A dye field knowledge enhanced large model intelligent agent system

Abstract

The application belongs to the technical field of cross between artificial intelligence and dye chemistry, and relates to a dye field knowledge enhanced large model intelligent agent system, which comprises a user input module, a mixed retrieval module, a reordering module and a no-illusion reasoning module. The system carries out keyword retrieval and semantic vector retrieval on a structured knowledge base in parallel through the mixed retrieval module to obtain candidate text blocks. The reordering module is used for deep semantic interaction scoring and reordering, and high-confidence knowledge segments are screened. Finally, the no-illusion reasoning module calls a bottom large language model based on preset constraint conditions, and generates traceable reasoning in combination with an original problem. The structured knowledge base is fused with academic literature and basic information in the dye field. The application can effectively inhibit the knowledge illusion of the large model, guarantee the rigorous traceability of the output content, generate an experimental scheme with reference value for an unknown dye system, and significantly improve the knowledge reuse efficiency and scheme development efficiency of research and development in the dye field.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and dye chemistry, and relates to a large-scale intelligent agent system for enhancing knowledge in the dye field. Background Technology

[0002] As core functional chemicals in textile printing and dyeing, functional materials, and environmental monitoring, dyes involve multiple key stages in their research and development and production, including molecular synthesis, separation and purification, structural characterization, performance testing, and kinetic studies. This process heavily relies on precise experimental parameters, rigorous mechanistic support, and efficient knowledge reuse. Currently, the lack of experimental parameters for unknown systems is a prominent issue in dye research and industrialization. For novel dyes or compound dye systems without publicly available literature, researchers typically rely on extensive manual, gradient experiments, resulting in high reagent and time costs. Furthermore, the success rate is highly dependent on the operator's professional experience. Simultaneously, dye knowledge is highly fragmented; structural information, physicochemical properties, separation conditions, and detection parameters are scattered across different data sources, making manual retrieval and screening extremely inefficient. This can easily lead to overlooking key literature or extracting incorrect parameters, hindering the improvement of knowledge reuse efficiency.

[0003] With the development of large language model technology, it has shown significant advantages in natural language interaction and information integration. However, general-purpose large language models are pre-trained on large-scale general corpora and lack professional knowledge in the vertical field of dye chemistry. When answering professional questions in the dye field, they only generate text by continuing the text based on statistical language patterns in the training corpus, which may output parameters that seem professional but are actually impractical, resulting in a serious knowledge illusion problem. More importantly, the reasoning process of general-purpose large language models is an unexplainable "black box" operation, and the source literature of its reasoning cannot be clearly identified. The output conclusions cannot be traced, which cannot meet the rigor requirements of scientific research scenarios. Retrieval-enhanced generation technology provides an effective technical path to alleviate the knowledge illusion of large models and inject external professional knowledge. By retrieving relevant literature fragments from external knowledge bases and injecting context before generating answers, the boundaries of model generation are constrained, which can significantly improve the factual accuracy and traceability of answers. It has already been applied in multiple vertical fields.

[0004] For example, patent CN120492598B discloses a legal domain knowledge base question-answering method and system based on a large language model intelligent agent. This system establishes a standard retrieval enhancement and generation technology architecture, including document segmentation, vector indexing, semantic retrieval, memory enhancement, and answer generation. It introduces a dynamic context window adjustment mechanism and a multi-turn dialogue memory module, enabling it to accurately understand complex legal provisions and judicial cases. However, this system is designed for the textual characteristics of the legal domain and is adapted to matching scenarios of legal provisions and judicial cases. Dye chemistry literature contains a large number of professional terms, complex chemical mechanism descriptions, and precise experimental parameters. This system cannot accurately identify professional knowledge or integrate cross-document related knowledge, nor does it possess the ability to transfer and deduce based on existing rules.

[0005] For example, the system disclosed in patent CN119274774B, and various search-enhanced generation frameworks in the medical field, are all based on standardized guidelines, medical records, and drug instructions to build a knowledge base, enabling functions such as matching diagnostic results based on symptoms and assisting in diagnosis. Their core solution is the accurate matching of existing knowledge. The knowledge base of such systems has a relatively fixed knowledge structure, and the output content is strictly limited to the scope of existing literature; essentially, it is based on matching and presenting existing information. When applied to the dye field, when faced with novel dye combinations not directly recorded in the knowledge base, the system cannot effectively integrate and reason with the retrieved similar dye schemes and established separation mechanisms within the field. It can only return scattered fragments of similar literature, failing to generate experimental schemes with practical reference value and making it difficult to support the development and optimization of dye experimental schemes.

[0006] Overall, existing vertical domain retrieval enhancement generation systems generally lack dedicated structured knowledge bases for the dye field, have limited knowledge sources, and cannot cover the multi-dimensional knowledge needs of basic dye information and academic literature. At the same time, existing systems generally only have knowledge matching and presentation capabilities, lacking reasoning constraint mechanisms for scientific experimental design. They cannot deduce experimental schemes for unknown dye systems based on existing knowledge and rules, while ensuring that conclusions are traceable and free of illusions. This makes it difficult to adapt to the research and development needs of the dye field.

[0007] Therefore, it is necessary to provide a dye domain knowledge-enhanced large model intelligent agent system to solve the technical problems existing in the prior art. Summary of the Invention

[0008] The purpose of this invention is to address the problems existing in the prior art and provide a knowledge-enhancing large-scale intelligent agent system in the field of dyes.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A dye domain knowledge-enhanced large-scale intelligent agent system includes:

[0011] The construction steps of a structured knowledge base are as follows:

[0012] (a) Obtain electronic documents of literature and scanned electronic documents of professional reference books;

[0013] The electronic documents were obtained by searching and downloading relevant literature in the field of dyes from academic databases such as Web of Science and CNKI.

[0014] The scanned electronic documents of professional reference books were obtained by collecting and scanning basic information on at least 4,000 dye varieties in 13 major categories from professional reference books such as "Complete Guide to Dye Varieties".

[0015] (b) Preprocess the electronic documents of the literature and the scanned electronic documents of professional reference books to obtain structured documents;

[0016] (c) Perform adaptive chunking on the structured document to obtain text blocks;

[0017] (d) Each text block is split into short sentences and a vector index is generated sentence by sentence using the Embedding model to obtain text block vectors. The generated text block vectors are stored in a locally deployed vector database for subsequent semantic vector retrieval similarity matching calculations.

[0018] The user input module is used to input the original question posed by the user in natural language and output the original question to the hybrid retrieval module, the reordering module, and the hallucination-free reasoning module.

[0019] The hybrid retrieval module is used to input the original question and perform keyword retrieval and semantic vector retrieval on the structured knowledge base to obtain candidate text blocks, and output the candidate text blocks to the re-ranking module. The two retrieval methods complement each other, taking into account both accurate matching of professional terms and deep semantic association mining, and balancing the breadth of recall coverage and ranking accuracy.

[0020] The reordering module is used to input the original question and candidate text blocks, and after performing deep semantic interactive scoring, it reorders all candidate text blocks from high to low according to their relevance scores, and outputs the top candidate text blocks (i.e., high-confidence knowledge fragments) to the hallucination-free reasoning module. The reordering module plays a secondary role in screening candidate text blocks. Unlike the coarse-grained matching in the mixed retrieval stage, this module can achieve more fine-grained semantic association comparison, ensuring the confidence of the knowledge used for reasoning from the input end.

[0021] The illusion-free reasoning module is used to input the original question and candidate text blocks provided by the reordering module. Based on preset constraints, it calls the underlying large language model to perform reasoning and generation, and outputs a structured answer that meets the constraints to the user interface or downstream application module. The preset constraints can be embedded into the underlying large language model in the form of prompt word framework to standardize the reasoning and generation boundaries of the model.

[0022] The system interacts in natural language, providing literature-based answers to questions about existing knowledge; for unknown dye systems, it can generate standardized initial experimental schemes through structural analysis, similarity retrieval, and mechanism matching, and output targeted optimization suggestions based on user feedback on experimental defects.

[0023] As a preferred technical solution:

[0024] As described above, in a dye-related knowledge-enhancing big model intelligent agent system, in step (a), the research directions of relevant literature in the dye field cover chromatographic separation and analysis techniques such as capillary electrophoresis (CE), spectroscopic characterization techniques, non-aqueous dyeing techniques, dye structure and dyeing properties, hydrolysis kinetics, wastewater treatment, and functional dyes.

[0025] Journal articles are preferred in the field of dyes, but dissertations and conference papers are also accepted.

[0026] The publication years of literature related to dyes are mainly within the last 20 years, while some classic literature can be traced back to the 1980s;

[0027] The 13 major categories of dyes are acid dyes, basic dyes, direct dyes, disperse dyes, reactive dyes, ice dyes, solvent dyes, sulfur dyes, vat dyes, organic pigments, fluorescent whitening agents, food dyes, and mordant dyes.

[0028] Basic dye information includes dye name, CAS number (or CI number, EC number), domestic trade name, chemical structure type, chemical structure diagram, molecular formula, molecular weight, hue, properties and uses, fastness performance table, manufacturing method, references, English trade name and manufacturer, and domestic manufacturer.

[0029] Based on the aforementioned basic dye information, the system can extract the corresponding key fields from the original question and match them with the database through hybrid retrieval. The returned matching results can be used for subsequent structural similarity comparison and transfer derivation.

[0030] In the dye domain knowledge enhancement big model intelligent agent system described above, step (b) involves preprocessing the electronic document as follows: first, filtering out information in the electronic document that is not related to the core technology content, then improving the integrity and standardization of the text, and converting it into a structured document;

[0031] Information unrelated to the core technical content in electronic documents includes redundant information such as headers and footers, advertising snippets, repeated paragraphs, and irrelevant appendices, as well as supplementary information such as reference lists, external links, and email addresses.

[0032] Improving the integrity and standardization of text means correcting issues such as garbled characters and incorrect formatting, and removing unnecessary line breaks and spaces that affect the continuity of text;

[0033] The conversion process is as follows: First, the layout structure is identified. Then, the electronic document is parsed page by page using an automatic parsing algorithm. This method can ignore the segmentation markers, accurately extract the text content, automatically identify and retain the core information such as the main text, charts and graphs, and experimental data of the electronic document, and merge the pictures and tables into the text paragraphs with the closest context to form segments containing complete knowledge units, thus obtaining a structured document.

[0034] In the dye domain knowledge-enhancing big model intelligent agent system described above, step (b) involves preprocessing the scanned electronic document of the professional reference book as follows: using an OCR engine to identify and extract the content of the scanned electronic document and convert it into two-column editable text; performing reconstruction processing on the recognition result from two-column layout to single-column continuous text; manually checking and correcting the recognition error and layout misalignment of the professional information against the original scanned document; and finally organizing it into a structured document.

[0035] This invention designs dedicated preprocessing workflows for two completely different types of knowledge carriers: electronic documents of literature and scanned electronic documents of professional reference books.

[0036] For electronic documents: Filter out redundant information such as headers, footers, and references, and merge the descriptions of figures and tables into the corresponding text paragraphs to ensure the integrity of knowledge units;

[0037] For scanned electronic documents of professional reference books: Design a complete process of "OCR recognition + reconstruction from two-column layout to single-column layout + manual verification of professional information", focusing on correcting the recognition errors of core professional information such as dye name, CAS number, molecular formula, and fastness performance parameters.

[0038] This design is not a conventional choice: core reference books in the dye field are mostly outdated scanned versions with a two-column layout, containing a large number of precise chemical parameters. General OCR recognition has an extremely high error rate, and parameter distortion will directly lead to the failure of subsequent experiments. This invention ensures the professional reliability of the knowledge base from the source through differentiated processing. It is a targeted optimization specific to the characteristics of knowledge carriers in the dye field and does not belong to the conventional approach of general domain transfer.

[0039] Existing technologies all employ a unified, general preprocessing workflow, lacking the format reconstruction and professional verification steps specific to scanned electronic documents of professional reference books. For example, patent application CN120296147A performs structured processing on knowledge base documents, generating a knowledge graph containing documents, chapters, knowledge points, knowledge entries, objects, activities, and attributes, and imports it into Elasticsearch to create an index. If this general solution is directly applied to the dyeing field, it can only achieve:

[0040] At the data level, general texts are replaced with literature and reference book texts in the dye field, and simple labels are made according to the major categories of dyes.

[0041] At the processing level, all documents undergo the same general cleaning, general OCR, and general structuring processes, without distinguishing between electronic documents and scanned electronic documents of professional reference books.

[0042] Therefore, in the field of dyes, this scheme essentially only replaces the data layer of the general search-enhanced generation architecture, lacks underlying adaptation for the knowledge forms in the dye field, cannot achieve efficient identification of professional knowledge and integration of cross-document related knowledge, and is also difficult to support the accurate retrieval and generation of dye experimental schemes.

[0043] As described above, in a dye domain knowledge-enhancing big model intelligent agent system, the specific process of step (c) is as follows: using sentences as basic units, the structured document is divided into text blocks of a certain block length, and a certain overlap rate is maintained between adjacent text blocks;

[0044] The specific values ​​for the block length and overlap rate mentioned above were obtained through iterative debugging, and the process is as follows:

[0045] (i) Build the system and initialize the block length (512 characters) and overlap rate (10%).

[0046] (ii) Use a test set containing at least 60 professional questions to verify the system’s performance, with retrieval recall and knowledge illusion rate as the core evaluation indicators;

[0047] (iii) Determine whether the retrieval recall rate is ≥90% and the knowledge illusion rate is ≤5%. If yes, end the process and output the block length and overlap rate of the last iteration. Otherwise, adjust the parameters according to the following rules and return to step (ii).

[0048] If the retrieval recall rate is less than 90%, then reduce the block length within the range of 256 to 512 characters and increase the overlap rate within the range of 10% to 20%.

[0049] If the retrieval recall rate is ≥90% and the knowledge illusion rate is >5%, then increase the block length within the range of 256~512 characters and reduce the overlap rate within the range of 10%~20%.

[0050] Through iterative debugging, the following parameter optimization rules can be obtained: when the block length is less than 256 characters, it is easy to cause context fragmentation, while when it is more than 512 characters, it will reduce retrieval accuracy; a reasonable overlap rate can achieve the best balance between semantic coherence and storage efficiency.

[0051] The execution logic of adaptive block segmentation is as follows: the algorithm reads sentences one by one and accumulates the character length until the length of the current block is close to the target length and the current sentence is read. The current block is saved and the content corresponding to the last proportion is retained as the starting point of the next block. This avoids key experimental parameters and related chemical mechanism descriptions being mechanically split because they happen to be located at the block boundary, thus ensuring the semantic coherence of subsequent retrieval and reasoning processes.

[0052] This invention uses sentences as the basic unit for segmentation, and the segment length and overlap rate are determined through iterative debugging. It explicitly uses [retrieval recall rate ≥ 90% and knowledge illusion rate ≤ 5%] as the dual-constraint termination condition, and provides the corresponding parameter adjustment rules.

[0053] The block segmentation method of this invention is significantly different from the block segmentation methods of existing technologies. For example, patent application CN120296147A uses header-level block segmentation without any parameter iteration optimization logic; patent application CN120387448A uses semantic block segmentation that transforms coarse-grained large models into fine-grained ones. Its optimization goal is to improve the semantic coherence of a single block, which is completely different from the technical path of this invention, which is "reverse iteration of block parameters based on the final system performance index".

[0054] Existing segmentation methods only pursue semantic coherence, while the segmentation method of this invention directly optimizes the final retrieval effect and illusion suppression. It is designed for the characteristic of dye literature where "experimental parameters and mechanistic descriptions are strongly bound"—too fine segmentation will break the complete experimental logic, while too coarse segmentation will dilute keywords and reduce retrieval accuracy. Its debugging logic, dual-index setting, and parameter range are not disclosed in the prior art and are not conventional choices for those skilled in the art.

[0055] The specific process of obtaining candidate text blocks by keyword retrieval in a dye domain knowledge-enhancing big model intelligent agent system as described above is as follows: the original question is used as the query statement, and the BM25 algorithm is used to score the relevance between each keyword in the original question and all text blocks in the structured knowledge base. After sorting all text blocks from high to low scores, the top 10 to 20 text blocks are selected. For each selected text block, context expansion is performed to recall the previous and next adjacent blocks in the structured document. After merging and deduplication, candidate text blocks are output.

[0056] The BM25 relevance score is calculated using three factors: first, the word frequency factor, where the higher the frequency of a keyword in the corresponding text block, the greater its contribution to the score; second, the inverse document frequency factor, where the narrower the range of documents in which the keyword appears in the entire knowledge base, the higher its weight; and third, the length normalization factor, which is used to offset the score bias caused by the greater number of natural words in long texts.

[0057] For example, when the original question contains 5 keywords, under the premise that the text length and terminology are similar, the text block that hits more keywords will have a higher overall score; while for rare professional terms, due to their high inverse document frequency weight, a single hit can significantly improve the ranking priority of the corresponding text block, thereby ensuring the accurate retrieval of professional information.

[0058] Keyword retrieval excels at precisely matching specific terms such as dye names, experimental parameters, and reagent types, ensuring accurate recall at the terminology level. Because an overlap rate is set during block segmentation, adjacent text blocks inherently possess content continuity; combined with contextual expansion, this further ensures the integrity of experimental parameter groups and mechanistic derivation chains distributed across paragraphs.

[0059] The specific process of obtaining candidate text blocks by semantic vector retrieval of a structured knowledge base in a dye domain knowledge enhancement big model intelligent agent system described above is as follows: First, the original question is converted into a query vector using an Embedding model. Then, the cosine similarity between the query vector and all text block vectors in the structured knowledge base is calculated. Next, a similarity threshold of 0.55~0.85 is set, and text blocks with cosine similarity higher than the similarity threshold are selected and sorted from high to low according to cosine similarity. The top 10~20 text blocks are selected. Finally, for each selected text block, context expansion is performed to recall the previous and next adjacent blocks in the structured document. After merging and deduplication, candidate text blocks are output.

[0060] Semantic vector retrieval excels at understanding semantic needs such as mechanism derivation and process optimization logic, and can cover synonymous expressions and implicit connections across paragraphs. The similarity threshold and recall number of semantic vector retrieval can be adjusted according to scenario requirements: in the initial state, a broad recall strategy with a lower similarity threshold and a larger number of recall segments can be used to avoid missed detections; after optimization, the similarity threshold can be increased and the number of recall segments can be reduced, at the cost of moderately sacrificing the recall rate of potentially relevant segments, to significantly improve the information density and accuracy of the context fed into subsequent inference stages, reduce semantically similar but topic-irrelevant noise, and thus effectively suppress the hallucinations generated by the underlying large language model when there is insufficient evidence or high evidence noise.

[0061] As described above, in a dye domain knowledge-enhancing large model intelligent agent system, the reordering module works as follows: First, each candidate text block is constructed with the original question to form a joint input sequence "[CLS] Original Question [SEP] Candidate Text Block [SEP]", which is then input into a deep Transformer model, which outputs a relevance score. Then, all candidate text blocks are reordered from high to low according to their relevance scores, and the top K candidate text blocks are truncated, with K ranging from 5 to 15. Finally, a truncation threshold of 0.3 to 0.7 is set, and candidate text blocks with relevance scores lower than the truncation threshold are removed. The remaining candidate text blocks from the top K candidate text blocks are then output to the hallucination-free reasoning module.

[0062] Setting a truncation threshold can prevent the forced inclusion of low-quality information when the structured knowledge base lacks sufficient relevant content. It also facilitates the subsequent hallucination-free reasoning module in determining the level of knowledge coverage and triggering supplementary searches or standardized responses. The reordering module effectively filters out semantically similar but technically irrelevant noise fragments, significantly reducing the probability of these noise fragments entering the underlying large language model context. This suppresses the knowledge illusion phenomenon caused by evidence noise or lack of evidence in the underlying large language model from the data source.

[0063] The reordering process of this invention is significantly different from existing technologies. For example, the reordering process in patent application CN120296147A involves: merging and optimizing the scores of the candidate search results based on the search source; calculating the comprehensive score of the search results according to preset calculation rules; and outputting each search result in descending order of the comprehensive score. Essentially, it is a multi-source result fusion and sorting process within the search stage. The only technical problem it solves is that the score scales of vector, ES, and graph searches are not uniform, making direct merging and sorting impossible. Its entire operation involves: normalizing the initial scores of the three search paths; adding a fixed score based on the "number of simultaneously hit search sources" (adding 0.05 / 0.1 for dual-source hits and 0.15 for triple-source hits); and then outputting the results in descending order of the total score. This mechanism has clear technical boundaries.

[0064] First, the ranking is based solely on "number of search sources + initial keyword matching score", which does not involve the deep semantic interaction between the question and the candidate text, and cannot determine content with "overlapping keywords but unrelated technical scenarios".

[0065] Secondly, by using fixed hard rules for scoring, the sorting accuracy can only solve the problem of score unification for results from multiple sources, and cannot further improve the semantic relevance of the results.

[0066] The reordering process of this invention is an independent semantic-level secondary fine ranking step after the initial retrieval screening. The technical essence is sentence-pair semantic interaction scoring based on the deep Transformer model: the original question and candidate text blocks are constructed as a joint input sequence, the relevance score is calculated through deep semantic interaction, and then a truncation threshold is used for secondary screening.

[0067] The core differences between the reordering process of this invention and patent application CN120296147A are reflected in three aspects:

[0068] First, the technical levels are different: Patent application CN120296147A is the initial screening and ranking within the retrieval process; this invention is an independent fine ranking process after the retrieval is completed, which belongs to two different processing flows.

[0069] Second, the sorting criteria are different: patent application CN120296147A relies on the regular numerical addition and subtraction of the retrieval source; this invention relies on the deep semantic matching degree between the question and the text;

[0070] Third, they address different technical problems: Patent application CN120296147A addresses the issue of inconsistent score scales in multiple search results; this invention addresses the semantic noise problem in the initial screening results where "technical terms are hit but the technical scenario is irrelevant".

[0071] The improvement of the reordering process in this invention compared to patent application CN120296147A is not a conventional technical choice in the art because:

[0072] On the one hand, the technical approach of patent application CN120296147A remains in the architecture of "multi-retrieval fusion + rule sorting", which already has sorting function. Those skilled in the art have no motivation to add an additional semantic reordering architecture based on the Transformer model on the basis of the existing sorting steps, and the existing technology has not provided corresponding improvement inspiration.

[0073] On the other hand, precisely because the dye field exhibits the technical characteristic of "the same dye corresponding to different media and different reaction systems with vastly different parameters," keyword matching is highly prone to producing false positive results, necessitating the introduction of semantic reordering to filter out specialized noise. In general domains, the required retrieval accuracy does not reach this level, and those skilled in the art have no explicit motivation to apply semantic reordering to RAG systems in the dye field. This design effectively filters out noise fragments that are semantically similar but technically irrelevant, significantly improving the confidence of knowledge fragments fed into the inference process, suppressing illusions from the data source, and representing a significant technological advancement for specialized dye scenarios.

[0074] As described above, the constraints of the illusion-free reasoning module of the dye domain knowledge-enhancing large model intelligent agent system include source tracing constraints, derivation constraints, translation constraints, and output constraints.

[0075] The traceability constraint stipulates that all output parameters and schemes must be accompanied by the identifier of the supporting source document. When the directly related source document cannot be found in the structured knowledge base, the user shall be clearly informed in a standardized format. It is strictly forbidden to make up any parameters. The source document includes electronic documents of literature and scanned electronic documents of professional reference books.

[0076] Derivation constraints allow for transfer derivation based on rules that are clearly recorded and verified in source documents, even when there are no directly related source documents. The rules and corresponding source documents are clearly marked.

[0077] Translation constraints require that, in the absence of any discernible pattern, forced derivation is prohibited. Instead, mature solutions with high structural similarity to similar substances or systems are selected from the structured knowledge base, and after adaptive parameter adjustments, they are used as reference outputs. At the same time, users are clearly informed that the solution is a translation reference for similar substances, with a lower confidence level than directly retrieved and deduced solutions, and it is recommended to further optimize it after experimental verification.

[0078] The output constraints standardize and regulate the specifications of all output schemes, and provide a link at the end of the output to jump to the original text in the structured knowledge base, ensuring that all conclusions are traceable.

[0079] The above four types of constraints can be embedded into the underlying large language model through the system prompt word framework, forcing the model's reasoning and generation process to strictly follow the rule boundaries, ensuring the rigor and traceability of the output.

[0080] The hallucination-free reasoning module of this invention is significantly different from existing reasoning modules. For example, the reasoning module in patent application CN120296147A only includes two steps: 1. Reorganizing the retrieved atomic content according to the title hierarchy to ensure contextual coherence; 2. Inputting the large language model to polish the aggregated content and generate the final answer. Its essence is the reorganization of the retrieved content layout + language polishing, with the core function being to optimize the fluency and contextual coherence of the answer, which belongs to the preprocessing optimization of the model input. It has no constraint mechanism on the large model's generation behavior; the model can still freely play its role and supplement information beyond the retrieved content. Hallucination suppression only relies on the indirect effect of "more complete input context," without rigid control from the generation logic.

[0081] The core of the hallucination-free reasoning module of this invention is a progressive reasoning paradigm under four levels of constraints. Through four types of rigid rules—source constraints, derivation constraints, translation constraints, and output constraints—it strictly controls the output boundary of the model from the perspective of generative logic.

[0082] When there is direct literature evidence, strictly base the output parameters on the original text and enforce source tracing;

[0083] When there are no direct documents but common patterns exist, we can only deduce based on verified patterns and indicate the basis for our conclusions.

[0084] When there is no discernible pattern, we can only translate similar schemes based on structural similarity, and mark the confidence level and risk warning;

[0085] All outputs must be structured and traceable; unfounded or fabricated parameters are strictly prohibited.

[0086] The hallucination-free reasoning module of the present invention achieves significantly better results than existing reasoning modules for the following reasons:

[0087] First, there is a fundamental difference in the hallucination suppression effect. The indirect suppression method of patent application CN120296147A cannot completely prevent the model from freely manipulating parameter details, while dye research scenarios have extremely high requirements for parameter accuracy. Any unfounded, fabricated parameters will lead to experimental failure. This invention completely prohibits unfounded outputs from the generation rules, achieving true "hallucination-free" results and fully meeting the rigorous requirements of scientific research scenarios.

[0088] Second, it represents a fundamental breakthrough in capability boundaries. Patent application CN120296147A can only reiterate the content of existing literature, essentially serving as a "data lookup" tool. This invention, while strictly ensuring traceability, expands upon two compliant reasoning paths: "law deduction" and "structural translation." It can generate scientifically based reference experimental schemes for unknown dye systems without direct literature documentation, upgrading the RAG system from a "knowledge retrieval tool" to a "research and development auxiliary tool"—a technological effect that existing technologies cannot achieve.

[0089] Patent application CN120296147A offers no technical inspiration for "hierarchical constraints on model generation behavior," and existing general-purpose RAG systems generally adopt an open paradigm of "input context + free model generation." This invention solidifies the rigorous methodology of dye research into the system's generation constraint rules, constructing a three-level progressive controllable reasoning framework of "matching-derivation-translation." This is a unique innovation specific to the dye research and development scenario, not a conventional approach in general software design, nor something that would be readily apparent to those skilled in the art.

[0090] Beneficial effects:

[0091] (1) By constructing a structured knowledge base in the field of dyes, this invention integrates academic literature in the field with basic information on dyes, and realizes the unified collection and structured storage of knowledge in the field of disperse dyes, which effectively improves the reuse efficiency of knowledge in the field.

[0092] (2) The present invention uses a hybrid retrieval module to perform keyword retrieval and semantic vector retrieval in parallel, and combines a re-ranking module to perform deep semantic interactive scoring and re-ranking of candidate text blocks, taking into account both the recall coverage and the accuracy of the results of knowledge retrieval, and can efficiently filter out high-confidence knowledge fragments.

[0093] (3) The present invention uses the non-illusion reasoning module to call the underlying large language model for reasoning generation based on preset constraints, which effectively constrains the generation boundary of the model, suppresses the phenomenon of knowledge illusion, and ensures the traceability and scientific rigor of the output content.

[0094] (4) Under the specification of preset constraints, the present invention supports the transfer derivation based on verified laws and the translation reference of similar substance schemes, and can generate scientifically based reference schemes for unknown dye systems without direct literature records, which greatly reduces the cost of manual trial and error in dye experimental research and development. Attached Figure Description

[0095] Figure 1 Overall workflow diagram of the large-scale intelligent agent system for enhancing knowledge in the dye field;

[0096] Figure 2 Layered architecture diagram of a large-scale intelligent agent system for enhancing knowledge in the dye domain;

[0097] Figure 3 A schematic diagram illustrating the structure of a knowledge base in the dye industry;

[0098] Figure 4 A schematic diagram of the workflow of the hybrid retrieval module in a large-scale intelligent agent system for enhancing knowledge in the dye domain;

[0099] Figure 5 Summary diagram of problems in the performance verification test set of the large-scale intelligent agent system for enhancing knowledge in the dye domain;

[0100] Figure 6 Schematic diagrams of the molecular structures of SES-RB5, VS-RB5, HES-RB5, SES-RB49, VS-RB49 and HES-RB49;

[0101] Figure 7 Capillary electrophoresis separation spectra of six analytes under different cyclodextrin conditions are shown; (a) is the separation spectrum under α-CD conditions, (b) is the separation spectrum under β-CD conditions, and (c) is the separation spectrum under γ-CD conditions. Detailed Implementation

[0102] The present invention will be further described below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0103] A dye domain knowledge-enhanced large-scale intelligent agent system, with a hierarchical architecture as follows: Figure 2 As shown, it includes:

[0104] (1) Construct a structured knowledge base;

[0105] The steps for building a structured knowledge base are as follows:

[0106] (a) Obtain electronic documents of literature and scanned electronic documents of professional reference books;

[0107] The electronic documents were obtained by searching and downloading relevant literature in the field of dyes from academic databases such as Web of Science and CNKI. The research directions of the relevant literature in the field of dyes cover chromatographic separation and analysis techniques such as capillary electrophoresis, spectroscopic characterization techniques, non-aqueous dyeing techniques, dye structure and dyeing properties, hydrolysis kinetics, wastewater treatment, and functional dyes. The types of relevant literature in the field of dyes are mainly journal articles, but also include dissertations and conference papers. The publication years of relevant literature in the field of dyes are mainly from the last 20 years, with some classic literature dating back to the 1980s. A total of more than 800 relevant literatures in the field of dyes were finally included in the knowledge base.

[0108] This scanned electronic document of a professional reference book was obtained by collecting and scanning basic information on at least 4,000 dyes across 13 major categories from professional reference books such as "Complete Guide to Dye Varieties." The 13 major categories of dyes are acid dyes, basic dyes, direct dyes, disperse dyes, reactive dyes, ice dyes, solvent dyes, sulfur dyes, vat dyes, organic pigments, fluorescent whitening agents, food dyes, and mordant dyes. The basic dye information includes the dye name, CAS number (or CI number, EC number), domestic trade name, chemical structure type, chemical structure diagram, molecular formula, molecular weight, hue, properties and uses, fastness performance table, manufacturing method, references, English trade name and manufacturer, and domestic manufacturer.

[0109] The data structure of the above-mentioned structured knowledge base is as follows: Figure 3 As shown;

[0110] (b) Preprocess the electronic documents of the literature and the scanned electronic documents of professional reference books to obtain structured documents;

[0111] The preprocessing process for electronic documents is as follows: First, information irrelevant to the core technical content is filtered out. Then, the integrity and standardization of the text are improved, transforming it into a structured document. This irrelevant information includes redundant information such as headers and footers, advertising snippets, repeated paragraphs, and irrelevant appendices, as well as supplementary information such as reference lists, external links, and email addresses. Improving the integrity and standardization of the text involves correcting issues such as garbled characters and incorrect formatting, and removing unnecessary line breaks and spaces that affect text continuity. The transformation process involves: first, page layout recognition; then, an automatic parsing algorithm is used to parse the electronic document page by page. This method can ignore segmentation markers, accurately extract text content, automatically identify and retain core information such as the main text, charts, and experimental data, and merge images and tables into the relevant text paragraphs to form segments containing complete knowledge units, resulting in a structured document.

[0112] Taking a typical PDF file of academic literature in the field of dyes as an example, such files contain two-column text, headers and footers, page breaks, figures and tables, and a list of references. During preprocessing, a conventional layout parsing tool is first used to read the PDF file, automatically identifying and distinguishing areas such as the main text, figure and table titles, headers and footers, and references. The main text content is extracted into continuous text in the normal reading order, while the figures and tables retain their position markers, and the output is a preliminary structured document. On this basis, redundant content such as headers and footers, journal advertisements, author information, list of references, and acknowledgments that are not related to the core technical content is further filtered out. Character garbled characters, extra line breaks and spaces caused by scanning or format conversion are corrected, and the impact of pagination on the continuity of text is eliminated. The main text content is spliced ​​into continuous text in the normal reading order of scientific and technological literature, and figure captions and table captions are merged into their corresponding main text paragraphs, so that each paragraph contains a complete knowledge unit, and finally a standardized and unified structured document is obtained.

[0113] The above preprocessing process can be implemented based on conventional document parsing tools in the field. For example, basic processing can be completed through the automatic document parsing and cleaning function built into the HiAgent large model application development platform.

[0114] The preprocessing process for scanned electronic documents of professional reference books is as follows: the OCR engine is used to recognize and extract the content of the scanned electronic documents and convert them into two-column editable text. The recognition results are then reconstructed from two-column layout to single-column continuous text. After manually checking and correcting the recognition errors and layout misalignments of the professional information by comparing them with the original scanned documents, the documents are organized into structured documents.

[0115] The scanned electronic document of the professional reference book was a low-resolution, two-column PDF file. The content was arranged by dye type, covering a large amount of professional information such as name, number, chemical parameters, and performance data tables, making it unsuitable for direct use in building a structured knowledge base. During preprocessing, the PDF file was first converted using a standard OCR tool to obtain an editable Word document that retained the original two-column layout. Then, the document formatting function was used to adjust the two-column layout to a single-column continuous text, ensuring the content on that page flowed smoothly in normal reading order. Issues such as cross-column breaks, text misalignment, and table row / column misalignment that occurred during the conversion process were addressed. Initial corrections were made to layout issues to ensure the content order was completely consistent with the original document. Finally, for professional content that is prone to identification errors, such as dye names, CI numbers, molecular formulas, and fastness performance parameters, each page was manually checked and corrected against the original scanned document to prevent erroneous data from entering the knowledge base and to ensure the accuracy of subsequent retrieval and reasoning. After the verification was completed, the text was further converted into a Markdown format structured document with a unified heading hierarchy and standardized table structure, so that different dye entries have a consistent knowledge unit structure, which facilitates the subsequent adaptive block segmentation algorithm to accurately identify knowledge boundaries and avoid the incorrect segmentation of information for different dyes.

[0116] (c) Perform adaptive chunking on the structured document to obtain text blocks;

[0117] The specific process is as follows: using sentences as the basic unit, the structured document is divided into text blocks of a certain length, and a certain overlap rate is maintained between adjacent text blocks;

[0118] The specific values ​​for the block length and overlap rate mentioned above were obtained through iterative debugging, and the process is as follows:

[0119] (i) Build the system and initialize the block length (512 characters) and overlap rate (10%).

[0120] (ii) Use a test set containing at least 60 professional questions to verify the system’s performance, with retrieval recall and knowledge illusion rate as the core evaluation indicators;

[0121] The retrieval recall rate is calculated using a question-by-question annotation method for the test set: For each question in the test set, dye professionals pre-mark all standard text blocks directly related to the question in the knowledge base as a baseline set; after the system runs, the number of text blocks that hit the baseline set in the recall results is counted, and the single-question recall rate = (number of relevant text blocks hit / total number of relevant text blocks in the baseline set for that question) × 100%. The average single-question recall rate for all test questions is taken as the overall retrieval recall rate of the system.

[0122] The knowledge illusion rate was calculated using a manual, question-by-question verification method: Professional staff checked the output of each question individually. If any of the following three scenarios were found—fabricated experimental parameters, fabricated source document conclusions, or unfounded mechanistic derivations—the question was deemed to contain knowledge illusion. The knowledge illusion rate was calculated as: (Number of questions containing knowledge illusion / Total number of questions in the test set) × 100%.

[0123] (iii) Determine whether the retrieval recall rate is ≥90% and the knowledge illusion rate is ≤5%. If yes, end the process and output the block length and overlap rate of the last iteration. Otherwise, adjust the parameters according to the following rules and return to step (ii).

[0124] If the retrieval recall rate is less than 90%, regardless of the knowledge illusion rate, it indicates that the core knowledge fragments have not been effectively recalled. This is usually caused by the truncation of core content at the block boundaries or the dilution of keywords due to excessively large blocks. In this case, the block length should be reduced within the range of 256 to 512 characters, and the overlap rate should be increased within the range of 10% to 20%. Reducing the block length can make the knowledge topic of a single text block more focused, with a higher proportion of keywords and semantic features, and more prominent feature weights during retrieval matching, thereby improving the recall rate. Increasing the overlap rate can ensure that the complete knowledge unit across blocks is completely covered by at least one text block, avoiding the loss of boundary information. The two work together to improve the retrieval recall rate.

[0125] If the retrieval recall rate is ≥90% and the knowledge illusion rate is >5%, it indicates that the contextual information fed into the large model is severely fragmented. Because the model lacks a complete logical chain and parameter association background, it is prone to automatically fill in the missing information, thus generating illusions. In this case, the block length should be increased within the range of 256~512 characters, and the overlap rate should be reduced within the range of 10%~20%. Increasing the block length allows a single text block to carry a complete set of experimental parameters and the corresponding theoretical description, providing sufficient reasoning basis for the model. Reducing the overlap rate can reduce redundant information and avoid logical conflicts and information interference caused by multiple similar content. The two work together to reduce the knowledge illusion rate.

[0126] The reason for limiting the adjustment range to 256-512 characters and 10%-20% is as follows: If the block length is less than 256 characters, even if the length is further reduced, the text block can no longer carry any complete dye domain knowledge unit, and instead each block will become a "knowledge fragment", making retrieval and reasoning meaningless; if the block length is greater than 512 characters, a single text block will inevitably contain multiple dye varieties or multiple independent knowledge points, keywords will be diluted in long texts, making it impossible to locate them accurately during retrieval, and a large amount of irrelevant content will enter the context as noise, which will exacerbate the illusion; if the overlap rate is less than 10%, the boundary protection effect will almost disappear, and it will be no different from no overlap; if the overlap rate is greater than 20%, there will be too much redundant content, storage and computing costs will increase significantly, and the same knowledge unit will appear repeatedly in three or more blocks, which will easily cause repeated references or logical conflicts after being fed into the model, which will damage the generation quality. Therefore, the above range limit is a necessary technical boundary to maintain the balance between the integrity of knowledge units and the efficiency of retrieval and reasoning.

[0127] (d) Split each text block into short sentences and use the Embedding model to generate vector indices for each sentence to obtain the text block vector;

[0128] (2) User input module;

[0129] The user input module is used to input the original question posed by the user in natural language and output the original question to the hybrid retrieval module, the reordering module, and the hallucination-free reasoning module;

[0130] (3) Hybrid retrieval module;

[0131] like Figure 4 As shown, the hybrid retrieval module is used to input the original question and perform keyword retrieval and semantic vector retrieval on the structured knowledge base respectively to obtain candidate text blocks, and output the candidate text blocks to the re-sorting module;

[0132] The specific process of obtaining candidate text blocks by keyword retrieval of the structured knowledge base is as follows: the original question is used as the query statement, and the BM25 algorithm is used to score the relevance of each keyword in the query statement to all text blocks in the structured knowledge base. All text blocks are sorted from high to low scores and the top 10 to 20 text blocks are selected (the top 10 are selected in this embodiment), which can cover key semantic information and avoid too much redundancy. For each selected text block, context expansion is performed to recall the previous and next adjacent blocks of the text block in the structured document. After merging and deduplication, the candidate text blocks are output.

[0133] The BM25 relevance score is calculated using three factors: first, the term frequency factor, where the higher the frequency of a keyword in a corresponding text block, the greater its contribution to the score; second, the inverse document frequency factor, where the narrower the range of documents in which a keyword appears in the entire knowledge base, the higher its weight; and third, the length normalization factor, which is used to offset the scoring bias caused by longer texts having a greater number of natural words. For example, when the original question contains 5 keywords, under the premise that the text length and terminology are similar, the text block that hits more keywords will have a higher overall score. For rare professional terms, due to their high inverse document frequency weight, a single hit can significantly improve the ranking priority of the corresponding text block, thereby ensuring the accurate recall of professional information.

[0134] The specific process of obtaining candidate text blocks by semantic vector retrieval of the structured knowledge base is as follows: First, the original question is converted into a query vector using the Embedding model. Then, the cosine similarity between the query vector and all text block vectors in the structured knowledge base is calculated. Next, the similarity threshold is set to 0.55~0.85 (the specific value in this embodiment is 0.65). While ensuring the recall rate, the retrieval accuracy is improved. Text blocks with cosine similarity higher than the similarity threshold are selected and sorted from high to low according to cosine similarity. The top 10~20 text blocks are selected (the top 10 are selected in this embodiment). Finally, for each selected text block, context expansion is performed to recall the previous and next adjacent blocks of the text block in the structured document. After merging and deduplication, the candidate text blocks are output.

[0135] (4) Reordering module;

[0136] The reordering module is used to input the original question and candidate text blocks, and after performing deep semantic interactive scoring, it reorders all candidate text blocks from high to low according to their relevance scores, and outputs the top candidate text blocks to the hallucination-free reasoning module.

[0137] The reordering module works as follows: First, each candidate text block is combined with the original question to form a joint input sequence "[CLS] Original Question [SEP] Candidate Text Block [SEP]", which is then input into the deep Transformer model, which outputs a relevance score. Then, all candidate text blocks are reordered from high to low according to their relevance scores, and the top K candidate text blocks are truncated. The value of K ranges from 5 to 15 (K=10 in this embodiment). Finally, the truncation threshold is set to 0.3 to 0.7 (0.65 in this embodiment). Candidate text blocks with relevance scores lower than the truncation threshold are removed from the top K candidate text blocks, and the remaining candidate text blocks from the top K candidate text blocks are output to the hallucination-free reasoning module.

[0138] (5) Non-hallucination reasoning module;

[0139] The illusion-free reasoning module is used to input the original question and the candidate text blocks sent by the reordering module. Based on the preset constraints, it calls the underlying large language model to perform reasoning and generation, and outputs the structured answer that meets the constraints to the user interface or downstream application module.

[0140] The constraints of the hallucination-free reasoning module include source tracing constraints, derivation constraints, translation constraints, and output constraints.

[0141] The traceability constraint stipulates that all output parameters and schemes must be accompanied by the identifier of the supporting source document. When the directly related source document cannot be found in the structured knowledge base, the user shall be clearly informed in a standardized format. It is strictly forbidden to make up any parameters. The source document includes electronic documents of literature and scanned electronic documents of professional reference books.

[0142] Derivation constraints allow for transfer derivation based on rules that are clearly recorded and verified in source documents, even when there are no directly related source documents. The rules and corresponding source documents are clearly marked.

[0143] Translation constraints require that, in the absence of any discernible pattern, forced derivation is prohibited. Instead, mature solutions with high structural similarity to similar substances or systems are selected from the structured knowledge base, and after adaptive parameter adjustments, they are used as reference outputs. At the same time, users are clearly informed that the solution is a translation reference for similar substances, with a lower confidence level than directly retrieved and deduced solutions, and it is recommended to further optimize it after experimental verification.

[0144] The output constraints standardize and regulate the specifications of all output schemes, and provide a link at the end of the output to jump to the original text in the structured knowledge base, ensuring that all conclusions are traceable.

[0145] The overall workflow of the aforementioned dye domain knowledge-enhanced large-scale intelligent agent system is as follows: Figure 1 As shown.

[0146] The aforementioned dye domain knowledge-enhanced large-scale model intelligent agent system can be built on the HiAgent large-scale model application development platform. The construction steps are as follows:

[0147] S1. Register and log in to the HiAgent large model application development platform to support visual orchestration of intelligent agents and knowledge bases;

[0148] S2, go to the “Agent” module in the left navigation bar of the homepage, click the blue “Create Agent” button in the upper right corner of the page, enter the agent name, and choose to upload or automatically generate an avatar; select “Conversational Agent” for the agent type. In this mode, the agent will think and plan tasks autonomously, which is suitable for natural language question answering and solution generation.

[0149] S3, enter the agent orchestration page, and in the "role setting" module on the left, fill in the system prompt words (including the specific content of tracing constraints, derivation constraints, translation constraints and output constraints) according to the multi-dimensional constraint framework designed in this invention.

[0150] S4, in the “Model Configuration” module, select Doubao-pro-32k as the underlying large language model;

[0151] In the parameter settings, the temperature parameter is set to 1.00, and the kernel sampling threshold (top-p) is set to 0.70;

[0152] In the input and output settings, the maximum output length for a single response is set to 4096 tokens;

[0153] S5, in the "Knowledge Base" module, click "Add Knowledge Base" and select the already built structured knowledge base;

[0154] S6, in the "Knowledge Base Recall Settings" module, set the recall mode to forced mode, that is, only use the structured knowledge base content to generate answers, and then use the structured knowledge base content as context input to the underlying large language model after matching.

[0155] Set the search method to hybrid search (keyword search and semantic vector search) and enable the Rerank re-ranking model;

[0156] Set the number of fragments to be retrieved, the similarity threshold, and enable the context expansion function;

[0157] S7, check the "References" and "Attribution" options. When enabled, the output content will automatically be marked with the corresponding source document identifier and jump link; check "Dialogue Suggestions". After the agent replies, it will automatically provide suggestions for the user's questions based on the dialogue content. Setting dialogue suggestions can help users have better conversations.

[0158] To verify the technical effectiveness of the aforementioned knowledge-enhanced large-scale intelligent agent system in the dye field, a test set containing 61 professional questions was used to verify the system's performance. The test questions covered various scenarios, including concept definition, mechanism derivation, process optimization, experimental schemes, and transfer derivation. Figure 5As shown in the test, the system can accurately retrieve and completely output experimental schemes that are clearly recorded in the structured knowledge base, with accompanying source document traceability links. For unknown systems not directly recorded in the structured knowledge base, it can perform rigorous summarization and deduction based on similar structural schemes and established mechanisms, and clearly mark the source documents on which the deduction is based. The overall retrieval recall rate is ≥90%, no knowledge illusion phenomenon was detected, and the consistency of the results of repeated answers to the same question is 100%. All output content can be traced back to the corresponding paragraphs of specific source documents, verifying the effectiveness and reliability of this system. Any combination of parameters within the previously disclosed range of block length 256~512 characters, overlap rate 10%~20%, similarity threshold 0.55~0.85, reordering K=5~15, and truncation threshold 0.3~0.7 will only slightly fluctuate the recall rate and illusion rate, and can still achieve the core technical effects of traceable structured knowledge base, suppression of knowledge illusion, and generation of dye experimental reference schemes, without changing the core innovation of this invention. The embedding model and reordering model can adopt conventional semantic matching models in this field. Those skilled in the art can determine the appropriate parameter combination through a limited number of conventional iterative experiments based on the debugging process disclosed above, with the target of the compliance criterion, without the need for innovative design of parameter debugging logic.

[0159] This invention also uses the simultaneous capillary electrophoretic separation and analysis of six analytes, including Reactive Black 5, Reactive Brown 49, and their respective activated and hydrolyzed forms, as an application scenario to verify the effectiveness of this system in generating experimental schemes and iteratively optimizing unknown dye systems. The separation method for this system is not directly recorded in the system's structured knowledge base. The verification process is as follows:

[0160] The target system consisted of two structurally similar isomers of reactive dyes, Reactive Black 5 and Reactive Brown 49, along with their respective activated and hydrolyzed forms, totaling six analytes: SES-RB5, VS-RB5, HES-RB5, SES-RB49, VS-RB49, and HES-RB49. Specifically, SES-RB5 represents Reactive Black 5, VS-RB5 represents the activated form of Reactive Black 5, HES-RB5 represents the hydrolyzed form of Reactive Black 5, SES-RB49 represents Reactive Brown 49, VS-RB49 represents the activated form of Reactive Brown 49, and HES-RB49 represents the hydrolyzed form of Reactive Brown 49. The chemical structures of the six analytes are as follows: Figure 6 As shown, it belongs to the "unknown system" as defined in this invention.

[0161] In the initial parameter generation phase, the user inputs the original question through the user input module. The original question is "Please provide an initial scheme for the capillary electrophoretic separation and analysis of six components, including two dyes, Reactive Black 5 and Reactive Brown 49, and their activated and hydrolyzed forms." The original question is simultaneously output to the hybrid retrieval module, the reordering module, and the hallucination-free reasoning module. After receiving the original question, the hybrid retrieval module performs keyword retrieval and semantic vector retrieval on the structured knowledge base to obtain candidate text blocks, and outputs the candidate text blocks to the reordering module. The reordering module receives the original question and candidate text blocks, performs deep semantic interaction scoring, and reorders all candidate text blocks according to their relevance scores from high to low, outputting the top high-confidence knowledge fragments to the hallucination-free reasoning module. After receiving the original question and high-confidence knowledge fragments, the hallucination-free reasoning module calls the underlying large language model based on preset constraints to initiate the structured summary reasoning process. First, the molecular structures of Reactive Black 5 and Reactive Brown 49 are analyzed. Both are divinyl sulfone reactive dyes, and their parent structures are large conjugated systems containing two naphthalene rings and one benzene ring, with four sulfonic acid groups in each. The only difference between the two lies in the slight variation in the distribution of substituents on the molecular backbone. This results in their physicochemical parameters, such as charge density, molecular size, and hydrophobicity, being almost identical, making them typical isomers with remarkably similar mobility and partitioning behavior. A structural similarity search was then performed in the structured knowledge base, revealing existing capillary electrophoresis separation methods for Active Black 5. The background electrolyte conditions were 15 mmol / L borate buffer (pH adjusted to 9.30 with 0.5 mol / L borate solution) and 15% (v / v) ACN (acetonitrile). Simultaneously, the agent retrieved source documents in the structured knowledge base regarding the mechanism of cyclodextrin separation of isomers, revealing that the inclusion interaction between the hydrophobic cavity of cyclodextrin and the aromatic ring of the dye molecule can distinguish isomers. Based on known schemes with similar structures and combined with the mechanism of cyclodextrin inclusion complexation, the agent generated an initial parameter scheme through summarization and reasoning. The complete output background electrolyte is 20 mmol / L borate buffer (pH adjusted to 9.30 with 0.5 mol / L borate solution), 15% (V / V) ACN, and the additive is 20 mmol / L α-CD (α-cyclodextrin).

[0162] Capillary electrophoresis experiments were conducted according to the above initial parameters. The capillaries used were uncoated fused silica capillaries with an inner diameter of 50 μm, a total length of 50 cm, and an effective length of 40 cm. Sample injection was performed using a pressure injection method with an injection pressure of 0.5 psi and an injection time of 15 s. The separation voltage was set to +20 kV, and the detection wavelength was 200 nm. The separation results under α-CD conditions are as follows: Figure 7As shown in (a), the chromatogram clearly shows that the peaks of each component are crowded within a narrow time window, with most peaks completely overlapping. Fewer than four peaks are identifiable, and the baseline fluctuates significantly. The activated forms VS-RB5 and VS-RB49, and the hydrolyzed forms HES-RB5 and HES-RB49, which have extremely similar structures, almost completely co-migrate under these conditions, resulting in severely insufficient resolution and failing to meet the requirements for accurate quantitative analysis.

[0163] During the optimization phase, the user reported experimental defects to the agent, describing them as "insufficient separation, with peaks of each component almost completely overlapping and unable to be separated." The user input module output this feedback as a new original question to the corresponding modules. After receiving the feedback, the agent's hybrid retrieval module conducts an in-depth search of relevant source documents in the structured knowledge base. After the re-sorting module filters high-confidence knowledge fragments, the information is transmitted to the hallucination-free reasoning module. Based on the retrieved source document knowledge, the agent outputs optimization directions, suggesting that the matching degree between the cavity size and the cross-sectional size of the dye molecule be improved by changing the cyclodextrin type, thereby enhancing inclusion selectivity. Specifically, the suggestion is to replace α-CD with β-CD (β-cyclodextrin) and γ-CD (γ-cyclodextrin) in sequence, perform gradient comparison, and keep other parameters unchanged.

[0164] Before conducting the experiment, professionals analyzed and specified the optimization suggestions for the intelligent agent's output. Cyclodextrin molecules are cyclic oligosaccharides composed of D-glucose units linked end-to-end by α-1,4-glycosidic bonds, with an overall truncated pyramidal shape. Their outer edges are hydrophilic due to the presence of numerous hydroxyl groups, while the internal cavity exhibits a hydrophobic environment due to the aggregation of glycosidic bond oxygen atoms and methylene groups. Dye molecules contain hydrophobic aromatic groups such as naphthalene and benzene rings, which can enter the cyclodextrin cavity through hydrophobic interactions, forming host-guest inclusion complexes. The difference between different types of cyclodextrin lies in the varying number of glucose units, resulting in significant differences in the cavity's inner diameter. Literature review confirmed the cavity inner diameter data for different cyclodextrins as follows: α-CD approximately 0.47 to 0.53 nm, β-CD approximately 0.60 to 0.65 nm, and γ-CD approximately 0.75 to 0.83 nm. Both Reactive Black 5 and Reactive Brown 49 molecules contain hydrophobic aromatic groups such as naphthalene and benzene rings. After molecular mechanics optimization and van der Waals surface calculations, the cross-sectional dimensions of their key hydrophobic aromatic backbone fragments are approximately 0.70 nm and 0.75 nm, respectively. This size difference, and its matching degree with the inner diameter of the hydrophobic cavity of cyclodextrin, is one of the key factors affecting the stability and selectivity of the host-guest inclusion complex, which may further affect the difference in their inclusion constants (Kf), ultimately manifesting in electrophoretic mobility. Based on these data, a structure-activity relationship analysis was performed. When the α-CD cavity is too small, the two dye molecules can only achieve partial inclusion, meaning only one end or side chain of the molecule enters the cavity, forming a shallow inclusion complex, unable to form a stable deep inclusion complex. In the shallow inclusion state, the 0.05 nm cross-sectional difference between the two molecules cannot be effectively identified, which is related to… Figure 7 The experimental results in (a) are in perfect agreement. While the β-CD cavity is larger, theoretically improving inclusion depth and selectivity, its inner diameter (0.60 to 0.65 nm) is still smaller than the cross-sectional dimensions of the two dye molecules, making it unlikely to form deep, stable inclusion complexes. The γ-CD cavity, between 0.75 and 0.83 nm, closely matches the cross-sectional dimensions of Reactive Black 5 (0.70 nm) and Reactive Brown 49 (0.75 nm). Within the γ-CD cavity, both dye molecules can achieve deep inclusion, accommodating the entire naphthalene ring or bicyclic fragment. Furthermore, in this case, the cross-sectional size of Reactive Brown 49 is closer to the upper limit of the γ-CD cavity's inner diameter, resulting in stronger spatial constraints on its conformational freedom and molecular motion within the cavity. This amplifies subtle differences in spatial configuration and binding modes between the two dyes, translating into observable electrophoretic mobility differences, making it the most promising candidate for baseline separation. This confirms that the scheme of sequentially replacing CDs is reasonable in terms of structure-activity relationship.

[0165] Based on the above analysis, the directional recommendations for the intelligent agent system are concretized into a specific experimental plan. While keeping other conditions constant, β-CD and γ-CD are used sequentially as separation additives for gradient comparison and verification. It is expected that the separation effect will improve progressively in the order of α-CD → β-CD → γ-CD. The experiments are conducted using the confirmed optimized parameters, specifically: a background electrolyte of 20 mmol / L borate buffer (pH adjusted to 9.30 with 0.5 mol / L borate solution), and 15% (v / v) ACN. The experiment attempts to use 20 mmol / L β-CD and γ-CD sequentially as separation additives.

[0166] The separation results under β-CD conditions are as follows: Figure 7 As shown in (b), compared with the α-CD condition, the migration time window of each component peak was widened, and the number of identifiable peaks increased to 4 to 5. Some previously overlapping peak pairs began to show a separation trend. However, some chromatographic peaks had weak responses and low peak heights, resulting in unsatisfactory sensitivity. Overall, the resolution under the β-CD condition was significantly improved compared to α-CD, preliminarily verifying that increasing the cavity size is beneficial for enhancing inclusion selectivity in analytical judgment. However, the overall separation quality still does not meet the requirements for accurate quantitative analysis and further optimization is needed.

[0167] Separation results under γ-CD conditions are as follows Figure 7 As shown in (c), six chromatographic peaks are clearly distinguishable, eluting in the following order: HES-RB5, HES-RB49, VS-RB5, VS-RB49, SES-RB5, and SES-RB49. Complete baseline separation was achieved between each adjacent peak. Figure 7 (a) Crowded overlapping peaks and Figure 7 Compared to the spectrum in (b) where some peaks still overlapped and sensitivity was poor, the chromatographic peaks under γ-CD conditions were sharp and symmetrical, with stable baselines, balanced response intensities, and a significantly improved signal-to-noise ratio. The resolution Rs between adjacent peaks all reached above 1.5, fully meeting the requirements for quantitative analysis.

[0168] The separation effects of the three cyclodextrins are summarized in Table 1.

[0169] Table 1: Comparison of the separation effects of three cyclodextrins

[0170] α-CD 0.47~0.53 <4 <1.0 The peaks overlapped so much that they were difficult to identify. β-CD 0.60~0.65 4~5 1.0~1.2 Some peak responses are weak, resulting in insufficient sensitivity. γ-CD 0.75~0.83 6 ≥1.5 The peaks have balanced responses and sharp peak shapes.

[0171] As shown in Table 1, with the increase of the cyclodextrin cavity inner diameter, the number of resolvable chromatographic peaks increases progressively, and the resolution improves progressively. β-CD, as an intermediate transition, shows a significant improvement in separation performance compared to α-CD, but still has shortcomings; the weak peak responses of some analytes suggest that inclusion selectivity has not yet reached its optimal state. γ-CD exhibits the best performance in terms of resolution, peak response uniformity, and baseline stability, ultimately achieving complete baseline separation of all six analytes. This progressive improvement trend is highly consistent with predictions made by professionals based on cavity size matching theory, providing a complete chain of experimental evidence for the separation mechanism that "cavity size-molecular cross-sectional size matching determines inclusion selectivity."

[0172] In this experimental sequence, the initial scheme generated by the intelligent agent system failed to achieve effective separation. However, after receiving experimental feedback, it quickly identified the optimization direction of "changing the type of cyclodextrin" based on source document retrieval. By reviewing specific structural parameter data, the system transformed the directional suggestion into a specific experimental scheme and successfully predicted the separation effect ranking of the three cyclodextrins through structure-activity relationship analysis.

[0173] The entire development process of this scheme only went through one round of optimization iteration, and completed three sets of experimental verifications of the initial α-CD scheme, β-CD scheme, and γ-CD scheme. Compared with the traditional manual trial and error method, which usually requires several weeks or even months of multiple rounds of gradient experiments, the efficiency has been greatly improved.

[0174] It should be noted that for novel separation mechanisms or the mechanisms of action of unconventional additives that are beyond the coverage of the structured knowledge base, the agent may give a hint of insufficient knowledge coverage due to the lack of source documents. In this case, manual intervention is required to supplement experimental exploration or to conduct a series of pre-experiments based on the suggestions provided by the model.

Claims

1. A knowledge-enhancing large-scale intelligent agent system for the dye domain, characterized in that, include: The construction steps of a structured knowledge base are as follows: (a) Obtain electronic documents of literature and scanned electronic documents of professional reference books; The electronic documents were obtained by retrieving and downloading relevant literature in the field of dyes from academic databases. The scanned electronic document of the professional reference book was obtained by collecting and scanning basic information on at least 4,000 dye varieties from 13 major categories of professional reference books. (b) Preprocess the electronic documents of the literature and the scanned electronic documents of professional reference books to obtain structured documents; (c) Perform adaptive chunking on the structured document to obtain text blocks; (d) Split each text block into short sentences and use the Embedding model to generate vector indices for each sentence to obtain the text block vector; The user input module is used to input the original question posed by the user in natural language and output the original question to the hybrid retrieval module, the reordering module, and the hallucination-free reasoning module. The hybrid retrieval module is used to input the original question and perform keyword retrieval and semantic vector retrieval on the structured knowledge base respectively to obtain candidate text blocks, and output the candidate text blocks to the re-sorting module; The reordering module is used to input the original question and candidate text blocks, and after performing deep semantic interactive scoring, it reorders all candidate text blocks from high to low according to their relevance scores, and outputs the top candidate text blocks to the hallucination-free reasoning module. The illusion-free reasoning module is used to input the original question and candidate text blocks provided by the reordering module. Based on preset constraints, it calls the underlying large language model to perform reasoning and generation, and outputs a structured answer that meets the constraints to the user interface or downstream application module.

2. The dye domain knowledge-enhancing large-scale intelligent agent system according to claim 1, characterized in that, In step (a), the research directions of relevant literature in the dye field cover chromatographic separation and analysis technology, spectroscopic characterization technology, non-aqueous dyeing technology, dye structure and dyeing properties, hydrolysis kinetics, wastewater treatment, and functional dyes; Journal articles are preferred in the field of dyes, but dissertations and conference papers are also accepted. The publication years of literature related to dyes are mainly from the last 20 years; The 13 major categories of dyes are acid dyes, basic dyes, direct dyes, disperse dyes, reactive dyes, ice dyes, solvent dyes, sulfur dyes, vat dyes, organic pigments, fluorescent whitening agents, food dyes, and mordant dyes. Basic information about dyes includes dye name, CAS number, domestic trade name, chemical structure type, chemical structure diagram, molecular formula, molecular weight, hue, properties and uses, fastness performance table, manufacturing method, references, English trade name and manufacturer, and domestic manufacturer.

3. The dye domain knowledge-enhancing large-scale intelligent agent system according to claim 1, characterized in that, In step (b), the preprocessing of the electronic document is as follows: first, filter out information in the electronic document that is not related to the core technical content, then improve the integrity and standardization of the text, and transform it into a structured document.

4. The dye domain knowledge-enhancing large-scale intelligent agent system according to claim 1, characterized in that, In step (b), the preprocessing of the scanned electronic document of the professional reference book is as follows: the OCR engine is used to identify and extract the content of the scanned electronic document and convert it into two-column editable text. The recognition result is reconstructed from two-column layout to single-column continuous text. After manually checking and correcting the recognition error and layout misalignment of the professional information by comparing it with the original scanned document, it is organized into a structured document.

5. The dye domain knowledge-enhancing large-scale intelligent agent system according to claim 1, characterized in that, The specific process of step (c) is as follows: using sentences as the basic unit, the structured document is divided into text blocks of a certain length, and a certain overlap rate is maintained between adjacent text blocks; The specific values ​​for the block length and overlap rate mentioned above were obtained through iterative debugging, and the process is as follows: (i) Build the system and initialize the block length and overlap rate; (ii) Use a test set containing at least 60 professional questions to verify the system’s performance, with retrieval recall and knowledge illusion rate as the core evaluation indicators; (iii) Determine whether the retrieval recall rate is ≥90% and the knowledge illusion rate is ≤5%. If yes, end the process and output the block length and overlap rate of the last iteration. Otherwise, adjust the parameters according to the following rules and return to step (ii). If the retrieval recall rate is less than 90%, then reduce the block length within the range of 256 to 512 characters and increase the overlap rate within the range of 10% to 20%. If the retrieval recall rate is ≥90% and the knowledge illusion rate is >5%, then increase the block length within the range of 256~512 characters and reduce the overlap rate within the range of 10%~20%.

6. The dye domain knowledge-enhancing large-scale intelligent agent system according to claim 1, characterized in that, The specific process of obtaining candidate text blocks by keyword retrieval of the structured knowledge base is as follows: the original question is used as the query statement, and the BM25 algorithm is used to score the relevance of each keyword in the original question to all text blocks in the structured knowledge base. After sorting all text blocks from high to low scores, the top 10 to 20 text blocks are selected. For each selected text block, context expansion is performed to recall the previous and next adjacent blocks in the structured document. After merging and deduplication, candidate text blocks are output.

7. The dye domain knowledge-enhancing large-scale intelligent agent system according to claim 1, characterized in that, The specific process of obtaining candidate text blocks by semantic vector retrieval of a structured knowledge base is as follows: First, the original question is converted into a query vector using an embedding model. Then, the cosine similarity between the query vector and all text block vectors in the structured knowledge base is calculated. Next, a similarity threshold of 0.55~0.85 is set, and text blocks with cosine similarity higher than the similarity threshold are selected and sorted from high to low according to cosine similarity. The top 10~20 text blocks are selected. Finally, for each selected text block, context expansion is performed to recall the previous and next adjacent blocks in the structured document. After merging and deduplication, candidate text blocks are output.

8. The dye domain knowledge-enhancing large-scale intelligent agent system according to claim 1, characterized in that, The reordering module works as follows: First, each candidate text block is combined with the original question to form a joint input sequence "[CLS] Original Question [SEP] Candidate Text Block [SEP]", which is then input into the deep Transformer model, which outputs a relevance score. Then, all candidate text blocks are reordered from high to low according to their relevance scores, and the top K candidate text blocks are truncated, with K ranging from 5 to 15. Finally, a truncation threshold of 0.3 to 0.7 is set, and candidate text blocks with relevance scores lower than the truncation threshold are removed. The remaining candidate text blocks from the top K candidate text blocks are then output to the hallucination-free reasoning module.

9. The dye domain knowledge-enhancing large-scale intelligent agent system according to claim 1, characterized in that, The constraints of the hallucination-free reasoning module include source tracing constraints, derivation constraints, translation constraints, and output constraints. The traceability constraint stipulates that all output parameters and schemes must be accompanied by the identifier of the supporting source document. When the directly related source document cannot be found in the structured knowledge base, the user shall be clearly informed in a standardized format. It is strictly forbidden to make up any parameters. The source document includes electronic documents of literature and scanned electronic documents of professional reference books. Derivation constraints allow for transfer derivation based on rules that are clearly recorded and verified in source documents, even when there are no directly related source documents. The rules and corresponding source documents are clearly marked. Translation constraints require that, in the absence of any discernible pattern, forced derivation is prohibited. Instead, mature solutions with high structural similarity to similar substances or systems are selected from the structured knowledge base, and after adaptive parameter adjustments, they are used as reference outputs. At the same time, users are clearly informed that the solution is a translation reference for similar substances, with a lower confidence level than directly retrieved and deduced solutions, and it is recommended to further optimize it after experimental verification. The output constraints standardize and regulate the specifications of all output schemes, and provide a link at the end of the output to jump to the original text in the structured knowledge base, ensuring that all conclusions are traceable.

Images