Semiconductor factory fault tracing method and system based on LLM and retrieval enhancement

Abstract

The invention discloses a semiconductor factory fault tracing method and system based on LLM and retrieval enhancement, belongs to the field of semiconductor manufacturing and intelligent troubleshooting, and aims to solve the problems that historical processes and defect report fields of a semiconductor factory are not uniform, a retrieval mode is single, the accuracy rate is low, and a fault tracing result lacks causal reasoning and interpretation. Firstly, report key information fields are extracted through combination of a large language model, rules and codes, and source identifiers are added; secondly, field semantic normalization is achieved through large language model cue word mapping, and field data structure specifications are unified; then writing the data into an elastic search database, defining text type fields and adding comprehensive content fields containing field combination information; when a user asks a question, the question is analyzed through a large language model to generate a selected statement for query, and two-stage retrieval is executed. According to the method, the retrieval accuracy and recall rate can be improved, the interpretability of fault tracing is realized, and the method has dynamic expansion capability and assists engineers in quickly removing faults and making decisions.

Description

TECHNICAL FIELD

[0001] The present application relates to the fields of semiconductor manufacturing technology and intelligent fault diagnosis, and in particular to a semiconductor factory fault traceability method and system based on LLM and retrieval enhancement. BACKGROUND

[0002] Semiconductor manufacturing is one of the industries with the highest technology intensity and the most complex process flow in the world. An advanced process wafer production line needs to cover hundreds of processes, involving thousands of high-precision equipment. Small fluctuations in process parameters or slight wear of equipment components of each process may cause defects on the wafer, directly affecting product yield and production efficiency. In order to achieve rapid positioning and solution of faults, a semiconductor factory will generate a large amount of process reports and defect analysis documents, including real-time logs of equipment, wafer detection reports, engineer troubleshooting records, supplier technical documents, etc. These documents are the core data support for fault traceability.

[0003] However, the existing fault traceability method of the semiconductor factory still has many technical bottlenecks, which is difficult to meet the production demand of high density and high rhythm. The data format is seriously heterogeneous, and the standardization is difficult. The documents of different sources and different types have significant format differences. From the department dimension, the naming habits of the report fields of each department are different, and there is no unified definition. From the equipment dimension, the output log formats of different manufacturers' equipment are not compatible, the field names are similar but the semantics are slightly different, and it is difficult for manual to unify the mapping. From the time dimension, during the process iteration of the factory, the field definition will be adjusted with the upgrading of technology, and the field semantics of the historical documents and the new documents cannot be directly associated, resulting in that the data cannot be reused across periods. Some factories try to establish a field mapping table by manual arrangement, but a large amount of manpower needs to be invested, and the updating speed is far behind the data generation speed, eventually forming a data island, and the documents of different departments, different equipment and different periods cannot be analyzed, so engineers need to check the documents of different systems one by one during fault traceability, which is extremely low in efficiency.

[0004] The existing retrieval tool for fault traceability mainly relies on keyword accurate matching or fixed field screening, which cannot adapt to the complexity requirements of semiconductor fault query. Keyword matching has the risk of missing detection, because the effective results are missed due to the great difference between synonyms or English expressions and keywords; fixed field screening has the problem of false detection, which will return irrelevant results at the same time, and needs to be excluded one by one by manual; and it lacks the ability of multi-condition semantic association retrieval, actual fault query often involves multi-dimensional conditions, the existing system needs engineers to input conditions one by one in multiple field boxes, and cannot handle fuzzy semantics, if a condition is input incorrectly, the retrieval result is empty.

[0005] Existing fault tracing solutions are mostly black box outputs, only providing simple conclusions without supporting historical evidence and data sources. Engineers cannot verify the reliability of the results, and if the model training data is biased, it may give incorrect conclusions. The system does not record the original document sources corresponding to the results, and engineers need to re-search all related documents to further review historical troubleshooting details, prolonging the troubleshooting cycle. Moreover, it is difficult to adapt to complex fault scenarios. For "defects caused by multiple factors", only a single root cause can be output, and the contribution of each factor cannot be distinguished, making it difficult for engineers to develop targeted comprehensive solutions.

[0006] Semiconductor process iteration speed is fast, and each process upgrade introduces new defect types, device parameters, and field definitions. However, the extensibility of existing fault tracing systems is severely insufficient. New fields need to be manually modified, and the adaptation period is long. During this period, new data cannot be included in the tracing system, leading to delayed fault localization. Different process defect semantics differ, and existing systems need to define semantic rules for each process type. They cannot achieve rule migration through adaptive methods, increasing maintenance costs. Moreover, they cannot be compatible with new data sources, requiring the development of dedicated data interfaces and analysis modules, further extending the system adaptation period.

[0007] Currently, some technologies have attempted to combine large language models with retrieval techniques in the field of industrial fault analysis or data processing, but they all have obvious limitations and cannot directly adapt to the needs of semiconductor factory fault tracing. For example, patent CN119127973A discloses an MRO fault intelligent analysis system and method based on a Text-to-SQL large model. Its core is to convert user fault analysis problems into SQL statements through a Text-to-SQL model, retrieve MRO fault data in a relational database, and generate natural language analysis results. Although this technology achieves natural language to SQL conversion, it only targets MRO field fault statistical analysis and does not solve the standardization problem of non-uniform field names in unstructured reports. Moreover, it relies on the fixed table structure of a relational database and cannot support flexible retrieval of unstructured process reports in a semiconductor factory, lacking the ability to reason about fault causes and explain results.

[0008] Another patent CN119396975A proposes an electric power operation and maintenance knowledge question and answer system based on a large language model. By constructing an electric power operation and maintenance knowledge graph and a semantic vector database, it realizes semantic retrieval and answer generation for user questions. This technology focuses on knowledge questions and answers in the electric power operation and maintenance field and introduces the Retrieval Augmented Generation (RAG) approach, but does not involve field semantic normalization processing. It cannot handle the field heterogeneity problem of different process reports and different device logs in a semiconductor factory, and its knowledge graph construction relies on specific data in the electric power field, making it difficult to dynamically adapt to new fields and new defect types brought by semiconductor process iteration, lacking scalability.

[0009] There is also a patent CN120012756A discloses a government data quality inspection rule generation method and system based on a large language model, which extracts field metadata vectors and retrieves historical knowledge fragments to construct dynamic prompt word generation data quality inspection rules. This technology is only used for quality inspection of government data, and the core goal is to generate inspection SQL statements, not fault traceability search and root cause analysis. It does not involve multi-dimensional search strategies for fault-related data, nor does it have reasoning and evidence capabilities for fault root causes. It is completely unsuitable for complex fault traceability scenarios in semiconductor factories.

[0010] In summary, the existing semiconductor factory fault traceability method has problems such as difficulty in data standardization, low search efficiency, poor result interpretability, and insufficient scalability. The existing technology in related fields cannot adapt to the process characteristics and fault traceability needs of semiconductor manufacturing, and has been unable to meet the efficient troubleshooting needs of advanced process wafer manufacturing. There is an urgent need for a technical solution that can automatically normalize data, accurately search, explain traceability, and have dynamic expansion capabilities to improve fault location efficiency and decision-making credibility and ensure stable production yield. SUMMARY

[0011] The present application aims to address the shortcomings of the prior art. The present application proposes a semiconductor factory fault traceability method and system based on LLM and search enhancement. By combining the semantic understanding ability of the large language model and the precise data recall ability generated by search enhancement, it can achieve automatic normalization of heterogeneous data, accurate search of fault-related information, and explainability and traceability of traceability results. It does not need to rely on manual field mapping and rule definition, thereby improving the fault location efficiency and decision-making credibility of semiconductor factories and reducing labor costs.

[0012] The technical solution to achieve the purpose of the present application is:

[0013] The semiconductor factory fault traceability method based on LLM and search enhancement includes the following steps:

[0014] Step S1: From the historical process reports and defect reports of the semiconductor factory, use a large language model LLM combined with rules and codes to identify and extract fields, and add a file name field to each extracted data to record its source, forming a set of data to be processed.

[0015] Step S2: To address the problem of non-uniform field names in the key information fields extracted in step S1, use a large language model prompt LLM mapping method: input the prompt containing the field name to be judged and the list of pre-set unified data structure field categories into the LLM, so that the LLM outputs the target field category corresponding to the field name to be judged, achieving semantic normalization of all key information fields.

[0016] Step S3: write the semantic normalized data of step S2 into the elastic search database, wherein each field of the normalized data is defined as text type, and a comprehensive content field is added in the elastic search database index;

[0017] Step S4: when a user inputs a fault traceability question in natural language form, a method combining LLM and a preset prompt word is used to analyze the key fields in the question and generate corresponding SELECT query conditions, which are selection statement query conditions;

[0018] Step S5: based on the SELECT query conditions generated in step S4, an ES SQL query method is used to perform accurate retrieval, wherein ES SQL is an elastic search database structured query language;

[0019] Step S6: comprehensive scoring is performed on the multiple candidate results obtained by step S5, the candidate results are sorted according to the scoring results, and a sorted result set is generated;

[0020] Step S7: an LLM is used to combine the sorted result set of step S6 to generate an interpretable root cause analysis (RCA) answer, and the answer can be traced back to its data source.

[0021] Further, the fields extracted in step S1 include but are not limited to process, defect code, defect name, process number, loop, layer identification, device module, suspected cause, and follow-up processing measures. The process includes FinFET process. The defect code conforms to the SEMI semiconductor defect standard code. The process number includes CMP process identification. The layer identification includes metal layer identification and dielectric layer identification. The device module includes CMP polishing head, photolithography light source module, and PVD sputtering target. The suspected cause includes semiconductor device slurry contamination and sputtering power fluctuation related content. The follow-up processing measures include PVD power sensor calibration and photolithography mask replacement related operations.

[0022] Further, the integrated content field in step S3 is a combination of the fields formed after the semantic normalization in step S2 and does not include the file name field used to record the data source, the range of the fields is consistent with the range of the fields extracted in step S1 of claim 2, and the combined content of the integrated content field is associated with the semiconductor process scene.

[0023] Further, the combined form of the integrated content field in step S3 is one of the following two forms: form one, sequentially concatenating all fields except the file name field in the key:value string format; form two, storing the key:value of all fields except the file name field in the JSON dictionary format, the key:value is a key-value pair, and the JSON is the JavaScript Object Notation.

[0024] Further, in step S3, when the user question is parsed into structured fields such as wafer batch identification and metal level identification, SQL retrieval is directly performed on the corresponding fields; when the structured retrieval result is empty, the system automatically triggers a degradation mechanism to perform full-text retrieval on the integrated content field.

[0025] Further, in step S5, if no results are found, the system automatically switches to keyword or vector similarity retrieval on the integrated content field to improve recall rate.

[0026] Further, in step S6, the factors considered in the integrated scoring include at least one of field matching degree, semantic similarity, historical data frequency, and data time weight.

[0027] Further, in step S7, the explainable RCA answer includes defect root cause, historical evidence data, and possible fault resolution measures.

[0028] The semiconductor factory fault tracing system based on LLM and retrieval enhancement is used to implement the semiconductor factory fault tracing method based on LLM and retrieval enhancement, and the system comprises:

[0029] The field extraction module is configured to extract fields from the historical process reports and defect reports of the semiconductor factory by combining LLM and rules and codes, and add a file name field to each data, the range of the extracted fields is consistent with the range of the fields extracted in step S1 of claim 2;

[0030] Mapping normalization module: configured to use the LLM Prompt mapping method to normalize fields with inconsistent field names extracted in step S1 to target field categories with a preset unified schema; Elastic search database storage module: configured to write the normalized data into the elastic search database, define each field as text type, and construct a comprehensive content field containing the combined information of the fields extracted in step S1 after semantic normalization in step S2;

[0031] Question parsing module: Configured to receive user questions in natural language, parse key fields using LLM and preset prompts, and generate corresponding SELECT query conditions;

[0032] Search module: Configured to first perform ES SQL exact search based on the SELECT query conditions generated in step S4, and automatically switch to keyword or vector similarity search on the comprehensive content field if the result is empty;

[0033] Results processing module: Configured to score and sort candidate search results based on at least one of the following factors: field matching degree, semantic similarity, frequency of occurrence of historical data, and time weight of data;

[0034] Reasoning and Response Module: Configured to generate interpretable RCA responses based on LLM and sorted results, including root causes of defects, historical supporting data, possible troubleshooting measures, and traceable data sources.

[0035] Compared with the prior art, the present invention, employing the above technical solution, has the following beneficial effects:

[0036] (1) The semiconductor factory fault tracing method and system based on LLM and retrieval enhancement proposed in this invention can realize the automatic standardization of heterogeneous data and greatly reduce the labor cost: Through the semantic mapping capability of the large language model, this invention can automatically normalize the process / defect report fields from different sources and in different formats to a unified schema without the need for manual organization of field mapping tables or definition of rules. This not only improves the efficiency of data standardization by several times, but also adapts to the changes in fields during process iteration, solving the problems of data silos and adaptation lag in the traditional method.

[0037] (2) The semiconductor factory fault tracing method and system based on LLM and retrieval enhancement proposed in this invention adopts a two-level strategy of precise retrieval combined with semantic retrieval. It ensures accurate matching of highly relevant results through structured query, avoids missed detection due to expression differences by using global semantic retrieval of comprehensive content fields, and filters invalid information through multi-factor comprehensive sorting. This significantly improves the recall rate and accuracy of fault-related data compared with traditional keyword matching methods, reducing the time cost for engineers to identify information.

[0038] (3) The root cause analysis answer generated by the LLM-based and search-enhanced semiconductor factory fault tracing method and system proposed in this invention is directly related to historical supporting data and source documents (traceable through the file name field), clearly explaining the correlation logic between the root cause and the defect and the historical processing effect. Engineers can intuitively verify the reliability of the results. For complex faults with multiple superimposed factors, it can distinguish the contribution of each factor and provide targeted measures, effectively shortening the fault investigation cycle and improving the scientific nature of troubleshooting decisions. Attached Figure Description

[0039] Figure 1 This is a flowchart of the semiconductor factory fault tracing method based on LLM and retrieval enhancement of the present invention;

[0040] Figure 2 This is a diagram illustrating the combined form of the comprehensive content fields;

[0041] Figure 3 A structural diagram illustrating the components of an interpretable RCA response;

[0042] Figure 4 This is a block diagram of a semiconductor factory fault tracing system based on LLM and enhanced retrieval. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] A semiconductor factory fault tracing method and system based on LLM and enhanced retrieval, the process of which is as follows: Figure 1 As shown, it includes the following steps:

[0045] Step S1: From the historical process reports and defect reports of the semiconductor factory, use the Large Language Model (LLM) combined with rules and codes to identify and extract fields. At the same time, add a field to each extracted data to record the file name of its source, forming a set of data to be processed.

[0046] Step S2: Addressing the issue of inconsistent field names among the key information fields extracted in Step S1, the Large Language Model (LLM) Prompt mapping method is employed: Prompt words containing the names of the fields to be judged and a pre-defined list of field categories with a unified data structure are input into the LLM. The LLM then outputs the target field category corresponding to the name of the field to be judged, achieving semantic normalization of all key information fields. For example:

[0047] Prompt: "Given the field name DefectTypeCode, which category should it belong to among [defect_code, op_no, process, module, suspect_cause]?"

[0048] LLM output: defect_code.

[0049] Using the above method, field names such as "DefectTypeCode" (given field name), "def_code" (defect code), and "code" in different reports can be automatically normalized to the standard field "defect_code".

[0050] Step S3: Write the semantically normalized data from step S2 into the Elastic Search Database. Each field of the normalized data is defined as a text type, and a comprehensive content field is added to the Elastic Search Database index.

[0051] Step S4: When the user inputs a fault tracing question in natural language, the method combining LLM and preset prompts is used to parse out the key fields in the question and generate corresponding SELECT query conditions. The SELECT query conditions are the selection statement query conditions. For example, the following prompt is used to generate SQL:

[0052] {

[0053] SQL table name: defect_library; Database table fields are as follows:

[0054] Process, Defect Code, Defect Name, Operation Number, Loop Count, Layer IDs, Module, Suspect Cause, Follow-up Actions

[0055] The question is as follows: What is the suspected cause for defect code 803?

[0056] Please write the SQL and output only the SQL, without any other explanations or text.

[0057] Requirements:

[0058] Only output a SELECT statement.

[0059] The SELECT statement must be in the following format:

[0060] select field_name1, field_name2, ... from table_name where ...

[0061] Do not use COUNT, DISTINCT, GROUP BY, ORDER BY, LIMIT, or any aggregate functions.

[0062] Do not add explanations or extra text, only the SQL query.

[0063] }

[0064] Step S5: Based on the SELECT query conditions generated in step S4, perform precise retrieval using ES SQL query method, where ES SQL is the Elastic Search Database Structured Query Language;

[0065] Step S6: Score the multiple candidate results obtained in step S5, sort the candidate results according to the scoring results, and generate a sorted result set.

[0066] Step S7: Using LLM combined with the sorted result set from Step S6, generate an interpretable root cause analysis (RCA) response, and the response can be traced back to its data source; the process includes FinFET process; the defect code conforms to the SEMI semiconductor defect standard coding; the operation number op_no includes CMP operation identifier; the layer identifiers layer_ids include metal layer identifier and dielectric layer identifier; the module includes CMP polishing head, lithography light source module, and PVD sputtering target; the suspected cause suspect_cause includes semiconductor equipment slurry contamination and sputtering power fluctuation related content; the follow-up measures follow_up include PVD power sensor calibration and lithography mask replacement related operations.

[0067] Furthermore, the fields extracted in step S1 include, but are not limited to, process, defect code, defect name, operation number, loop count, layer ID, module, suspected cause, and follow-up.

[0068] Furthermore, the comprehensive content field in step S3 is a field formed by semantic normalization of the fields extracted in step S1 in step S2, and does not include a combination of fields used to record the file name field of the data source. The range of the field is consistent with the range of fields extracted in step S1 of claim 2. The combined content of the comprehensive content field is associated with the semiconductor process scenario.

[0069] Furthermore, the combination of the comprehensive content fields in step S3 can be one of the following two forms, such as... Figure 2 As shown: Format 1 involves concatenating all fields except the filename field in the key:value string format, for example:

[0070] {

[0071] Process: 28nm

[0072] Defect code: 803

[0073] Process number (op_no): 1V.ECU10

[0074] Equipment module: Chemical Mechanical Polishing (CMP)

[0075] Suspected cause: Slurry contamination

[0076] }

[0077] Format two: Store the key:value pairs of all fields except the file name field in JSON dictionary format, for example:

[0078] {

[0079] Process: 28nm

[0080] Defect code: 803

[0081] Process number (op_no): 1V.ECU10

[0082] Equipment module: Chemical Mechanical Polishing (CMP)

[0083] Suspected cause: Slurry contamination

[0084] }

[0085] Furthermore, in step S3, when the user's question is parsed into structured fields, such as wafer batch identifier or metal level identifier, an SQL search is performed directly on the corresponding field; when the structured search result is empty, the system automatically triggers a degradation mechanism to perform a full-text search on the comprehensive content field.

[0086] Furthermore, in step S5, if no results are found, the search automatically switches to keyword or vector similarity retrieval on the comprehensive content field to improve the recall rate.

[0087] Furthermore, the factors considered in the comprehensive scoring in step S6 include at least one of field matching degree, semantic similarity, frequency of occurrence of historical data, and data time weight.

[0088] Furthermore, the interpretable RCA response in step S7 includes the root cause of the defect, historical supporting data, and possible troubleshooting measures, such as... Figure 3 As shown.

[0089] A semiconductor factory fault tracing system based on LLM and retrieval enhancement is proposed to implement a semiconductor factory fault tracing method based on LLM and retrieval enhancement, such as... Figure 4 As shown, the system includes:

[0090] Field extraction module: configured to extract fields from historical process reports and defect reports of semiconductor factories using a combination of LLM, rules, and codes, and add a file name field to each data. The range of extracted fields is consistent with the range of fields extracted in step S1 of claim 2.

[0091] Mapping normalization module: configured to use the LLM Prompt mapping method to normalize fields with inconsistent field names extracted in step S1 to target field categories with a preset unified schema; Elastic search database storage module: configured to write the normalized data into the elastic search database, define each field as text type, and construct a comprehensive content field containing the combined information of the fields extracted in step S1 after semantic normalization in step S2;

[0092] Question parsing module: Configured to receive user questions in natural language, parse key fields using LLM and preset prompts, and generate corresponding SELECT query conditions;

[0093] Search module: Configured to first perform ES SQL exact search based on the SELECT query conditions generated in step S4, and automatically switch to keyword or vector similarity search on the comprehensive content field if the result is empty;

[0094] Results processing module: Configured to score and sort candidate search results based on at least one of the following factors: field matching degree, semantic similarity, frequency of occurrence of historical data, and time weight of data;

[0095] Reasoning and Response Module: Configured to generate interpretable RCA responses based on LLM and sorted results, including root causes of defects, historical supporting data, possible troubleshooting measures, and traceable data sources.

[0096] Example 1:

[0097] The system extracts fields from the report, such as "DefectTypeCode: 804" and "operation_number:1V.ECU10".

[0098] Using the LLM Prompt mapping, "DefectTypeCode" is normalized to "defect_code" and "operation_number" is normalized to "op_no".

[0099] Example of normalization results:

[0100] JSON

[0101] {

[0102] Defect code: "804"

[0103] "Process Number (op_no)": "1V.ECU10"

[0104] }

[0105] Example 2:

[0106] A user entered the question: "What are the possible causes of defect 804 in the CU CMP stage of the 28nm process?"

[0107] The system parsed the following fields: process=28nm, module=CMP, defect_code=804.

[0108] First, execute the SQL search; if the result is empty, then execute the keyword search.

[0109] The search result, when entered as LLM, generates the following answer: "This defect has occurred multiple times in similar historical processes. The main cause is slurry contamination. It is recommended to check the slurry supply and filtration system."

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A semiconductor factory fault tracing method based on LLM and retrieval enhancement, characterized in that, Includes the following steps: Step S1: From the historical process reports and defect reports of the semiconductor factory, use the Large Language Model (LLM) combined with rules and codes to identify and extract fields. At the same time, add a field to each extracted data to record the file name of its source, forming a set of data to be processed. Step S2: To address the issue of inconsistent field names among the key information fields extracted in Step S1, the Large Language Model Prompt Mapping Method (LLM Prompt Mapping) is adopted: Prompt words containing the field names to be judged and a list of pre-defined unified data structure field categories are input into the LLM, so that the LLM outputs the target field category corresponding to the field name to be judged, thereby achieving semantic normalization of all key information fields. Step S3: Write the semantically normalized data from step S2 into the Elastic Search Database. Each field of the normalized data is defined as a text type, and a comprehensive content field is added to the Elastic Search Database index. Step S4: When the user inputs a fault tracing question in natural language, the key fields in the question are parsed using a combination of LLM and preset prompt words, and corresponding SELECT query conditions are generated. The SELECT query conditions are the selection statement query conditions. Step S5: Based on the SELECT query conditions generated in step S4, perform precise retrieval using ES SQL query method, where ES SQL is the Elastic Search Database Structured Query Language; Step S6: Score the multiple candidate results obtained in step S5, sort the candidate results according to the scoring results, and generate a sorted result set. Step S7: Using LLM combined with the sorted result set from step S6, generate an interpretable root cause analysis (RCA) response, and the response can be traced back to its data source.

2. The semiconductor factory fault tracing method based on LLM and retrieval enhancement according to claim 1, characterized in that, The fields extracted in step S1 include, but are not limited to, process, defect code, defect name, operation number, loop number, layer ID, module, suspected cause, and follow-up.

3. The semiconductor factory fault tracing method based on LLM and retrieval enhancement according to claim 1, characterized in that, The comprehensive content field in step S3 is the field formed by semantic normalization of the field extracted in step S1 in step S2, and does not include the combination form of the file name field used to record the data source. The scope of the field is consistent with the scope of the field extracted in step S1 of claim 2.

4. The semiconductor factory fault tracing method based on LLM and retrieval enhancement according to claim 1, characterized in that, The combination of the comprehensive content fields in step S3 can be in one of the following two forms: Form 1, concatenating all fields except the file name field in the key:value string format; Form 2, storing the key:value of all fields except the file name field in JSON dictionary format, where key:value is a key-value pair and JSON is JavaScript object representation.

5. The semiconductor factory fault tracing method based on LLM and retrieval enhancement according to claim 1, characterized in that, In step S3, when a user's question is parsed into a structured field, an SQL search is performed directly on the corresponding field; if the structured search result is empty, the system automatically triggers a degradation mechanism to perform a full-text search on the comprehensive content field.

6. The semiconductor factory fault tracing method based on LLM and retrieval enhancement according to claim 1, characterized in that, In step S5, if no results are found, the search will automatically switch to keyword or vector similarity retrieval on the comprehensive content field to improve the recall rate.

7. The semiconductor factory fault tracing method based on LLM and retrieval enhancement according to claim 1, characterized in that, The factors considered in the comprehensive scoring in step S6 include at least one of the following: field matching degree, semantic similarity, frequency of occurrence of historical data, and data time weight.

8. The semiconductor factory fault tracing method based on LLM and retrieval enhancement according to claim 1, characterized in that, The interpretable RCA response in step S7 includes the root cause of the defect, historical supporting data, and possible troubleshooting measures.

9. A semiconductor factory fault tracing system based on LLM and enhanced retrieval, characterized in that, The system for implementing the LLM-based and retrieval-enhanced semiconductor factory fault tracing method as described in any one of claims 1-8, the system comprising: Field extraction module: configured to extract fields from historical process reports and defect reports of semiconductor factories using a combination of LLM, rules, and codes, and add a file name field to each data. The range of extracted fields is consistent with the range of fields extracted in step S1 of claim 2. Mapping normalization module: configured to use the LLM Prompt mapping method to normalize fields with inconsistent field names extracted in step S1 to target field categories with a preset unified schema; Elastic search database storage module: configured to write the normalized data into the elastic search database, define each field as text type, and construct a comprehensive content field containing the combined information of the fields extracted in step S1 after semantic normalization in step S2; Question parsing module: Configured to receive user questions in natural language, parse key fields using LLM and preset prompts, and generate corresponding SELECT query conditions; Search module: Configured to first perform ES SQL exact search based on the SELECT query conditions generated in step S4, and automatically switch to keyword or vector similarity search on the comprehensive content field if the result is empty; Results processing module: Configured to score and sort candidate search results based on at least one of the following factors: field matching degree, semantic similarity, frequency of occurrence of historical data, and time weight of data; Reasoning and Response Module: Configured to generate interpretable RCA responses based on LLM and sorted results, including root causes of defects, historical supporting data, possible troubleshooting measures, and traceable data sources.

Images