Introduction
In the high-stakes environment of semiconductor manufacturing, test data analytics is critical for maintaining yield and identifying process fluctuations. However, building an "Intelligent Engineering Assistant" (IEA) capable of navigating these complexities has historically faced a significant scalability bottleneck. Previous iterations, such as IEA-Plot, relied on mapping user instructions to a static set of pre-defined API calls. This approach suffered from a "Knowledge Acquisition Gap," where developers could not anticipate every possible engineering query, and a "Scalability Wall," where every new analysis type required manual backend updates.
IEA-Plugin addresses these challenges by shifting from a static mapping model to an agentic reasoning framework. By leveraging Large Language Models (LLMs) like GPT-o3-mini and the LangGraph framework, IEA-Plugin transforms vague, multi-step engineering queries into structured workflows and automatically generates the backend API specifications required for execution.
Methodology
Agentic Programming and Iterative Reasoning
IEA-Plugin is built using the LangGraph framework, which treats the reasoning process as a state machine rather than a single linear prompt. This allows the agent to "think," retrieve relevant examples, and refine its plan iteratively. The system employs a ReAct (Reasoning and Acting) combined with Retrieval-Augmented Generation (RAG) strategy. When a query is received, the agent generates an internal "thought," uses that thought to search a database of past (query, workflow) pairs, and continues to deliberate until the reasoning stabilizes into a coherent plan.
In-Context Learning (ICL) and API Distillation
A core component of the system is its ICL-centric strategy. Instead of hard-coding logic, the system maintains a repository of examples to guide the LLM. Once a multi-step workflow is generated, the system uses GPT-4o to classify individual steps into categories such as "Data," "Analysis," or "Output."
Crucially, the system then performs Automated API Distillation. It generates the formal API specifications or specific code (such as Cypher queries for Neo4j databases) needed to fulfill the workflow. This effectively decouples the user's intent from the software's implementation, allowing the system to build its own backend capabilities dynamically.
Key Insights
The Power of "Thought"
One of the most significant findings in the development of IEA-Plugin is the impact of explicit reasoning steps. Forcing the LLM to generate a "thought" before constructing a workflow significantly improves the quality and specificity of the output. While "thoughtless" workflows tended to be generic, "thoughtful" workflows recognized nuances—such as the ambiguity in a phrase like "multiple weeks"—and added specific steps to define time intervals and parameters, leading to more robust engineering plans.
API Stabilization and Reuse
The research demonstrates that as the system encounters more queries, its need to generate entirely new API functions decreases. The system begins to reuse "Analysis" functions (e.g., calculate_yield) across different workflows. This suggests that a comprehensive, stable API can be distilled automatically from a diverse set of user requirements, proving that the agentic approach scales more efficiently than manual development.
Results and Performance
Scalability and Efficiency
The IEA-Plugin was realized in under three months, demonstrating the rapid development cycle enabled by advanced reasoning models. In testing, the system showed a high rate of function reuse. For an initial set of 80 queries, the model derived 141 API functions. When 40 additional, more complex queries were introduced, the system only required 47 new functions. This decreasing ratio of new functions per step confirms that the system effectively builds a reusable library of analytical tools.
Workflow Differentiation
Semantic similarity tests confirmed that workflows generated with internal reasoning were substantially different from those generated without it (similarity scores < 0.8). The reasoning-enabled workflows were consistently more tailored to the specific goals of the user, successfully prioritizing primary goals (like correlation analysis) while condensing secondary tasks (like classification) into efficient single steps.
Limitations and Future Directions
Despite its success, IEA-Plugin faces challenges regarding the subjectivity of "correctness." In engineering, there is rarely a single correct workflow for a complex problem; different LLM prompts may yield different but equally valid plans. Additionally, the inherent statistical variability of LLMs means that while the first response is usually intuitive, results can occasionally vary in optimality.
Future work will focus on moving toward a "closed-loop" execution model. In this vision, the agent will not only plan the workflow and generate the API but also execute the code and validate the results against live databases in real-time, further reducing the manual burden on semiconductor engineers.
Summary updated