Reasoning module, experimental feedback module, new technology product AI development system and related methods and platforms, equipment

Abstract

This invention, entitled "Inference Module, Experimental Feedback Module, AI Development System for New Technology Products, and Related Methods, Platforms, and Equipment," belongs to the field of artificial intelligence technology. The technical problem it aims to solve is one where existing large-scale language reasoning models cannot directly guide the development of new technology products. The key points of the system's technical solution are as follows: The reasoning module includes: a vector database routing module for selecting the target vector database most relevant to the problem; specifically, based on a comprehensive score (Scorei), selecting the vector database with the highest Scorei score as the target vector database; the vector database for storing structured literature from various fields; and a controlled reasoning execution module for executing the controlled reasoning stage.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to inference modules, experimental feedback modules, AI development systems for new technology products, and related methods, platforms, and equipment. Background Technology

[0002] With the rapid development of artificial intelligence (AI) technology, AI intelligent models are gradually being introduced into new product development to improve development speed and avoid problems such as reliance on empiricism, long testing cycles, and poor data transfer. This is leading to the emergence of a large-scale language reasoning model based on private experimental, production, and literature databases. This model can infer the direction of the next step in product development based on existing knowledge in the relevant field and the existing experimental structure. Through a "human-machine co-creation" development path, it significantly improves the development efficiency and innovation depth of new technology products, enabling collaborative development and industrialization of new technology products by small teams, across multiple categories, with multiple centers, and across generations.

[0003] This system, designed for reasoning development and incorporating a large-scale language reasoning model, structures literature, experimental, and production data to create a vectorized knowledge space (private database). Based on this, it invokes the large-scale language model for controlled reasoning, generating structured reports that directly guide the experimental and process development required for new technology product projects. Simultaneously, the system processes experimental data generated during experiment execution under researcher-defined relational constraints, allowing experimental data to participate as supplementary input in subsequent controlled reasoning and flowing back into the vectorized knowledge space in structured form. This supports iterative reasoning and experimental design optimization based on experimental results. Thus, it realizes a research and development process encompassing problem definition, knowledge retrieval, reasoning generation, experimental verification, and experimental feedback.

[0004] Relevant patent documents retrieved:

[0005] Chinese invention patent document, publication number CN116932717A, entitled "A Knowledge Base Retrieval Method and Apparatus Based on Self-Evaluation and Self-Feedback of a Large Language Model," describes a knowledge base retrieval method comprising the following steps: domain knowledge collection; data preprocessing according to preprocessing logic; generation of an evaluation dataset using a large language model based on evaluation data; retrieval of relevant knowledge information; generation of relevant answers using a large reasoning model, followed by relevance and correctness evaluation; and output of optimized answers. This technology proposes a knowledge base retrieval method and apparatus based on self-evaluation and self-feedback of a large language model, utilizing a large language model for knowledge questioning, automatic evaluation, and automatic feedback. Based on external knowledge, it generates appropriate evaluation data to automatically evaluate the entire method chain; the evaluation results are fed back to the method chain for automatic optimization and iteration.

[0006] Chinese invention patent, publication number CN116932708A, entitled "Open Domain Natural Language Reasoning Question Answering System and Method Driven by a Large Language Model," comprises: a question rewriting module that rewrites user questions into rewritten questions; a central computing and management module that manages the computing and knowledge resources of the large language model, and outputs the rewritten questions and the computing and knowledge resources of the large language model required by the question core engine module to one or more sub-question answering modules in the question answering core engine module according to the type of the rewritten questions; the question answering core engine module reasoning based on the rewritten questions and the computing and knowledge resources of the large language model to obtain one or more candidate answers to the rewritten questions and interpretability descriptions of the candidate answers; and an aggregation reasoning module aggregating reasoning based on the candidate answers and interpretability descriptions of the candidate answers to obtain the final answer to the rewritten questions and interpretability descriptions of the final answer. The system is supported by a large language model, supports comprehensive question types, is easily expandable, interpretable, and highly versatile.

[0007] Chinese invention patent, publication number CN118333169A, entitled "Knowledge-Driven Large Language Model Reasoning Method, Apparatus, Device, and Medium," describes a reasoning method comprising the following steps: Subgraph generation: Extracting main entities from a large language model and generating a queryable subgraph in an external database; Subgraph search reasoning: Constructing a large language model capable of converting natural language questions into Cypher queries and performing subgraph retrieval in the external database; Joint graph reasoning: If the subgraph search reasoning fails to retrieve an answer, using the previous hop knowledge of the question as input to obtain the most relevant knowledge from the external database, performing joint reasoning through the large language model, and generating the final answer. This technology provides ample knowledge support for large language models, making their reasoning in specific knowledge more comprehensive and accurate, improving the transparency and interpretability of reasoning, enhancing the accuracy of reasoning, reducing the cost of knowledge updates, and promoting the development and application of artificial intelligence technology.

[0008] Chinese invention patent, publication number CN118798368A, entitled "A Reasoning Method for Complex Problems Based on Dynamic Collaboration of a Large Language Model and a Domain Knowledge Base," constructs a multi-round collaborative reasoning framework that seamlessly integrates a pre-trained large-scale language model with a specially built domain knowledge base. The language model is responsible for semantic understanding, task decomposition, and knowledge extraction of the research problem; the domain knowledge base, constructed from massive amounts of research literature through fine-grained semantic indexing, provides accurate and comprehensive external knowledge support for the language model's reasoning. During the reasoning process, the language model actively and repeatedly queries relevant literature from the knowledge base according to task requirements, dynamically integrating the acquired new knowledge into subsequent reasoning steps, achieving synergistic enhancement of cognitive reasoning and knowledge retrieval.

[0009] Existing knowledge retrieval and reasoning technologies based on large language models have not formed a closed-loop system for the "full-process R&D of new technology products" and generally suffer from the following technical defects: 1. It is only designed for general question answering and knowledge retrieval, and is not customized for new technology product development scenarios. Existing technologies focus on natural language question answering, open domain reasoning, and general knowledge base retrieval. They are not designed for new product development tasks such as experimental design, process development, product iteration, and industrialization. They cannot directly generate executable experimental plans, development routes, and structured R&D reports, and cannot directly guide real R&D projects.

[0010] 2. Lack of deep integration of private experimental data, production data, and domain-specific data; existing technologies mostly rely on public literature / general knowledge bases, and have not built a dedicated vectorized knowledge space based on the company's private experimental data, production data, and internal literature. This makes it impossible to use the company's unique knowledge for accurate reasoning and easily deviates from the company's actual R&D conditions and process foundation.

[0011] 3. There is no closed-loop R&D mechanism of "experiment execution - data feedback - iterative optimization"; the existing technology only realizes knowledge retrieval → answer generation → self-evaluation and optimization, without taking the actual experimental execution results as feedback input, and cannot dynamically correct the subsequent reasoning direction and optimize the experimental design based on experimental data. It does not support the whole process of closed-loop R&D from problem definition to knowledge retrieval, reasoning generation, experimental verification and data feedback.

[0012] 4. The reasoning process lacks "controlled constraints" in the R&D scenario; existing large language model reasoning is mostly open-domain free reasoning, without setting relationship constraints, experimental logic constraints, and process feasibility constraints that match product development. This easily leads to problems such as reasoning results being detached from reality, unapplicable, and conflicting with existing experimental data, resulting in insufficient innovation depth and engineering practicality.

[0013] Furthermore, existing large-scale language reasoning models typically integrate directly with other knowledge bases, including structured and semi-structured data. However, to ensure the accuracy of the reasoning results, extensive knowledge fusion is often required, which consequently impacts their processing power and efficiency. Additionally, the large-scale knowledge fusion places higher demands on user question-and-answer responses, requiring precise input to obtain the appropriate reasoning results, which also affects the stability of the reasoning process.

[0014] Based on this, the present invention provides a reasoning module, an experimental feedback module, an AI development system for new technology products, and related methods, platforms, and equipment. Summary of the Invention

[0015] The purpose of this invention is to provide: The invention relates to a reasoning module, an experimental feedback module, an AI development system for new technology products, and related methods, platforms, and equipment, in order to address the technical problem that existing large language reasoning models cannot directly guide the development of new technology products.

[0016] Terminology Explanation: Unless otherwise defined, all technical terms used herein have the same meanings as commonly understood by one of ordinary skill in the art to which this subject matter pertains. Unless otherwise stated, all patents, patent inventions, and disclosures cited throughout this document are incorporated herein by reference in their entirety. Where multiple definitions exist for terms herein, the definitions provided in this chapter shall prevail.

[0017] It should be understood that the above brief description and the following detailed description are exemplary and for illustrative purposes only, and do not limit the subject matter of the invention in any way. In this invention, the singular is used in conjunction with the plural unless otherwise specifically stated. It should also be noted that, unless otherwise stated, the use of “or” or “or” means “and / or”. Furthermore, the use of the term “comprising” and other forms such as “including,” “containing,” and “contains” are not limiting.

[0018] Unless specifically defined herein, the use of various commercially available products herein employs standard techniques. For example, they may be implemented using the manufacturer's instructions for use, or in accordance with methods known in the art or the description of this invention. The techniques and methods described herein can generally be implemented according to conventional methods well known in the art, based on the descriptions in the various general and more specific documents cited and discussed in this specification.

[0019] The terms “optional / arbitrary” or “optionally / arbitrarily” mean that the event or situation described below may or may not occur, including both the occurrence and non-occurrence of the event or situation.

[0020] The term "Web of Science" as used in this article refers to a citation indexing database developed by Clarivate Analytics.

[0021] The term "Google Scholar" used in this article refers to Google Scholar.

[0022] The term “TF-IDF” used in this article refers to a method of converting text into numerical vectors to measure the importance of a word in a document.

[0023] The term "OpenAI File Interface" used in this article refers to the API used for uploading and managing local files. The core of it is to specify the purpose, with different purposes corresponding to different formats and size limits.

[0024] The term "SCAMPER" used in this article refers to the SCAMPER creative thinking method.

[0025] The term "TRIZ" used in this article refers to the TRIZ innovative thinking method.

[0026] In a first aspect, the present invention provides: an inference module for an AI development system for a new technology product, comprising: The vector database routing module is used to automatically select the target vector database most relevant to the user's input question when multiple vector databases exist. Specifically, it includes: Based on the comprehensive score Scori, the vector database with the highest Scori score is selected as the target vector database: Scorei = Simi + Bonusi Simi represents the basic score for semantic relevance, and Bonusi represents the bonus score for keyword hit. Vector databases are used to store structured literature from various fields; The controlled reasoning execution module is used to execute the controlled reasoning phase. In the controlled reasoning phase, the large language model accesses the vector database for retrieval and integration, and at the same time accesses public domain knowledge. The integration results of the vector database and the integration of public domain knowledge are combined and fused to obtain the reasoning report.

[0027] First preferred option: The semantic relevance baseline score Simi is obtained in the following way: The text set is mapped to the feature vector space by using the TF-IDF text vectorization method to combine the question text with the summary of all vector database entries. Extract the TF-IDF vector corresponding to the question text; Retrieve the TF-IDF vector corresponding to the summary of each vector database entry; Calculate the cosine similarity between the TF-IDF vector corresponding to the question text and the TF-IDF vector corresponding to the summary of each vector database entry, and use it as the Simi.

[0028] Second preferred option: Bonusi keyword hit bonus points are obtained in the following ways: For each entry in the vector database, determine whether the question text matches any keywords in the entry's pre-defined keywords list; if it does, add a pre-defined bonus value, Bonusi, to the overall score of the entry. Bonusi = 0.15 × Hiti Hiti represents the number of keywords that were hit.

[0029] Third preferred option: The maximum value of Bonusi is limited to 1. When the Bonusi calculated based on the number of keyword hits is less than 1, its actual calculated value is used; when the calculated result is greater than or equal to 1, the Bonusi value is limited to 1.

[0030] Fourth preferred option: The reasoning module also includes: The inference template setting module is used to set the inference rules and the output format of the inference results; in the controlled inference execution module, the controlled inference stage is executed, inference is performed according to the inference rules, and the inference results are output according to the output format.

[0031] Fifth preferred option: The reasoning rules include Prompt and innovation analysis and novelty constraints; The Prompt includes: 1) Role and Ability Constraints: Clearly define the role and access permissions of the reasoning module; 2) Task objective definition: It is explicitly stated that it cannot be a literature review, and it is mandatory to integrate, reason, and generate new documents; 3) Citation and Evidence Rules: Mandate evidence grading to prevent the model from disguising "inferences" as "documentary facts"; 4) Forced partitioning of output structure: The output reasoning results need to be partitioned into three parts: content in the vector database, domain knowledge, and synthesis and inference. Innovation analysis and novelty constraints include: In addition to answering the questions, you will also need to provide the following: 1) At least three novel approaches that can substantially change the way problems are solved; 2) Minimum feasible experiment; 3) An alternative system architecture that differs significantly from the existing paradigm; 4) Provide a labeled diagram; 5) Creative driving factors; 6) Conflict between novelty and prior art.

[0032] Secondly, this invention provides: an experimental feedback module for a new technology product AI development system, comprising: The association submodule is used to contextually associate reasoning questions with lab reports; The Real-Time Inference Feedback submodule is used to execute the real-time inference feedback process; When the association submodule associates the reasoning problem with the experimental report in context, the experimental report participates in the reasoning process of the large language model as supplementary information. The experimental report is only valid for the associated reasoning problem and is not required to maintain global relevance for subsequent different reasoning problem scenarios in the system.

[0033] First preferred option: During the execution of real-time reasoning feedback, the real-time reasoning feedback submodule uses the reasoning problem and the associated experimental report as a unified reasoning input: the current reasoning problem is used as the core input, and the experimental report file identifier is used as an auxiliary input. These are encapsulated together into a unified reasoning input structure, and preset reasoning rules are associated in this input structure, so that the model reasoning process is constrained by control instructions.

[0034] Thirdly, this invention provides: a new technology product AI development system, comprising: The database module is used to perform structured parsing of collected PDF documents and generated experimental reports and convert them into JSON files to form a vector database. The reasoning module is used to parse the problem based on a pre-designed Prompt, and under the constraints of the Prompt, access the vector database and combine it with public domain knowledge to perform retrieval and integration, and output a reasoning report. The experiment feedback module is used to feed back the experimental results to the AI ​​development system for new technology products under controlled conditions, based on the experiment report or experimental results, and to participate in subsequent reasoning and knowledge accumulation. The reasoning module adopts the reasoning module described above, and the experimental feedback module adopts the experimental feedback module described above.

[0035] Fourthly, the present invention provides: a method for generating inference reports for an AI development system for a new technology product, comprising: Step s1: Select and collect literature in a specific technical field, extract JSON format data from the literature, form a JSON file, and upload the JSON file containing the JSON format data to the vector database; Step s2: Set the reasoning rules; Step s3: Based on the reasoning problem and the experimental report, construct a unified reasoning input structure; Step s4: Input the unified inference input structure into the large language model and judge the target vector database; Step s5: Based on the reasoning rules, within the scope of the target vector database and in conjunction with publicly available domain knowledge, execute the reasoning process and output a reasoning report.

[0036] First preferred option: Step s1 involves extracting JSON format data from the literature to form a JSON file, specifically including: The selected and collected documents are downloaded and stored in a pre-defined document catalog, with the file name serving as the unique identifier for each document. Define the structured extraction rules and JSON template; Upload all PDF documents in the bibliography to the large language model and set the large language model calling strategy; The uploaded PDF documents are parsed using a large language model to generate JSON files with JSON structured data.

[0037] Second preferred option: Step s3 specifically includes: Step s31: Contextualize the questions and experimental reports; Step s32: Construct a unified inference input structure: First, initialize the inference service by calling the client; Then, the lab report attachments are used as temporary input resources, uploaded to the file management component, and the file identifier corresponding to the lab report attachments is obtained; Then, configure the inference toolset, including vector retrieval tools and attachment parsing tools; Next, a unified reasoning input is constructed. The current reasoning problem is taken as the core input, and the experimental report file identifier is taken as an auxiliary input. They are encapsulated together into a unified reasoning input structure, and the preset reasoning rules are associated in this input structure so that the model reasoning process is constrained by control instructions.

[0038] Third preferred option: The unified inference input structure includes at least: Reasoning question text; Lab report attachment reference (File ID); Inference control command (Prompt or its identifier); Search tool and attachment parsing tool configuration parameters.

[0039] Fourth preferred option: Step s4 specifically includes: Step s41: Load the registry of the vector database; Step s42, Query preprocessing: Perform normalization on the query request text; perform the same normalization on the topic description text in each vector database entry; Step s43: Calculate the semantic relevance base score; Step s44: Calculate keyword hit bonus points; Step s45: Calculate and rank the overall scores; The vector database corresponding to the highest score among all comprehensive scores is selected as the target vector database.

[0040] Fifth preferred option: Step s5 specifically includes: Step s51: Inference call initialization; Step s52, Controlled reasoning execution: Step s521: Semantic retrieval and result filtering; Step s522, multi-step logical reasoning; Step s53: Generate structured results.

[0041] Fifthly, the present invention provides: an Internet of Things platform, including a display screen and the aforementioned new technology product AI development system, wherein after the new technology product AI development system outputs a reasoning report related to the development of the new technology product, the reasoning report is displayed on the display screen.

[0042] Sixthly, the present invention provides: a product performance testing device, including a user interface and the aforementioned new technology product AI development system. The user interface is used to input reasoning questions, experimental data, experimental reports, etc., and based on the reasoning questions and experimental data or experimental reports, the new technology product AI development system generates a next-step test plan or product improvement plan for product performance testing.

[0043] The present invention has at least the following beneficial effects: Compared with existing technologies, the present invention has better technical effects in terms of the logic, stability and consistency of reasoning results.

[0044] This invention scores multiple vector databases based on both text similarity and keyword hit rate, selecting the highest-scoring database as the target database. This avoids the irrelevant information retrieval problems caused by using a single vector database, reduces retrieval noise, and improves the relevance of search results, thereby enhancing the overall retrieval efficiency and accuracy of the analysis results. Furthermore, experimental verification shows that this invention, utilizing a JSON structured data entry scheme, maintains high retrieval stability and inference consistency even with large-scale document processing due to its field-based, hierarchical organization and unified processing of multimodal information text, demonstrating significant comprehensive advantages.

[0045] Compared with existing technologies, the present invention has better technical effects in terms of the validity, logic and stability of reasoning results.

[0046] This invention links experimental reports with reasoning problems and inputs them into the model, ensuring a correlation between the two. Experimental verification shows that the final reasoning results have a more convergent range of values ​​and are more accurate. Furthermore, it provides effective reasoning results and offers valuable suggestions for future development. Attached Figure Description

[0047] Figure 1 This is a structural diagram of the reasoning module of the present invention; Figure 2 This is a structural diagram of the AI ​​development system for the new technology product of this invention; Figure 3 This is a flowchart of the reasoning report generation method of the present invention. Detailed Implementation

[0048] The following non-limiting embodiments are intended to enable those skilled in the art to gain a more comprehensive understanding of the present invention, but do not limit the invention in any way. The following content is merely an exemplary description of the scope of protection claimed by the present invention, and those skilled in the art can make various changes and modifications to the present invention based on the disclosed content, and such changes should also fall within the scope of protection claimed by the present invention.

[0049] The present invention will be further described below by way of specific embodiments. Unless otherwise specified, all instruments, devices, equipment and other items used in the embodiments of the present invention are obtained through conventional commercial means.

[0050] Example 1 This embodiment provides an inference module for an AI development system for a new technology product. This module analyzes the problem based on a pre-designed Prompt, accesses a vector database under the constraints of the Prompt, and retrieves relevant content. Simultaneously, it combines publicly available domain knowledge to complete any missing information in the vector database, generating a final output report through comprehensive reasoning. Specifically, as shown... Figure 1 As shown, the inference module includes: M1, Vector Database Routing Module The vector database routing module is used to automatically select the target vector database most relevant to the question based on the user's input when multiple vector databases exist, for subsequent retrieval and reasoning.

[0051] Specifically, methods for automatically selecting the target vector database most relevant to the problem include: Step 1: Load the registry of the vector database; read the registry file and parse it into a list of entries. Optionally, perform field validation on the list of entries. If vector_store_id or summary is missing, mark it as an invalid entry and skip it.

[0052] The registry for the vector database is stored as a JSON file, containing key information for each vector database, specifically including: name: Library name; vector_store_id: A unique identifier for the library; Summary: Library topic description text; keywords: a list of keywords.

[0053] The summary is used to express the research topics and typical issues covered by the vector database; the keywords are used to enhance the ability to hit key terms.

[0054] Step 2: Query preprocessing; Perform normalization on the query request text (the text content of the question), including removing leading and trailing whitespace characters and unifying it to lowercase; Perform the same normalization on the topic description text in each vector database entry, and ensure the consistency between the question text and the text representation in the vector database.

[0055] Step 3: Calculate the semantic relevance base score; The preprocessed question text (query request text) and the topic description texts of all vector database entries are combined to form a text set. For each vector database entry i, the basic relevance score Simi between the question text q and the summary of vector database entry i is calculated.

[0056] The relevance score Simi can be obtained by combining "text representation + similarity calculation": The question text q and the summaries of all vector database entries are mapped to the feature vector space using the TF-IDF text vectorization method; where the TF-IDF text vectorization method is a method of turning text into numerical vectors, used to measure the importance of a word to a document.

[0057] Then, the TF-IDF vector corresponding to the question text and the TF-IDF vector corresponding to the topic description text of each vector database entry are taken respectively, and the cosine similarity is calculated as Simi; if the inference model contains n vector databases, then a question text will eventually obtain n Simi.

[0058] Step 4: Calculate keyword hit bonus points; For each vector database entry i, calculate the keyword hit reward score Bonusi. The specific method is as follows: For each entry in the vector database, determine whether the question text matches any keywords in the entry's pre-defined keywords list; if it does, add a pre-defined bonus value, Bonusi, to the overall score of the entry.

[0059] Bonusi = 0.15 × Hiti Hiti represents the number of keywords that were hit.

[0060] In addition, the maximum value of Bonusi is limited to 1. When the Bonusi calculated based on the number of keyword hits is less than 1, its actual calculated value is used; when the calculated result is greater than or equal to 1, the Bonusi value is limited to 1.

[0061] Step 5: Calculate the overall score and rank them; For each vector database entry i, calculate the comprehensive score Scores i, i.e.: Scorei = Simi + Bonusi.

[0062] Then, the vector database with the highest Scori score among all vector databases was selected as the target vector database.

[0063] The vector database routing module is located before vector retrieval. Its purpose is to avoid the problems of increased retrieval noise and reduced relevance caused by using a single vector database, thereby improving the overall retrieval efficiency and accuracy of the analysis results.

[0064] M2, Vector Database The process of creating a vector database includes: The search results, including target documents, references, and cited documents, are manually filtered to form a literature database for the relevant research field. All documents in the database are downloaded as PDFs. The acquired PDF documents are then stored in a pre-defined bibliographic directory, with the filename serving as a unique identifier for each document.

[0065] Pre-defined system prompt rules (System Prompt) are used to guide the large language model in parsing documents. Additionally, the system prompt rules (System Prompt) specify chapter types and abstract strategies. 1) For non-result sections or non-conclusion sections, such as abstract, introduction, methods, etc., only generate a summary and set detail_summary to an empty string; 2) For result-type or conclusion-type sections, such as the results and conclusion sections, only generate detail_summary and set summary to an empty string; 3) Pure reference lists and chapters that only list reference entries are ignored and not included in the sections array.

[0066] Iterate through all PDF files in the bibliography and perform the following operations on each PDF file: 1) Open the file and upload it to the large language model via the OpenAI file interface to obtain the corresponding file identifier (file_id). Each file has a unique file identifier.

[0067] 2) Construct a user prompt for this PDF file, including at least: The model indicates that "a PDF document has been attached as input"; Specifies the id used in the JSON for the document, where id = the PDF filename without the file extension.

[0068] 3) Call the Responses interface of the large language model: The input field contains both System Prompt and user prompt, as well as file_id. Set an appropriate max_output_tokens to limit the length of the output JSON and avoid exceeding the model limit; usually, max_output_tokens is set to 20,000 tokens.

[0069] On the client side, set a limited retry strategy for network exceptions such as APIConnectionError.

[0070] The input PDF file is parsed as a whole, and the following document information in the PDF is comprehensively utilized: Main text content; Chapter titles, hierarchical structure, and page number information; Image content and its figure caption; Table sections, headers, and cell text.

[0071] The parsed JSON objects are written to a predefined output directory as .json files, following the original PDF filenames. If necessary, the generated JSON files can be further imported into a vector database.

[0072] M3, Reasoning Template Setting Module The inference template setting module is used to set the inference rules and the output format of the inference results, including: M31, Prompt settings module The Prompt setting module is used to set and store the Prompt of the inference module. The Prompt includes the following: M311, Role and Ability Constraints Role and capability constraints are used to clarify the identity and access permissions of the reasoning module, such as the prompt text: You are an expert-level research collaborator in the field of materials science. You have access to a vector database of uploaded papers.

[0073] M312, Task Objective Definition The task objective is defined to clarify that it is not a "literature review" and mandates integration, reasoning, and the generation of new documents. For example, the prompt might state: Integrate relevant literature and common knowledge, answer the following questions, and explain your reasoning process, rather than simply summarizing the literature.

[0074] M313, Rules of Citation and Evidence Citation and evidence rules are used to enforce evidence grading, preventing models from disguising "inferences" as "documentary facts" and providing auditability for patents, papers, technical reports, etc. For example, the original promotion states: Citation rules cannot be changed; citations are only allowed in sections A and B; fabricated references are prohibited in section C. Here, A refers to content in the vector database, B refers to the publicly available domain knowledge, and C refers to the comprehensive inference.

[0075] M314, Forced Partitioning of Output Structure The forced output structure partitioning is used to force the inference results to be partitioned according to the following criteria: A. Contents of the vector database This section can only use already uploaded papers and is only allowed to include: facts, experimental results, formulas, methods, and figures, and must be labeled with [1][2].

[0076] B. Public Domain Knowledge This part of the content does not belong to the "acknowledged knowledge" in the vector database. It must come from an authoritative source and have a valid URL. Use [a][b] to distinguish it from part A (the content in the vector database).

[0077] C. Synthesis and Inference This section only allows: logical reasoning based on Part A and Part B, which can only be new hypotheses, new designs, or new experiments, and must be rigorously cited, with this part clearly defined as "model inference".

[0078] M315, Writing and Expression Constraints Writing and expression constraints are used to standardize the style of academic or technical documents and prevent models from being "fabricated for the sake of appearance." For example, the original prompt text should avoid colloquial language, use chapter headings and numbered lists, retain formulas, avoid forced citations, and strictly control the placement of citations.

[0079] M316, Data Gap and Planning Statement The data gap and plan declaration is used to force the model to admit "I don't know" and provide a next experimental design when the requested content requires new data and a suitable final inference result cannot be generated. For example, the prompt text reads: If the requested content requires new data, please specify exactly what data is missing and provide a data collection plan.

[0080] M32, Innovation Analysis and Novelty Constraint Submodule Provided that the requirements of basic analysis and evidence reasoning are met, the reasoning control instruction template may further include the following innovation analysis and novelty constraint instructions, the contents of which are as follows: In addition to answering the questions, you also need to provide the following: 1) At least three novel approaches that can substantially change the way problems are solved. For each approach, please include: Core mechanism (in short); Why it might be superior to existing methods (mechanisms and principles related to materials chemistry / physics); Prerequisites include the materials, instruments, and data used; Key adjustable parameters and their reasonable ranges based on known orders of magnitude; Predicted trade-offs and failure modes.

[0081] 2) Minimum feasible experiment (MVE), and the minimum feasible experiment must include: specific measurement plan, sample preparation, indicators and success criteria.

[0082] 3) An alternative system architecture that differs significantly from existing paradigms (e.g., different adsorption pathways, layered materials, processing routes, flow configurations, sensor-in-the-loop controllers, or data-driven active design loops).

[0083] 4) Provide a labeled block diagram (text description is also acceptable), interface definitions, and an end-to-end test plan.

[0084] 5) Creative Drivers: Apply at least two structured ideation tools and explain how they guide the development of the solution, for example: Morphology Matrix: Lists orthogonal design dimensions and options.

[0085] SCAMPER or TRIZ: Describes which operators or principles were used and where they were used.

[0086] If cross-domain analogies reveal overlooked mechanisms, these analogies may be included selectively (with clear annotation).

[0087] 6) Conflict between novelty and prior art: For each novel method and alternative system, please include: Previous work that is closest to the content in the vector database A or the knowledge in domain B, and the differences between the schemes at the mechanism or system level.

[0088] Assume novelty scores (1-5) and risk scores (1-5), and explain each in one sentence. The novelty score and risk score are automatically determined by the large language model. The novelty score indicates the degree of novelty of the given solution, and the risk score indicates the degree of risk of the given solution.

[0089] Rigor and evaluation: Experimental design diagrams are required: tables that link each design variable to measurable outputs, instruments, sampling plan, confounding factors, and statistical tests.

[0090] Provide ablation strategies: remove or change which assumptions to falsify the assumption.

[0091] Provide a decision table to compare the baselines in the literature with each new method in terms of performance, cost, scalability, robustness, and security (qualitative comparisons are acceptable, but must be well-founded).

[0092] Clearly state all assumptions. If the evidence is weak or missing, mark it as "assumption".

[0093] The innovation analysis and novelty constraint submodule is used to impose structured constraints on innovation generation behavior during the reasoning process, avoid unfounded free play, and ensure that the generated novel solutions have clear mechanism hypotheses, experimental feasibility, and control evaluation basis.

[0094] M4, Controlled Inference Execution Module The controlled inference execution module is used to perform inference based on the vector database and the vector database routing module, based on the proposed question. After obtaining the target vector database identifier, the question structure, and the corresponding inference control instructions (Prompt), the system enters the controlled inference execution phase. This phase is completed by the controlled inference execution module, and its specific process includes the following steps: X1, Inference call initialization; By calling the API, the problem description, the inference rules set in the inference template setting module, and the retrieval interface parameters of the target vector database are passed to the large language model inference engine to initialize the controlled inference process.

[0095] X2, Controlled reasoning execution; Under the constraints of the inference rules, the large language model performs inference tasks according to the preset inference process, rather than generating free text. The inference process includes searching the target vector database, combining public domain knowledge, evidence integration, and logical inference.

[0096] X21. Semantic retrieval and result filtering; Semantic retrieval is performed based on the question text and target vector database, and the relevance of the retrieval results is evaluated and filtered (the large language model evaluates and filters automatically) to obtain a set of candidate information that is highly relevant to the current question.

[0097] X22, Multi-step logical reasoning; Under the constraints of vector database and domain knowledge, multi-step logical reasoning is performed on the filtered candidate information set to form intermediate reasoning results consistent with the research objectives.

[0098] X3. Generation of structured results; The reasoning results are organized and generated according to a predefined output format, and output experimental design suggestions, material development conclusions or other structured reasoning results.

[0099] By implementing the above-mentioned controlled reasoning execution mechanism, irrelevant information is prevented from entering the reasoning process, thereby improving the interpretability, consistency, and traceability of the reasoning results.

[0100] Example 2 This embodiment provides an experimental feedback module for a new technology product AI development system. Based on the experimental results, the module feeds back the experimental results to the new technology product AI development system under controlled conditions and participates in subsequent reasoning and knowledge accumulation.

[0101] To ensure that experimental data can be effectively used in subsequent reasoning processes and support iterative analysis based on experimental results, the experimental reports received by the experimental feedback module must meet the following requirements: 1) Basic requirements for the composition of a lab report Experiment reports should include at least a description of the experimental subjects, a description of the experimental conditions, experimental results data, and an explanation of the experimental conclusions or observations. Each part should have a clear logical distinction and hierarchical structure to support subsequent processing and citation.

[0102] 2) Requirements for describing experimental subjects and conditions The experimental report should clearly describe the experimental subjects and their corresponding experimental conditions so that the experimental results can be understood as results obtained under specific conditions.

[0103] The experimental conditions include, but are not limited to, the experimental environment, the setting of key variables and the control relationship, but an exhaustive description of all potential influencing factors is not required.

[0104] 3) Requirements for the presentation and structuring of experimental data Experimental results in the lab report should be presented in an extractable and structured format, including quantitative data, parameter lists, or explicit trend descriptions.

[0105] Experimental data should have clear variable definitions and unit descriptions to avoid ambiguity and support subsequent inference input and vectorized storage.

[0106] 4) Requirements for the correspondence between experimental results and conclusions The experimental conclusions, phenomenon descriptions, or empirical judgments given in the experimental report should be able to establish a clear correspondence with the experimental data, avoiding the situation where only conclusive descriptions are given without traceable data evidence.

[0107] 5) Usability requirements for inference input The key information in the experimental report should be usable as supplementary input to the large language model along with the research question, including clear variable meanings, expression of experimental results, and directional descriptions of conclusions, to support the controlled inference process.

[0108] Then, the experimental feedback steps are executed, specifically including: 1) Contextualize the questions with the lab report; During the experimental feedback process, researchers, based on their understanding of the research objectives and the problem content, predefine the correlation between the experimental data and the current reasoning problem, and use this correlation as a constraint on the experimental data feedback. This predefining involves, after the researchers are familiar with the experimental report content, posing reasoning questions related to the experimental data; thus, the correlation between the experimental data and the current reasoning problem is predefined.

[0109] The aforementioned correlation is used to limit the applicability of experimental data in the current reasoning process, ensuring that the experimental data participates in the reasoning process of the large language model only as supplementary information within the context in which the researcher deems it relevant to the current reasoning problem. This correlation holds only for the currently proposed reasoning problem and does not require the experimental data to maintain global relevance to subsequent different reasoning problem scenarios in the system. That is, the correlation between the experimental data and the currently proposed reasoning problem is exclusively defined.

[0110] By employing the above methods, it can be ensured that the experimental data passed to the large language model as supplementary input maintains consistency with the current reasoning problem in terms of research variables, experimental conditions, or target evaluation indicators. This avoids irrelevant experimental data from interfering with the reasoning process and improves the relevance and interpretability of the reasoning results in specific problem scenarios.

[0111] 2) Perform the immediate reasoning feedback process First, the inference service call client is initialized. The system initializes the call client within the inference module, configuring the model identifier, inference rule identifier (Prompt ID), vector database identifier (Vector Store ID), and multi-turn session identifier (Conversation ID). It also sets call timeout thresholds and retry strategies to avoid multiple inference outputs caused by repeated calls.

[0112] Then, the experimental report attachment is used as a temporary input resource; the prepared experimental report is uploaded to the file management component as a file (e.g., a PDF document) as an auxiliary input, and a file identifier (File ID) corresponding to the file is obtained. The file identifier is used to reference the experimental report file in subsequent model calls; the attachment is a one-time inference context input and does not need to be written to the vector database.

[0113] Then, configure the inference toolset (retrieval and attachment parsing in tandem); enable the following tool capabilities in inference calls: Vector retrieval tool: Specifies the vector database identifier, enabling the reasoning module to retrieve relevant literature or knowledge fragments from the vector database based on the current reasoning problem, which is used to provide historical background information and evidence support for the reasoning process; Attachment parsing tool: Enables file parsing and computation environment, allowing the large language model to read the contents of the experimental report attachments and extract and summarize the text, tables, or key data therein.

[0114] By combining the above tools, the AI ​​development system for new technology products with an experimental feedback module can simultaneously possess "historical knowledge retrieval capability" and "latest experimental attachment parsing capability".

[0115] Next, a unified inference input is constructed. The current inference problem is used as the core input, while the experimental report file identifier is used as a secondary input. These are encapsulated together into a unified inference input structure, and preset inference rules are associated with this structure, thus constraining the model's inference process with control commands. The unified inference input structure includes at least: Reasoning question text; Lab report attachment reference (File ID); Inference control command (Prompt or its identifier); Search tool and attachment parsing tool configuration parameters.

[0116] Next, the model is invoked and controlled inference is performed; the unified inference input structure is passed to the model via an interface call to execute the controlled inference process. During the inference process: The model triggers a vector retrieval tool based on the reasoning question, identifies the target vector database, and retrieves historical documents or knowledge fragments related to the question from the target vector database. The model simultaneously reads the contents of the experimental report attachments and extracts and summarizes the key experimental conditions, structural data, and trends of change. Under the constraints of the inference rules, the model comprehensively analyzes the "retrieved historical evidence" and the "latest evidence in the experimental report" to form inference conclusions for the current problem and generate suggestions for improving the experiment.

[0117] Then, the inference output can be obtained and generated as a report; the output text is extracted from the model call results as the inference output content. The system generates an archiveable document (e.g., a Word report) based on a predetermined report template. The report includes at least: a report title, generation time, and the main body of the inference output; and the model output is written into the document in paragraph or item format to form an inference report that can be reviewed, referenced, and used for iterative experimental design.

[0118] In the process of developing new product designs, multiple reasoning processes are required. After the experimental data is input as an auxiliary file and a reasoning report is generated, the experimental report can be structured into a JSON format file. The structured experimental data is then uploaded to a vector database to form searchable experimental knowledge entries, which can support semantic retrieval, evidence citation, and historical comparison analysis for subsequent reasoning tasks.

[0119] Based on the correspondence between the inference output and the experimental results (experimental data), the subsequent experimental direction or new problem description is generated, and this information is fed back to the inference module to trigger the next round of problem construction and controlled inference execution, thus forming an iterative closed loop of inference-experiment-re-inference.

[0120] Effect verification experiment To verify that in this embodiment, experimental data must be input together with the research question and maintain a high degree of relevance in order to obtain reasoning results that meet actual needs, the following comparative verification experiment was conducted.

[0121] The research question selected for the controlled verification experiment is: We are developing a carbon block with an ultra-high trichloromethane removal life (a microporous adsorption membrane material formed by sintering high molecular weight PE binder and powdered activated carbon PAC at about 200 degrees Celsius). What characteristics of the powdered activated carbon PAC are required? What are the optimal values ​​for ultramicropore / total pore volume, ultramicropore / micropore, and mesopore / total pore volume? While maintaining consistency in the formulation of the research question, various input scenarios are constructed, including: Experiment 1: No experimental data is provided; Experiment 2: Provide experimental data that does not directly correspond to the problem; Experiment 3: Provide experimental data on powdered activated carbon that is highly relevant to the problem, and pass it along with the research question as input to the large language model; Each experimental condition was repeated three times.

[0122] The following are some of the reasoning results for Experiment 1: Executive Summary (≤150 words) For polyethylene sintered activated carbon blocks designed to achieve ultra-high trihalomethane (chloroform) removal rates, activated carbon with extremely high microporosity and a high proportion of narrow micropores, combined with hydrophobic surface properties and low ash content, should be selected. Activated carbon with predominantly narrow micropores should be used, but a moderate amount of mesopores should be retained to ensure pore accessibility after polyethylene impregnation. The initial design range is: (narrow micropores / V) ≈ 0.45–0.70, (narrow micropores / V) ≈ 0.70–0.90, (mesopores / V) ≈ 0.10–0.25 (to be optimized through experimental design). Hydraulically, bypass defects (cracks) should be eliminated, and a hole transport time (EBCT) ≥ 10 s should be used as a positive benchmark. Radial flow geometry should be used whenever possible. A median through-pore diameter of ~1–5 μm should be set, and porosity should be determined through volumetric / skeletal density, porosity determination / mercury intrusion porosimetry, and permeability testing.

[0123] For Experiment 2, no direct answer was given, but some of the reasoning results are as follows: C. Your synthesis and reasoning (novelty) 1. Novelty - Using new PDF to set PAC selection principles: Since the block lifetime ranking follows the PAC static capacity ranking under fixed EBCT / sintering conditions [1], the PAC static capacity is used as the main screening KPI for rod development; the rod design is mainly aimed at maximizing the use of micropores (rather than “creating new adsorption sites”).

[0124] 2. Innovation – Designing a rod-shaped macroporous network for utilization rather than storage: Your desorption validation shows that the interparticle voids are not additional chloroform reservoirs [1]. Therefore, the rod-shaped through-pore diameter and porosity should be set to (i) prevent channeling / bypassing effects and (ii) minimize ΔP while maintaining EBCT homogeneity.

[0125] 3. Novel – Mesopore “just right” rule: You observe that mesopores are associated with faster desorption behavior (peak shift to the left)[1], which means that mesopores help with mass transfer; add a controllable proportion of mesopores to PAC, but limit their size to avoid sacrificing the narrow micropore volume that sets the equilibrium capacity.

[0126] 4. Innovation – How to obtain the “optimal” V ratio through experiments: Construct a response surface model, with input being ((V narrow micro / V,;V narrow micro / V micro ,;V meso / V,;ε open ,;k)), the output is ((q b The optimal solution is the Pareto front, not a single ratio, where ΔP is the volume of the treatment at 95% removal rate.

[0127] 5. Innovation – Practical Quality Assurance / Quality Control Checkpoints for Mass Production: Each production batch should pass (i) bubble point integrity test (rejecting low bubble point = potential for large defects / bypasses), (ii) penetration window, and (iii) short-term EBCT chloroform challenge test with 95% removal threshold.

[0128] The following are some of the reasoning results for Experiment 3: C. Your comprehensive analysis and reasoning (novel; based on the experimental data you provided + A / B) Note: Each of the following points is derived based on the PACA / B / C data and conclusions you provided in the message. Since the PDF could not be retrieved and cited in the provided library, all of them are marked as novel.

[0129] C1) Deduce the pore structure target of "high school grade three foam landfill PAC" from your data. The given chloroform capacities are: PAC A 1.91 mg / g > PAC B 1.28 mg / g > PAC C 0.92 mg / g. The "≤1 nm cumulative pore capacity" is also in the same order, which is equivalent to the following: the chloroform capacity is mainly controlled by the "≤1 nm narrow volume," rather than by the simple SSA or total micropore capacity.

[0130] The structural ratios calculated using your Table 1 (based on your given Vtot / Vmic and cumulative pore volumes of ≤1nm and 2–50nm): oPAC A: Narrowband (≤1nm) / V ≈ 0.695; Narrowband / Midband ≈ 0.811; Midband (2–50) / V ≈ 0.0042 oPAC B: Vnarrow / V ≈ 0.605; Vnarrow / Vmic ≈ 0.738; Vmeso / V ≈ 0.0188 oPAC C: Vnarrow / V ≈ 0.595; Vnarrow / Vmic ≈ 0.621; Vmeso / V ≈ 0.0617 Therefore, a PAC that prioritizes capacity should at least satisfy the following: oV narrow range / V≥0.65 (preferably 0.68–0.72) oVnarrow micro / Vmicro ≥ 0.78 (preferably 0.80–0.85) oVmeso / V should be as low as possible (<0.02), but considering that you have also observed that "mesopores still allow PAC B / C to reach equilibrium faster", if the carbon rod needs to make full use of micropores under short EBCT, at least 0.01–0.03 "small mesopores" should be used as mass transfer pores (capacity-rate trade-off).

[0131] The reasoning results under the combined conditions of Experiment 1, Experiment 2, and Experiment 3 are shown in Table 1: Table 1

[0132] Experimental results show that when no experimental data is provided or only data that does not directly correspond to the research question is provided, the parameter range given by the large language model is relatively scattered, and in some cases, no clear conclusion can be drawn, or the obtained parameter range deviates significantly from the actual carbon rod process requirements.

[0133] In contrast, when experimental data of powdered activated carbon, which is highly relevant to the research question, is used as an auxiliary input and passed to the large language model for controlled inference along with the question, the range of values ​​for ultramicropore / total pore volume, ultramicropore / micropore, and mesopore / total pore volume output by the model is significantly narrowed. Moreover, actual verification shows that the relevant structural parameters of powdered activated carbon with high trichloromethane removal lifetime all fall within the above parameter range.

[0134] When experimental data that is not directly related to the research question is provided to the large language model, the inference results not only fail to converge, but also deviate further from the actual engineering requirements in some parameter dimensions, and even fail to provide a direct answer.

[0135] The above comparison results show that the contextual relevance between experimental data and research questions not only determines the validity of the reasoning results, but also misleads the reasoning process when the input experimental data is irrelevant to the question, thus causing the reasoning results to deviate further from the actual engineering requirements compared to the case without experimental data input.

[0136] Example 3 This embodiment provides a new technology product AI development system, such as Figure 2 As shown, it includes a database module, an inference module, and an experimental feedback module.

[0137] The database module is used to perform structured parsing of collected PDF documents and generated experimental reports and convert them into JSON files. It refines and reorganizes the document content, effectively reducing the overall number of tokens while achieving a unified textual representation of multimodal information (text, charts, and images).

[0138] The reasoning module adopts the reasoning module provided in Embodiment 1. It analyzes the problem based on a pre-designed Prompt and accesses the vector database under the constraints of the Prompt to retrieve relevant content. At the same time, it combines domain knowledge to complete the information that may be missing in the vector database and generates the final output report through reasoning.

[0139] The experimental feedback module adopts the experimental feedback module provided in Embodiment 2. Based on the experimental results, it feeds back the experimental results to the AI ​​development system of new technology products under controlled conditions, and participates in subsequent reasoning and knowledge accumulation.

[0140] Example 4 This embodiment provides a method for generating inference reports for an AI development system for a new technology product, as described above. Figure 3 As shown, it includes: Step s1: Select and collect literature in a specific technical field, extract JSON format data from the literature, form a JSON file, and upload the JSON file containing the JSON format data to the vector database; Step s1 is implemented through the database module, and the specific implementation process includes: The search results, including target documents, references, and cited documents, are manually filtered to form a literature database for the relevant research field. All documents in the database are downloaded as PDFs. The acquired PDF documents are then stored in a pre-defined bibliographic directory, with the filename serving as a unique identifier for each document.

[0141] Pre-defined system prompt rules (System Prompt) are used to guide the large language model in parsing documents. Additionally, the system prompt rules (System Prompt) specify chapter types and abstract strategies. 1) For non-result sections or non-conclusion sections, such as abstract, introduction, methods, etc., only generate a summary and set detail_summary to an empty string; 2) For result-type or conclusion-type sections, such as the results and conclusion sections, only generate detail_summary and set summary to an empty string; 3) Pure reference lists and chapters that only list reference entries are ignored and not included in the sections array.

[0142] Iterate through all PDF files in the bibliography and perform the following operations on each PDF file: 1) Open the file and upload it to the large language model via the OpenAI file interface to obtain the corresponding file identifier (file_id). Each file has a unique file identifier.

[0143] 2) Construct a user prompt for this PDF file, including at least: The model indicates that "a PDF document has been attached as input"; Specifies the id used in the JSON for the document, where id = the PDF filename without the file extension.

[0144] 3) Call the Responses interface of the large language model: The input field contains both System Prompt and user prompt, as well as file_id. Set an appropriate max_output_tokens to limit the length of the output JSON and avoid exceeding the model limit; usually, max_output_tokens is set to 20,000 tokens.

[0145] On the client side, set a limited retry strategy for network exceptions such as APIConnectionError.

[0146] The input PDF file is parsed as a whole, and the following document information in the PDF is comprehensively utilized: Main text content; Chapter titles, hierarchical structure, and page number information; Image content and its figure caption; Table sections, headers, and cell text.

[0147] The parsed JSON objects are written to a predefined output directory as .json files, following the original PDF filenames. If necessary, the generated JSON files can be further imported into a vector database.

[0148] Step s2: Set the inference rules; this step is implemented through the inference template setting module in the inference module, including: Step s21, Prompt settings; Prompt includes the following: Role and Ability Constraints: Role and capability constraints are used to clarify the identity and access permissions of the reasoning module, such as the prompt text: You are an expert-level research collaborator in the field of materials science. You have access to a vector database of uploaded papers.

[0149] Task objective definition: The task objective is defined to clarify that it is not a "literature review" and mandates integration, reasoning, and the generation of new documents. For example, the prompt might state: Integrate relevant literature and common knowledge, answer the following questions, and explain your reasoning process, rather than simply summarizing the literature.

[0150] Citation and Evidence Rules: Citation and evidence rules are used to enforce evidence grading, preventing models from disguising "inferences" as "documentary facts" and providing auditability for patents, papers, technical reports, etc. For example, the original promotion states: Citation rules cannot be changed; citations are only allowed in sections A and B; fabricated references are prohibited in section C. Here, A refers to content in the vector database, B refers to the publicly available domain knowledge, and C refers to the comprehensive inference.

[0151] Forced partitioning of output structure: The forced output structure partitioning is used to force the inference results to be partitioned according to the following criteria: A. Contents of the vector database This section can only use already uploaded papers and is only allowed to include: facts, experimental results, formulas, methods, and figures, and must be labeled with [1][2].

[0152] B. Public Domain Knowledge This part of the content does not belong to the "acknowledged knowledge" in the vector database. It must come from an authoritative source and have a valid URL. Use [a][b] to distinguish it from part A (the content in the vector database).

[0153] C. Synthesis and Inference This section only allows: logical reasoning based on Part A and Part B, which can only be new hypotheses, new designs, or new experiments, and must be rigorously cited, with this part clearly defined as "model inference".

[0154] Writing and expression constraints: Writing and expression constraints are used to standardize the style of academic or technical documents and prevent models from being "fabricated for the sake of appearance." For example, the original prompt text should avoid colloquial language, use chapter headings and numbered lists, retain formulas, avoid forced citations, and strictly control the placement of citations.

[0155] Data Gap and Plan Statement: The data gap and plan declaration is used to force the model to admit "I don't know" and provide a next experimental design when the requested content requires new data and a suitable final inference result cannot be generated. For example, the prompt text reads: If the requested content requires new data, please specify exactly what data is missing and provide a data collection plan.

[0156] Step s22, Innovation Analysis and Novelty Constraints; Provided that the requirements of basic analysis and evidence reasoning are met, the reasoning control instruction template may further include the following innovation analysis and novelty constraint instructions, the contents of which are as follows: In addition to answering the questions, you also need to provide the following: 1) At least three novel approaches that can substantially change the way problems are solved. For each approach, please include: Core mechanism (in short); Why it might be superior to existing methods (mechanisms and principles related to materials chemistry / physics); Prerequisites include the materials, instruments, and data used; Key adjustable parameters and their reasonable ranges based on known orders of magnitude; Predicted trade-offs and failure modes.

[0157] 2) Minimum feasible experiment (MVE), and the minimum feasible experiment must include: specific measurement plan, sample preparation, indicators and success criteria.

[0158] 3) An alternative system architecture that differs significantly from existing paradigms (e.g., different adsorption pathways, layered materials, processing routes, flow configurations, sensor-in-the-loop controllers, or data-driven active design loops).

[0159] 4) Provide a labeled block diagram (text description is also acceptable), interface definitions, and an end-to-end test plan.

[0160] 5) Creative Drivers: Apply at least two structured ideation tools and explain how they guide the development of the solution, for example: Morphology Matrix: Lists orthogonal design dimensions and options.

[0161] SCAMPER or TRIZ: Describes which operators or principles were used and where they were used.

[0162] If cross-domain analogies reveal overlooked mechanisms, these analogies may be included selectively (with clear annotation).

[0163] 6) Conflict between novelty and prior art: For each novel method and alternative system, please include: Previous work that is closest to the content in the vector database A or the knowledge in domain B, and the differences between the schemes at the mechanism or system level.

[0164] Assume novelty scores (1-5) and risk scores (1-5), and explain each in one sentence. The novelty score and risk score are automatically determined by the large language model. The novelty score indicates the degree of novelty of the given solution, and the risk score indicates the degree of risk of the given solution.

[0165] Rigor and evaluation: Experimental design diagrams are required: tables that link each design variable to measurable outputs, instruments, sampling plan, confounding factors, and statistical tests.

[0166] Provide ablation strategies: remove or change which assumptions to falsify the assumption.

[0167] Provide a decision table to compare the baselines in the literature with each new method in terms of performance, cost, scalability, robustness, and security (qualitative comparisons are acceptable, but must be well-founded).

[0168] Clearly state all assumptions. If the evidence is weak or missing, mark it as "assumption".

[0169] Step s3: Based on the reasoning problem and the experiment report, construct a unified reasoning input structure; specifically including: Step s31: Contextualize the questions and experimental reports; During the experimental feedback process, researchers, based on their understanding of the research objectives and the problem content, predefine the correlation between the experimental data and the current reasoning problem, and use this correlation as a constraint on the experimental data feedback. This predefining involves, after the researchers are familiar with the experimental report content, posing reasoning questions related to the experimental data; thus, the correlation between the experimental data and the current reasoning problem is predefined.

[0170] The aforementioned correlation is used to limit the applicability of experimental data in the current reasoning process, ensuring that the experimental data participates in the reasoning process of the large language model only as supplementary information within the context in which the researcher deems it relevant to the current reasoning problem. This correlation holds only for the currently proposed reasoning problem and does not require the experimental data to maintain global relevance to subsequent different reasoning problem scenarios in the system. That is, the correlation between the experimental data and the currently proposed reasoning problem is exclusively defined.

[0171] By employing the above methods, it can be ensured that the experimental data passed to the large language model as supplementary input maintains consistency with the current reasoning problem in terms of research variables, experimental conditions, or target evaluation indicators. This avoids irrelevant experimental data from interfering with the reasoning process and improves the relevance and interpretability of the reasoning results in specific problem scenarios.

[0172] Step s32: Construct a unified inference input structure; First, the inference service call client is initialized. The system initializes the call client within the inference module, configuring the model identifier, inference rule identifier (Prompt ID), vector database identifier (Vector Store ID), and multi-turn session identifier (Conversation ID). It also sets call timeout thresholds and retry strategies to avoid multiple inference outputs caused by repeated calls.

[0173] Then, the experimental report attachment is used as a temporary input resource; the prepared experimental report is uploaded to the file management component as a file (e.g., a PDF document) as an auxiliary input, and a file identifier (File ID) corresponding to the file is obtained. The file identifier is used to reference the experimental report file in subsequent model calls; the attachment is a one-time inference context input and does not need to be written to the vector database.

[0174] Then, configure the inference toolset (retrieval and attachment parsing in tandem); enable the following tool capabilities in inference calls: Vector retrieval tool: Specifies the vector database identifier, enabling the reasoning module to retrieve relevant literature or knowledge fragments from the vector database based on the current reasoning problem, which is used to provide historical background information and evidence support for the reasoning process; Attachment parsing tool: Enables file parsing and computation environment, allowing the large language model to read the contents of the experimental report attachments and extract and summarize the text, tables, or key data therein.

[0175] By combining the above tools, the AI ​​development system for new technology products with an experimental feedback module can simultaneously possess "historical knowledge retrieval capability" and "latest experimental attachment parsing capability".

[0176] Next, a unified inference input is constructed. The current inference problem is used as the core input, while the experimental report file identifier is used as a secondary input. These are encapsulated together into a unified inference input structure, and preset inference rules are associated with this structure, thus constraining the model's inference process with control commands. The unified inference input structure includes at least: Reasoning question text; Lab report attachment reference (File ID); Inference control command (Prompt or its identifier); Search tool and attachment parsing tool configuration parameters.

[0177] Step s4: Input the unified inference input structure into the large language model to determine the target vector database; specifically including: Step s41: Load the registry of the vector database; read the registry file and parse it into a list of entries. Optionally, perform field validation on the list of entries. If vector_store_id or summary is missing, mark it as an invalid entry and skip it.

[0178] Step s42: Query preprocessing; Perform normalization on the query request text (the text content of the question), including removing leading and trailing whitespace characters and unifying it to lowercase; Perform the same normalization on the topic description text in each vector database entry, and ensure the consistency between the question text and the text representation in the vector database.

[0179] Step s43: Calculate the semantic relevance base score; The preprocessed question text (query request text) and the topic description texts of all vector database entries are combined to form a text set. For each vector database entry i, the basic relevance score Simi between the question text q and the summary of vector database entry i is calculated.

[0180] The relevance score Simi can be obtained by combining "text representation + similarity calculation": The question text q and the summaries of all vector database entries are mapped to the feature vector space using the TF-IDF text vectorization method; where the TF-IDF text vectorization method is a method of turning text into numerical vectors, used to measure the importance of a word to a document.

[0181] Then, the TF-IDF vector corresponding to the question text and the TF-IDF vector corresponding to the topic description text of each vector database entry are taken respectively, and the cosine similarity is calculated as Simi; if the inference model contains n vector databases, then a question text will eventually obtain n Simi.

[0182] Step s44: Calculate keyword hit bonus points; For each vector database entry i, calculate the keyword hit reward score Bonusi. The specific method is as follows: For each entry in the vector database, determine whether the question text matches any keywords in the entry's pre-defined keywords list; if it does, add a pre-defined bonus value, Bonusi, to the overall score of the entry.

[0183] Bonusi = 0.15 × Hiti Hiti represents the number of keywords that were hit.

[0184] The maximum value of Bonusi is limited to 1. When the Bonusi calculated based on the number of keyword hits is less than 1, its actual calculated value is used; when the calculated result is greater than or equal to 1, the Bonusi value is limited to 1.

[0185] Step s45: Calculate and rank the overall scores; For each vector database entry i, calculate the comprehensive score Scores i, i.e.: Scorei = Simi + Bonusi Then, the vector database with the highest Scori score among all vector databases was selected as the target vector database.

[0186] Step s5: Based on the inference rules, perform the inference process within the target vector database and output an inference report. This specifically includes: Step s51: Inference call initialization; By calling the API, the problem description, the inference rules set in the inference template setting module, and the retrieval interface parameters of the target vector database are passed to the large language model inference engine to initialize the controlled inference process.

[0187] Step s52, controlled inference execution; Under the constraints of the inference rules, the large language model performs inference tasks according to the preset inference process, rather than generating free text. The inference process includes retrieval of the target vector database, evidence integration, and logical inference.

[0188] Step s521: Semantic retrieval and result filtering; Semantic retrieval is performed based on the question text and target vector database. The retrieval results are then evaluated for relevance, filtered, and denoised to obtain a set of candidate information that is highly relevant to the current question.

[0189] Step s522, multi-step logical reasoning; Under the constraints of vector database and domain knowledge, multi-step logical reasoning is performed on the filtered candidate information set to form intermediate reasoning results consistent with the research objectives.

[0190] The model simultaneously reads the contents of the experimental report attachments and extracts and summarizes the key experimental conditions, structural data, and trends of change. Under the constraints of the inference rules, the model comprehensively analyzes the "retrieved historical evidence" and the "latest evidence in the experimental report" to form inference conclusions for the current problem and generate suggestions for improving the experiment.

[0191] Step s53: Generation of structured results; The inference results are organized and generated according to a predefined output format, outputting experimental design suggestions, material development conclusions, or other structured inference results. The inference report should include at least: report title, generation time, and the main body of the inference output; and the model output should be written into a document in paragraph or item format to form an inference report that can be reviewed, referenced, and used for iterative experimental design.

[0192] Example 5 This embodiment provides an Internet of Things (IoT) platform, including a display screen and the aforementioned new technology product AI development system. After the new technology product AI development system outputs a reasoning report related to the development of the new technology product, the reasoning report is displayed on the display screen.

[0193] Example 6 This embodiment provides a product performance testing device, including a user interface and the aforementioned new technology product AI development system. The user interface is used to input reasoning questions, experimental data, experimental reports, etc. Based on the reasoning questions and experimental data or experimental reports, the new technology product AI development system generates the next test plan or product improvement plan for product performance testing.

[0194] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention do not depart from the essence and scope of the technical solution of the present invention.

Claims

1. The inference module of a new technology product AI development system, characterized in that: include: The vector database routing module is used to automatically select the target vector database most relevant to the user's input question when multiple vector databases exist. Specifically, it includes: Based on the comprehensive score Scori, the vector database with the highest Scori score is selected as the target vector database: Scorei = Simi + Bonusi Simi represents the basic score for semantic relevance, and Bonusi represents the bonus score for keyword hit. Vector databases are used to store structured literature from various fields; The controlled reasoning execution module is used to execute the controlled reasoning phase. In the controlled reasoning phase, the large language model accesses the vector database for retrieval and integration, and at the same time accesses public domain knowledge. The integration results of the vector database and the integration of public domain knowledge are combined and fused to obtain the reasoning report.

2. The reasoning module according to claim 1, characterized in that, The semantic relevance baseline score Simi is obtained in the following way: The text set is mapped to the feature vector space by using the TF-IDF text vectorization method to combine the question text with the summary of all vector database entries. Extract the TF-IDF vector corresponding to the question text; Retrieve the TF-IDF vector corresponding to the summary of each vector database entry; Calculate the cosine similarity between the TF-IDF vector corresponding to the question text and the TF-IDF vector corresponding to the summary of each vector database entry, and use it as the Simi.

3. The reasoning module according to claim 1, characterized in that, Bonusi keyword hit bonus points are obtained in the following ways: For each entry in the vector database, determine whether the question text matches any keywords in the entry's pre-defined keywords list; if it does, add a pre-defined bonus value, Bonusi, to the entry's overall score. Bonusi = 0.15 × Hiti Hiti represents the number of keywords that were hit.

4. The reasoning module according to claim 1, characterized in that, When calculating Bonusi, the maximum value of Bonusi is set to 1: when the Bonusi calculated based on the number of keyword hits is less than 1, the actual calculated value is used; If the calculation result is greater than or equal to 1, then Bonusi will be set to 1.

5. The reasoning module according to claim 1, characterized in that, The reasoning module also includes: The inference template setting module is used to set the inference rules and the output format of the inference results; in the controlled inference execution module, the controlled inference stage is executed, inference is performed according to the inference rules, and the inference results are output according to the output format.

6. The reasoning module according to claim 1, characterized in that, The reasoning rules include Prompt and innovation analysis and novelty constraints; The Prompt includes: 1) Role and Ability Constraints: Clearly define the role and access permissions of the reasoning module; 2) Task objective definition: It is explicitly stated that it cannot be a literature review, and it is mandatory to integrate, reason, and generate new documents; 3) Citation and Evidence Rules: Mandate evidence grading to prevent the model from disguising "inferences" as "documentary facts"; 4) Forced partitioning of output structure: The output reasoning results need to be partitioned into three parts: content in the vector database, domain knowledge, and synthesis and inference. Innovation analysis and novelty constraints include: In addition to answering the questions, you will also need to provide the following: 1) At least three novel approaches that can substantially change the way problems are solved; 2) Minimum feasible experiment; 3) An alternative system architecture that differs significantly from the existing paradigm; 4) Provide a labeled diagram; 5) Creative driving factors; 6) Conflict between novelty and prior art.

7. The experimental feedback module of the AI ​​development system for new technology products, characterized in that, include: The association submodule is used to contextually associate reasoning questions with lab reports; The Real-Time Inference Feedback submodule is used to execute the real-time inference feedback process; When the association submodule associates the reasoning problem with the experimental report in context, the experimental report participates in the reasoning process of the large language model as supplementary information. The experimental report is only valid for the associated reasoning problem and is not required to maintain global relevance for subsequent different reasoning problem scenarios in the system.

8. The experimental feedback module according to claim 7, characterized in that, During the execution of real-time reasoning feedback, the real-time reasoning feedback submodule uses the reasoning problem and the associated experimental report as a unified reasoning input: the current reasoning problem is used as the core input, and the experimental report file identifier is used as an auxiliary input. These are encapsulated together into a unified reasoning input structure, and preset reasoning rules are associated in this input structure, so that the model reasoning process is constrained by control instructions.

9. A new technology product AI development system, characterized in that, include: The database module is used to perform structured parsing of collected PDF documents and generated experimental reports and convert them into JSON files to form a vector database. The reasoning module is used to parse the problem based on a pre-designed Prompt, and under the constraints of the Prompt, access the vector database and combine it with public domain knowledge to perform retrieval and integration, and output a reasoning report. The experiment feedback module is used to feed back the experimental results to the AI ​​development system for new technology products under controlled conditions, based on the experiment report or experimental results, and to participate in subsequent reasoning and knowledge accumulation. The reasoning module adopts the reasoning module according to any one of claims 1-6, and the experimental feedback module adopts the experimental feedback module according to claim 7 or 8.

10. The method for generating reasoning reports in the AI ​​development system for new technology products as described in claim 9, characterized in that, include: Step s1: Select and collect literature in a specific technical field, extract JSON format data from the literature, form a JSON file, and upload the JSON file containing the JSON format data to the vector database; Step s2: Set the reasoning rules; Step s3: Based on the reasoning problem and the experimental report, construct a unified reasoning input structure; Step s4: Input the unified inference input structure into the large language model and judge the target vector database; Step s5: Based on the reasoning rules, within the scope of the target vector database and in conjunction with publicly available domain knowledge, execute the reasoning process and output a reasoning report.

11. The reasoning report generation method according to claim 10, characterized in that, Step s1 involves extracting JSON format data from the literature to form a JSON file, specifically including: The selected and collected documents are downloaded and stored in a pre-defined document catalog, with the file name serving as the unique identifier for each document. Define the structured extraction rules and JSON template; Upload all PDF documents in the bibliography to the large language model and set the large language model calling strategy; The uploaded PDF documents are parsed using a large language model to generate JSON files with JSON structured data.

12. The reasoning report generation method according to claim 10, characterized in that, Step s3 specifically includes: Step s31: Contextualize the questions and experimental reports; Step s32: Construct a unified inference input structure: First, initialize the inference service by calling the client; Then, the lab report attachments are used as temporary input resources, uploaded to the file management component, and the file identifier corresponding to the lab report attachments is obtained; Then, configure the inference toolset, including vector retrieval tools and attachment parsing tools; Next, a unified reasoning input is constructed. The current reasoning problem is taken as the core input, and the experimental report file identifier is taken as an auxiliary input. They are encapsulated together into a unified reasoning input structure, and the preset reasoning rules are associated in this input structure so that the model reasoning process is constrained by control instructions.

13. The reasoning report generation method according to claim 12, characterized in that, The unified inference input structure includes at least: Reasoning question text; References to the lab report attachments; Inference control commands; Search tool and attachment parsing tool configuration parameters.

14. The reasoning report generation method according to claim 10, characterized in that, Step s4 specifically includes: Step s41: Load the registry of the vector database; Step s42, Query preprocessing: Perform normalization on the query request text; perform the same normalization on the topic description text in each vector database entry; Step s43: Calculate the semantic relevance base score; Step s44: Calculate keyword hit bonus points; Step s45: Calculate and rank the overall scores; The vector database corresponding to the highest score among all comprehensive scores is selected as the target vector database.

15. The reasoning report generation method according to claim 10, characterized in that, Step s5 specifically includes: Step s51: Inference call initialization; Step s52, Controlled reasoning execution: Step s521: Semantic retrieval and result filtering; Step s522, multi-step logical reasoning; Step s53: Generate structured results.

16. An Internet of Things (IoT) platform, characterized in that, The system includes a display screen and the AI ​​development system for new technology products as described in claim 9. After the AI ​​development system for new technology products outputs a reasoning report related to the development of new technology products, the reasoning report is displayed on the display screen.

17. Product performance testing equipment, characterized in that, The system includes a user interface and the AI ​​development system for new technology products as described in claim 9. The user interface is used to input reasoning questions, experimental data, and experimental reports. Based on the reasoning questions, experimental data, or experimental reports, the AI ​​development system for new technology products generates a next-step test plan or product improvement plan for product performance testing.

Images