DU-AI: A Privacy-Preserving Korean Infertility-Counseling Chatbot with Efficient On-Premise Retrieval-Augmented Generation

The Republic of Korea now records one of the lowest total fertility rates in the world. In this context, the number of couples turning to assisted-reproduction care has grown year after year [1]. For these couples, treatment is particularly demanding because it places two different kinds of strain on the same person at the same time. On one side, there is a continuous stream of clinical details to manage, such as hormone schedules, the timing of embryo transfer, and the interpretation of β-hCG measurements. Each of these has a direct clinical consequence. On the other side, there is a prolonged emotional load. It accumulates from uncertainty, from repeated treatment cycles, and from social and family pressure that the sensitivity of the subject makes difficult to discuss. A counselor who can address both aspects simultaneously, with accuracy and empathy, is precisely what such patients require, yet such support remains scarce. Moreover, because the topic is highly private, many patients are reluctant to contact public services at all. This gap is precisely what the conference theme, Machinanity for Humanity, addresses. A technology in this domain must be both technically reliable and emotionally supportive. A conversational assistant built on a large language model is an obvious candidate to fill that gap. A general-purpose model applied without safeguards, however, exhibits four critical failure modes in this domain. The model hallucinates numbers. Obstetric questions frequently depend on exact calendar arithmetic, and language models are known to reason unreliably about dates and quantities [12]. A due date or a hormone doubling time computed by the model therefore cannot be trusted. The model also drifts across languages. A model serving Korean users tends to leak Chinese characters, stray English, its own reasoning scaffolding, and even fragments of retrieved documents into the reply the user reads. Distress is another failure mode. A counseling agent must recognize expressions of suicidal thinking and respond with real crisis resources, and it must not fabricate sympathy for feelings the user never expressed. Privacy is the final failure mode. A cloud API transmits intimate reproductive-health data outside the device and cannot be deployed inside a clinic or a welfare center. Hosting the model locally appears to solve this problem. A naive local deployment, however, corrupts its own output as soon as multiple users connect concurrently. We designed DU-AI to address these failures. It is a Korean infertility-counseling chatbot that runs entirely on-premise and does not depend on the language model at every stage. Its guiding principle is to invoke the model only when no deterministic stage can produce the answer. Tasks the model performs poorly, namely exact computation, safety judgments, and language hygiene, are handled deterministically before the model is consulted. Tasks the model performs well, namely fluent Korean explanation and empathy, are grounded in retrieved evidence. First, we propose a deterministic clinical-calculator layer that answers the numeric questions of reproductive medicine with symbolic formulas instead of learned approximations. Second, we propose a context-aware safety-and-hygiene pipeline. It routes crises, manages conversational register and follow-up, decides in both directions when to answer or decline, and cleans the model's Korean before it is shown. Third, we propose a serving architecture that keeps the whole system on a single consumer GPU while handling concurrent users without token corruption. We validate each contribution with its own quantitative experiment. To our knowledge, DU-AI is the first Korean infertility-counseling agent to combine deterministic clinical computation, crisis-aware bidirectional safety with Korean output hygiene, and corruption-free concurrent on-premise serving in one system. The remainder of this paper reviews related work, describes the proposed architecture and our three contributions, and reports the experiments that quantify each of them.

Research on medical conversational agents has produced capable systems, but they target a different setting from ours. ChatDoctor [7] fine-tunes an open LLM on medical dialogue, and clinical-knowledge models in the Med-PaLM line [8] demonstrate strong general medical question answering. These systems, however, are built for English and assume cloud-scale inference. They make no guarantee about the correctness of a computed number, and they carry no explicit layer for handling a user in psychological crisis. None of these properties transfers directly to a private, Korean, welfare-oriented deployment. Retrieval-augmented generation is the standard remedy for factual hallucination, and it underlies our grounding of empathetic explanation. By conditioning the generator on retrieved passages, RAG measurably reduces fabricated content on knowledge-intensive tasks [2], [6]. Its limitations, however, correspond directly to two of our contributions. Retrieval can supply a fact that exists in some document, but it cannot supply the result of a calculation that appears in no document, such as a due date. It also does nothing to stop the generator itself from drifting out of Korean. A parallel line of work adds programmable guardrails around a model to keep a conversation safe and on topic [9]. These frameworks are valuable, but in practice they tend to be one-directional. They block disallowed topics. They do not balance the opposite risks of turning away a legitimate but unusual question and of drifting off purpose. They are also not adapted to the hygiene problems of Korean generation, or to the high-stakes case of routing a reproductive-health crisis. Our topic gate and crisis layer are designed to address exactly these gaps. Finally, the components we build on for Korean understanding and efficient local inference are individually mature. High-quality Korean dense retrievers make accurate in-language retrieval practical [3], [4], [5]. Quantized local runtimes with continuous batching make it feasible to run a capable model on a single GPU [10]. What remains largely unaddressed in published counseling prototypes is the integration. Korean retrieval, symbolic clinical reasoning, and a safety layer must be combined in one on-premise stack. Crucially, that stack must also serve concurrent users without the token-stream corruption that a shared model context produces. Table 1 places the proposed system against representative systems along seven capability axes and shows that ours is the only design to satisfy all of them.

3.1 System Overview The proposed DU-AI is a single-node service composed of three parts. It combines a FastAPI orchestrator, a local dense-retrieval index built with the KURE-v1 Korean encoder [4] over a curated corpus, and a locally hosted quantized language model (Qwen3-14B in Q5_K_M form [3]) served by a batching local runtime [10]. Figure 1 traces a user turn through the proposed system and illustrates our design principle. The language model sits at the very end of the pipeline. A turn reaches the model only after a series of deterministic stages has had the opportunity to answer or redirect it. Each turn is first examined for signs of crisis and, failing that, for simple social speech. It is then offered to the clinical-calculator router. If no calculator applies, the turn passes through the bidirectional topic gate. Only then is it embedded, matched against the retrieval index, and handed to the batched generator. Finally, the raw token stream is passed through the output-hygiene filters before any text is displayed to the user. Our three contributions correspond to the highlighted stages of Fig. 1, namely the calculator layer (C1), the safety and hygiene stages (C2), and the thread-bridge over batched generation (C3). Because every stage runs on one on-premise GPU, no part of the exchange leaves the device. Fig. 1. DU-AI system architecture. The novel contributions are tagged C1 (clinical calculators), C2 (safety and output hygiene), and C3 (concurrent on-premise serving). 3.2 Deterministic Clinical Calculators (C1) Many questions in infertility counseling are computational rather than factual. When a patient asks for her due date after transferring a five-day blastocyst on a particular date, she is not asking for a fact that any document contains. She is asking for the result of a calculation. Computing that result correctly requires several exact steps. The transfer date must be treated as an estimated point of conception, and the embryo's culture age must be included. The standard 266-day obstetric rule must then be applied, and the total must be mapped back onto the calendar. Language models are unreliable at this chain of date operations. They do not reliably know the current date, and they do not perform multi-month arithmetic dependably. Our system removes this computation from the model entirely. When a turn is recognized as one of the reproductive-medicine computations, it is routed to a small symbolic function. The function performs the arithmetic exactly and then expresses the answer as an ordinary Korean sentence. The proposed layer covers the five computations most frequently required in routine counseling. It estimates the expected date of delivery with Naegele's rule [11], derives the gestational age from the last menstrual period or the transfer date, predicts the next menstruation from the cycle length, and counts the days since an embryo transfer. Given two β-hCG values and their interval, it also reports the hormone's doubling time with a plain-language interpretation. A key design choice in our system is conservative routing. A calculator may act only when the current turn expresses the intent and provides the quantities the formula requires, either directly or recoverably from the immediate context. When a question is ambiguous or under-specified, it is passed on to retrieval rather than answered with a fabricated value. Two properties follow from our design. Because every value is produced by an explicit formula and not by the model, each answer is correct by construction and can be audited against the formula. In addition, our system openly states its default assumptions, such as a five-day blastocyst for an unstated embryo age or a twenty-eight-day cycle for an omitted cycle length. The deterministic core therefore maintains a natural conversational tone. 3.3 Context-Aware Safety and Output Hygiene (C2) Safety in a counseling setting cannot be left to the model's discretion. The proposed system therefore enforces it in several deterministic layers that act before any text is generated. The most important layer concerns crisis. When a turn expresses suicidal or self-harming thoughts, our system does not rely on the model to generate the response. It returns a fixed message that acknowledges the user's pain and provides the national crisis-line numbers. The principle is that a response of this gravity must be exact and must never depend on a stochastic generator. Another layer manages conversational register. A brief acknowledgment, such as a short word of agreement or thanks, is met with a fixed, register-consistent reply. This prevents the model from fabricating context or shifting between polite and casual registers. A turn that expresses distress without acute danger is instead guided toward grounded, empathetic counseling rather than a clinical fact. The same layer resolves follow-up questions by attaching them to the last substantive answer, not to a trivial acknowledgment that may occur between them. It also exploits the head-final structure of Korean sentences, so that a modifier carried over from the previous answer does not divert retrieval from the intended topic. A further layer is a bidirectional topic gate, and its bidirectionality is its defining property. Rejecting every turn that is not strictly medical would exclude legitimate users. Accepting every turn would allow the assistant to deviate from its intended scope. Our gate instead admits a turn whenever retrieval is confident, or whenever the turn is otherwise related to infertility, emotional well-being, or everyday health. It issues a polite refusal only when a turn is clearly unrelated to any of these. Our final layer addresses language itself. A Korean-serving model still leaks foreign script, stray English words, its own reasoning scaffolding, and traces of the retrieved documents into what it writes. The generated stream is therefore passed through a chain of filters that removes each artifact in turn. The chain also restores stray Chinese characters embedded in Korean words to their Hangul equivalents. The person on the other side reads clean and consistent Korean rather than the model's unpolished output. 3.4 Privacy-Preserving Concurrent Serving (C3) Serving a local language model to multiple concurrent users from an asynchronous web server raises two problems that a naive implementation cannot handle. One is corruption. If every request streams its tokens from one shared model context, the interleaved streams overwrite the shared key-value cache. The model then begins to emit degenerate and repetitive tokens. The other is serialization. If generation is allowed to block the server's event loop, then requests are effectively processed sequentially regardless of the number of concurrent users. We answer the serialization problem with a thread-bridge. Each request's synchronous work, which covers generation and hygiene filtering, runs inside a worker thread. The worker thread hands finished tokens back to the request's coroutine through a queue. The event loop is never blocked, and many requests can be generating at the same instant. We answer the corruption problem by delegating generation to a local runtime that performs continuous batching with an independent slot for each request. Giving every request its own context removes the shared-cache corruption outright, and processing the concurrent requests together is what raises throughput. Query embedding is delegated to a thread pool for the same non-blocking reason. Because our entire stack runs on a single consumer GPU with no external service in the loop, the intimate reproductive-health information a user shares never leaves the machine. This property is a prerequisite for deployment inside a clinic or a welfare center.

4.1 Experimental Setup All experiments run on one node with an NVIDIA RTX 4090 (24 GB). The generator is Qwen3-14B in Q5_K_M quantization, served by Ollama with flash attention and a q8_0 key-value cache. Retrieval uses KURE-v1 embeddings over a corpus of 4,839 Korean infertility FAQ chunks compiled from 826 and 4,790 curated items. The pure-LLM baseline is the same Qwen3-14B model answering directly, without the calculator, gate, or filter layers. This ablation-style baseline isolates the effect of the proposed orchestration, because the baseline and our system share the identical underlying model. A comparison against proprietary cloud LLMs is excluded by design, because such services violate the on-premise privacy requirement of our deployment target. We measure every number reported below on this hardware, and we design each of the three experiments to isolate exactly one of our contributions. 4.2 Clinical-Numeric Reliability (C1) We built a suite of 150 Korean numeric queries. It contains forty on the due date, forty on gestational age, twenty-five on the next menstruation, twenty-five on days since transfer, and twenty on β-hCG doubling. Each query is paired with a ground truth derived independently from the governing formula. To ensure a fair comparison, the Qwen3-14B baseline is given the current date in its prompt. Without it, the date-relative tasks would be impossible for any model. For reproducibility, the evaluation harness fixes the reference date and uses deterministic decoding with temperature zero. Figure 2 and Table 2 report exact-match accuracy. The baseline results reveal a consistent pattern. Even when given today's date, the model is essentially unable to compute a due date. It scores zero and deviates by an average of approximately 218 days. It also performs poorly on the logarithmic β-hCG calculation, with an accuracy of ten percent. It succeeds only on the short same-month date additions required for the next-menstruation and days-since-transfer tasks. Our deterministic layer, by contrast, is correct on every task, lifting overall exact-match from 30.7% to 100%. Its routing is also perfect. Our calculators fire on all 150 numeric turns, and never once on a 25-turn control set of ordinary counseling questions. The experiment thus confirms C1 directly. The quantities that matter most in this domain are precisely those the baseline computes incorrectly, and our symbolic layer computes all of them correctly. Fig. 2. Clinical-numeric exact-match accuracy per task (n = 150) for the Qwen3-14B baseline and the proposed DU-AI. TABLE 2 Clinical-numeric accuracy of the Qwen3-14B baseline and DU-AI (n = 150) Task n Pure-LLM exact-match Pure-LLM MAE (days) DU-AI exact-match EDD (due date) 40 0.0% 217.5 100% Gestational age 40 10.0% — 100% Next menstruation 25 76.0% 7.0 100% Days-post-transfer 25 84.0% — 100% β-hCG doubling 20 10.0% — 100% Overall 150 30.7% — 100% Calculator routing precision and recall are both 1.00. Correctness on routed turns is 100% by construction. 4.3 Safety and Output Hygiene (C2) This experiment has three parts. For crisis, we assembled thirty turns expressing ideation, both explicitly and in paraphrase, alongside ninety-four non-crisis turns that include ordinary emotional distress. For the topic gate, fifty on-topic and thirty off-topic turns are judged against a similarity-threshold-only baseline. For hygiene, thirty generated responses are compared before and after the filter chain, on content language-purity and on the rate of any visible artifact. Figure 3 and Table 3 collect the results. On crisis, our system routes at 0.93 recall with perfect precision, producing no false alarms across the ninety-four non-crisis turns. The unmodified Qwen3-14B baseline provides a crisis-hotline number in only 0.03 of cases and is therefore unsafe in practice. On topic control, our gate answers every legitimate counseling question. Its recall is 1.00, against 0.82 for a threshold baseline that wrongly refuses nearly a fifth of real users. This property is essential in a welfare service, where rejecting a distressed user carries a far higher cost than tolerating a mildly off-topic answer. The accompanying drop in precision is a deliberate, recall-oriented trade-off, and the full precision, recall, and F1 values are reported in Table 3. On hygiene, our filter chain delivers perfectly clean Korean and zero user-visible artifacts on the thirty-response set. On an extended sample of sixty responses, content purity remains 98.3% with a 1.7% artifact rate. The raw stream, by contrast, always carries reasoning scaffolding, and it leaks foreign script or English in a minority of answers. Together these results establish C2. Fig. 3. Safety and output-hygiene metrics for the Qwen3-14B baseline and the proposed DU-AI. TABLE 3 Safety and output hygiene (measured) Metric Baseline (Qwen3-14B) DU-AI (ours) Crisis hotline recall 0.03 0.93 Crisis precision (no false alarm, 94 non-crisis) — 1.00 Topic-gate recall (real questions not refused) 0.82 1.00 Topic-gate precision / F1 0.95 / 0.88 0.75 / 0.86 Korean content purity (no drift) ≈90–97% 100% User-visible artifact rate 100% 0.0% 4.4 Concurrent Serving (C3) We issued eight chat requests simultaneously, with a fixed generation length and identical prompts. A serial configuration with one slot is equivalent to naive shared-context serialization. We compare it against the batched configuration with four parallel slots on the same GPU. Figure 4 and Table 4 report the outcome. Batching with our thread-bridge reduces the wall-clock time to complete all eight requests from 74.1 to 33.0 seconds. It raises aggregate throughput from 12.9 to 28.4 tokens per second, a 2.24× improvement. At the same time, it removes the token-interleaving corruption entirely. None of the forty concurrent responses degenerates, whereas the shared-context baseline visibly does. This confirms C3 on commodity hardware, which is the deployment target. Fig. 4. Concurrency for eight simultaneous requests on one RTX 4090 under serial and batched serving. TABLE 4 Concurrency for eight simultaneous requests on a single RTX 4090 Configuration Wall-clock Aggregate throughput Mean latency Degenerate repetition Serial / naive shared-context 74.1 s 12.9 tok/s 43.5 s present (interleaving) DU-AI (batched + thread-bridge) 33.0 s 28.4 tok/s 23.5 s 0 / 40 4.5 End-to-End Composition We evaluated the proposed system with all three layers enabled end to end. Retrieval-grounded answers over a held-out set of conversations were clean Korean, numerically correct, and appropriately declined or crisis-routed where required. The contributions therefore compose without regressing one another. Figure 5 shows our deployed system in operation. A user enters through a landing screen (a) and completes a two-step onboarding. A client profile that personalizes counseling is collected and kept strictly on the device (b), and a knowledge-base panel confirms the 4,839 indexed chunks used in our experiments (c). In the counseling view (d), the assistant responds to emotional distress with grounded Korean empathy, and retrieved educational media open in a full-screen viewer (e). The interface thus reflects the properties quantified above, namely on-device privacy, clean grounded generation, and retrieval that enriches the counseling dialogue.

We have presented DU-AI, an on-premise Korean infertility-counseling chatbot whose neuro-symbolic orchestration converts a general language model into an assistant that is both clinically reliable and empathetic. Our three contributions are deterministic clinical calculators, a context-aware safety-and-hygiene pipeline, and corruption-free concurrent on-premise serving. We support each contribution with a targeted, quantitative experiment. Exact-match on clinical numerics rises from 30.7% for a date-aware baseline to 100%. Our safety pipeline reaches perfect crisis precision at 0.93 recall, refuses no legitimate question, and produces perfectly clean Korean. Our serving design more than doubles concurrent throughput while eliminating token corruption. The unifying principle of our design is a deliberate division of labor. Computation is delegated to symbolic code, safety to deterministic routing, and language hygiene to filters. The language model is thereby reserved for grounded, empathetic explanation, the task for which it is best suited. In this sense, DU-AI is a concrete realization of the conference theme, Machinanity for Humanity. It is machine intelligence engineered, from the interface down to the GPU, around human welfare. Future work will extend the symbolic layer beyond its present five computations, add automatic checks on the freshness of the knowledge corpus, and carry out a clinician-in-the-loop study of counseling quality with real users.

This work was supported by the National Research Foundation of Korea (NRF) grant funded by Korean Government [Ministry of Science and ICT (MSIT)] under Grant RS-2026-25482392.