LeRAAT: Real-Time Aviation Advisor

• LeRAAT is an LLM-enabled framework that integrates live flight data, weather, and aircraft documentation to offer precise, context-specific pilot guidance.
• It employs a Retrieval-Augmented Generation pipeline over manuals and regulatory materials to ensure safety-aligned recommendations and reduced hallucinations.
• The modular architecture combines an X-Plane plugin, Relay Server, and GPT-4o backend to achieve sub-second response times in both VR and on-screen simulation environments.

Searching arXiv for the LeRAAT paper to ground the article in the cited preprint. LeRAAT is an LLM-Enabled Real-Time Aviation Advisory Tool that integrates a LLM with the X-Plane flight simulator to deliver real-time, context-aware pilot assistance during both routine and emergency operations. It is presented as a modular framework that combines live flight data, weather conditions, and aircraft documentation to generate recommendations aligned with aviation best practices and tailored to the particular situation. The framework employs a Retrieval-Augmented Generation (RAG) pipeline over aircraft type-specific manuals, performance specifications, emergency procedures, and aviation regulatory materials, and is demonstrated in both virtual reality and traditional on-screen simulation for pilot training, human factors research, and operational decision support (Schlichting et al., 5 Mar 2025).

1. Definition and scope

LeRAAT is described as a modular framework that embeds a LLM within a high-fidelity flight simulation environment to deliver context-specific guidance to pilots during both routine and emergency operations. Its stated objective is to support high-stakes decision-making in aviation emergencies by providing quick access to precise, context-specific information.

The framework is explicitly organized around three tightly coupled components: an X-Plane plugin serving as the user interface, a Relay Server for data processing, retrieval, and prompt assembly, and an LLM backend using GPT-4o. These components ingest live flight data, weather, and aircraft documentation, perform retrieval-augmented generation, and present concise, safety-guided recommendations with latencies on the order of one second. The paper characterizes this as a first step toward safely integrating LLM-based advisories into the cockpit (Schlichting et al., 5 Mar 2025).

A plausible implication is that LeRAAT is not positioned as an autonomous control system, but as a decision-support layer whose outputs are bounded by retrieved documentation and simulator-derived context.

2. System architecture and data flow

LeRAAT’s architecture comprises an X-Plane plugin, a Relay Server, and an LLM backend. The X-Plane plugin is integrated via XPPython3 into the simulator and supports both on-screen overlay and VR cockpit displays. It monitors master warnings and cautions, extracts ECAM messages, autopilot and autothrottle states, and transmits these parameters to the Relay Server. It also renders LLM responses in a paginated, readable format following Airbus’s “dark cockpit” philosophy.

The Relay Server receives flight-state messages from the plugin together with weather and airport-database queries. It preprocesses flight data and ECAM outputs, indexes and retrieves relevant document chunks, computes alternate-airport options, and constructs the final LLM prompt. It also maintains per-session conversation context to support both stateless and stateful LLM backends. The LLM backend currently uses GPT-4o as the generation decoder and is described as swappable with other LLMs or local models without changes to plugin or relay logic.

The operational data flow proceeds through an armed-to-active trigger, either pilot-initiated via a “Query” button or automatically upon a master warning or caution. The plugin packages aircraft state and ECAM messages and sends them to the Relay Server; the server then pulls weather and airport data, performs RAG against indexed manuals, assembles the prompt, forwards it to the LLM, and returns the response to the plugin for display. This architecture emphasizes modular separation between simulator integration, retrieval logic, and generation (Schlichting et al., 5 Mar 2025).

3. Retrieval-augmented generation pipeline

LeRAAT’s RAG module is described as grounding LLM outputs by retrieving the top 10 relevant text chunks from an external corpus. That corpus consists of aircraft type-specific manuals, performance specifications and charts, emergency-procedure checklists, and regulatory materials such as FAA directives and SOPs.

During document ingestion, PDFs are converted to structured text using pymupdf4llm. Overlapping chunks of approximately 500 words are created to preserve context. Each chunk $d_i$ is embedded with OpenAI’s text-embedding-3 to produce a vector $\mathbf{e}_i$, and all embeddings are stored in a FAISS vector index. At query time, the assembled metadata—formatted flight state, ECAM messages, and pilot query—is embedded into $\mathbf{e}_q$. Each document chunk is scored by cosine similarity,

$$s(q, d_i) = \cos(\mathbf{e}_q, \mathbf{e}_i) = \frac{\mathbf{e}_q \cdot \mathbf{e}_i}{\|\mathbf{e}_q\|\;\|\mathbf{e}_i\|}\,.$$

The top 10 chunks by $s(q,d_i)$ are concatenated into the generation prompt, with headers indicating document source and section. The paper’s wording that the module “ground-truts” LLM outputs indicates an explicit attempt to constrain generation to retrieved evidence and thereby reduce unsupported completions (Schlichting et al., 5 Mar 2025).

This suggests that the retrieval stage is not merely informational context injection, but the principal mechanism used to align advisory generation with aircraft-specific and regulatory source material.

4. Prompting, response structure, and timing

LeRAAT uses GPT-4o with no additional fine-tuning; conversation state is held by the Relay Server. The system prompt template defines the model as “a real-time aviation safety advisor” and injects structured context including flight level, airspeed, ECAM messages, weather at nearest alternates, and the retrieved manual excerpts. The prompt directs the LLM to ground responses exclusively on the retrieved excerpts, follow established Airbus check-and-action structures, and keep replies succinct, ideally under 100 words per bullet.

A worked example is given for a hydraulic-pump failure on final approach. At time $t_0$, the simulator emits a master warning and ECAM “HYD PUMP 2 FAIL” to the plugin. At $t_0 + 50$ ms, the plugin packages flight state and ECAM data and sends them to the Relay Server. At $t_0 + 150$ ms, the Relay Server embeds the query and retrieves the top 10 chunks, such as emergency hydraulic procedures and a degraded-braking chart. At $t_0 + 170$ ms, the prompt is assembled and sent to GPT-4o. At $t_0 + 1.1$ s, the LLM returns a recommendation; at $\mathbf{e}_i$0 s, the plugin displays advice in Active mode.

The reported typical latency breakdown is approximately 50 ms for data extraction, 100 ms for document retrieval, 20 ms for prompt assembly, 900 ms for LLM inference, and an end-to-end latency of approximately 1.07 s. The response-time distribution is expressed as

$\mathbf{e}_i$1

These timings place the advisory loop in the real-time regime claimed by the framework, while still making the LLM inference stage the dominant contributor to latency (Schlichting et al., 5 Mar 2025).

5. Evaluation and observed results

The initial user study involved three professional A320 pilots across two emergency scenarios: fuel imbalance and dual-hydraulic failure. The reported Time-to-Resolution metric is defined as $\mathbf{e}_i$2, and pilots following LeRAAT guidance closed out emergencies on average 25 % faster. Qualitative feedback includes the statement that “Providing synthesis of where you are, where you can go, and for which reasons… is something that’s missing in other systems but was provided here.” The details also state that all SMEs valued the automated diversion-airport ranking by runway length, weather, and distance.

The paper reports that no instances of hazardous or ungrounded advice were recorded and attributes this to the system’s RAG grounding, which prevented hallucinations observed in baseline LLM-only trials. It also provides an evaluation expression for retrieval accuracy,

$\mathbf{e}_i$3

Within the limits of the described study, these findings connect retrieval quality, advisory latency, and operational usefulness. A plausible implication is that the framework’s most salient contribution lies in synthesizing aircraft state, alternate options, and procedure-grounded guidance into a single advisory surface rather than merely retrieving isolated checklist fragments (Schlichting et al., 5 Mar 2025).

6. Applications, limitations, and future directions

The demonstrated applications are traditional on-screen simulation for pilot training and human-factors studies, fully immersive VR cockpit overlays for advanced crew-resource-management experiments, and ongoing research into real-time AI support in general aviation. The inclusion of both standard display and VR modes indicates that LeRAAT is framed not only as an operational prototype but also as a research instrument for simulator-based studies.

Several limitations are explicitly identified. The system lacks NOTAM integration and runway-direction specifics, and pilots requested dynamic NOTAM ingestion. The interface is text-only, and all SMEs recommended transitioning to voice input and output for reduced workload. The paper also notes remaining hallucination risk under rare or conflicting documentation, with planned guardrails including real-time confidence scoring and cross-validation against multiple document sources. Future corpus extensions are stated to include live NOTAM streams and live NOTAM-chart overlays (Schlichting et al., 5 Mar 2025).

These limitations delimit the current scope of the framework. This suggests that LeRAAT’s present design emphasizes document-grounded textual advisories under simulator conditions, while future work is expected to expand both the operational data feeds and the interaction modality.

7. Position within safety-critical decision support

LeRAAT is explicitly framed as a decision-support tool for safety-critical aviation contexts. Its modular RAG architecture, sub-second performance, and positive pilot feedback are presented as the basis for that framing. The system is designed to generate concise, prioritized advisories, to follow Airbus check-and-action structures, and to rely on retrieved excerpts from manuals and regulatory materials rather than unconstrained generation.

The most important technical distinction made in the source description is between the grounded RAG configuration and baseline LLM-only trials in which hallucinations were observed. In that sense, LeRAAT treats retrieval not as an auxiliary enhancement but as a safety mechanism intended to bound the advisory space to aircraft-specific manuals, performance charts, emergency procedures, FAA directives, and SOPs. A plausible implication is that the framework’s research significance lies in coupling live simulator telemetry with retrieval-constrained language generation, thereby enabling study of real-time AI assistance under controlled but operationally realistic conditions (Schlichting et al., 5 Mar 2025).