Untrained intelligent interaction system targeting based on deterministic features and self-reflection

Abstract

This invention discloses a training-free intelligent interaction system based on deterministic feature targeting and self-reflection. The core is deployed on consumer-grade terminals such as mobile phones, tablets, and computers, and is compatible with embedded chip deployment. The system replaces "recognition" and "understanding" with multimodal deterministic retrieval. All inputs are converted into deterministic feature identifiers, accurately compared with a local feature library, and output corresponding results or trigger preset rules. The system integrates core capabilities such as context memory, referential resolution, dual-mode interaction, dynamic reasoning path generation, deterministic deduction, and self-reflection, supporting multi-turn dialogue, cross-turn reasoning chain continuation, operation execution, and visual output. This invention transforms unstructured text in documents into logical constraint equations, achieves O(1) matching of massive fingerprint databases through Trie tree indexing, replaces linear calculations with rule boundary judgment, and achieves O(1) complexity logical contradiction detection and automatic correction through a reverse hash index tree. The global feature library supports offline construction from multiple channels such as public datasets, industry databases, customer-owned data, internet crawlers, manual input, and user feedback, without involving model training. The system solves the technical problems of existing intelligent interaction systems, such as reliance on cloud recognition, privacy leaks, uncontrollable output, and misleading multimodal functions.

Description

Technical Field

[0001] This invention belongs to the field of edge-side deterministic computing and intelligent interaction technology, specifically relating to a training-free intelligent interaction system based on deterministic feature targeting and self-reflection. The system core is deployed on consumer-grade terminals such as mobile phones, tablets, and computers. It replaces probabilistic recognition with deterministic retrieval, and all processing is completed locally, requiring no network, no training, and no hallucination output. Background Technology

[0002] The existing technology has the following problems: 1. The large-scale multimodal model solution claims to be able to "understand" images and "comprehend" documents, but it relies on cloud uploads, which poses a high risk of privacy leaks; the output is based on probability, which can be misleading; and the accuracy of recognition cannot be guaranteed, so it is actually "guessing".

[0003] 2. Traditional image recognition solutions are based on models such as SVM and CNN, which require a large amount of training data and cannot achieve "training-free" operation; moreover, model updates require retraining, resulting in high maintenance costs.

[0004] 3. Most document processing solutions only extract text and cannot understand the rule-based content in documents, such as contract terms and drug instructions; or they rely on cloud-based semantic understanding, which poses a high risk to privacy.

[0005] 4. Edge-side intelligent solutions have limited functionality, such as only having a voice assistant, and cannot integrate multimodal inputs such as images, documents, and voice, nor can they achieve deterministic interaction across scenarios.

[0006] 5. Existing rule engines lack self-reflection capabilities and cannot detect logical contradictions in the reasoning process, which may result in inconsistent output results.

[0007] 6. The existing feature library construction relies on manual annotation or model training, which is costly to maintain and cannot achieve rapid offline updates.

[0008] This invention proposes a deterministic retrieval and deterministic reasoning scheme: all multimodal inputs are converted into deterministic feature identifiers, accurately compared with a local feature database, and then multi-step reasoning is performed through dynamic reasoning path generation and deterministic deduction to output results or trigger preset rules. This is "search plus reasoning" rather than "guessing," and "matching plus reasoning" rather than "recognition," fundamentally solving the problem of multimodal functionality being misnamed. Summary of the Invention

[0009] 1. Core Definition 1.1 Deterministic Retrieval This invention does not perform "image recognition," "semantic understanding," or "speech recognition." Instead, it converts the input into deterministic feature identifiers and performs precise comparisons with a local feature library. These feature identifiers include image-aware hash fingerprints, video keyframe fingerprints, document field key-value pairs and logical constraints, and system speech-to-text instruction keywords.

[0010] 1.2 Dynamic Reasoning Path Generation When no direct feature match is found, the system automatically completes multi-step reasoning through structural decomposition, rule block matching, and path generation. The structural decomposition unit identifies the combinatorial structure of the problem, including "X's Y's Z", "X is greater than Y, Y is greater than Z", and "A causes B, B causes C", dynamically decomposing them into sub-target sequences. The rule block library stores the smallest combinable deterministic rule units, including "X's father causes father's name", "X's mother causes mother's name", and "A causes B and B causes C causes A causes C". The path generator automatically generates deterministic reasoning paths based on the sub-target sequences and the rule block library, using semantic structure alignment.

[0011] 1.3 Deterministic Deduction Based on an existing deterministic rule base, new deterministic knowledge, solutions, or entities are automatically generated through rule combination, parameter replacement, rule triggering based on time-series data, and counterfactual inference. The rule combiner combines multiple rule blocks to generate new composite rules. The parameter replacer replaces parameters within a known rule framework to generate new solutions. The time-series data-based rule trigger compares the changing trends of sensor or input data with preset threshold rules to trigger warnings or actions. This trigger compares the calculation results with preset rule boundaries to determine whether to trigger preset logic. The counterfactual inferencer changes conditions to deduce different results, such as the conclusion that "if a cooling fan had been installed at the time, the equipment would not have stopped."

[0012] 1.4 Self-reflection The system performs secondary verification, error correction, and optimization of its own reasoning process. This includes reasoning chain recording, reverse verification, rule consistency checking, path optimization, and self-correction. Reverse verification uses a pre-generated "conclusion-condition" reverse hash index tree, achieving logical contradiction detection with O(1) time complexity through hash lookup.

[0013] 1.5 Deterministic Matching All matches are exact comparisons and do not involve similarity calculations, probability ranking, or statistical inference.

[0014] 1.6 No training The feature library is built once before system deployment, requiring no model training, iterative optimization, or online learning. Feature library updates use an offline replacement mode, including downloading the new fingerprint library and importing the new rule package, without involving dynamic learning.

[0015] 2. Construction of the global feature library and information sources The global feature library is a static, pre-built library. All feature libraries are built offline, without involving model training, probability calculations, or manual annotation (except during the review process). Information sources include at least one of the following: 2.1 Public Datasets This includes publicly available authoritative data such as dictionaries, encyclopedias, idiom databases, legal provisions, drug instructions, national standards, and industry norms. The system extracts structured content through deterministic rules and automatically generates feature templates.

[0016] 2.2 Industry Database This includes specialized data from vertical sectors such as medical clinical guidelines, industrial equipment manuals, financial product specifications, and aircraft maintenance manuals. The system analyzes the document structure and extracts key fields and rules.

[0017] 2.3 Customer-owned data This includes authorized data such as internal company documents, historical work orders, product drawings, equipment parameters, and operating procedures. The system connects via an interface and automatically converts the data into a feature library format.

[0018] 2.4 Publicly Available Internet Data The system collects publicly available content such as encyclopedia entries, technical blogs, product pages, video subtitles, and image descriptions through automated web crawlers. Feature-answer mappings are extracted, and the content is then stored in the database after review.

[0019] 2.5 Manual data entry Feature-answer mappings, inference rules, chat templates, and fingerprint database samples were entered by domain experts or annotation personnel. The entry process does not involve model training and serves only as a data source.

[0020] 2.6 User Feedback During system operation, user-corrected or supplemented features and answers are reviewed, verified, and then stored in the database. User feedback data is not used for online training but only as offline update material.

[0021] All information sources undergo data cleaning, feature decomposition, and standardization preprocessing during collection and storage to ultimately generate deterministic feature units, which are then written into the global feature library. During storage, both a reverse hash index tree and a Trie tree index are constructed.

[0022] 3. System Architecture The system includes a multimodal deterministic retrieval unit, a multi-domain feature library, a training-free inference engine core, a context memory module, a referential resolution unit, a dual-mode interaction module, and a multi-interface output module.

[0023] 3.1 Multimodal Deterministic Retrieval Unit The multimodal deterministic retrieval unit is deployed locally on the terminal, and all processing is not uploaded to the cloud.

[0024] The image retrieval subunit uses a perceptual hashing algorithm, including pHash or dHash, to convert images into deterministic hash fingerprints, which are then compared with the local fingerprint database in O(1) precision using a Trie tree index.

[0025] The video retrieval subunit extracts keyframes from the video, generates hash fingerprint sequences for the keyframes using a perceptual hashing algorithm, and compares them with the local video fingerprint database.

[0026] The document retrieval subunit parses documents, including PDF, Word, PPT, and Excel, extracts structured fields, including table key-value pairs, clause numbers plus content, and monetary values, and converts the field values ​​into logical constraint equations, such as "if the condition value is greater than the threshold, then execute the action", and compares them with the local rule base for rule triggering.

[0027] The voice retrieval subunit is configured to receive the speech-to-text output by the terminal operating system, extract keywords, and perform precise matching with the local instruction set.

[0028] 3.2 Multi-domain feature library Multi-domain feature libraries are deployed on the terminal's local storage medium and are categorized by domain, including at least one of the following: drug packaging fingerprint library, product packaging fingerprint library, contract terms rule library, logical constraint rule library, instruction keyword set, and inference rule block library. All feature libraries are built offline and do not involve model training. Feature library information sources include at least one of the following: public datasets, industry databases, customer-owned data, web crawlers, manual input, and user feedback.

[0029] 3.3 Core of the Training-Free Inference Engine The core of the training-free inference engine includes a deterministic feature targeting unit, a dynamic inference path generation module, a deterministic inference module, a reverse hash index tree verification unit, a Trie tree optimized index unit, and a self-reflection module.

[0030] The deterministic feature targeting unit converts the input into deterministic feature identifiers, including fingerprints, logical constraints, and keywords, to achieve O(1) matching.

[0031] The dynamic reasoning path generation module includes a structure decomposition unit, a rule block library, and a path generator, which are used to identify the combinatorial structure of the problem and dynamically generate deterministic reasoning paths.

[0032] The deterministic inference module includes at least one of a rule combiner, a parameter substitute, a time-series data-based rule trigger, and a counterfactual inference unit. The time-series data-based rule trigger is used to compare the calculation results with preset rule boundaries to determine whether to trigger preset logic.

[0033] The reverse hash index tree verification unit implements logical verification with O(1) complexity.

[0034] Trie tree optimization index unit realizes dynamic pagination loading of feature library with loading latency of less than 2 milliseconds. Trie tree index realizes O(1) fast scheduling of massive fingerprint library.

[0035] The self-reflection module includes a reasoning chain recording unit, a contradiction detection unit, and an automatic correction unit. It uses a reverse hash index tree to verify the logic of the reasoning process and correct contradictions.

[0036] 3.4 Context Memory Module The context memory module is used to store entities, relationships, and intermediate inference results from the most recent N rounds of dialogue, and supports the continuation of the inference chain across multiple rounds of dialogue.

[0037] 3.5 Refers to the digestion unit The pronoun resolution unit adopts a nearest neighbor priority plus type verification strategy to achieve accurate identification and replacement of pronouns, with a computation latency of less than 0.1 milliseconds.

[0038] 3.6 Dual-mode interaction module The dual-mode interaction module supports on-demand switching between emotional value mode and deterministic mode, with a switching latency of less than 1 millisecond.

[0039] 3.7 Multi-interface output module The multi-interface output module supports at least one of the following: text output, voice output, image output, chart output, file output, and operation execution. Operation execution includes setting an alarm, adding a schedule, opening the app, and sending control commands.

[0040] 4. Technical Effects In terms of image processing, this invention uses fingerprint matching, which is a deterministic solution, while the large-scale multimodal solution uses probabilistic recognition, which may lead to errors.

[0041] In terms of video processing, this invention uses keyframe fingerprint matching, which is a deterministic solution, while the large-model multimodal solution uses probabilistic recognition, which may lead to errors.

[0042] In terms of document processing, this invention uses field extraction, logical constraint translation, and rule triggering, which is a deterministic solution. In contrast, the large-scale multimodal solution uses semantic understanding, which may lead to errors.

[0043] In terms of speech processing, the present invention uses system APIs and keyword matching, which is a deterministic solution, while the large-scale multimodal solution uses cloud-based recognition and semantic understanding.

[0044] In terms of multi-step reasoning, this invention uses dynamic path generation plus deterministic deduction plus rule boundary triggering, which is a deterministic scheme, while the large-scale multimodal scheme uses probabilistic generation, which may lead to errors.

[0045] In terms of self-reflection, this invention uses a reverse hash index tree to achieve O(1) logical contradiction detection and automatic correction, which existing rule engines do not have.

[0046] In terms of feature library construction, this invention supports offline construction from multiple channels such as public datasets, industry databases, customer-owned data, web crawlers, manual input, and user feedback, without the need for model training, while existing solutions rely on manual annotation or cloud training.

[0047] In terms of privacy and security, this invention processes all data locally without uploading it to the cloud, whereas large-scale multimodal solutions require uploading to the cloud.

[0048] Regarding training dependencies, this invention requires no training, while large-scale multimodal models require massive amounts of data for training.

[0049] In terms of output reliability, this invention achieves 100% accuracy within the library, while the large-scale multimodal scheme is uncertain.

[0050] In terms of interpretability, this invention is fully traceable, while the large-scale multimodal scheme is a black box.

[0051] In terms of system architecture, this invention achieves O(1) scheduling of massive fingerprint databases through Trie tree indexing, while the large model multimodal solution relies on cloud computing power. Detailed Implementation

[0052] Example 1: Drug verification using image fingerprint matching and Trie tree indexing The user takes a photo of the medicine box with their mobile phone. The system calculates the perceptual hash fingerprint locally, and locates the matching item in the drug fingerprint database in O(1) through Trie tree indexing. When the drug fingerprint is matched, the drug information is output. The output is: This is amoxicillin capsule, one capsule three times a day, please use within the expiration date.

[0053] Example 2: Contract risk control, document field extraction, logical constraint translation, and rule triggering. The user uploads a contract PDF. The system parses the PDF locally, extracts fields, and sets the penalty for breach of contract to 50%. This field is then converted into a logical constraint: if the penalty percentage exceeds 30%, the risk level is high. This is compared with the rule base, triggering the high-risk rule. The output is: "The penalty percentage for breach of contract is 50%, exceeding the 30% threshold. It is recommended to negotiate a reduction with the other party."

[0054] Example 3: Dynamic reasoning path generation and complex problem decomposition The user asks, "Who is Zhang San's father's mother?" The system deconstructs and identifies the "X of Y of Z" structure, breaking it down into sub-target one, "Zhang San's father," and sub-target two, "the mother of the result of sub-target one." The path generation matches the rule blocks "X's father leads to father's name" and "X's mother leads to mother's name." The path is executed as follows: Zhang San leads to father Li Si, Li Si leads to mother Wang Wu. The output is: "Zhang San's father's mother is Wang Wu."

[0055] Example 4: Rule triggering based on time series data, trend determination plus rule boundaries The user input motor vibration values ​​for the past three days as 0.1 mm, 0.12 mm, and 0.15 mm, respectively. The system calculates the trend, determining the rate of increase to be 0.05 mm per day, and compares it to the rule boundary. The rule library pre-sets an alert if the vibration value exceeds 0.25 mm or if it continuously increases by more than 0.05 mm per day for three consecutive days. The alert rule is triggered. The output is: "Vibration value is continuously increasing, expected to reach 0.30 mm in three days, exceeding the safety threshold. It is recommended to check the bearing in advance."

[0056] Example 5: Counterfactual deduction, deduction based on assumptions The user asked, "If a cooling fan had been installed at the time, would the device still have stopped?" The system invoked the counterfactual inference tool, changed the condition to "install a cooling fan," and re-analyzed the scenario, concluding that if a cooling fan had been installed at the time, the temperature would not have exceeded the threshold and the device would not have stopped.

[0057] Example 6: Self-reflection, logical contradiction detection and correction The user asked, "Why is the motor overheating?" The initial reasoning was that vibration combined with temperature caused bearing wear. Self-reflection was initiated: reverse verification detected the rule "normal current rules out bearing wear," but contradiction detection revealed that the current was normal, contradicting the bearing wear conclusion. Automatic correction and backtracking were then performed, and the cause was re-identified as poor heat dissipation. The corrected conclusion was output.

[0058] Example 7: Multi-turn dialogue plus reference resolution plus reasoning chain continuation In the first round, the user asks, "Who is Zhang San's father?" The system outputs "Li Si," and the context is stored. In the second round, the user asks, "How old is he this year?" The system resolves the issue by using pronouns to determine that "he" equals "Li Si," and outputs "Li Si is 45 years old this year." In the third round, the user asks, "Who is his mother?" The system resolves the issue by using pronouns to determine that "he" equals "Li Si," and infers "Li Si's mother," outputting "Wang Wu."

[0059] Example 8: Dual-mode switching When a user says, "I'm in a bad mood today and want to chat," the system automatically switches to emotional value mode and provides a warm response. When a user asks, "What should I do if the motor is overheating?" the system automatically switches to deterministic mode and provides a concise diagnosis.

[0060] Example 9: Visual Output When a user queries the vibration trend of the equipment, the system outputs a textual conclusion and generates a trend graph, displaying the vibration value change curve on the terminal screen, allowing the user to intuitively view the data trend.

[0061] Example 10: File Output After users complete the contract risk control analysis, the system supports exporting the analysis report as a PDF file and saving it locally on the terminal. Users can share, print, or archive the report.

[0062] Example 11: Offline Construction of Feature Library The system runs an automated build tool on the PC to extract drug name, usage, dosage, and expiration date information from drug instruction manual PDFs, generate perceptual hash fingerprints, and build a drug packaging fingerprint database. Simultaneously, it extracts penalty clauses from a contract template library to build a logical constraint rule library. After completion, the feature library is packaged into a binary file and burned to the terminal's local storage.

Claims

1. A training-free intelligent interaction system based on deterministic feature targeting and self-reflection, characterized in that, Deployed in consumer-grade terminals such as mobile phones, tablets, and computers, and compatible with embedded chip deployment, including: The multimodal deterministic retrieval unit is deployed locally on the terminal, and all processing is not uploaded to the cloud, including: The image retrieval subunit uses a perceptual hash algorithm (at least one of pHash and dHash) to convert images into deterministic hash fingerprints, and performs an O(1) precise comparison with the local fingerprint database through a Trie tree index. The video retrieval subunit extracts key frames from the video, generates hash fingerprint sequences for the key frames using a perceptual hashing algorithm, and compares them with the local video fingerprint database. The document retrieval subunit parses documents (at least one of PDF, Word, PPT, and Excel), extracts structured fields (table key-value pairs, clause numbers plus content, and monetary values), and converts the field values ​​into logical constraint equations (such as executing an action if the IF condition value is greater than a threshold), and compares the rule triggers with the local rule base. The voice retrieval subunit is configured to receive the speech-to-text output by the terminal operating system, extract keywords, and perform precise matching with the local instruction set; A multi-domain feature library is deployed on the local storage medium of the terminal. It is classified by domain and includes at least one of the following: drug packaging fingerprint library, product packaging fingerprint library, contract terms rule library, logical constraint rule library, instruction keyword set, and reasoning rule block library. All feature libraries are built offline and do not involve model training. The information sources of the feature libraries include at least one of the following: public datasets, industry databases, customer-owned data, publicly available Internet data, manual input, and user feedback. The core of the training-free inference engine includes: The deterministic feature targeting unit converts the input into deterministic feature identifiers (fingerprints, logical constraints, keywords) to achieve O(1) matching; The dynamic reasoning path generation module includes a structure decomposition unit, a rule block library, and a path generator, which is used to identify the combinatorial structure of the problem and dynamically generate deterministic reasoning paths. The deterministic deduction module includes at least one of a rule combiner, a parameter substitute, a rule trigger based on time-series data, and a counterfactual deduction device; the rule trigger based on time-series data is used to compare the calculation result with a preset rule boundary to determine whether to trigger preset logic; the counterfactual deduction device is used to change the conditions to deduce different results; The reverse hash index tree verification unit implements logical verification with O(1) complexity; Trie tree optimization index unit realizes dynamic pagination loading of feature library with loading latency of less than 2 milliseconds. Trie tree index realizes O(1) fast scheduling of massive fingerprint library; The self-reflection module includes a reasoning chain recording unit, a contradiction detection unit, and an automatic correction unit. It uses a reverse hash index tree to verify the logic of the reasoning process and correct contradictions. The context memory module is used to store entities, relationships, and intermediate inference results from the most recent N rounds of dialogue, and supports the continuation of the inference chain in multi-round dialogues; The pronoun resolution unit adopts a nearest neighbor priority plus type verification strategy to achieve accurate identification and replacement of pronouns with a computation latency of less than 0.1 milliseconds; The dual-mode interaction module supports switching between emotional value mode and deterministic mode as needed; The multi-interface output module supports at least one of the following: text output, voice output, image output, chart output, file output, and operation execution; the operation execution includes setting an alarm, adding a schedule, opening an APP, and sending control commands. The system has a single retrieval latency of less than 5 milliseconds, requires no network dependency, and outputs results based on deterministic matching and deterministic reasoning, which are interpretable, reproducible, and free of illusions.

2. The system according to claim 1, characterized in that, The structure decomposition unit of the dynamic reasoning path generation module identifies the combination patterns of the problem, including "X of Y of Z", "X is greater than Y, Y is greater than Z", and "A causes B, B causes C", and dynamically decomposes them into sub-target sequences; the path generator automatically generates reasoning paths based on the sub-target sequences and the rule block library through semantic structure alignment.

3. The system according to claim 1, characterized in that, The rule combiner of the deterministic inference module is used to combine multiple rule blocks to generate new composite rules; the parameter replacer is used to replace parameters to generate new schemes within a known rule framework; and the rule trigger based on time-series data is used to compare the changing trend of input data with the preset rule boundary to trigger an early warning or execute an action. Counterfactual inference tools are used to deduce different outcomes by changing the conditions.

4. The system according to claim 1, characterized in that, The reasoning chain recording unit of the self-reflection module is used to record reasoning steps and intermediate results; the contradiction detection unit compares the consistency between the reasoning conclusion and the conditions through a reverse hash index tree to detect logical contradictions; the automatic correction unit automatically backtracks to the most recent contradiction-free intermediate result and regenerates the response when a logical contradiction is detected.

5. The system according to claim 1, characterized in that, The document retrieval subunit is used to convert unstructured text in a document into logical constraint equations based on preset semantic anchors (including "penalty", "term", and "amount"). The logical constraint equations are then matched and triggered with rule blocks in the rule base.

6. The system according to claim 1, characterized in that, The image retrieval subunit is used for drug packaging verification. It accurately compares the perceived hash fingerprint with the genuine product fingerprint database and outputs at least one of the following: drug name, usage and dosage, and expiration date.

7. The system according to claim 1, characterized in that, The system supports multi-turn dialogue and referential resolution: the context memory module stores the entities and relationships of the most recent N rounds of dialogue, and the referential resolution unit identifies "he, she, it, this, that" and replaces them with the corresponding entities to achieve cross-turn reasoning chain continuation.

8. The system according to claim 1, characterized in that, The system is compatible with embedded chip (STM32H7 series) deployment. In embedded scenarios, the running memory is less than or equal to 2MB, and in consumer terminal scenarios, the running memory is greater than or equal to 4GB.

9. The system according to claim 1, characterized in that, The system can operate immediately upon power-up, without the need for loading models, training, or iterative optimization; all outputs are deterministic results that are interpretable, auditable, and free of illusions.

10. The system according to claim 1, characterized in that, The multi-domain feature library supports offline updates, which can be replaced by downloading a new version of the fingerprint library or rule package, without involving online training or dynamic learning.