Managing context transmissions to backend services

Abstract

A device, system, and method are provided for managing context among multiple backend services. In examples, context management includes maintaining a plurality of context storages, each corresponding to a different backend service such as a different artificial intelligence (AI) service; receiving a user input from a client device via a conversation interface, the user input associated with a target backend service; determining a target context data store, the target context data store corresponding to the target backend service; transmitting the user input and a first context data from the target context data store to the target backend service, without transmitting a second context data corresponding to a context data store other than the target context data store to the target backend service; receiving a response to the user input from the target backend service; and outputting the response to the client device via the conversation interface.

Description

86353596MANAGING CONTEXT TRANSMISSIONS TO BACKEND SERVICESBACKGROUND

[0001] Artificial Intelligence (Al) models may be used as backend services. For example, generative Al modalities such as large language models (LLMs) may receive user input (e.g., commands or other entries input by a user of a computing device) and associated data and generate and return a response to the user, for example in a conversation-thread-type environment. In generating the output, the LLM may receive and use context, which may be data and metadata that is transmitted along with the user input and used by the LLM to generate responsive content.BRIEF DESCRIPTION OF THE DRAWINGS

[0002] The following drawings are provided to help illustrate various features of examples of the disclosure and are not intended to limit the scope of the disclosure or exclude alternative implementations.|0003| FIG. 1 illustrates an example user interface according to some aspects of the present disclosure.

[0004] FIG. 2 illustrates an example context management architecture according to some aspects of the present disclosure.

[0005] FIG. 3 illustrates an example context management categorization flow according to some aspects of the present disclosure.

[0006] FIG. 4 illustrates an example context management method according to some aspects of the present disclosure.

[0007] FIG. 5 illustrates an example context management system according to some aspects of the present disclosure.DETAILED DESCRIPTION

[0008] In some implementations, a single user interface may be presented to allow a user to interact with multiple different backend services (e.g., multiple different LLMs or other Al models) concurrently. In examples, the computing device presenting the user interface may route the user input and context to the appropriate destination backend service. As used herein, “context” refers to the information that the model uses to understand and generate responses. Generally, context may be any data or metadata that is transmitted to the backend service in1QB\93945883.286353596addition to the user input, and is usable by the backend service to generate a response to the user input. For example, and without limitation, context may include text surrounding the query in a user input, previous user inputs or conversational exchanges, current or previous information regarding the user or the user’s computing device (e.g., location data, user ID data, and the like), and / or any other relevant background information (e.g., attached documents, fdes the user currently has open, etc.) provided to the model. To avoid occupying memory and / or computing resources of the device operating the backend service (e.g., a server implementing the Al model), context may be transmitted along with each user input rather than stored within the device operating the backend service itself.

[0009] Context is relevant to the operation of LLMs for several reasons. For one, context affects the accuracy and relevance of responses generated by the model. By maintaining a rich context, LLMs can better understand the nuances of a conversation or text exchange, which leads to more accurate and contextually appropriate responses. This may be particularly relevant in scenarios such as customer support, where understanding the scope of a user’s issue helps provide effective solution. Moreover, context allows LLMs to tailor their responses to the individual needs and preferences of users. In interactive applications such as digital assistants, personalized responses based on the user’s history (e.g., prior interactions) can enhance user experience. Additionally, in dialogue systems, maintaining context across multiple exchanges may ensure continuity in the conversation even if the conversation includes multiple concurrent threads or topics. This enhances human-like interactions and maintains the flow of dialogue, making it more natural and engaging for the user. Furthermore, with a well-understood context, LLMs can generate responses more efficiently, reducing the need for back-and-forth clarification with the user. This speeds up interaction time and improves the overall efficiency of the communication process.

[0010] In comparative implementations, the computing device presenting the user interface (or an intermediate computing device) maintains a context store (e.g., a storage area where context is stored) and transmits context from the context store to the backend service along with each user input. Thus, the backend service invoked receives the user input and uses the context to generate a response. However, if the user is interacting with multiple different backend services (e.g., multiple different LLMs), the destination backend service for a particular user input receives context related to the destination backend service but also context related to other backend services that may have been previously invoked by the user earlier in the interaction. For example, if a user is interacting with the user interface to conduct a2QB\93945883.286353596conversation related to one specialized topic (e.g., a health-related conversation) and then inputs something related to another specialized topic (e.g., a tech-related conversation), then according to comparative examples the computing device transmits context that is related to both specialized topics to the destination backend service. These comparative examples suffer from several technical disadvantages.

[0011] One non-limiting example of such a disadvantage is that, because the transmitted context includes content that is unrelated to the topic at hand, the receiving backend service is more likely to provide inaccurate, misleading, or false responses. Another non-limiting example is that a particular backend service may receive information that should be kept private and / or secure (e.g., health or banking information). In some instances, such a privacy breach may run counter to laws or regulations governing data security. Moreover, because the transmitted context includes more content, the interaction between the user and the backend services utilizes more communication overhead (e.g., transmission resources). The increased amount of context received by the backend services results in an increase in the amount of processing and / or memory resources that the backend service uses in responding to the user input. Where the different backend services have diverse operational requirements and / or data formats, comparative examples often necessitate extensive customization or configuration. Moreover, the scalability of the backend services is reduced, for example due to an increased need for the backend services to clear their respective caches and / or reinitialize conversational threads. This is especially true when managing many services and / or handling a high volume of interactions.

[0012] The present disclosure addresses these and other issues in the comparative examples by providing for systems, methods, and media for context management between a user device and multiple backend services. In examples, the context management modalities set forth herein provide for isolation of context between different backend services, such that each backend service receives only the context related to the interaction between the user and that particular backend service and does not receive context related to the interaction(s) between the user and other backend services. In some examples, the present disclosure sets forth a context manager (e.g., a context management system, method, or medium) that maintains isolated context windows for each backend service and routes each user input to the appropriate backend service along with the context corresponding to that service and only that service. This provides several technical advantages over the comparative examples, including improved accuracy of responses from the backend services, improved data privacy and security,3QB\93945883.286353596improved legal and / or regulatory compliance, increased scalability, reduced communication overhead for transmissions to the backend services (and an associated reduction in cost for invoking the models), and reduced processing and / or memory overhead by the backend services themselves.

[0013] FIG. 1 illustrates an example user interface 100 according to various aspects of the present disclosure. The user interface 100 may be presented to a user of a computing device, for example by being displayed in a graphical user interface (GUI) of the computing device. As used herein, a “computing device” may be any device having a processor and a memory, including but not limited to a desktop computer, a laptop computer, a tablet computer, a notebook computer, a smartphone, a smartwatch, a wearable computer (e.g., a virtual reality headset), an in-vehicle computer, and the like. The user interface 100 may represent a standalone application (“app”) installed on the computing device, a web browser interface on the computing device, or any other application structure. For simplicity, the user interface 100 is referred to herein as being an “application.” The user interface 100 presents a conversation interface to the user, which includes a text entry box 110 by which the user may enter text¬ based user inputs. While not expressly illustrated in FIG. 1, the conversation interface may additionally or alternatively include other input modalities, such as a speech input interface, a gesture input interface, and the like, in which case the user interface 100 may include voice capture, image capture, and / or video capture capabilities. When a user enters an input into the text entry box 110, it populates a conversation thread which, as shown, includes several past user inputs 120-124. Responsive to the user inputs 120-124, the conversation thread includes several backend service responses 130-132. The backend service responses 130-132 may correspond to the same or different backend services. Moreover, while FIG. 1 illustrates three user inputs 120-124 and two backend service responses 130-132, in practice any number of user inputs and / or backend service responses may be present.

[0014] As used herein, a “conversation thread” refers to the continuous sequence of exchanges or messages that occur between a model or models and a user within a single interaction session. This thread comprises the inputs from the user (e.g., inputs 120-124) and the corresponding responses from the model(s) (e.g., responses 130-132), linked together to form a coherent dialogue. Each part of the thread builds upon previous messages, allowing the model to maintain a continuous and context-aware conversation. For example, one user input may introduce a person or topic, and a later user input may simply use a pronoun such as “she” or “it” to refer back to the person or topic. Unlike comparative systems where a single service4QB\93945883.286353596handles all exchanges within a thread (which may introduce ambiguity where multiple people / topics are present), the present disclosure allows for different segments of a conversation thread to be managed by distinct backend services. Each service independently processes and responds to user inputs while maintaining its own isolated context. This approach enables a single conversation thread to span multiple backend services, each preserving a separate and secure context.

[0015] FIG. 2 illustrates an example high-level architecture 200 for context management. For example, FIG. 2 may illustrate a logical flow between computing elements underlying the user interface 100. As shown in FIG. 2, the architecture 200 includes an Al interface application 210 (which may be the same as or similar to the application presenting the user interface 100 of FIG. 1), a context orchestrator 220, and a plurality of backend services 230. For illustration purposes, the plurality of backend services 230 includes an Al-based online search service 232, an Al-based tech support service 234, an Al-based image generation service 236, and an AI-based virtual health assistant service 238. It should be understood that more, fewer, and / or different backend services may be present. Moreover, while the plurality of backend services 230 are illustrated as a single group, it should be understood that at least some of the individual backend services 230, in any combination, may be local to the computing device presenting the Al interface application 210 or may be remote from the computing device presenting the Al interface application 210 (e.g., cloud-based). Additionally, the present disclosure presents examples in which the plurality of backend services 230 are Al -based backend services, in practical implementations some of the backend services 230 may not be Al-based (e.g., may be a simple math backend service or the like).

[0016] The Al interface application 210 permits an interface by which the user may ask questions around many different topics, such as questions related to online information, questions related to local fdes, questions related to cloud-based files, and the like. The application 210 presents a single interface through which the user may invoke many different backend services 230. The backend services 230 are, for examples, specialized solutions that are activated based on the user input (the “prompt”). As an example, if the user asks a question that may be answered using an online search, the online search backend service 232 will be activated; if the user asks for a custom image to be created, the image generation backend service 236 will be activated; if the user asks a question about a particular product, the tech support backend service 234 will be activated, and so on.

[0017] The architecture 200 takes the user input and provides it to the context orchestrator5QB\93945883.286353596220 to select which backend service 230 should be used. The context orchestrator 220 determines the target backend service 230 based on the content of the user input (and, in some implementations, other information related to the user input) and routes the user input accordingly. In the illustrated example, the context orchestrator 220 itself sends the user input to the appropriate backend service 230; however, in other implementations the context orchestrator 220 may instead return the user input (and, in some implementations, corresponding context or other information) to another computing element and the other computing element may send the user input to the appropriate backend service.

[0018] In one example, the context orchestrator 220 may receive a user input and a context (e.g., a user context and / or a device context). Based on the user input and the context, the context orchestrator 220 may determine an intent (or domain) of the user input (e.g., via an intent mapping component of the orchestration component). The context orchestrator 220 may determine which backend service is most suited to handle the user input based on the intent (e.g., via an agent dispatch component). For instance, a math-focused backend service is better suited to handle a user input related to math while a web search backend service is better suited to handle a user input related to generally available information. The context orchestrator 220 may dispatch the user input to the best suited backend service for processing and / or may maintain context storage(s) accordingly. In some examples, the context orchestrator 220may detect a change in context (or intent) and call the intent mapping component again and choose a different backend service based on the changed context. The context orchestrator 220 also manages overall chat history and memory such that seamless transitions between agents may be achieved.

[0019] Table 1 presents a conversation thread for an example interaction between a hypothetical user and the Al-based backend services, mediated by a single application interface.Table 1 Hypothetical Conversation Thread with Multiple Services Used to Fulfdl Requests6QB\93945883.286353596

[0020] In the example shown in Table 1, the user engages with different topics, ranging from online searches and tech issues to image generation, in the same conversation thread. The prompts in this sample conversation thread can be tagged with different backend services, for example via the context orchestrator 220, as shown in Table 2.Table 2 - Backend Services Activated Based on the User PromptQB\93945883.286353596

[0021] FIG. 3 illustrates an example logical flow for categorizing the hypothetical conversation of Tables 1 and 2 to further explain the context management modalities of the present disclosure. In the example, a user input stream 310 comprising a plurality of different user prompts is input by a user into an Al interface application 320. The Al interface application 320 may be the same as or similar to the Al interface application 210 illustrated in FIG. 2. In this example, the user input stream 310 includes eight prompts related to three different backend services, as illustrated by the upward-facing pentagon, octagon, and downward-facing pentagon symbols. The Al interface application 320 determines which user prompt corresponds to which backend service (for example, by invoking a context orchestrator such as the context orchestrator 220 of FIG. 2). The Al interface application 320 interacts with a storage 330 (e.g., a memory of the computer device operating the Al interface application 320 and / or a cloudbased memory) and maintains a plurality of context stores 332-336, each of which corresponds to a different backend service. Thus, the user inputs corresponding to a first backend service (e.g., inputs 1, 2, and 4 of Tables 1 and 2, corresponding to an online search service) are allocated to a first context store 332, user inputs corresponding to a second backend service (e.g., inputs 3, 5, 6, and 8 of Tables 1 and 2, corresponding to a tech support service) are allocated to a second context store 334, and so on. The context stores 332-336 may further include other types of context such as, as noted above, information regarding the user or the user’s computing device, other relevant background information, and the like. In some cases, some or all of the storage 330 (e.g., some or all of the context stores 332-336) may be secure or encrypted. Each conversation thread might have many backend services activated, depending on the user inputs and the services available. Each backend service activated will have its own context storage, and each context storage is isolated and pertains only to the specific backend service. Context storages are not shared among different backend services. While FIG. 3 illustrates each of the context stores 332-336 within the same storage 330, a plurality of storages may be utilized and the context stores 332-336 may be distributed among the storages in any combination (e.g., context store 332 in a first storage, context stores 334 and 336 in a second storage, etc.). Moreover, the context stores 332-336 need not be physically8QB\93945883.286353596distinct areas in the storage 330. In some implementations, a single physical context store may be implemented and the context allocated to each backend service may be identified through the use of metadata tags or other identifiers thereby to logically differentiate the context.

[0022] FIG. 4 illustrates an example context management method 400 according to example implementations of the present disclosure. The method 400 may be performed by at least one electronic processor, such as at least one electronic processor of a client device running an Al conversation application. At operation 402, the method includes maintaining a plurality of different context data stores. Each of the plurality of context data stores may correspond to a different backend service of a plurality of backend services. The plurality of backend services may respectively correspond to different Al-based services.

[0023] At operation 404, a user input is received from a client device via a conversation interface (e.g., as illustrated in FIG. 1). The user input may be associated with a particular backend service from among the plurality of backend services, referred to as a “target” backend service. The user input is analyzed to determine a target context data store (that is, the context data store corresponding to the target backend service) from among the plurality of context data stores. Operation 406 may be performed by, under the control of, or in coordination with a context orchestrator service (e.g., the context orchestrator 220 of FIG. 2) in some examples. Operation 406 may include receiving a context data store indicator that is based on a content of the user input, and selecting the target context data store from among the plurality of context data stores based on the context data store indicator.

[0024] After the target context data store is determined, at operation 408 the user input and a first context data (e.g., context data from the target context data store) is transmitted to the target backend service. To ensure complete isolation of context, only the first context data store is transmitted and context data corresponding to other (non-target) context data stores are not sent to the target backend service. The target backend service may generate a response based on the user input and the first context data, which is received at operation 410. Then, at operation 512, the response is presented to the user (e.g., by being displayed on the client device via the conversation interface.

[0025] While FIG. 4 illustrates the operations 402-412 being performed one after another in a particular order, the present disclosure is not so limiting. In examples, at least a portion of the operations 402-412 may be performed in another order and / or two or more of the operations 402412 may be performed simultaneously or substantially simultaneously (i.e., in parallel). In9QB\93945883.286353596one particular example, operation 402 may be performed after and in response to operation 404 and / or 406, such that a context data store is maintained (e.g., created, modified, updated, closed, etc.) when implicated by a particular user input. For example, the target context data store (and, in implementations, only the target context data store) may be updated based on the user input. As another example, if the target data store does not yet exist (i.e., if at operation 406 it is determined that none of the context data stores corresponds to the target backend service), the target context data store may be initiated. The target context data store may be initiated by, for example, designating an area of storage for the target context data store, resetting the area of storage for the target context data store, partially resetting the area of storage for the context data store, and the like.

[0026] In some examples, the method 400 may include (e.g., as part of operation 402 or as another operation) determining whether a size of a context data store exceeds a threshold, which may be a predetermined storage size. If a context data store does exceed the threshold, the method 400 may include reinitializing the interaction; for example, outputting a request to the client device requesting the user to reinitialize an interaction corresponding to the context data store, generating an auto-summary, and / or resetting the context data store. The user may be able to reinitialize the interaction by entering a command to clear the conversation (in which case the corresponding context data store may be partially or totally erased). In implementations, the user may be able to control and / or modify settings to determine which actions are perfomied when the size of the context data store exceeds the threshold.

[0027] FIG. 5 illustrates an example context management system 500 according to example implementations of the present disclosure. The system 500 includes at least one electronic processor 510 and a memory 520 operatively connected to the at least one electronic processor 510. The at least one electronic processor 510 is to cause the context management system 500 to perform several operations. For purposes of illustration, the memory 520 is illustrated as having sets of instructions 522-530 that are executable by the at least one electronic processor 510 to perform the operations; however, in implementations a single set of instructions may be present. Thus, the memory 520 may be, include, and / or implement a non-transitory computer-readable medium that stores instructions executable by the at least one processor 510.

[0028] As illustrated, the memory 520 includes instructions 522, such that the at least one electronic processor 502 causes the system 500 to maintain a plurality of context data stores in the memory 520 (and / or in another memory), each of the plurality of context data stores corresponding to a different backend service of a plurality of backend computing services. The10QB\93945883.286353596memory' 520 includes instructions 524. such that the at least one electronic processor 502 causes the system 500 to receive a user input via a conversation interface (e.g., a text- and / or voice-based interface), the user input being associated with a target backend service from among the plurality of backend computing services (e.g., Al services). The memory 520 further includes instructions 526, such that the at least one electronic processor 502 causes the system 500 to determine a target context data store from among the plurality’ of context data stores, the target context data store corresponding to the target backend computing services. For example, the system 500 may receive a context data store indicator, the context data store indicator being based on a content of the user input, and determine the target context data store based on the context data store indicator.

[0029] The memory 520 as illustrated further includes instructions 528, such that the at least one electronic processor 502 causes the system 500 to transmit the user input and a first context data from the target context data store to the target backend computing service without transmitting a second context data corresponding to a context data store other than the target context data store to the target backend computing service. The memory' 520 includes instructions 530, such that the at least one electronic processor 502 causes the system 500 to receive a response to the user input from the target backend service and / or to output the response via the conversation interface. The memory' 520 may further include instructions such that the at least one electronic processor causes the system 500 to update the target data store based on the user input; and / or causes the system 500 to determine whether a size of the context data store exceeds a threshold and, if so, output a request to reinitialize an interaction corresponding to the target context data store and / or reset the target context data store. If none of the plurality' of context data stores corresponds to the target backend computing service, instructions may cause the system 500 to initiate the target context data store.

[0030] Certain operations of methods according to the technology7, or of systems executing those methods, can be represented schematically in the figures or otherwise discussed herein. Unless otherwise specified or limited, representation in the figures of particular operations in particular spatial order can not necessarily require those operations to be executed in a particular sequence corresponding to the particular spatial order. Correspondingly, certain operations represented in the figures, or otherwise disclosed herein, can be executed in different orders than are expressly illustrated or described, as appropriate for particular examples of the technology. Further, in some examples, certain operations can be executed in parallel,11QB\93945883.286353596including by dedicated parallel processing devices, or separate computing devices that interoperate as part of a large system.

[0031] The disclosed technology is not limited in its application to the details of construction and the arrangement of components set forth in the following description or illustrated in the following drawings. Other examples of the disclosed technology are possible and examples described and / or illustrated here are capable of being practiced or of being carried out in various ways.

[0032] A plurality7of hardware and software-based devices, as well as a plurality' of different structural components can be used to implement the disclosed technology’. In addition, examples of the disclosed technology can include hardware, software, and electronic components or modules that, for purposes of discussion, can be illustrated and described as if the majority of the components were implemented solely in hardware. However, in one example, the electronic based aspects of the disclosed technology can be implemented in software (for example, stored on non-transitory computer-readable medium) executable by a processor. Although certain drawings illustrate hardware and software located within particular devices, these depictions are for illustrative purposes. In some examples, the illustrated components can be combined or divided into separate software, firmware, hardware, or combinations thereof. As used herein, “software” and “firmware” are examples of machine-readable instructions. As one example, instead of being located within and performed by a single electronic processor, logic and processing can be distributed among multiple electronic processors. Regardless of how they are combined or divided, hardware and software components can be located on the same computing device or can be distributed among different computing devices connected by a network or other suitable communication links.

[0033] Any suitable non-transitory computer usable or computer readable medium may be utilized. The computer-usable or computer-readable medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device. More specific examples (anon-exhaustive list) of the computer-readable medium would include the following: a portable computer diskette, a hard disk, a randomaccess memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a portable compact disc read-only memory (CD-ROM), an optical storage device, or a magnetic storage device. In the context of this disclosure, a computer-usable or computer-readable medium may be any medium that can contain, store,12QB\93945883.286353596communicate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device.

[0034] As used herein in the context of computer implementation, unless otherwise specified or limited, the terms “component,” “system.” “module,” “block,” and the like are intended to encompass part or all of computer-related systems that include hardware, software, a combination of hardware and software, or software in execution. For example, a component can be, but is not limited to being, a processor device, a process being executed (or executable) by a processor device, an object, an executable, a thread of execution, a computer program, or a computer. By way of illustration, both an application running on a computer and the computer can be a component. Components (or system, module, and so on) can reside within a process or thread of execution, can be localized on one computer, can be distributed between two or more computers or other processor devices, or can be included within another component (or system, module, and so on).13QB\93945883.2

Claims

86353596CLAIMS1. A context management method, comprising:by at least one processor:maintaining a plurality of context data stores, each of the plurality of context data stores corresponding to a different backend service of a plurality of backend services, the plurality of backend services respectively corresponding to different artificial intelligence (Al) sendees;receiving a user input from a client device via a conversation interface, the user input associated with a target backend service from among the plurality of backend services;determining a target context data store from among the plurality of context data stores, the target context data store corresponding to the target backend service;transmitting the user input and a first context data from the target context data store to the target backend senice. without transmitting a second context data corresponding to a context data store other than the target context data store to the target backend service;receiving a response to the user input from the target backend service; and outputting the response to the client device via the conversation interface.

2. The method of claim 1, wherein determining the target context data store includes:receiving a context data store indicator, the context data store indicator being based on a content of the user input; andselecting the target context data store from among the plurality of context data stores based on the context data store indicator.

3. The method of claim 1. further comprising, by the at least one processor: updating the target context data store based on the user input.

4. The method of claim 1, further comprising, by the at least one processor: updating the target context data store based on the user input without updating ones of the plurality of context data stores other than the target context data store.14QB\93945883.2863535965. The method of claim 1. further comprising, by the at least one processor: determining whether a size of the target context data store exceeds a threshold; and in response to a determination that the size of the target context data store exceeds the threshold, outputting to the client device a request to reinitialize an interaction corresponding to the target context data store.

6. The method of claim 5, further comprising, by the at least one processor: resetting the target context data store.

7. The method of claim 1. further comprising, by the at least one processor: determining that none of the plurality of context data stores corresponds to the target backend service; andinitiating the target context data store.

8. A context management system, comprising:at least one processor; anda memory operatively connected to the at least one processor, the at least one processor to cause the context management system to:maintain a plurality of context data stores in the memory, each of the plurality of context data stores corresponding to a different backend service of a plurality of backend computing services;receive a user input via a conversation interface, the user input associated with a target backend service from among the plurality of backend computing services; determine a target context data store from among the plurality of context data stores, the target context data store corresponding to the target backend computing service;transmit the user input and a first context data from the target context data store to the target backend computing service, without transmitting a second context data corresponding to a context data store other than the target context data store to the target backend computing service;receive a response to the user input from the target backend service; and output the response via the conversation interface.15QB\93945883.2863535969. The system of claim 8, the plurality of backend services respectively corresponding to different artificial intelligence (Al) services.

10. The system of claim 8, the at least one processor to cause the context management system to:receive a context data store indicator, the context data store indicator being based on a content of the user input; anddetermine the target context data store based on the context data store indicator.

11. The system of claim 8, wherein the conversation interface is at least one of a text-based interface, an image-based interface, a video-based interface, or a voice-based interface.

12. The system of claim 8, the at least one processor to cause the context management system to, in response to a determination that none of the plurality of context data stores corresponds to the target backend computing service, initiate the target context data store.

13. The system of claim 8, wherein at least one of the plurality of backend computing services is remote from the client device.

14. A non-transitory computer-readable medium containing instructions that when executed cause a processor to:maintain, in a memory operatively connected to the processor, a plurality of context data stores, each of the plurality of context data stores corresponding to a different backend artificial intelligence (Al) service of a plurality of backend Al services;receive a user input from an input device via a conversation interface, the user input associated with a target backend Al service from among the plurality of backend Al services, the conversation interface corresponding to each of the plurality of backend Al services; determine a target context data store from among the plurality of context data stores, the target context data store corresponding to the target backend Al service;transmit the user input and a first context data from the target context data store to the target backend service, without transmitting a second context data corresponding to any context data store other than the target context data store to the target backend Al service;16QB\93945883.286353596receive a response to the user input from the target backend Al service; and outputting the response to the client device via the conversation interface.

15. The computer-readable medium of claim 14, the instructions when executed causing the processor to:receive a context data store indicator, the context data store indicator being based on a content of the user input; anddetermine the target context data store based on the context data store indicator.QB\93945883.2

Images