IM-Chat: Injection Molding Conversational AI

Updated 7 July 2026

IM-Chat is a multi-agent conversational interface for injection molding that preserves and operationalizes both documented and field knowledge.
It employs retrieval-augmented generation and a modular workflow with specialized agents for planning, execution, and translation to support troubleshooting and process optimization.
The framework leverages a data-driven diffusion model and cosine similarity-based retrieval to deliver precise process condition recommendations and multilingual support.

IM-Chat, short for Injection Molding Chatting interface, is a multi-agent framework based on LLMs for knowledge transfer in the injection molding industry. It is designed to preserve and operationalize both limited documented knowledge—such as troubleshooting tables, manuals, interviews, and technical references—and extensive field knowledge represented by production data and a data-driven process condition generator. The framework adopts retrieval-augmented generation (RAG), tool-calling agents, and a modular architecture so that it can support multilingual interaction, routine troubleshooting, and quantitative process recommendation without the need for fine-tuning (Lee et al., 21 Jul 2025).

1. Industrial problem setting and rationale

The problem addressed by IM-Chat is the preservation and transfer of field knowledge in injection molding. The paper states that the industry depends heavily on experienced operators who know how to interpret defects, adjust machine settings, and respond to changing environmental conditions in real time. Much of this knowledge is described as tacit and therefore difficult to capture in manuals alone. The paper further identifies three pressures that intensify the problem: veteran workers are retiring, new workers are fewer, and factories often employ multilingual teams, creating communication barriers when expertise must be shared quickly and accurately (Lee et al., 21 Jul 2025).

IM-Chat is therefore framed as a conversational AI system for industrial decision support rather than as a general-purpose chatbot. Its purpose is to preserve knowledge, translate it across languages, and make it accessible in a form that can support both documented troubleshooting and data-driven process decisions. The framework is explicitly intended to support questions about defects, settings, manuals, and environmental operating conditions within a single conversational workflow (Lee et al., 21 Jul 2025).

A broader communication rationale can be inferred from instant-messaging research. Workplace IM has been described as less intrusive than a phone call, more immediate than email, and supported by presence detection (Maina, 2013). This suggests why a chat interface is a plausible front end for factory knowledge transfer: it combines low-friction interaction with the possibility of rapid clarification and incremental problem formulation.

2. Agentic architecture and control flow

IM-Chat is organized into three stages: input formatting, task solving, and output formatting. The system is not a single model; it is a coordinated set of specialized agents with narrow roles. The paper emphasizes this modularity because it makes the framework easier to extend with new tools and easier to adapt to other industrial domains (Lee et al., 21 Jul 2025).

Component	Role
Task Formatter	Condenses current user message together with relevant chat history
Translator	Normalizes the task into English and tracks the original language
Classifier	Decides whether the request is injection-molding-related
ReAct agent	Handles non-injection questions using internet search only when needed
Planner	Decomposes an injection-molding query into subtasks
Executor	Carries out the first subtask using the appropriate tool
Supervisor	Checks whether current information is sufficient
Replanner	Revises the plan if the answer is still incomplete
Reporter	Summarizes the reasoning and translates the answer back

For non-injection-molding queries, the system routes the request to a lightweight ReAct agent that handles general questions and uses internet search only when needed. For injection-molding-related queries, the system switches to a more structured plan-and-execute workflow. The Planner decomposes the query into subtasks; the Executor carries out the first subtask with an appropriate tool; the Supervisor decides whether the accumulated evidence is sufficient; and, if it is not, the Replanner updates the plan and the loop continues. The Reporter then summarizes the reasoning and returns the answer in the user’s language (Lee et al., 21 Jul 2025).

The architecture is deliberately constrained. The Executor acts on only the first planned subtask at a time, which is intended to keep the workflow controlled and prevent it from jumping ahead. The paper also notes that some subtasks can be solved without any tool, relying purely on internal reasoning when appropriate (Lee et al., 21 Jul 2025).

3. Knowledge sources, retrieval, and process-condition generation

The framework’s central knowledge strategy is retrieval-augmented generation. Rather than relying only on parametric memory, IM-Chat grounds its responses in curated injection-molding resources. The paper distinguishes between limited documented knowledge and extensive field knowledge, and the system handles these two knowledge types through different mechanisms (Lee et al., 21 Jul 2025).

Documented knowledge includes senior-worker troubleshooting tables, manuals, interviews, and technical references. The troubleshooting table captures defect-to-parameter adjustment rules from experienced workers. The manual component includes a 227-page machine manual containing material settings, operating procedures, and inspection guidelines. The troubleshooting table is chunked and embedded, then queried via cosine similarity. The manual is parsed page by page into structured Markdown/JSON, embedded, and searched using similarity plus MMR reranking to improve relevance and diversity (Lee et al., 21 Jul 2025).

This retrieval path supports concrete tasks. The paper gives a troubleshooting example asking how to reduce burr defects, for which the system recommends decreasing injection speed in priority order. A manual example asks for the recommended mold temperature range for ABS, and the manual retriever answers 40–60°C (Lee et al., 21 Jul 2025).

The more distinctive mechanism is the handling of field knowledge through a data-driven process condition generator based on a diffusion model trained on real manufacturing data. Its inputs are five structured variables: product class, factory temperature, factory humidity, machine temperature, and machine humidity. From these, it generates ten process parameters for a defect-free or acceptable setup: three injection speeds, three injection pressures, three injection positions, and hold time. A Diffusion Input Formatter extracts numeric values from natural-language user input and prompts for missing information when necessary. The generator then produces 64 candidate parameter sets, and a CatBoost-based surrogate model ranks them by predicted probability of yielding a good product. The best candidate is returned as the recommendation (Lee et al., 21 Jul 2025).

This dual structure—retrieval for documents and generation for production conditions—distinguishes IM-Chat from systems that only retrieve text. It allows the framework to answer both rule-oriented questions and environment-conditioned process questions within the same dialogue.

4. Formal and algorithmic components

The paper gives explicit formulas for both the retrieval pipeline and the diffusion model. For manual retrieval, it uses Maximal Marginal Relevance (MMR) to balance relevance and diversity among retrieved manual chunks. The paper states the MMR form as:

$MMR(D_i) = argmax\ |2 \cdot sim(D_i, Q)-(1-2)\cdot max\ sim(D_i, D_j)|\ \ D_i \in R \setminus S,\ D_i \in S$

The stated intent is to avoid returning several nearly identical manual pages and instead balance similarity to the query with diversity relative to already selected results (Lee et al., 21 Jul 2025).

The diffusion model is presented as learning a reverse denoising process from noisy samples back to data. The reverse conditional is given as:

$q(X_{t-1}|X_t) = q(X_t|X_{t-1})q(X_{t-1}) / q(X_t)$

and the approximate learned reverse process is stated as:

$p_\theta(X_{t-1}|X_t) \sim q(X_{t-1}|X_t)$

The forward noising process is described by:

$q(X_t|X_{t-1}) = \mathcal{N}(X_t; \sqrt{1-\beta_t}X_{t-1}, \beta_t I)$

with closed-form sampling:

$X_t = \sqrt{\bar{\alpha}_t}X_0 + \sqrt{1-\bar{\alpha}_t}\epsilon$

where $\epsilon \sim \mathcal{N}(0,I)$ . The practical loss is simplified to a noise-prediction objective:

$L_t = \mathbb{E}_{X_0,\epsilon}\left\|\epsilon - \epsilon_\theta(\sqrt{\bar{\alpha}_t}X_0 + \sqrt{1-\bar{\alpha}_t}\epsilon, t)\right\|^2$

The classifier-free guidance mechanism is expressed as:

$\hat{\epsilon}(Z_t, y, t) = w\epsilon(Z_t, y, t) + (1-w)\epsilon(Z_t, t)$

with a drop rate of 0.1 and guidance weight $w=3$ (Lee et al., 21 Jul 2025).

These formal components matter because IM-Chat does not treat industrial expertise as exclusively textual. The retrieval side represents documented procedural knowledge, while the diffusion side represents operational knowledge encoded in production data. A plausible implication is that the framework attempts to unify symbolic guidance, textual evidence, and quantitative recommendation within one conversational layer.

5. Evaluation protocol and empirical findings

The paper evaluates IM-Chat in two settings: single-tool tasks and hybrid multi-tool tasks. The single-tool benchmark contains 100 English-language tasks, divided into 25 each for troubleshooting-table retrieval, manual retrieval, diffusion-model generation, and ordinary reasoning tasks that may or may not require web search. The hybrid benchmark contains 60 tasks total, split into three categories of 20 each: diffusion model + troubleshooting table, diffusion model + manual retrieval, and diffusion model + internet search (Lee et al., 21 Jul 2025).

Three models are compared: GPT-4o, GPT-4o-mini, and GPT-3.5-turbo. Human domain experts score each response on a 10-point rubric focused on relevance and correctness. The paper also reports an automated evaluation using GPT-4o as a judge, prompted to score relevance and accuracy on the same 0–10 scale (Lee et al., 21 Jul 2025).

The qualitative performance pattern is consistent across the reported results. More capable models perform better, and the gap widens in complex, tool-integrated settings. In the single-tool evaluation, GPT-4o consistently scores highest in human evaluation, especially on troubleshooting-table and diffusion-model tasks. GPT-3.5-turbo performs worst, often because it misselects tools or fails to parse long manual content and tables correctly. GPT-4o-mini occupies an intermediate position but sometimes suffers from inefficient planning (Lee et al., 21 Jul 2025).

The automated GPT-4o evaluator broadly follows the same trend, but its agreement with experts is weak, especially for numerically grounded diffusion outputs. The reported Pearson correlations between expert and LLM evaluation are -0.14 for GPT-4o, 0.25 for GPT-4o-mini, and 0.35 for GPT-3.5-turbo. The paper therefore treats automated judging as useful for scalable benchmarking but not as a substitute for expert review in manufacturing (Lee et al., 21 Jul 2025).

Latency and cost reveal additional tradeoffs. For single-tool tasks, GPT-4o achieved a strong accuracy-latency balance at 24.4 seconds per query; GPT-3.5-turbo was fastest at 17.2 seconds but less reliable; GPT-4o-mini was slowest at 32.4 seconds because of inefficient re-planning. In hybrid tasks, latency increased substantially because diffusion inference dominates runtime: GPT-4o took 66.0 seconds, GPT-4o-mini 63.2 seconds, and GPT-3.5-turbo 33.3 seconds. The paper adds that the apparent speed of GPT-3.5-turbo was partly misleading because it often failed to invoke the required tool and therefore produced incomplete outputs (Lee et al., 21 Jul 2025).

These findings support the paper’s conclusion that multi-agent LLM systems are viable for industrial knowledge workflows, while also showing that evaluation in such settings remains tightly coupled to domain expertise.

The paper identifies several limitations. IM-Chat still struggles with visual and semi-structured information such as images, diagrams, engineering drawings, and complex tables, even though the manual parser uses a vision-LLM. It depends on commercial LLM APIs like GPT-4o, raising concerns about cost, privacy, and vendor dependence. Its current tool ecosystem is limited to troubleshooting tables, manuals, internet search, and the diffusion model, rather than broader industrial tools such as simulators, equation solvers, CAD systems, or optimization modules. The framework is also sensitive to vague inputs; the paper describes this with the familiar “garbage in, garbage out” principle (Lee et al., 21 Jul 2025).

Future work is directed toward multimodal parsing, more diverse tools, local lightweight models, and better input-refinement mechanisms such as clarification agents or guided templates. This trajectory is consistent with a broader movement in chat-based systems toward richer context handling and modality integration (Lee et al., 21 Jul 2025).

The term “IM-Chat” is not exclusive to injection molding. In the paper “Chatting Makes Perfect: Chat-based Image Retrieval”, the setting is described as IM-Chat: an image retrieval task in which the search intent is refined through a dialogue (Levy et al., 2023). In later work on Memory-Augmented Query Intent Understanding (MAQIU), chat-based image retrieval is further formalized through progressive intent memory and historical recall, with a reported 86.4% reduction in dialogue encoding FLOPs compared with the prior baseline ChatIR (Chen et al., 17 May 2026). Other related systems explore adjacent design spaces: MIRIAM pairs chat with mission visualization for autonomous systems and is explicitly mixed initiative (Hastie et al., 2018); HuixiangDou treats IM group chat as a technical-assistance setting requiring aggressive rejection of irrelevant chatter (Kong et al., 2024); and M²Chat targets interleaved text-image conversations rather than industrial troubleshooting (Chi et al., 2023).

This suggests that “IM-Chat” now functions less as a single canonical architecture than as a family of chat-mediated interaction paradigms. Within that family, the injection-molding IM-Chat framework is notable for combining multilingual conversation, RAG over industrial documents, and diffusion-based process recommendation in a single modular system (Lee et al., 21 Jul 2025).