to broadcast an old, incorrect response that is still present in the cache, due to the lack of an effective invalidation strategy.
[0010] Another challenge lies in handling requests whose meaning subtly evolves: very similar prompts—that is, the instructions a user gives the model to generate a response—can require different answers depending on details of wording or context. A prompt can be a question, a command, an incomplete sentence, or even a long document. The model analyzes the provided prompt and attempts to produce the most coherent or relevant response possible by generating a new one and / or using data from the cache.
[0011] A cache management system that is too rigid and indifferent to the relevance of a response to a query risks overlooking differences between very similar prompts and producing results that may be poorly suited to user queries. State-of-the-art solutions therefore have the drawback of producing unsuitable responses to queries if a cache is used, or requiring significant data processing to satisfactorily answer a query if responses are generated without using a cache.
[0012] Thus, there is a need for new solutions to improve the management of cache mechanisms.
[0013] To this end, and according to a first aspect, the invention relates to a method for managing a temporary data storage, called a cache, in which responses generated by a large LLM language model implemented by an entity of a communication network are stored, the method comprising: receiving feedback information relating to a first response generated by the LLM and transmitted to a user equipment that originated the first request; applying, based on the feedback information received, a processing of the first response intended to be applied by the cache.
[0014] Correspondingly, the invention also relates to an entity of a communication network implementing a large LLM language model, said entity comprising: at least one temporary data storage, called a cache, in which responses generated by the LLM are stored, and at least one processor configured to: receive feedback information relating to a first response generated by the LLM and transmitted to a user equipment that originated the first request; and apply, based on the feedback information received, a processing of the first response intended to be applied by the cache.
[0015] The underlying idea on which the invention is based is to use feedback information to apply processing to responses stored in the cache of an LLM.
[0016] Taking this feedback into account allows us to decide, for example, how long responses are kept in the cache and according to what criteria they should be updated or invalidated. This contributes to the implementation of a high-performing, reliable, and relevant caching system.
[0017] Feedback information allows for the qualification of information present in the cache, potentially taking into account a temporal characteristic. Thus, if information from a cache, once transmitted or used to formulate a response to the cache, receives one or more positive feedback messages, for example from one or more users, then this cache information can be considered sufficiently relevant to generate a response to a new request, without resorting to a full query of one or more LLMs and one or more databases. This process therefore ensures that a response is appropriate to a request while conserving the software, transmission, and energy resources of a communication network responsible for managing requests and their associated responses.
[0018] The invention applies equally to classic LLMs receiving text as input and to multimodal LLMs which can receive images, videos or audio content as input, the cache information being able to be adapted to these different types of input data.
[0019] According to embodiments of the process of the invention, feedback information is received from the user equipment that originally issued the first request.
[0020] The user who formulated a request can advantageously give an opinion on information received from an LLM and thus intervene on the caching of the response received or part of the response received, thus allowing to influence its use or not when receiving a similar or similar request.
[0021] According to embodiments of the process of the invention, the processing consists of storing in the cache the first response, and a confidence score value representing the relevance of the first response.
[0022] Thanks to this confidence score, the LLM has contextual information allowing it to know whether this answer was deemed satisfactory or not by the different users to whom it was offered and thus to determine whether or not it should be offered.
[0023] According to embodiments of the method of the invention, when feedback information indicates that the first response conforms to the expectations of a user of the user equipment that initiated the first request, the value of the confidence score increases.
[0024] Receiving positive "feedback", that is, "feedback" indicating that the answer obtained is either correct or meets the expectations of the user who made the request, helps to increase the value of the confidence score associated with the first answer.
[0025] Since the first response is considered satisfactory by LLM users, it is not necessary to ask the LLM to generate a new response to a given query. This optimizes the storage space used, which already contains information that can be effectively used when generating a response to a future query, whether from the same user or a different user.
[0026] According to embodiments of the method of the invention, when feedback information indicates that the first response does not conform to the expectations of a user of the user equipment that initiated the first request, the value of the confidence score decreases.
[0027] Receiving negative feedback can indicate that the response received is incorrect, incomplete, outdated, or does not meet the needs of the user who made the request. Such negative feedback can stem from a poorly worded request by the user, the use of cached information that is inappropriate for the question asked, or information that is no longer valid because it is outdated or ill-suited to the specific context of the submitted request.
[0028] Thus, when the confidence score value is below a threshold, the first response is removed from the cache.
[0029] By removing answers that have received negative "feedback", not only do we free up storage space but we also ensure that these answers will no longer be offered to the user, thus improving their user experience.
[0030] According to embodiments of the process of the invention, the processing consists of: modifying the first response according to the feedback information received; and storing the first modified response in the cache.
[0031] Using feedback information, it is possible to determine that the generated response, regardless of its accuracy, can be modified to improve its relevance or incorporate updated knowledge. Thus, a cached response element can be progressively and continuously updated based on the various feedback received in response to one or more queries, thereby improving the relevance of the information stored in the cache.
[0032] The changes to the response can be included in the feedback information. Both unmodified and modified responses can be cached. This allows for a history to be maintained for each response. Indeed, while the initial response may not be satisfactory for an initial query, it could be relevant for future queries.
[0033] In other cases, the first response, i.e. the unmodified response, is removed from the cache, and only the modified response remains stored in the cache.
[0034] This makes it easier to trace the generated response back to the query that led to the generation of that response.
[0035] Therefore, it is possible to provide the user with, in addition to the generated response, the previous queries that led to that response. This allows the user to determine how to modify a previous query or how to submit a new query in order to increase their chances of obtaining the most relevant answer possible.
[0036] In one particular embodiment, the process also includes memorizing the first request in association with the first response.
[0037] This facilitates traceability between the generated response and the request that led to its generation. Furthermore, this record-keeping of a request and its response allows cached information to be associated with one or more outgoing requests, thus improving the relevance of the response to a contextual element. Upon receiving a subsequent request, cached information can then be used based on the contextual proximity of that subsequent request to a previous request that generated feedback specific to qualifying the cached information.
[0038] Optionally, in addition to the generated response, the user can be shown the previous queries that led to it. This allows the user to determine how to modify a previous query or how to submit a new one to increase their chances of getting the most relevant answer possible.
[0039] According to embodiments of the process which is the subject of the invention, the feedback information is stored in: a cache buffer, and / or a cache vector database.
[0040] The feedback information can be stored with the corresponding response in a cache buffer in which the responses generated by the LLM are stored, and / or with the query(ies) that led to the generation of the response in the cache vector database.
[0041] The cache vector base is the part of the cache in which the requests processed by the LLM are stored.
[0042] According to embodiments of the process which is the subject of the invention, the feedback information belongs to a group comprising at least: text data; audio data; an image; a video; an icon.
[0043] According to embodiments of the method of the invention, it comprises: receiving a second request similar to the first request; selecting, based on the value of the confidence score relating to the first response, a second response to the second request; transmitting the second response to a user device that originated the second request.
[0044] In this scenario, the feedback information allows the cache to select the most appropriate response, knowing that the second request is similar to the first. This allows the cache to benefit from experience when selecting the second response.
[0045] Therefore, if a response to the second request already exists with positive feedback and is stored in the cache, there is no need to request the LLM to generate a new response. This reduces the waiting time for the user who initiated the second request and also reduces the resource usage of the entity hosting the LLM.
[0046] According to a second aspect, the invention relates to a method of communication between a user equipment of a first communication network and a temporary data storage, called a cache, in which are stored responses generated by a large LLM language model implemented by an entity of a second communication network, the method comprising: the reception of a response generated by the LLM following the transmission of a request sent to the cache; the sending, to the cache, of feedback information relating to the response received.
[0047] Correspondingly, the invention also relates to a user equipment of a first communication network configured to communicate with at least one temporary data storage, called a cache, in which are stored responses generated by a large LLM language model implemented by an entity of a second communication network, the user equipment comprising at least one processor configured to: receive a response generated by the LLM following the transmission of a request sent to the cache; send, to the cache, feedback information relating to the response received.
[0048] In a particular embodiment, the management process that is the subject of the invention is implemented by a computer.
[0049] In this respect, the invention also relates to a computer program on a recording medium, this program being capable of being implemented in a computer or more generally in the entity of the communication network according to the invention and comprising instructions adapted to the implementation of a management process as described above.
[0050] In a particular embodiment, the communication method which is the subject of the invention is implemented by a computer or user equipment such as a mobile phone, for example a smartphone.
[0051] In this respect, the invention also relates to a computer program on a recording medium, this program being capable of being implemented in a computer or more generally in the entity of the communication network according to the invention and comprising instructions adapted to the implementation of a communication process as described above.
[0052] These programs can use any programming language, and be in the form of source code, object code, or code somewhere between source code and object code, such as in a partially compiled form, or in any other desirable form.
[0053] The invention also relates to an information carrier or recording medium readable by a computer, and comprising instructions for computer programs as mentioned above.
[0054] The information or recording medium can be any entity or device capable of storing programs. For example, the medium may include a storage means, such as a ROM, for example a CD-ROM or a microelectronic circuit ROM, or a magnetic recording means, for example a hard drive, or a flash memory.
[0055] On the other hand, the information or recording medium can be a transmissible medium such as an electrical or optical signal, which can be carried via an electrical or optical cable, by radio link, by wireless optical link or by other means.
[0056] The programs according to the invention can in particular be downloaded onto an Internet-type network.
[0057] Alternatively, the information or recording medium may be an integrated circuit in which programs are incorporated, the circuit being adapted to execute or to be used in the execution of the processes according to the invention.
[0058] It can also be envisaged, in other embodiments, that the processes according to the invention, as well as the communication network entity and the user equipment according to the invention, present in combination all or part of the aforementioned characteristics.
[0059] Other features and advantages of the present invention will become apparent from the description below, with reference to the accompanying drawings, which illustrate an example of an embodiment without being limiting in any way. In the figures:
[0060] lare represents a communication system 1, conforming to the invention in a particular embodiment;
[0061] lare represents schematically the hardware architecture of a computer on which the different entities conforming to the invention, belonging to the system of the, are based;
[0062] larepresents schematically the architecture of the entity conforming to the invention, belonging to the system of the ;
[0063] The diagram represents, in the form of a flowchart, the main steps of the processes that are the subject of the invention, as implemented by the different entities conforming to the invention belonging to the system. Description of the invention
[0064] Lare represents an example of a communication system, conforming to the invention in a particular embodiment.
[0065] In this embodiment, system 1 comprises: at least one entity 10 of a telecommunications network NW in which a large language model LLM, such as a multimodal LLM, is hosted. Such entity 10 may be, for example, a standalone server or one belonging to a cluster of servers. Thus, entity 10 may be a software entity, a hardware entity, or a combination of both; and at least one user device UE1 interacting with the LLM hosted within entity 10. A user device may be any device connected to a telecommunications network, including, in particular, a smartphone or a computer.
[0066] In the following description and on the, by way of example, for the sake of simplification, we consider a single user equipment UE1, and a single entity 10. Of course, a higher number of user equipment and entities 10 can be considered.
[0067] Finally, it matters little whether the user equipment UE1 is attached to the NW telecommunications network hosting entity 10.
[0068] The interactions between the different components of system 1 are described in more detail with reference to the.
[0069] In the embodiment described herein, the user equipment UE1 and entity 10 exhibit the hardware architecture of a computer 2 as illustrated in Figure 1. This hardware architecture includes, in particular, a processor (PROC), random access memory (MEM), read-only memory (ROM), non-volatile memory (NVM), and communication means (COM) enabling the various entities to communicate with each other and with one or more other devices (not shown in the figures). The non-volatile memory (NVM) constitutes a storage medium according to the invention, readable by the processor (PROC), on which one or more programs according to the invention are stored. The non-volatile memory (NVM) also includes a storage space dedicated to storing cached data according to the invention.
[0070] A first program, denoted PROG1, is stored in the non-volatile memory NVM and includes instructions defining the main steps of a communication method according to the invention as implemented by the user equipment UE1. More specifically, it defines the functional modules of the user equipment UE1, which relies on and / or controls all or part of the PROC, MEM, ROM, NVM, and COM elements of the computer 2 mentioned above.
[0071] In the embodiment described here, this first program PROG1 defines in particular the following functional modules of the UE, which are activated during the implementation of the invention: a transmission module configured to send requests and messages including feedback information to entity 10; and a reception module configured to receive, from entity 10, a response to the requests sent.
[0072] The operation of these modules of the UE1 user equipment is detailed further later with reference to the steps of the management and communication processes according to the invention.
[0073] The second program, PROG2, defines in particular the following functional modules of entity 10 as represented with reference to the diagram and which are activated during the implementation of the invention: a communication interface 101 configured to receive requests and messages from the user equipment UE1, including feedback information, and to send responses to these requests to the user equipment; a vectorization module, or "embedding" module, 102 configured to generate vectors representing text, an image, audio, or video content included in one or more requests sent by the user equipment UE1; a temporary data storage space 103, also referred to as a cache in the remainder of this document. This cache 103 is divided into two separate storage spaces which may or may not be the same size.A first storage space 103-1, referred to in the rest of the document as vector base 103-1, is intended to store the vectors generated by the vectorization module 102.The second storage space 103-2, referred to in the rest of the document as the cache buffer 103-2, is intended to store responses to requests generated by a large multimodal language model LLM 105 which will be presented below; a comparison module 104 configured to compare the vectors generated by the vectorization module 102 with vectors stored in the vector database 103-1; a large multimodal language model LLM 105 configured to generate one or more responses to requests received by the communication interface 101, these responses being intended to be transmitted to the user equipment UE1 and to be stored in the cache buffer 103-2; and a management module 106 configured to process feedback information emitted by the user equipment UE1 and then transmit it to the cache 103.
[0074] The operation of these modules of entity 10 is also detailed further later with reference to the steps of the management and communication processes according to the invention.
[0075] The document describes the main steps of the cache management process 103 and the communication process according to the invention, in a particular embodiment.
[0076] A user of the UE1 user equipment wishes to query the LLM 105. To this end, the user enters an RQT request using a human-machine interface (not shown in the figures) such as a touch screen or a microphone.
[0077] An example of such an RQT request might be, "Can you give me a list of monuments to visit in Beijing?" Another example of an RQT request might be, "Can you find me an image similar to the following image?", this RQT request also including that image.
[0078] The RQT request can be in the form of a text message and / or an audio message. If it includes an image, this can also be in the form of a video or an animated image such as a GIF (Graphics Interchange Format).
[0079] Once the RQT request is entered, it is transmitted (S010) by the user equipment UE1 to entity 10.
[0080] Upon receipt of the RQT request, the communication interface 101 transmits it to the vectorization module 102 which then generates (S020) one or more Vect-RQT vectors representing the RQT request.
[0081] Vectorization refers to the digital representation of an element, such as a group of words, an image, audio and / or video content, in the form of a vector, i.e. a sequence of numbers, so that these vectors can be compared with each other, even if they come from different sources.
[0082] For example, using this technique, an image of a cat and the word "cat" can be placed close together in the same vector space because they refer to the same concept. This allows a computer system to understand that they are related, even though they do not belong to the same data type.
[0083] To achieve this, each input is first processed by a specific encoder: a natural language processing model for text, a convolutional neural network or another visual model for images, and so on, depending on the nature of the content. These encoders transform each input into a vector in a fixed-dimensional space. Then, a training process aligns these vectors in a shared space, learning to group together those that describe the same object (such as a picture of a dog and the caption "a running dog") and to separate those that are not relevant to each other. This type of representation makes it possible, in particular, to search for an image based on a text query, or conversely, to generate a text description from an image.
[0084] The resulting Vect-RQT vector is then transmitted to the comparison module 104, which then performs the comparison (S030) of the Vect-RQT vector with the set of Vect-Prev-RQT vectors stored in the vector base 103-1 of cache 103. These Vect-Prev-RQT vectors correspond to Prev-RQT requests previously received and processed by entity 10. Such Prev-RQT requests may have been issued by user equipment UE1 or by other user equipment.
[0085] Such a comparison involves, for example, the calculation, using comparison module 104, of a semantic similarity score. This score reflects the similarity between two concepts, words, phrases, or texts represented by the vectors being compared, based on their meanings. Thus, the higher the value of this similarity score, the closer the meanings of the phrases or words represented by the compared vectors.
[0086] If one or more Vect-Prev-RQT vectors have a high semantic similarity score with the Vec-RQT vector—for example, if this score is greater than or equal to a Thresh threshold—it means that an identical or similar query has already been processed by entity 10; this is referred to as a "Hit." Such a Prev-RQT-Sim query could be, for example, "Which monuments are must-sees when visiting Beijing?"
[0087] Since a similar request has already been processed by entity 10, the comparison module 104 requires (S040) the Resp-Prev-RQT response associated with the Vect-Prev-RQT vector identified during step S030 stored in buffer 103-2 of cache 103.
[0088] Indeed, each Resp-RQT or Resp-Prev-RQT response generated by LLM 105 as part of the processing of an RQT, Prev-RQT request is stored in buffer 103-2 of cache 103. This Resp-RQT or Resp-Prev-RQT response is also associated with the vector Vect-RQT, Vect-Prev-RQT representing the RTQ, Prev-RQT request that originally generated the Resp-RQT or Resp-Prev-RQT response.
[0089] In order to allow the association between two elements stored in separate sections of a memory, such as cache 103, various mechanisms can be implemented.
[0090] An example of such a mechanism is the use of pointers. In such a mechanism, a first element stored in a first section of memory can contain an address or identifier that allows the location of a second element stored in a second section of memory. This principle is commonly used in dynamic data structures, where each element contains an explicit reference to another.
[0091] Another example of such a mechanism is the implementation of associative structures such as hash tables, which allow for the establishment of correspondences between keys and values. Each key can identify a first element stored in a first section of memory, while the corresponding value can contain a reference or data from another memory section.
[0092] Finally, such a mechanism can also rely on the use of shared unique identifiers such as primary and secondary keys. Thus, two elements stored in independent memory segments can be associated if they possess a common or related identifier, enabling a logical link between them.
[0093] Once the Resp-Prev-RQT response is obtained, the comparison module 104 transmits it (S050) to the communication interface 101 which in turn transmits it (S060) to the user equipment UE1.
[0094] If no Vect-Prev-RQT vector has a sufficiently high semantic similarity score with the Vec-RQT vector, for example if this score is below the Thresh threshold, this means that no query indicating or similar to the RQT query has already been processed by entity 10, we then speak of "No Hit".
[0095] Since no similar request has already been processed by entity 10, the comparison module 104 (S070) transmits the Vect-RQT vector identified during step S030 to the LLM 105 so that the latter generates (S080) a Resp-RQT response to the RQT request represented by the Vect-RQT vector.
[0096] Once the Resp-RQT response is generated by the LLM 105, the latter transmits it (S090) to the communication interface 101 which in turn transmits it (S100) to the user equipment UE1.
[0097] Whether the response received by the user equipment UE1 follows the RQT request in the form of the Resp-Prev-RQT response obtained in step S040 or the Resp-RQT response generated in step S080, then steps S110 to S130 described below are implemented.
[0098] Upon receipt of the response provided by entity 10 to the RQT request, the user of the user equipment UE1 is prompted to enter feedback information or "feedback" FBck relating to the response received.
[0099] Such Facebook feedback can take various forms. For example, it can be icons such as a thumbs up, to indicate that the response received meets the user's expectations, or a thumbs down, to indicate that the response received does not meet, or only partially meets, the user's expectations.
[0100] These icons are presented to the user by means of a display device on the user equipment (not shown in the figures) such as a touch screen displaying a graphical interface designed to facilitate exchanges with entity 10. The user can then select the icon that corresponds to their perception of the response received via the human-machine interface of the user equipment UE1.
[0101] Facebook feedback can also take the form of a text or voice message in which the user expresses their perception of the response received. Such a message may also include corrections or additional information intended to improve the relevance of the response.
[0102] Once the retroactive information relating to the response received has been entered, it is transmitted (S110) to entity 10.
[0103] Once received, the FBck retroactivity information is processed (S120) by management module 106.
[0104] Three scenarios are possible: the FBck feedback information is negative, meaning that the response received does not fully meet the user's expectations. Negative feedback can mean that the response is incorrect, for example, if entity 10 answers "5" to the question "What is 2+2?", that the response is outdated, for example, if entity 10 answers "The current Pope is Francis" to the question "As of April 25, 2025," when he has just passed away, or that the provided response is incomplete or does not truly address the problem the user of the UE1 user equipment is facing; the FBck feedback information is positive, meaning that the response received meets the user's expectations.Positive feedback information can mean that the response is correct, not outdated, appears complete, or addresses the issue the user of the UE1 user equipment is facing; FBck feedback information includes information enabling the response provided to be updated, regardless of whether such feedback information is positive or negative.
[0105] Facebook feedback can be positive or negative, in one of the forms indicated above (icon, text message or voice message), and may also include information allowing the response provided to be updated, regardless of whether such feedback is positive or negative.
[0106] When the feedback information FBck is negative, the management module 106 transmits information relating to the memory of the response to which the feedback information applies to cache 103.
[0107] When the response to which the positive feedback information FBck applies is the Resp-RQT response, the management module requests cache 103 to store the latter in its buffer 103-2.
[0108] In one particular implementation, the Resp-RQT response is stored with the FBck feedback information.
[0109] In a second implementation, the RQT request that generated the response can be stored in parallel in the 103-1 vector database along with the FBck feedback information.
[0110] In a third implementation, a confidence score can also be associated with the stored Resp-RQT response. The score value increases when the feedback is positive and decreases when it is negative. Using such a confidence score is particularly useful for answers to frequently asked questions. This implementation allows for weighting the relevance of the cached responses and for adjusting the relevance of a response based on feedback received from one or more users.When the score is below a certain value, this response may be removed from the cache so that it is not transmitted in response to a new request, which may require generating a new response in accordance with the "no-hit" procedure described above or determining a response present in the cache and adapted to the received request.
[0111] These three implementations can be implemented jointly or combined in pairs.
[0112] When the negative response to which the FBck feedback information applies is the Resp-Prev-RQT response, the management module can request that cache 103 store this response in its 103-2 buffer and decrease the value of the corresponding confidence score. When the value of the associated confidence score is less than a given value, the Resp-Prev-RQT response can be removed from cache 103.
[0113] When the feedback information FBck is positive, the management module 106 transmits information relating to the memory of the response to which the feedback information applies to cache 103.
[0114] When the response to which the positive feedback information FBck applies is the Resp-RQT response, the management module requests cache 103 to store the latter in its buffer 103-2.
[0115] In one particular implementation, the Resp-RQT response is stored with the FBck feedback information.
[0116] In a second implementation, the RQT request that generated the response can be stored in parallel in the 103-1 vector database along with the FBck feedback information.
[0117] In a third implementation, a confidence score can also be associated with the memorized Resp-RQT response.
[0118] These three implementations can be implemented jointly or combined in pairs.
[0119] When the response to which the positive feedback information FBck applies is the Resp-Prev-RQT response, the management module can request cache 103 to increase the value of the corresponding confidence score.
[0120] Finally, when the FBck feedback information includes information enabling the update of the provided response, the management module 106 generates an update information for the response which it transmits to cache 103.
[0121] Cache 103 can then modify the response based on the information transmitted by management module 106 and retain only the modified response or choose to retain the response previously transmitted to user equipment UE1 and the modified response in accordance with the instructions of management module 106.
Claims
Method of managing a temporary data storage, called a cache (103), in which responses (Resp-RQT, Resp-Prev-RQT) generated by a large LLM language model (105) implemented by an entity (10) of a communication network (NW) are stored, the method comprising: receiving feedback information (FBck) relating to a first response (Resp-RQT, Resp-Prev-RQT) generated by the LLM (105) and transmitted to a user equipment (UE1) which originated the first request (RQT); applying, based on the feedback information (FBck) received, a processing of the first response intended to be applied by the cache (103). Management method according to claim 1, wherein the feedback information (FBck) is received from the user equipment (UE1) that originated the first request (RQT). Management method according to claim 1 or claim 2, wherein the processing consists of storing in the cache (103): the first response (Resp-RQT), and a confidence score value representing the relevance of the first response. Management method according to claim 3, wherein, when the feedback information (FBck) indicates that the first response conforms to the expectations of a user of the user equipment (UE1) originating the first request (RQT), the value of the confidence score increases. Management method according to claim 4, wherein, when the value of the confidence score is less than a threshold, the first response is removed from the cache (103). Management method according to any one of claims 3 or 4, wherein the processing consists of: modifying the first response based on the feedback information (FBck) received; and storing the first modified response in the cache (103). Management method according to claim 6, wherein the first unmodified response is removed from the cache (103). Management method according to any one of claims 1 to 7, further comprising the storage of the first request (RQT) in association with the first response (Resp-RQT). Management method according to any one of claims 1 to 8, wherein the first response (Resp-RQT, Resp-Prev-RQT) is relative to a vector (Vect--RQT) associated with the first request (RQT), resulting from a comparison with a set of vectors (Vect-Prev-RQT) corresponding to requests (prev-RQT) previously received by the entity (10) of the communication network (NW). Management method according to any one of claims 1 to 9, wherein the feedback information (FBck) is stored in: a cache buffer (103-2), and / or a cache vector base (103-1). A management method according to any one of claims 1 to 10, wherein the feedback information (FBck) belongs to a group comprising at least: text data; audio data; an image; a video; an icon. A management method according to any one of claims 3 to 11, comprising: receiving a second request similar to the first request (RQT), selecting, based on the value of the confidence score relating to the first response, a second response to the second request; transmitting the second response to a user equipment (UE1) that originated the second request. A method of communication between a user equipment (UE1) of a first communication network (NW) and a temporary data storage, called a cache (103), in which are stored responses generated by a large language model LLM (105) implemented by an entity (10) of a second communication network (NW), the method comprising: receiving a response generated by the LLM (105) following the transmission of a request sent to the cache (103); sending, to the cache (103), feedback information (FBck) relating to the response received. Entity (10) of a communication network (NW) implementing a large LLM language model (105), said entity comprising: at least one temporary data storage, called a cache (103), in which responses generated by the LLM (105) are stored, and at least one processor (PROC) configured to: receive feedback information (FBck) relating to a first response (Resp-RQT, Resp-Prev-RQT) generated by the LLM (105) and transmitted to a user equipment (UE1) originating the issuance of a first request (RQT); and apply, based on the feedback information (FBck) received, a processing of the first response intended to be applied by the cache (103). User equipment (UE1) of a first communication network (NW) configured to communicate with at least one temporary data storage, called a cache (103), in which are stored responses generated by a large language model LLM (105) implemented by an entity (10) of a second communication network (NW), the user equipment (UE1) comprising at least one processor (PROC) configured to: receive a response generated by the LLM (105) following the transmission of a request (RQT) sent to the cache (103); send, to the cache (103), feedback information (FBck) relating to the response received. Computer program (PROG2) comprising instructions for the implementation of a management method according to any one of claims 1 to 12, when said program is executed by a computer (2). Computer program (PROG1) comprising instructions for the implementation of a communication method according to claim 13, when said program is executed by a computer (2).
