An intelligent care conversation method based on intent routing

Abstract

The application discloses an intelligent care conversation method and system based on intention routing, and belongs to the technical field of artificial intelligence conversation systems. The method identifies five care intention categories through deterministic substring matching of a care field keyword library, routes to a visual language model, a medical reasoning model, an emergency treatment model or a general conversation model according to a fixed priority rule of a four-model routing table, uniformly degrades to a backup general model when the main model fails, and pushes a reply in real time through a streaming transmission protocol. The application avoids the misjudgment risk of a probability classifier and a state transition matrix in a care scene, and ensures zero misjudgment identification and reliable routing of safety-critical intentions.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence dialogue system technology, specifically to an intelligent care dialogue method and system based on intent routing, applicable to application scenarios such as elderly care, health consultation, and medication guidance. Background Technology

[0002] In intelligent care AI dialogue systems, users have diverse needs, including daily care consultations, care assessments, medication advice, image recognition, and emergency detection. Existing AI dialogue systems face the following technical challenges when handling these diverse needs:

[0003] First, existing systems typically use a single language model to process all types of user messages, failing to select the most suitable model based on message intent and type. For example, image recognition requires a visual language model, assessments and emergencies require medical-specific models, while routine consultations can use a general dialogue model. Using a single model leads to insufficient response quality in some scenarios.

[0004] Second, some existing systems use probabilistic intent classifiers or dialogue state tracking techniques based on state transition matrices to identify user intents and then route them to a large language model. These approaches have the following drawbacks: probabilistic classifiers output ranking results with confidence levels, which may lead to misinterpretation of intents in the caregiving domain (e.g., misinterpreting an emergency message as a routine consultation), and require a large amount of labeled data for training; state transition matrix approaches model dialogues as state machines, which lack flexibility and cannot handle situations where multiple intents coexist.

[0005] Third, existing systems lack precise identification mechanisms for specific intentions in the care domain. Intention categories in the care domain (such as care assessment, emergency detection, and medication consultation) have clear keyword triggering characteristics, do not require probabilistic reasoning, but require deterministic matching to ensure safety and reliability.

[0006] Fourth, the existing system lacks a degradation mechanism when the main model is unavailable, resulting in complete system unavailability when a model fails. Furthermore, the existing system's model routing scheme does not have a unified fallback capability, making it unable to maintain basic services when multiple specialized models, such as visual models and medical models, are simultaneously unavailable.

[0007] Fifth, the existing system lacks a mechanism to perform business boundary security checks on user messages before model routing, which may cause AI assistants to answer questions unrelated to business. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides an intelligent care dialogue method and system based on intent routing. By using deterministic keyword matching and recognition of specific intents in the care domain, fixed priority routing of a four-model routing table, and a unified fallback mechanism for multiple models, an intelligent dialogue scheme is achieved that selects the most suitable model for processing based on the type of user needs.

[0009] The core difference between this invention and existing technologies lies in the following: intent recognition uses deterministic substring matching based on a keyword database in the care domain instead of a probabilistic classifier or state transition matrix; model routing uses a four-model routing table with fixed priority instead of a dynamic decision tree; and degradation uses a unified backup model instead of multi-level cascading.

[0010] The technical solution of this invention to solve the above-mentioned technical problems is as follows: an intelligent care dialogue method based on intent routing, comprising: receiving user messages; performing three-layer content security checks; identifying five intent categories through deterministic matching of a care domain keyword library; routing to the corresponding language model (visual model, medical reasoning model, emergency treatment model, or general model) according to the fixed priority rules of the four-model routing table; automatically downgrading to a unified backup model when the main model fails; and pushing responses in real time through a streaming protocol.

[0011] The beneficial effects of this invention are as follows:

[0012] 1) Deterministic intent recognition: Five intent categories are identified by matching deterministic substrings from a keyword library in the care domain. The matching result is either a hit or a miss, which does not involve probability ranking and avoids the risk of misjudgment by probability classifiers in care scenarios (such as misjudging emergency messages as routine consultations). No labeled data training is required.

[0013] 2) Fixed priority routing: The four-model routing table adopts a fixed priority sorting rule, with emergency cases taking priority over images, images taking priority over assessments, assessments taking priority over medications, and medications taking priority over default general dialogues. This ensures the absolute priority of safety-critical scenarios and avoids uncertain routing results generated by the dynamic decision tree in boundary situations.

[0014] 3) Unified degradation fallback: All main models (visual model, medical reasoning model, emergency treatment model, general dialogue model) share the same backup general text generation model as the degradation target, ensuring that the system can still provide basic dialogue services when any or more dedicated models are unavailable, and that the dialogue context is not lost during degradation.

[0015] 4) Three-layer security protection: Through a three-layer mechanism of whitelist priority access, blacklist blocking, and default access, the AI ​​assistant is ensured to only answer questions within the scope of business. Attached Figure Description

[0016] Figure 1A flowchart of an intent-based intelligent care dialogue method provided in an embodiment of the present invention;

[0017] Figure 2 This is an architecture diagram of an intelligent care dialogue system provided in an embodiment of the present invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, this embodiment provides an intelligent care dialogue method based on intent routing, including the following steps:

[0020] Step S1: Receive dialogue messages input by the user through the terminal. Messages include at least one of text messages and image attachments. The user inputs messages through the dialogue interface of the terminal device, and the system receives the messages through the HTTP interface, while simultaneously obtaining the user identifier, session identifier, and list of image attachments.

[0021] Step S2: Perform content security checks on user messages. The checks employ a three-layer mechanism: First, check the whitelist, which contains business terms related to the care industry (such as "elderly care," "care," "nursing," "rehabilitation," "blood pressure," "blood sugar," "elderly," "parents," etc.). If a message contains a term from the whitelist, it is allowed to pass directly without checking the blacklist. Second, check the blacklist, which contains keywords completely unrelated to care services (such as games, programming, politics, etc.). If a message matches the blacklist but does not contain any terms from the whitelist, it is blocked and a rejection message is returned. If a message does not match either the whitelist or the blacklist, it is allowed to pass by default, leaving the decision to the language model.

[0022] Step S3: Identify the user's dialogue intent into one of five preset care domain intent categories. The five intent categories are: Intent A: Care Assessment, where the user message contains assessment initiation words (e.g., "Help me assess" or "Do an assessment"), and the processing method is to initiate the assessment process by asking questions dimension by dimension; Intent B: Routine Care Consultation, where the user directly asks care-related questions, and the processing method is to directly provide specific suggestions; Intent C: Image Recognition, where the user uploads image attachments, and the processing method is to call a visual language model for recognition and analysis; Intent D: Emergency Detection, where the user message matches a preset emergency keyword library (e.g., "chest pain," "stroke," or "coma"), and the processing method is to interrupt the process for emergency treatment; Intent E: Service Recommendation, triggered when the user explicitly expresses a need that can be met by care services in the dialogue, and the processing method is to proactively recommend corresponding services after providing suggestions. The intent recognition uses deterministic substring matching from the care domain keyword library, rather than a probability-based intent classifier or dialogue state tracking based on a state transition matrix. The matching result is either a hit or a miss, without involving probability ranking, ensuring zero misjudgment in intent recognition for safety-critical scenarios such as emergencies.

[0023] Step S4: Based on the intent recognition result and input type, the message is routed to the corresponding language model according to the fixed priority rules of the preset four-model routing table. The four-model routing table defines the correspondence between four language models and intent types: the visual language model is used for image attachment messages (recognizing photos of drug instructions, medicine boxes, prescriptions, wounds, blood glucose records, etc.); the medical reasoning language model is used for evaluation processes and medication consultation scenarios (possessing medical professional reasoning capabilities, asking questions dimension by dimension during evaluation); the emergency treatment language model is used for emergency detection hit scenarios (emergency instruction patterns of the same medical-specific model, generating first aid guidance); and the general dialogue model is used for daily consultation and service recommendation scenarios. The routing priority follows a fixed sorting rule: emergency keyword hits take precedence over image input, image input takes precedence over evaluation initiation words, evaluation initiation words take precedence over medication consultation keywords, and medication consultation keywords take precedence over the default general dialogue model. This fixed priority ensures the absolute priority of safety-critical scenarios and avoids uncertain routing results generated by the dynamic decision tree in boundary situations. Once the assessment process has started and the conversation has gone through multiple rounds, the medical reasoning language model is maintained instead of switching back to the general dialogue model. This is conversation-level stickiness rather than message-level routing, ensuring consistency of medical expertise throughout the assessment process.

[0024] Step S5: When the primary model fails to be invoked, automatically switch to the backup general text generation model for degradation processing. The backup model serves as the unified degradation target for all primary models (visual model, medical reasoning model, emergency treatment model, and general dialogue model), ensuring that the system can still provide basic dialogue services when one or more dedicated models are unavailable. During degradation, the dialogue context is not lost, and the user seamlessly switches from the dedicated model to the backup model.

[0025] Step S6: Push the responses generated by the language model to the user terminal in real time via a streaming protocol. The server sends events, and the responses are pushed incrementally in real time. During the push process, model identification information (informing the user which model is currently being used) and thinking status information (for models with reasoning capabilities, notifying the user that the model is thinking) are sent simultaneously.

[0026] This embodiment also includes a proactive service recommendation step: when intent E is detected, contact information for the corresponding service is appended to the end of the response. This contact information includes a phone number (clickable) and an online customer service link. Recommendations only appear after the user has explicitly expressed their need; there is no proactive sales pitch.

[0027] This embodiment also includes a post-processing step: the response content generated by the language model is detected, and when non-compliant negative brand statements are detected, they are automatically replaced with standard scripts that guide users to contact customer service.

[0028] like Figure 2As shown, this embodiment also provides an intelligent care dialogue system based on intent routing, including: a message receiving module, a content security module, an intent recognition module (deterministic keyword matching), a model routing module (four-model routing table with fixed priority), a degradation processing module (unified backup model), and a streaming response module. These modules work collaboratively to achieve intent recognition, multi-model routing, security checks, and unified degradation processing.

[0029] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent care dialogue method based on intent routing, characterized in that, Includes the following steps: S1: Receive a dialogue message input by the user through the terminal, the message including at least one of text message and image attachment; S2: Perform content security checks on user messages. The content security checks adopt a three-layer mechanism of whitelist priority access, blacklist blocking, and default access. The whitelist contains business terms in the care field, and the blacklist contains keywords unrelated to care business. S3: Identify user dialogue intent to one of five preset care field intent categories. The five intent categories are care assessment, daily care consultation, image recognition, emergency detection, and service recommendation. Care assessment is when the user message contains an assessment initiation word, emergency detection is when the user message hits a preset emergency keyword library, and service recommendation is triggered when the user explicitly expresses a need that can be met by care services in the dialogue. The intent identification is a deterministic matching based on the care field keyword library rather than a probability-based intent classifier. S4: Based on the intent recognition result and input type, the message is routed to the corresponding language model according to the preset four-model routing table. The four-model routing table defines the correspondence between four language models and intent types. The four language models are visual language model, medical reasoning language model, general dialogue model, and emergency treatment language model. Among them, the medical reasoning language model and the emergency treatment language model are different instruction modes of the same medical-specific model. S5: When the main model call fails, it automatically switches to the backup general text generation model for degradation processing. The backup model serves as the unified degradation target for all main models. S6: The response generated by the language model is pushed to the user terminal in real time through a streaming protocol.

2. The method according to claim 1, characterized in that, In step S3, the identification of the five intent categories uses a keyword library in the care domain for deterministic substring matching, rather than intent classification based on machine learning probability models or dialogue state tracking based on state transition matrices. The matching result is either a hit or a miss, and does not involve probability ranking.

3. The method according to claim 1, characterized in that, In step S4, the priority rules of the four-model routing table are as follows: emergency keyword hits take priority over image input, image input takes priority over evaluation start words, evaluation start words take priority over medication consultation keywords, and medication consultation keywords take priority over the default general dialogue model. The priority is a fixed sorting rule rather than a dynamic decision tree.

4. The method according to claim 1, characterized in that, In step S3, once the evaluation process has started and the conversation has gone through multiple rounds of dialogue, the medical reasoning language model is maintained without switching back to the general dialogue model, so that the dialogue in the evaluation process maintains the consistency of medical professional competence. This behavior is session-level stickiness rather than message-level routing.

5. The method according to claim 1, characterized in that, In step S2, the whitelist contains business terms related to care services. When a user message contains a term from the whitelist, it is allowed to pass directly without checking the blacklist. The blacklist contains keywords that are completely unrelated to care services. When a user message matches the blacklist but does not contain a term from the whitelist, it is blocked and a rejection message is returned. When a user message matches neither the whitelist nor the blacklist, it is allowed to pass by default.

6. The method according to claim 1, characterized in that, In step S4, the visual language model is used to process messages containing image attachments, including the recognition and analysis of photos of drug instructions, medicine boxes, prescriptions, wounds, and blood glucose records; the medical reasoning language model is used to process assessment processes and medication consultation scenarios; the emergency treatment language model is used to process emergency detection scenarios; and the general dialogue model is used to process daily care consultation and service recommendation scenarios.

7. The method according to claim 1, characterized in that, In step S5, the backup general text generation model serves as the downgrade target for all main models. When the visual language model, medical reasoning language model, or general dialogue model fails to be invoked, the backup model is automatically switched to, and the dialogue context is not lost during the downgrade.

8. The method according to claim 1, characterized in that, In step S6, the streaming protocol uses a server-sending event method, and the response content generated by the language model is pushed to the user terminal in real time in an incremental manner. During the push process, model identification information and thinking status identification information are sent simultaneously.

9. The method according to claim 1, characterized in that, It also includes a proactive service recommendation step: when the intent is identified as a service recommendation, the corresponding care service contact information is appended to the end of the language model response. The contact information includes a phone number and an online customer service link. The recommendation only appears after the user has clearly expressed their needs and is not proactively promoted.

10. The method according to claim 1, characterized in that, It also includes a post-output processing step: the response content generated by the language model is inspected, and when non-compliant negative brand statements are detected, they are automatically replaced with standard scripts that guide users to contact customer service.

11. An intelligent care dialogue system based on intent routing, characterized in that, include: The message receiving module is used to receive dialogue messages input by the user through the terminal; The content security module is used to perform three-layer content security checks on user messages; The intent recognition module is used to identify five intent categories through deterministic matching of a keyword database in the care field; The model routing module is used to route messages to the corresponding language model according to the fixed priority rules of the four-model routing table; the degradation processing module is used to switch to the backup general model when the main model fails; and the streaming reply module is used to push reply content in real time through the streaming transmission protocol.

Images