214 results about "Context window" patented technology

An ESG-specific multimodal AI foundation model is disclosed, featuring a Transformer-based architecture with approximately 30 billion parameters, designed explicitly for environmental, social, and governance (ESG) domain applications. This model uniquely supports extremely long context windows (up to 128,000 tokens), critical for comprehensive ESG analyses of lengthy documents such as sustainability reports and policies. It integrates textual and visual data through gated cross-attention and a Mixture-of-Experts (MoE) architecture, achieving precise multimodal context comprehension. The invention employs Group Relative Policy Optimization (GRPO) reinforcement learning strategy, refining model outputs based on group-relative advantages computed from multiple candidate generations, thus significantly enhancing ESG-specific reasoning and output quality.

A system and method for efficient natural language processing combines large and small language models with a thought caching architecture. The system includes a router that directs prompts either to a large language model for thought generation or to a thought cache containing previously generated thoughts. When using the large model, generated thoughts are combined with the original prompt and routed through a smaller language model to produce responses. The thought cache stores reasoning patterns that can be retrieved and reused, eliminating the need to regenerate similar thoughts for related prompts. The system supports both local and cloud-based caching, enabling personal and enterprise-wide thought storage and retrieval. This architecture reduces computational overhead while maintaining reasoning capabilities, effectively extends context windows beyond traditional limits, and enables efficient scaling across different deployment scenarios. The system can operate with reduced resources by leveraging cached thoughts without requiring constant access to the large model.

The invention discloses a dynamic spectrum autonomous collaborative optimization method based on a large language model, and the method comprises the steps: converting a channel state, a historical operation track and a conflict feedback signal into text-numerical value mixed tensors, and carrying out the splicing of the tensors to form a multi-dimensional environment state mixed representation; a distributed semantic inference engine is designed, a sparse Transform architecture and an adaptive conflict prediction mechanism are constructed in the distributed semantic inference engine, and a candidate strategy set of spectrum access and power modes is generated through a dynamic context window based on a sliding window local attention mechanism. The strategy generation module is used for performing real-time strategy generation and risk pre-judgment in a complex spectrum environment; through a conflict prediction module, performing similarity comparison on feature vectors extracted based on a multi-head attention mechanism and a learnable conflict prototype, and dynamically evaluating an interference risk coefficient of a candidate strategy; and dynamic spectrum autonomous collaborative optimization is carried out based on multi-dimensional environment state mixed characterization and an interference risk coefficient. The invention further discloses a corresponding system, electronic equipment and a computer readable storage medium.

The invention provides a long document processing method and system based on multi-agent collaborative reasoning. The method specifically comprises the steps that semantic coherence is reserved by adopting a sliding window overlapping partitioning strategy; a structured reasoning chain is generated by analyzing the agents, the agents are integrated to eliminate cross-block contradictions based on vector similarity and triple conflict detection, and a paragraph-level unified conclusion is output; and the summarizing agent constructs a global logic relationship by utilizing thinking chain prompt, reconstructs an answer framework according to user questions and binds a source identifier, and generates a traceable complete abstract or answer. According to the method, the technical problems of global reasoning failure caused by context window length limitation, semantic splitting and conclusion conflict caused by block processing, low single-model serial efficiency and the like in the prior art can be solved.

A method for managing conversation history in an Al system is disclosed. A user message is received and full conversation history payload is retrieved from a memory service. The full conversation history payload includes one or more messages, summaries, or moments. A prompt budget is dynamically allocated, based on an available context window. The prompt budget determines token allocations for the messages, summaries, or moments from the full conversation history payload. A prompt is assembled for a language model by selecting conversation history elements to fit within the allocated prompt budget. The selecting balances between recent verbatim messages, summarized content of older messages, and relevant older moments. The prompt budget is iteratively adjusted and the prompt is iteratively reassembled as new information is added, from one or more language model outputs or tool calls, while maintaining the full conversation history payload in the memory service.

The invention discloses a large-model context control system and method based on hierarchical memory and an association graph, and the method comprises the steps: constructing a tree graph mixed memory structure and a logic hierarchical memory model for a context control unit between a user and an LLM, and carrying out the self-adaptive processing of a super-long context through a self-adaptive processing engine, a single-theme persistent dialogue control mechanism, and the like. According to the method, the problem that an LLM context window is limited (including introduction of a mechanism similar to an Agent) is solved, just like an operating system helps developers / users to solve the problems that hardware standards hidden behind are not uniform, and complex resource control is caused, a reliable LLM memory (context) control service with multiple details hidden is provided for upper developers, and the method has the advantages of being high in practicability and easy to popularize and use. The final objective of the method is to approximately relieve the limitation of an LLM context window through a set of self-consistent system mechanism, so that a developer does not need to pay too much attention to the details of a bottom layer, and more energy can be focused on the implementation of an application function.

The invention provides a dialogue processing method and system based on a large model, and the method comprises the steps: obtaining a dialogue text length value, employing a pre-established semantic boundary identification set to detect a topic turning frequency value according to the dialogue text length value, and combining a historical statement sequence sorted according to time in a context window, generating an initial semantic segmentation vector containing feature extraction dimensions; according to the initial semantic segmentation vector, calculating a topic coherence score by adopting a dependency weight distribution table, and separating sub-vectors of which the orthogonality degree is higher than a preset threshold value through a vector decomposition precision value to generate a preliminary structured representation matrix; extracting emotional intensity fluctuation features and knowledge density distribution features from the preliminary structured representation matrix, and generating a refined structured representation matrix after adjusting the rank number of the matrix; and calculating the Euclidean distance between the topic coherence sub-vector and the logical reasoning sub-vector in the refined structured representation matrix.

The invention relates to the technical field of large language model application, in particular to a fault-tolerant processing method and system for a super-long text in large model service and a storage medium, and the method comprises the following steps: obtaining a super-long text to be processed and a maximum Token threshold of a large model context window, completing Token processing and judging whether the super-long text exceeds the limit or not; starting a multi-level alternative scheme including rapid compression, dynamic compression and chat history compression; starting an error recovery mechanism including hierarchical exception processing, intelligent degradation and parameter verification; outputting the target text of which the Token number is compliant; the system comprises a text acquisition module, a multi-level alternative scheme execution module, an error recovery module and an output module. The storage medium stores a computer program, realizes the method during execution, can adapt to multiple models, and meets industrial-grade super-long text processing requirements; the problems of Token overrun errors and inference service instability caused by lack of fault-tolerant mechanisms and insufficient boundary processing in the prior art are solved.

A computer-implemented method comprising: receiving, as input, a query and a source document intended for a content-grounded question-answering or multi-turn conversation task by a specified large language model (LLM) which has a context window size limit, wherein the source document has a size which exceeds the context window size limit; dividing the source document into a plurality of segments; applying a language model to each of the segments, to assign to each of the segments a relevance score; selecting the k-top segments having the highest the relevance scores; combining the selected k-top segments into a virtual document having a size which complies with the context window size limit; and feeding the virtual document as input to the specified LLM, to generate a response that is grounded in the content of the virtual document.

The invention discloses a data security classification and grading method and system based on a large language model, and relates to the technical field of language analysis, and the method comprises the steps: converting bank internal business data of each layer into text data through text conversion, unifying the data form of multi-layer or multi-modal data into text data, and storing the text data in a database; and a reliable basis is provided for subsequent language semantic recognition. According to the method, the text data is partitioned and sorted, subsequent context language understanding is facilitated, the context window length under each service scene is set, the context semantic understanding requirement of each service scene is met, and the reliability and accuracy of semantic association are improved. The internal business data of the bank is classified and graded based on the risk value of the business scene and the sensitive value of the composite field, the accuracy and adaptability of data security classification and grading are improved, security protection strategies of different degrees are carried out for different levels of different types of data, and the requirements of bank data storage and prevention are met.

The invention relates to the technical field of big language model dialogue system optimization, in particular to a big language model dynamic dialogue history compression method and system.The method comprises the steps that the maximum length of a context window matched with a target big language model, the maximum number of newly-generated lexical elements and the size of a safety buffer area are set, and then dialogue history is loaded; initial compression and verification are carried out through a keyword and TF-IDF mixed scoring system, multiple times of dynamic compression are carried out according to gradients if the conditions are not met, and finally, parameters are adjusted to adapt to the residual space when a model is called to generate response. The system comprises a dialogue history loading module, a parameter configuration module, a dynamic compression engine module and a large language model integration module. Through a multi-stage compression verification mechanism and a progressive multi-stage dynamic compression strategy, super-long texts such as engineering technology documents can be processed, service interruption is reduced, the compression efficiency is improved on the premise that key semantics are reserved, and multi-language dynamic compression is supported. The problems of system token overrun and service instability in the prior art are solved.

The invention discloses an intelligent education robot question answering system based on voice recognition and a knowledge graph, and particularly relates to the technical field of artificial intelligence. The method comprises the following steps: performing feature extraction on a user voice signal to obtain a text sequence and a multi-modal context feature; recognizing subject domain judgment and question answering intentions based on supervised classification and keyword rule fusion, and outputting a subject domain prior probability and a question answering target vector; determining polysemy words in the text sequence by using a context window, generating a semantic item sequence and semantic item confidence, constructing a subject domain sub-graph based on a subject domain prior probability, and obtaining a candidate reasoning path and a path scoring vector; determining a target teaching concept and an optimal reasoning path by combining Bayesian inference and consistency verification; generating a personalized question answering result aiming at the question asked by the user through fact retrieval and knowledge derivation in combination with the question answering target vector; accurate and efficient intelligent teaching question answering can be realized, and question answering accuracy and intelligent interaction capability of the education robot are effectively improved.

The invention relates to the field of intelligent report generation, in particular to an intelligent insight report generation method and system based on an AI large model, and the method comprises the steps: inputting an insight demand, and generating an insight data package comprising insight contents, associated data and industry labels; extracting a basic keyword set of the insight data packet to form a mixed feature code; after mixed feature coding preprocessing, weight distribution is carried out; and after weight distribution of the AI large model, injecting a high-weight feature vector into a semantic understanding core layer, injecting a low-weight feature vector into a logical reasoning layer, and outputting analysis data to form an intelligent insight report. According to the method, word embedding parameters are optimized according to field semantic characteristics, a parameter verification mechanism is introduced, field text characteristics are adapted by adjusting vector dimensions, context windows and low-frequency vocabulary filtering threshold values, window parameter validity is verified through cosine similarity, word frequency threshold value reasonability is verified through standard deviation, and field text characteristic matching is achieved. And it is ensured that the feature vectors can accurately capture domain-specific semantics.

The invention relates to the technical field of law informatization, and particularly discloses a law decision semantic deviation recognition method based on artificial intelligence, and the method comprises the steps: data collection and preprocessing, semantic unit segmentation, multi-level semantic analysis, semantic deviation recognition, decision consistency evaluation, and result interpretation and application. According to the scheme, semantic features are extracted from the three levels of vocabularies, syntax and context, the overall semantic representation of the judgment document is enhanced through an attention mechanism and a dynamic context window mechanism, it is ensured that semantic connection of all parts is close, the long-term dependency relationship is captured, misjudgment caused by the fact that a single semantic unit is separated from the context is avoided, and the detection accuracy is improved; a semantic deviation recognition model is constructed on the basis of a Transform architecture, hidden semantic deviation is effectively recognized through a deep semantic feature matrix, deviation scores of all semantic units are calculated through an adaptive weight distribution mechanism, and the objectivity and consistency of scores are improved.

The present disclosure provides a system for pre-training a large language model. The system includes an electronic health record (EHR) database storing patient-specific data, a processor, and a memory storing instructions. The instructions, when executed by the processor, cause the system to retrieve patient data from the EHR database, generate a clinical text corpus from the retrieved patient data, pre-append metadata to each note in the clinical text corpus to create pre-appended notes, optimize context windows for the pre-appended notes, tokenize the pre-appended notes to create a tokenized dataset, and pre-train the large language model using the tokenized dataset. The pre-trained large language model is configured to generate patient-specific responses to medical queries.

The invention discloses a long text reading understanding method based on dynamic partitioning and selection, which is characterized in that the method adopts dynamic partitioning to dynamically divide long text input into discrete text blocks, selects and screens out irrelevant text segments by using selection partitioning, and splices the remaining text segments according to an original sequence, so as to obtain a long text reading understanding result. In order to conform to context window limitation predefined by a large language model, the method specifically comprises the steps of text preprocessing, dynamic blocking, block selection, large model output and the like. Compared with the prior art, the method has the advantages that the semantic coherence and the understanding accuracy are improved, the processing capability of the model on the super-long text is enhanced, the internal semantic integrity of each block is ensured, the semantic ambiguity caused by the block is reduced, the semantic coherence damage caused by the block with the fixed length is avoided, and the semantic coherence and the understanding accuracy are improved; and the processing capability of the model on the super-long text is enhanced, the data utilization efficiency is high, and the method has an important value and a good application prospect in practical application.

The invention discloses a graph-driven self-attention compressible memory management method, and belongs to the technical field of penetration testing. According to the graph-driven self-attention compressible memory management method, an attention graph is constructed, semantic dependence and attention flow directions among multiple Agent nodes are accurately described by utilizing edge weights, the relevance and the accessibility of information are improved, and in the aspects of screening and loading, the expandability of the memory is improved. According to the method, key memory fragments are screened on the basis of concerned graph edge weights and task related features, only the loaded information is loaded, redundancy is effectively eliminated, the data volume processed by a system is reduced, and the system operation efficiency is improved; the semantic compression module can perform abstract processing on historical information based on various elements, and on the premise of ensuring that key semantic information is not lost, the abstract length is controlled within the window limit, so that the model can normally process information, and the adaptability and stability of the system to different data volumes are improved.

The invention discloses a dynamic context window customer service quality inspection method based on a large model, and the method comprises the following steps: S1, generating a dynamic window: scoring the dialogue round of a customer service and a user, dynamically updating and maintaining a context window according to a scoring result, and carrying out the context reconstruction of a dialogue in the window; s2, multi-dimensional quality inspection: inputting the dialogue text in the dynamic window into a large model, and outputting a preset field through a customized prompt trigger model; s3, clustering attribution: carrying out semantic clustering on related contents output by the large model by adopting a kmeans algorithm and BERT vectorization, and then generating a general description and an operable suggestion for each clustering result through prompt; and S4, result output and application: generating a structured json result containing a multi-dimensional quality inspection result and a clustering result, wherein the structured json result is used for api calling or visual platform display. The method has the advantages that efficient, accurate and multi-dimensional customer service quality inspection can be achieved, and the service quality can be improved and overcome.

Arrangements for intelligent content generation for process automation are provided. A domain model, being structured into tasks according to a defined schema, may be exported for processing by a large language model. A prompt and a context window associated with a task of the domain model may be received. A task template associated with the task may be modified. The modified task template may be enriched with data from a backend system. Content validation may be performed on content of the modified task template enriched with the data from the backend system. Schema validation may be performed for validating the modified task template enriched with the data from the backend system against the defined schema. Correction of invalid tasks may be performed in an iterative loop until the modified task template enriched with the data from the backend system is validated. Then, changes to the domain model may be applied.

The invention relates to a knowledge enhancement method and system based on a context structure. The method comprises the following steps: acquiring a query request input by a user; screening candidate evidences related to the query request from a pre-constructed case knowledge base based on multi-view retrieval; performing weighted calculation on a retrieval result of each view to generate an evidence score of each candidate evidence; screening based on the evidence scores to obtain a candidate evidence set and preprocessing the candidate evidence set; calculating the value and length of each candidate evidence in the pre-processed candidate evidence set; constructing an evidence selection problem based on a preset context window value and the value and length of each candidate evidence; solving the evidence selection problem to obtain an evidence subset with maximum value; generating a context structure based on the evidence subset with the maximum value and a preset prompt template; and inputting the context structure into a general language model for knowledge enhancement, and generating a knowledge enhancement result.

Certain aspects of the disclosure provide a method for code generation, including: selecting a code generation use case type; identifying the target entity in a graph knowledge base; extracting a graph segment of the graph knowledge base; identifying a set of methods in the graph segment; identifying a first method of the set of methods that has a largest context size of the set of methods; determining an overlapping context size between the first method and each other method of the set of methods, and selecting a maximum number of other methods of the set of methods in order of decreasing overlapping context size with respect to the first method, such that a total context size of the maximum number of other methods of the set of methods and the first method does not exceed a maximum context window size associated with a machine learning model.

The invention belongs to the field of text questioning and answering, and relates to a large language model long text questioning and answering method and system based on a hybrid context compression technology. The method comprises the following steps: analyzing and preprocessing a document, and converting an unstructured original text into a normalized text paragraph set; classifying questions of the long text question and answer scene based on a large language model; according to the question type and the preprocessed text, adaptively selecting the most suitable context compression method to compress the long text to obtain a compressed context; and generating an answer to the question based on a large language model by using the compressed context. According to the method, the advantages of two context compression technologies are integrated, the problem of context window limitation of processing a long text by a large language model is effectively solved, high-quality question and answer performance is kept while computing resource consumption and processing delay of the compression technologies are reduced, and the limitation of a single compression method on different types of problems is relieved.

The invention relates to the technical field of translation, and discloses a translation calibration method based on big data, which can accurately identify semantic deviation in a multi-meaning scene, for example, accurately judge that'cell 'should be translated into'cell' instead of'house 'in a biological text. The system can detect the language field mismatching problem of spoken expression in the official file, and provides term correction suggestions conforming to industry standards. Aiming at technical terms with multiple meanings in legal clauses, the calibration process can be combined with context window analysis to select an optimal translation scheme, the omission ratio of manual review is remarkably reduced, continuous optimization and cross-language knowledge sharing of a translation calibration model are realized, and the accuracy of translation calibration is improved. The problems that a traditional system is lagged in updating and low-resource language performance is insufficient are solved.

In some embodiments provided herein is a generative foundation model trained over millions of health system-scale electronic health records along with web-scale medical text corpora to acquire knowledge of both medical practices and theories, and use of the generative model for rare disease diagnosis (including rare ophthalmic, diseases and rare systemic diseases), emergency condition identification (including ophthalmic emergencies and systemic emergencies), complex disease solving ("diagnostic puzzles"), or generating multimodal medical imaging reports (including ophthalmic images and radiology images such as X-rays and CT scans). In some embodiments, the generative model involves the use of language data, for pre-training, language data for supervised finetuning using a instruction tuning approach (e.g., QA pairs), and a human-machine hybrid evaluation strategy. In some embodiments, both the pre-training and supervised finetuning phases involve the use of a particular method of scaling to extend the context window. In some embodiments, the human-machine hybrid evaluation strategy involves language data for automated evaluations, as well as evaluations by generalists and by different specialists (e.g., ophthalmologists and radiologists) of varying levels of experience. In some embodiments, the generative foundation model, MetaGP, is used for unmet clinical needs through integration of medical and multimodal imaging data.

A query is received from a client device. A subset of documents relevant to the query is determined in part by determining an optimal configuration for the query. The subset of documents is inputted in a context window for a response generator and the query is inputted as a prompt for a query response. The query response received from the response generator is outputted to the client device.

Systems and methods are provided for semantic clustering for arbitrarily long context windows for machine-learned sequence processing models. A computing system can obtain a context sequence. The computing system can determine a plurality of subsequences of the context sequence. The computing system can determine, using a machine-learned semantic embedding model, a semantic embedding for each subsequence. The computing system can determine, based on the semantic embedding, a plurality of semantic clusters. The computing system can generate, using a machine-learned sequence generation model and based at least in part on the semantic clusters, an output sequence.

The invention provides a time sequence prediction task-oriented basic model construction method, which comprises the following steps of: constructing a Patch division layer which is used for cutting an input time sequence by adopting a context window adaptive mechanism combining time domain and frequency domain statistical characteristics to generate overlapped time periods; an embedded layer is constructed and used for converting the processed time sequence into embedded representation and generating an embedded Token through linear projection processing; constructing a forward and reverse frequency domain enhanced bidirectional state space module which is used for outputting forward and reverse modeling results by taking the input sequence and the time reversal sequence thereof as input, and fusing the forward and reverse modeling results to obtain a final fusion result; constructing a feed-forward layer which is used for taking the final fusion result as input to encode a time dependency relationship; and constructing an output and projection layer for mapping the coded time sequence to a predicted future value through linear projection. The method can adapt to various time series data under zero sample setting, and has good modeling robustness and cross-variable modeling capability.

The invention belongs to the technical field of computer information security, and particularly relates to a log auditing technology based on multi-agent cooperation and an implementation framework thereof, which comprises the following steps that: a task decomposition agent performs structured modeling and logic slicing on original log data by utilizing a thinking chain reasoning mechanism to obtain a log auditing result; a complex internal threat detection task is decomposed into a plurality of subtasks with clear logic and independent semantics, and each subtask corresponds to a specific log analysis problem; and the tool generation agent automatically generates a reusable exclusive analysis tool by adopting an example-based programming normal form based on the semantic target and the input structure of each subtask. By constructing a three-layer collaborative architecture of a task decomposition agent, a tool generation agent and a task execution agent, the context window limitation of a single language model during processing of an ultra-long log sequence is effectively broken through, and structured and modular analysis of complex internal threat behaviors is realized.

The invention relates to the technical field of artificial intelligence, in particular to an adaptive context processing method and system for a large language model. The method comprises the following steps: collecting multi-source candidate context segments related to current user query; performing utility evaluation on each candidate fragment, outputting a quantized utility score, and performing multi-granularity compression on each fragment based on the utility score and a preset quality constraint to obtain a compressed fragment set; taking the maximization of the total utility score as a target, taking the condition that the total token cost is less than or equal to the global token budget as a constraint, and carrying out optimization selection to obtain an optimal context combination; and inputting the optimal context combination into a large language model to generate an output result for the current query. By applying the method, the context window management of the large language model can be optimized, the adaptive context processing capability of the information utilization efficiency is improved, and the problems of context window limitation, value judgment and the like are solved.

The invention discloses a text-oriented violent word abbreviation detection method, device and equipment and a storage medium, and relates to the technical field of natural language process.The method comprises the steps that preprocessing, candidate abbreviation dynamic positioning and context modeling are conducted on an input text, and a semantic relation graph is constructed; fusing the text semantic feature and the graph structure feature of the graph to generate a joint feature vector; performing multi-dimensional alignment on the vector and a preset violent knowledge graph, and calculating cosine similarity between the vector and each graph node; if the maximum cosine similarity exceeds a preset similarity threshold value and the hazard level of the corresponding node is higher than a preset hazard level, marking the candidate abbreviation corresponding to the maximum value as a violent word abbreviation; and if the predefined security counter-example word exists in the dynamic context window, clearing the mark. According to the method, the novel abbreviation variant of the violent word in the text can be accurately identified, the violent intention in an ambiguous context is distinguished, and the zero sample scene adaptability and the attack and defense resisting capability are improved.