Conversational system with artificial intelligence for digital commerce

Abstract

A dialogue-based artificial intelligence system for interaction management in digital commerce, consisting of: a housing that encloses an arrangement of interconnected printed circuit boards; a conversation processing unit arranged on a first circuit board, comprising a syntactic parsing processor for generating hierarchical parse structures from user input texts, a semantic embedding processor for generating high-dimensional semantic vectors from the parsed structures, and a dialogue state calculation circuit for managing a multi-level conversation memory using clocked recurrent memory buffers; a retrieval augmentation unit arranged on a second circuit board, comprising a vector indexing processor for storing and retrieving domain-specific embedding vectors, a similarity calculation processor for calculating cosine-based semantic distances between the semantic vectors and vectors stored in a knowledge store, and a document reconstruction structure for merging retrieved text segments into a unified context representation; a generative response unit arranged on a third circuit board, comprising a transformer execution processor for executing a quantized transformer architecture for natural language generation, a context fusion circuit for merging the unified context representation with the semantic vectors to generate response conditioning data, and a hallucination suppression circuit for comparing the generated text with retrieved text segments and attenuating content divergence beyond a predetermined semantic threshold; a domain integrity check unit that is communicatively connected to the generative response unit, wherein the domain integrity check unit includes a regulatory restriction store that stores domain-driven rules, and a conformance assessment processor configured to check generated responses for conformance with the domain-driven rules, blocking non-conforming segments or replacing them using verified text content from the retrieval extension unit; a knowledge storage array comprising at least one static domain repository for storing commercial specifications, regulatory guidelines, product descriptions, and domain documents, a dynamic interaction buffer for storing anonymized conversation state data, and a vector storage partition for storing embedding vectors that can be retrieved by the vector indexing processor; and a multi-interaction device consisting of a touchscreen panel, an array microphone, a speaker structure and a contactless communication interface, each configured to receive multimodal user inputs or generate multimodal system outputs.

Description

TECHNICAL FIELD

[0001] The invention relates to intelligent computer systems and, in particular, a dialogue-based AI device for automated customer interaction, for generating query-based responses, for multimodal query processing, and for decision support in digital commerce ecosystems. The device performs dialogue processing in real time using embedded processors, memory structures, sensor-based user interfaces, and a query-based language model execution pipeline. BACKGROUND OF THE INVENTION

[0002] Digital commerce has evolved from simple web-based interfaces to highly complex ecosystems requiring real-time customer support, personalized interaction, and the automation of multi-step tasks. Existing systems primarily rely on cloud-based chatbots or rule-based agents, which lack contextual information, emotional analysis, and domain-specific semantic intelligence.

[0003] Conventional dialogue systems are limited in their capabilities: 1. Maintain the context of a longer conversation; 2. Gain verifiable knowledge from domain-specific data sources; 3. Dynamic classification of user intent across heterogeneous domains; 4. Integrate physical or embedded devices with AI-based thinking; 5. Perform secure preprocessing and response generation directly on the device; and 6. offer deterministic interaction quality and explainability.

[0004] The cited work underscores the growing importance of dialogue-based artificial intelligence in digital commerce, the weaknesses of existing chatbots, and the need for contextual understanding, query-based reasoning, high accuracy, reduced misinterpretation, and increased customer trust. However, the current state of the art does not reveal a dedicated physical device with embedded processors, specialized memory units, dialogue control circuitry, and structured RAG-LLM hardware acceleration that enables deterministic, secure, and domain-specific dialogue output.

[0005] The present invention overcomes these limitations.

[0006] The rapid expansion of digital commerce has fundamentally changed how customers discover products, search for information, make financial decisions, and interact with service providers. As interactions increasingly shift from physical locations to online platforms, expectations for immediate, accurate, and personalized support have risen sharply. Companies in industries such as insurance, banking, retail, and logistics are now increasingly relying on conversational technologies to manage the multitude of customer interactions. This development has created a demand for sophisticated systems that can mimic human dialogue, process complex, multi-stage interactions, and accurately retrieve domain-specific information.Traditional user interfaces, such as graphical dashboards and menu-driven portals, often reach their limits when it comes to accommodating diverse linguistic expressions, ambiguous queries, and the necessary contextual understanding. Therefore, dialogue-based AI systems have proven to be a promising solution for bridging the gap between automated service delivery and user-centric communication, as demonstrated by fundamental research on dialogue-based systems in digital commerce.

[0007] Despite their growing popularity, existing chatbot and dialogue-based AI solutions still suffer from fundamental limitations that hinder their scalability and reliability in critical business environments. Early chatbot systems relied on deterministic pattern recognition approaches, generating responses from fixed templates linked to keywords. These systems struggled to generalize linguistic variation and understand context, resulting in unreliable conversational experiences that deteriorated as queries deviated from predefined structures. Even with the use of more advanced statistical methods for natural language processing, traditional chatbots have had difficulty capturing nuances, intents, sentiments, and ambiguities—elements essential for high-risk areas such as insurance or financial advice.Their inability to maintain continuity in multi-part conversations led to users frequently repeating information or clarifying previous statements, resulting in frustration and loss of trust.

[0008] Thanks to advances in deep learning, particularly through transformer-based architectures and large language models, conversational systems have improved in terms of fluency and semantic understanding. However, these models have introduced new challenges. Purely generative systems often produce responses that sound coherent but are not based on facts or domain-specific knowledge. This phenomenon, known as hallucination, poses serious risks when the system is tasked with providing regulated or sensitive information. For example, incorrect insurance policy information, faulty premium calculations, or misleading health advice can expose companies to compliance violations and customer losses.The cited study underlines hallucination as a recurring problem and highlights the need for systems that ensure factual accuracy and domain conformity, rather than relying solely on probabilistic language generation.

[0009] Traditional chatbots lack robust mechanisms for retrieving and synthesizing current and reliable knowledge. Many existing solutions rely on static decision trees or hard-coded responses that fail to adapt to changes in business rules, product offerings, or regulatory requirements. Even when using retrieval-based architectures, they often rely on simple TF-IDF similarity analyses or keyword extraction, failing to capture semantic relationships across large document repositories. These approaches result in low accuracy and unstable results when users phrase similar queries differently. This research underscores the importance of search stability and accuracy, demonstrating that many industry chatbots fail to find consistently relevant information for semantically equivalent queries.This inconsistency leads to unpredictable behavior and reduces users' trust in automated systems.

[0010] Another significant drawback of traditional dialogue systems is their inability to effectively perform domain classification. Many chatbots respond incorrectly to queries outside their subject area, attempt to fabricate information, or fail to redirect users appropriately. For example, an insurance chatbot might try to answer questions about unrelated topics such as stock prices or medical benefits, instead of indicating that the query is outside its scope. The evaluation of existing insurance chatbots presented in this document reveals poor domain classification accuracy. Some systems provide inappropriate or irrelevant answers, leading to misinterpretations and a lower perceived intelligence.

[0011] Contextual memory is another area where existing systems have weaknesses. Many commercial chatbots fail to maintain continuity between related queries, resulting in fragmented or repetitive conversations. This limitation impairs the system's ability to facilitate human-like interactions, especially when users ask follow-up questions or implicitly refer to previous statements. Without contextual memory, chatbots don't understand cues like "Compare this to the previous plan" or "Is this suitable for my age?" The study demonstrates that contextual memory is crucial for sustained interaction and shows that traditional systems perform poorly in this regard compared to more advanced approaches with integrated multi-round memory management.

[0012] Accuracy in intent recognition remains a challenge. Many chatbots rely on narrow intent categorizations or insufficiently trained models that cannot generalize to different language styles, dialects, or variations in phrasing. Misclassifications of intent lead to irrelevant responses, incomplete task completion, or unnecessary escalation of requests. In customer-critical environments, such errors undermine user trust and force them to contact human support, thus defeating the purpose of automation. The results of the cited study show that existing chatbots from major insurance providers have limited accuracy in interpreting user intent, particularly with ambiguous or complex requests.

[0013] Furthermore, most existing solutions lack a systematic evaluation system for real-time performance measurement. They do not quantify metrics such as hallucination frequency, Precision@k query accuracy, sentiment analysis, or response usefulness. Without integrated evaluation capabilities, companies struggle to monitor system quality or detect performance degradation. The study underscores the importance of multidimensional evaluation frameworks and demonstrates why conventional systems fail to meet the benchmarks required for reliable use in regulated industries.

[0014] Scalability and adaptability are also problematic with older systems. Many chatbots cannot absorb new information or adapt to updated product offerings without extensive manual reprogramming. Integrating external APIs, heterogeneous data environments, or architectures with advanced data querying is also difficult. As business ecosystems become increasingly complex, chatbots lacking dynamic adaptability quickly become obsolete, forcing companies to repeatedly redevelop their systems.

[0015] Another limitation is the lack of emotional intelligence and personalization. While large language models can mimic human tone of voice, most existing chatbots offer little personalization beyond greeting the user by name. They cannot detect moods, recognize frustration, or tailor their responses to demographic or behavioral patterns. This deficit reduces user satisfaction and impairs the potential of dialogue systems to serve as effective digital assistants. Research shows that the personalization scores of existing insurance chatbots fall significantly short of desirable thresholds, and emotional attunement remains underdeveloped in current systems.

[0016] Concerns regarding security and data privacy persist. Cloud-based chatbots expose sensitive customer data to external security vulnerabilities, are heavily dependent on network stability, and offer limited offline functionality. In sectors like the insurance industry, where personal data, policy numbers, and financial information are regularly exchanged, reliance on external data centers increases risk and latency. Existing chatbots lack secure on-device storage, data query isolation mechanisms, or policy-level verification mechanisms to prevent unauthorized or erroneous disclosures.

[0017] While dialogue-based AI has advanced significantly overall, the existing landscape of digital commerce chatbots exhibits shortcomings in context understanding, knowledge retrieval, compliance enforcement, personalization, evaluation, explainability, and domain accuracy. These deficiencies highlight the need for next-generation dialogue-based devices and architectures that enable robust, context-aware, and knowledge-based dialogues with deterministic accuracy. Fundamental research identifies these systemic gaps and lays the foundation for innovative hardware-software solutions that deliver reliable, secure, and human-like dialogue intelligence in commercial environments. SUMMARY OF THE INVENTION

[0018] The invention relates to a Conversational AI Transaction Device (CAITD) with a multi-layered hardware-software architecture for executing dialogue-based workflows in digital commerce. The device integrates: 1. A conversation processing unit (CPU-X) equipped with a processor for intent detection, a circuit for calculating the dialogue state, and logic for semantic classification. 2. A Retrieval Augmentation Unit (RAU), consisting of a vector indexing processor and a similarity search accelerator for knowledge retrieval on the device. 3. A Generative Response Unit (GRU) that executes a compact large language model (LLM) optimized for generation on the device with hallucination suppression mechanisms. 4. A Domain Integrity Verification Module (DIVM) that ensures generated content complies with domain policies, regulatory restrictions, and filtered knowledge sources. 5. A Multi-Interface Interaction Panel (MIIP) that enables user interaction via voice, text, touch and NFC. 6. A knowledge store array (KMA) containing indexed, encrypted commercial datasets that are retrieved for use in RAG workflows. 7. An Evaluation and Adaptation Engine (EAE) for real-time assessment of response quality, context preservation, hallucination detection, and accuracy of internal task completion.

[0019] The invention enables a hardware-based conversational engine that can replace human agents in commercial sectors such as insurance, banking, retail, and logistics. It supports highly accurate information retrieval, low-latency response generation, autonomous dialogue routing, and task completion verification.

[0020] The present invention aims to provide an advanced, dialogue-based AI system that enables highly precise, context-sensitive, and domain-specific interactions for digital commerce applications. The invention overcomes the limitations of existing chatbot systems through an integrated hardware-software solution that ensures deterministic, verifiable, and reliable dialogue outputs. A key objective is the development of a system that seamlessly interprets user intent across various linguistic expressions, maintains context in multi-stage dialogues, and retrieves domain-specific information with high precision. This increases user satisfaction and reduces the operational effort for commercial service providers. A further objective of the invention is the integration of a retrieval-optimized generative architecture that combines authoritative knowledge with natural language responses.This prevents misinterpretations and ensures that every generated output is based on authenticated data. This objective also includes an integrated misinterpretation suppression mechanism that continuously validates generated content against verified sources, thus ensuring compliance with regulatory requirements in sensitive sectors such as insurance, banking, and financial consulting.

[0021] The invention further aims to introduce a self-evaluating conversation engine that autonomously monitors its own performance based on metrics such as search relevance, precision, intent recognition accuracy, sentiment analysis, and contextual fidelity. By integrating an evaluation engine into the device, the invention ensures continuous optimization of conversation quality and stable long-term performance without constant external monitoring. Another objective is to provide an adaptive and scalable architecture that allows the device to integrate new commercial data, regulatory updates, or product information without fundamental redesign or manual reprogramming. This adaptability ensures longevity, reduces maintenance requirements, and enables companies to deploy the device in evolving digital ecosystems.

[0022] Another important objective of the invention is to provide a multimodal interaction interface that supports voice, text, gesture, and contactless identity recognition, thereby improving accessibility for various user groups. Thanks to integrated microphones, speakers, a touchscreen, sensors, and NFC components, the device offers a natural, intuitive, and user-friendly conversational experience suitable for busy business environments. Furthermore, the invention ensures data security and privacy by executing key conversational processes—including intent recognition, data retrieval, and response generation—directly on the device. This reduces reliance on external networks and minimizes risks posed by cloud-based security vulnerabilities.

[0023] A primary objective is to provide a domain integrity verification mechanism that ensures each conversation output complies with industry-specific guidelines and prevents the dissemination of unauthorized, inaccurate, or misleading information. This objective is particularly important for regulated industries where user trust and strict adherence to regulations are essential. By directly integrating regulatory frameworks into the device's verification circuitry, the invention ensures that responses always meet approved and current standards. A further objective is to support efficient deployment in various commercial environments, including insurance branches, retail stores, financial service centers, logistics centers, and customer kiosks, through a compact, robust, and energy-efficient hardware architecture optimized for continuous operation.

[0024] The invention ultimately aims to restore customer-centric interaction in digital commerce by providing an empathetic, coherent, and personalized dialogue that closely resembles human support while ensuring the accuracy, consistency, and scalability of a machine-controlled system. With these elements, the invention is intended to provide a pioneering solution that sets new standards in automated customer engagement, significantly increases operational efficiency, and establishes a reliable dialogue infrastructure for the future of digital commerce. BRIEF DESCRIPTION OF THE IMAGE

[0025] These and other features, aspects and advantages of the present invention will be better understood if the following detailed description is read with reference to the accompanying drawing, in which the same symbols represent the same parts: Fig. Figure 1 shows a block diagram of a dialogue-based AI system for interaction management in digital commerce.

[0026] Furthermore, those skilled in the art will recognize that the elements in the drawing are simplified and not necessarily drawn to scale. For example, the flowcharts illustrate the process by highlighting the main steps to facilitate understanding of the present disclosure. With regard to the construction of the device, one or more components may be represented in the drawing by conventional symbols. The drawing may show only those specific details relevant to understanding the embodiments of the present disclosure, so as not to clutter the drawing with details that are already apparent to those skilled in the art from the description contained herein. Detailed description of the invention

[0027] To facilitate understanding of the principles of the invention, reference is made below to the embodiment shown in the drawing, which is described using specific terms. It is understood, however, that this does not limit the scope of protection of the invention. Rather, modifications and further developments of the depicted system, as well as further applications of the inventive principles shown therein, are conceivable, insofar as they would normally occur to a person skilled in the art in the field of the invention.

[0028] It will be clear to those skilled in the art that the foregoing general description and the following detailed description are exemplary and explanatory of the invention and are not to be understood as a limitation of it.

[0029] References to “an aspect”, “another aspect”, or similar phrases in this description mean that a particular feature, structure, or property described in connection with the embodiment is included in at least one embodiment of the present disclosure. Therefore, phrases such as “in one embodiment”, “in another embodiment”, and similar expressions in this description may, but do not necessarily, all refer to the same embodiment.

[0030] The terms "includes," "comprehensive," or similar expressions denote non-exclusive inclusion. Thus, a procedure or method containing a list of steps does not only include those steps but may also include further steps not explicitly listed or inherent in the procedure or method. Likewise, the statement "includes..." for one or more devices, subsystems, elements, structures, or components, without further limitations, does not preclude the existence of other devices, subsystems, elements, structures, or components.

[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meanings generally known to those skilled in the art in the field to which this invention belongs. The systems, methods, and examples described herein serve only for illustration and are not to be understood as limiting.

[0032] Embodiments of the present disclosure are described in detail below with reference to the attached drawing.

[0033] Fig.Figure 1 shows a block diagram of a dialogue-based AI system for interaction management in digital commerce. The system 100 comprises a housing (102) enclosing an array of interconnected printed circuit boards; a dialogue processing unit (104) arranged on a first printed circuit board, comprising a syntactic parsing processor (104a) for generating hierarchical parse structures from user input, a semantic embedding processor (104b) for generating high-dimensional semantic vectors from the parsed structures, and a dialogue state calculation circuit (104c) for managing the multi-level dialogue memory using clocked recurrent memory buffers;a retrieval augmentation unit (106) arranged on a second circuit board, comprising a vector indexing processor (106a) for storing and retrieving domain-specific embedding vectors, a similarity calculation processor (106b) for calculating cosine-based semantic distances between the semantic vectors and vectors stored in a knowledge store, and a document reconstruction structure (106c) for merging retrieved text segments into a unified context representation;a generative response unit (108) arranged on a third circuit board, comprising a transformer execution processor (108a) for executing a quantized transformer architecture for natural language generation, a context fusion circuit for merging the unified context representation with the semantic vectors to generate response conditioning data, and a hallucination suppression circuit (108b) for comparing the generated text with retrieved text segments and attenuating content divergence beyond a predetermined semantic threshold;a domain integrity check unit (110) communicatively connected to the generative response unit, wherein the domain integrity check unit comprises a regulatory restriction store for storing domain-specific rules and a conformance assessment processor configured to check generated responses for conformity with the domain-specific rules, blocking non-conforming segments or replacing them with verified text content from the retrieval extension unit; a knowledge store array (112) comprising at least one static domain repository for storing commercial specifications, regulatory guidelines, product descriptions, and domain documents, a dynamic interaction buffer for storing anonymized conversation state data, and a vector storage partition for storing embedding vectors that can be retrieved by the vector indexing processor;and a multi-interaction device (114) comprising a touchscreen panel, an array microphone, a speaker structure, and a contactless communication interface, each configured to accept multimodal user inputs or generate multimodal system outputs; wherein the conversation processing unit, the retrieval extension unit, the generative response unit, the domain integrity check unit, and the knowledge store array work together in a coordinated manner to generate a verified, context-sensitive, retrieval-based conversation response in real time.

[0034] In one embodiment, the syntactic parsing processor (104a) comprises a deterministic tree construction circuit that assigns word class designations, dependency links, and sentence boundaries without relying on cloud-based language services, thereby enabling offline operation and reducing latency in the linguistic segmentation of conversational texts.

[0035] In one embodiment, the semantic embedding processor (104b) is configured to convert multimodal inputs, including text, speech-derived transcripts, and graphical inputs, into embedding vectors using an on-device transformer layer processor. The embedding vectors are normalized using a reference dictionary stored in the knowledge store array to ensure cross-session retrieval stability despite variations in query formulation.

[0036] In one embodiment, the dialogue state calculation circuit (104c) comprises a plurality of controlled memory registers arranged such that referential continuity is maintained across successive conversation contributions, the circuit applying temporal weighting factors to earlier conversation segments to preserve the context while preventing an excessive accumulation of outdated dialogue information.

[0037] In one embodiment, the vector indexing processor (106a) uses a multi-level index structure consisting of coarse-grained clustering for domain separation and fine-grained subclustering for document-level identification, thereby enabling the rapid retrieval of domain-specific embeddings in the presence of high conversation volume.

[0038] In one embodiment, the similarity calculation processor (106b) comprises several parallel similarity calculation blocks arranged to evaluate semantic distance metrics independently, the calculated distances being aggregated by a weighted distance harmonization circuit which increases the retrieval accuracy for ambiguous or multi-intent user requests.

[0039] In one embodiment, the document reconstruction structure (106c) comprises a selective aggregation circuit configured to extract highly relevant text segments, discard contradictory or outdated segments, and combine the remaining segments into a coherent context corpus used to condition the generative response unit.

[0040] In one embodiment, the transformer execution processor (108a) comprises a quantized multi-head attention array, a token-wise feedforward circuit, and a position dependency processor, each optimized for a reduced computational load to enable energy-efficient execution of a large language model on the device without dependence on external servers.

[0041] In one embodiment, the hallucination suppression circuit (108b) comprises a content attribution comparator configured to generate an attribution trust score for each generated token, and a divergence correction processor configured to replace low-trust tokens with retrievable alternatives, thereby limiting unsupported generative extension.

[0042] In one embodiment, the domain integrity checking unit (110) further comprises a rule propagation matrix configured to establish dependencies between regulatory rules and a multi-level constraint evaluator configured to check generated responses against both primary regulatory constraints and secondary derived constraints, thereby ensuring domain fidelity at multiple levels of granularity.

[0043] The system is implemented using discrete, hardware-based computing structures distributed across multiple circuit boards. This ensures that each functional block corresponds to a physically instantiated processing circuit and not a purely software-defined routine. The conversation processing unit is implemented as a hardware element with separate silicon processors for syntactic analysis, semantic embedding generation, and dialog state calculation. Each of these processors consists of integrated logic arrays, fixed-function arithmetic circuits, and register banks configured to operate without virtualized computations. The query extension unit is also implemented as a hardware subsystem.It comprises a vector indexing processor with dedicated memory address traversal circuitry, a similarity calculation processor with parallel cosine distance calculation blocks etched into the substrate, and a document reconstruction structure implemented through selective aggregation circuitry and address-mapped buffer controllers. The generative response unit is integrated into a separate chipset, which includes a transformer execution processor with quantized multi-head attention arrays and token-wise feedforward logic paths, as well as a hallucination suppression circuit with comparator elements, divergence correction logic, and a hardware-level embedded token attribution scoring circuit.The domain integrity check unit is implemented as a separate hardware processor and includes a constraint evaluation engine, a hard-wired rule propagation matrix, and conformance checking logic, all implemented as interconnected combinational and sequential circuits. The knowledge storage array consists of physically segmented, non-volatile memory banks and vector memory partitions on dedicated circuit boards. Address decoding logic enables direct hardware retrieval of domain documents and embedding vectors. The multi-interaction device also includes physical transducers such as touch-sensitive conductive layers, microphone arrays, and speaker coils, all electrically connected via interface controllers printed on the communication bus structure.

[0044] During operation, the system receives a multimodal user request via the multi-interaction device and forwards the input to the conversation processing unit. There, the syntactic parsing processor performs deterministic linguistic preprocessing. This includes token boundary detection, removal of non-informative lexical elements, morphological normalization, and the creation of hierarchical dependency structures to ensure that subsequent analysis is based on a consistent linguistic framework. The processed representation is then fed to the semantic embedding processor, which generates stable, high-dimensional embedding vectors that reflect both syntactic and semantic features of the input. These embeddings are further integrated with previously stored conversational context from the registers of the dialog state calculation circuit.This allows for a precise interpretation of the user's intent and determines whether the system needs to initiate a query operation or proceed directly with the generative construction.

[0045] If the intent analysis determines that the user request requires access to stored domain knowledge, the retrieval augmentation unit becomes active. At this stage, the request embedding is passed to the vector indexing processor, which performs a multi-stage semantic search in the vector memory area of ​​the knowledge storage array. Both the request and the stored domain documents reside in the same vector space, allowing the similarity calculation processor to perform cosine-based distance calculations using multiple parallel computation blocks. The most semantically relevant documents are extracted and, after removing redundant, outdated, or unreliable text segments, are combined into a consolidated context corpus using the document reconstruction structure.This consolidated corpus is compared with the multi-round interaction state managed by the dialog state calculation circuit, generating a context framework suitable for verified generative conditioning.

[0046] The context framework is then sent to the generative response unit, where the transformer execution processor performs controlled natural language generation. During generation, each token produced by the transformer is evaluated by the hallucination suppression circuit, which calculates an attribution confidence score by comparing the generated content to the retrieval-based corpus. Generative segments lacking sufficient foundation are intercepted and replaced using divergence correction logic. This ensures that the output is consistent with authoritative domain knowledge and avoids unsupported extensions. The resulting response is then sent to the domain integrity check unit, where the conformance assessment processor checks the text against the regulatory rules stored in the rule memory.Dependencies between rule sets are resolved using the rule propagation matrix, and non-compliant segments are blocked or replaced with compliant alternatives from the retrieved corpus. Once verification is complete, the response is returned to the user via the multi-interaction device.

[0047] Throughout the entire interaction cycle, the evaluation and adaptation engine continuously monitors system performance. It quantifies the stability of search results by checking whether semantically similar but syntactically different queries lead to identical search results from the vector indexing processor. The relevance of the search results is assessed using Precision@K metrics, which are calculated using manually validated datasets from regulated insurance documents. This allows for a direct measurement of the accuracy with which the similarity calculation processor classifies domain-specific information. The engine also evaluates the quality of the generated responses by assessing task completion. The system verifies whether requested actions—such as identifying branch offices, retrieving policy attributes, or creating checklists—were successfully performed.Furthermore, the intelligence of the responses is measured by evaluating the clarity and accuracy of explanations of regulated insurance procedures and product structures. Relevance is assessed by verifying whether each response is factually correct, stylistically appropriate, and procedurally sound for domain-specific requests such as policy reinstatement or claims initiation. The hallucination rate is measured by quantifying instances in which the generative response unit attempts to introduce non-existent political terms or financial values, with such attempts being flagged by the hallucination suppression circuit.

[0048] In scenarios where the system cannot generate a reliable answer due to insufficient search information or unclear intent, the evaluation engine logs unanswered events to uncover gaps in the knowledge base. Apology responses are also recorded to identify misinterpretation patterns in the semantic embedding processor or ambiguities in the domain intent classification logic. The system's natural language understanding is further examined by categorizing queries according to predefined insurance-related intent classes, such as policy status inquiries, document requests, branch location inquiries, or agent identification. Deviations or misclassifications within these categories contribute to optimizing the weighting mechanisms of the embedding and classification processes.

[0049] To ensure that the evaluation reflects user expectations in practice, empirical reference data from customer satisfaction surveys of various insurance providers are incorporated during system testing. This dataset forms the basis for identifying the factors that contribute to user trust, understanding, and satisfaction with regulated financial services. The insights gained from this dataset serve as the basis for adjustments to retrieval accuracy, conversation tone, compliance filtering, and interaction continuity. This enables the system to guarantee consistent, context-aware, and legally compliant communication across a wide range of digital trading environments.

[0050] The present invention relates to a dialogue-based AI system for highly precise, context-sensitive, and domain-specific interactions in digital commerce. The system is housed in a casing containing several interconnected printed circuit boards. Each circuit board supports specialized processing units that perform computationally intensive tasks related to natural language understanding, generating responses with search optimization, and verifying compliance with legal regulations. The system is designed to operate independently of remote servers, thereby ensuring low-latency, secure, and deterministic dialogue behavior. It is therefore suitable for use in financial service centers, insurance branches, vending machines, and other customer-facing environments that require high security and the dissemination of regulated information.

[0051] The dialogue flow begins with user input via a multi-interaction device, such as a touchscreen, array microphone, or contactless communication interface. This multimodal input is forwarded to the dialogue processing unit, where the syntactic parsing processor decomposes it into hierarchical linguistic structures. These include word class assignments, dependency relationships, and sentence boundaries. This deterministic syntactic decomposition ensures that the subsequent semantic representation is based on a stable linguistic framework. This is essential for the consistency of queries with syntactic variations but semantic equivalence. The parsed structure is then processed by the semantic embedding processor. This processor uses an in-device, multi-layered transformer architecture to convert linguistic elements into high-dimensional embedding vectors.These vectors are normalized using a reference dictionary stored in the knowledge base to minimize embedding variations and ensure query stability across different user sessions.

[0052] Once the embedding vectors are formed, the dialogue state calculation circuit evaluates the relationship of the current input to the existing conversational context. The circuit manages a dynamic sequence of context embeddings stored in clocked memory registers. Each memory register applies a temporal weighting function that prioritizes newer contributions to the conversation while preserving important information from earlier interaction phases. The dialogue state calculation circuit integrates the new embedding with the stored context vectors, creating a composite context embedding that reflects the multi-part nature of human dialogues. This mechanism allows the system to maintain continuity in follow-up questions and ensures that subsequent responses incorporate relevant details from previous contributions.

[0053] The composite embedding is then transferred to the retrieval augmentation unit, where the vector indexing processor performs a multi-stage search in the vector memory area of ​​the knowledge storage array. First, the processor applies coarse clustering to identify the domain-specific subspace relevant to the user query, such as insurance benefits, policy comparisons, or product specifications. After identifying the coarse domain, fine-grained subclustering allows the processor to find semantically aligned document embeddings. The similarity calculation processor evaluates the semantic proximity between the composite embedding and the indexed vectors using several parallel cosine distance calculation blocks.These parallel results are aggregated by a weighted distance harmonization circuit, which corrects ambiguities or overlapping query intents by amplifying the semantically dominant signal. The resulting high-level vectors are passed to the document reconstruction structure, which retrieves the corresponding text segments from the static domain repository.

[0054] The document reconstruction structure employs a technical filtering process to create a coherent and relevant contextual corpus. Segments flagged as outdated or inconsistent based on metadata and regulatory markers are suppressed. Segments with high confidence are merged into a consolidated text body, which serves as the factual basis for generative processing. The reconstructed corpus, along with the composite embedding, is sent to the generative response unit. There, the transformer execution processor operates with a quantized transformer architecture optimized for device performance. The transformer execution processor applies self-attention mechanisms, feedforward transformations, and positional encoding adjustments to extend the composite embedding into a preliminary natural language response.

[0055] Unlike conventional generative systems, the invention incorporates a hallucination suppression circuit to ensure factual accuracy. This circuit performs token-based attribution analysis by comparing each generated token to the reconstructed corpus using a content attribution comparator. For each token, an attribution confidence score is calculated based on semantic similarity, positional relevance, and regulatory compliance metadata. Tokens whose confidence score falls below a threshold are intercepted by the divergence correction processor and replaced with retrievable alternatives. This design ensures that the system avoids unsupported extensions and is based on verified domain content. The corrected response is then submitted to the regulatory compliance unit for domain integrity verification.

[0056] The domain integrity check unit operates using a rule propagation matrix stored in the regulatory constraint store. The conformance assessment processor analyzes the generated response by checking both explicit rule constraints—such as allowed policy descriptions and pricing guidelines—and inferred constraints from the rule propagation matrix. The conformance assessment processor identifies deviations from regulatory or domain-specific requirements and either suppresses non-compliant segments or replaces them with compliant alternatives from the query extension unit. This ensures that every output is not only factually correct but also complies with industry regulations and internal company policies, which is particularly critical for insurance and financial applications.

[0057] After verification, the system prepares the final response for user output. During speech-to-speech interactions, the sentiment recognition processor analyzes acoustic features of the previous user speech to adjust the tone, speech rate, and emphasis of the synthesized response delivered through the speaker. The multi-interaction device ensures that the user receives fluent, natural, and personalized conversational output, regardless of the input modality. During and after each interaction, the internal performance evaluation unit assesses various technical parameters, including retrieval accuracy, hallucination frequency, context retention stability, and domain classification accuracy. The unit then recalibrates the weighting factors in the semantic embedding processor and the transformer execution processor based on these performance metrics.This creates a continuous adaptive feedback loop that optimizes system behavior over time without external updates or retraining.

[0058] The drawing and the preceding description illustrate embodiments. Those skilled in the art will recognize that one or more of the described elements can be combined to form a single functional element. Alternatively, certain elements can be divided into several functional elements. Elements of one embodiment can be added to another. For example, the process flows described here can be modified and are not limited to the manner described herein. Furthermore, the actions of a flowchart need not be performed in the sequence shown; nor do all actions necessarily need to be carried out. Actions that do not depend on other actions can be performed in parallel with the other actions. The scope of protection of the embodiments is in no way limited by these specific examples. Numerous variations, whether explicitly stated in the description or not, such as...Differences in structure, dimensions, and materials are possible. The scope of protection of the embodiments is at least as comprehensive as described by the following claims.

[0059] The advantages, other benefits, and problem solutions have been described above with reference to specific embodiments. However, the advantages, benefits, problem solutions, and any components that can effect or enhance an advantage, benefit, or solution are not to be construed as critical, necessary, or essential features or components of the claims. REFERENCES 100 A Dialogue-Oriented Artificial Intelligence System for Interaction Management in Digital Commerce. 102 Housing structure 104 Conversation Processing Unit 104a Syntactic Parsing Processor 104b Semantic Embedding Processor 104c Dialog state calculation circuit 106 Call-up expansion unit 106a Vector Indexing Processor 106b Similarity Calculation Processor 106c Structure of Document Reconstruction 108 Generative Response Unit 108a Transformer execution processor 108b Hallucination suppression circuit 110 Domain Integrity Check Unit 112 Knowledge storage array 114 Multi-interaction device

Claims

[1] A dialogue-based artificial intelligence system for interaction management in digital commerce, consisting of: a housing that encloses an arrangement of interconnected printed circuit boards; a conversation processing unit arranged on a first circuit board, comprising a syntactic parsing processor for generating hierarchical parse structures from user input texts, a semantic embedding processor for generating high-dimensional semantic vectors from the parsed structures, and a dialogue state calculation circuit for managing a multi-level conversation memory using clocked recurrent memory buffers; a retrieval augmentation unit arranged on a second circuit board, comprising a vector indexing processor for storing and retrieving domain-specific embedding vectors, a similarity calculation processor for calculating cosine-based semantic distances between the semantic vectors and vectors stored in a knowledge store, and a document reconstruction structure for merging retrieved text segments into a unified context representation; a generative response unit arranged on a third circuit board, comprising a transformer execution processor for executing a quantized transformer architecture for natural language generation, a context fusion circuit for merging the unified context representation with the semantic vectors to generate response conditioning data, and a hallucination suppression circuit for comparing the generated text with retrieved text segments and attenuating content divergence beyond a predetermined semantic threshold; a domain integrity check unit that is communicatively connected to the generative response unit, wherein the domain integrity check unit includes a regulatory restriction store that stores domain-driven rules, and a conformance assessment processor configured to check generated responses for conformance with the domain-driven rules, blocking non-conforming segments or replacing them using verified text content from the retrieval extension unit; a knowledge storage array comprising at least one static domain repository for storing commercial specifications, regulatory guidelines, product descriptions, and domain documents, a dynamic interaction buffer for storing anonymized conversation state data, and a vector storage partition for storing embedding vectors that can be retrieved by the vector indexing processor; and a multi-interaction device consisting of a touchscreen panel, an array microphone, a speaker structure and a contactless communication interface, each configured to receive multimodal user inputs or generate multimodal system outputs. [2] System according to claim 1, wherein the syntactic parsing processor comprises a deterministic tree construction circuit that assigns word class designations, dependency links and sentence boundaries without relying on cloud-based linguistic services, thereby enabling offline operation and reducing latency in the linguistic segmentation of conversation texts. [3] System according to claim 1, wherein the semantic embedding processor is configured to convert multimodal inputs, including text, speech-derived transcripts and graphical inputs, into embedding vectors using an in-device transformer layer processor, and wherein the embedding vectors are normalized using a reference dictionary stored in the knowledge store array to ensure cross-session retrieval stability despite variations in query formulation. [4] System according to claim 1, wherein the dialogue state calculation circuit comprises a plurality of controlled memory registers arranged such that referential continuity is maintained across successive conversation contributions, and wherein the circuit applies temporal weighting factors to earlier conversation segments to preserve the context while preventing an excessive accumulation of outdated dialogue information. [5] System according to claim 1, wherein the vector indexing processor uses a multi-level index structure comprising coarse-grained clustering for domain separation and fine-grained subclustering for document-level identification, thereby enabling fast retrieval of domain-specific embeddings during periods of high conversation volume. [6] System according to claim 1, wherein the similarity calculation processor comprises several parallel similarity calculation blocks arranged to evaluate semantic distance metrics independently of each other, and wherein the calculated distances are aggregated by a weighted distance harmonization circuit which increases the retrieval accuracy for ambiguous or multi-intent user requests. [7] System according to claim 1, wherein the document reconstruction structure comprises a selective aggregation circuit configured to extract highly relevant text segments, discard contradictory or outdated segments, and combine the remaining segments into a coherent context corpus used to condition the generative response unit. [8] System according to claim 1, wherein the transformer execution processor comprises a quantized multi-head attention array, a token-wise feedforward circuit and a position dependency processor, each optimized for a reduced computational load to enable energy-efficient execution of a large language model on the device without dependence on external servers. [9] System according to claim 1, wherein the hallucination suppression circuit comprises a content mapping comparator configured to generate a mapping confidence value for each generated token, and a divergence correction processor configured to replace low-confidence tokens with retrievable alternatives, thereby limiting unsupported generative extension. [10] System according to claim 1, wherein the domain integrity checking unit further comprises a rule propagation matrix configured to establish dependencies between regulatory rules and a multi-stage constraint evaluator configured to check generated responses against both primary regulatory constraints and secondary derived constraints, thereby ensuring domain fidelity at multiple levels of granularity.

Images