Understanding AI in Vecteur

EU AI Act Art. 4 — AI Literacy

How Vecteur uses AI, what it can and cannot do, and when to verify outputs.

How Vecteur Works

Vecteur combines a physics computation engine with a large language model (LLM). When you ask a question, the AI assistant reasons about your request and calls specialized engineering tools that run validated physics computations. Results are cross-checked against reference data.

The AI model is provided by third-party cloud services selected in Settings (Anthropic or Mistral). It orchestrates the analysis but does not perform physics calculations itself — those are handled by Vecteur's own validated engine.

The AI assistant is a tool to augment your engineering judgment, not a replacement for it. All AI outputs are advisory. The engineer retains full decision authority.

What the AI Assistant Can Do

Mission Design

Design orbits, constellations, and ground networks through natural language. The AI calls validated physics tools to compute results.

  • Orbit selection and propagation
  • Constellation architecture
  • Ground station and target placement

Engineering Analysis

Run trade studies, system studies, and subsystem analyses. Results are validated against reference data.

  • Power, thermal, and link budgets
  • Coverage and access analysis
  • Requirements traceability

Reasoning & Exploration

Explore design alternatives, evaluate trade-offs, and understand engineering concepts conversationally.

  • Design space exploration
  • Pros/cons analysis
  • Concept explanations

Documentation

Generate design justifications, compliance narratives, and technical summaries from your workspace data.

  • Design rationale
  • Compliance documentation
  • Technical summaries

Important Limitations

Knowledge Cutoff

The AI model's training data has a cutoff (February 2025). It does not access live satellite databases, current debris catalogs, or real-time weather data. Recent standards or missions published after this date may not be known.

Hallucination Risk

The AI may generate plausible-sounding but incorrect information — citing non-existent papers, inventing parameters, or stating false facts with confidence. Always verify critical outputs independently.

Physics Results Are Advisory

While Vecteur's physics engine is validated against reference tools (SPICE, Orekit), the AI's reasoning and interpretation of results can introduce errors. Cross-check mission-critical outputs against independent references.

Workspace Access Is Sandboxed

The AI can read and write files within your project workspace, but access is isolated to your project. It cannot access other users' data, external systems, or files outside the sandbox.

Complex Problems

The AI performs well on focused tasks but may struggle with very large multi-step problems. Break complex analyses into smaller, verifiable steps.

Not a Substitute for Review

AI-generated designs, analyses, and documentation should always undergo engineering review. The AI cannot provide legal, regulatory, or safety-critical assurance.

When to Verify AI Outputs

Output TypeTrust LevelVerification Required
Explanations & ConceptsHighCross-check for gaps; ask follow-up questions
Code & ScriptsMediumAlways test with real data. Check logic, edge cases, error handling.
Numerical EstimatesLowAlways validate independently. Use reference tools or hand calculations.
Design RecommendationsMediumEvaluate trade-offs; cross-check against standards (ECSS, NASA-STD-7001, etc.)
Historical Facts & ReferencesLowAlways verify citations. The AI may invent plausible-sounding references.
Complex Multi-Step AnalysisLowBreak into steps; verify each step independently.

Verification Checklist

  • Does it make sense physically? Apply dimensional analysis, order-of-magnitude checks.
  • Can you trace the reasoning? Ask the AI to explain each step; verify logic.
  • Are there conflicting sources? If the AI contradicts a reference (NASA, ECSS, your standards), investigate.
  • Is the confidence justified? Watch for hedging language: "likely", "probably", "typically". These signal uncertainty.
  • What data isn't visible? Ask what assumptions the AI made. Challenge questionable ones.
  • Can you reproduce it? If the AI generates a result, can you recalculate it independently?

How Your Data is Handled

1

Your Query

Vecteur sends your question and relevant engineering context to the AI provider. Email addresses are stripped before transmission.

2

AI Processing

The AI model reasons about your request and calls Vecteur's physics tools. Results flow back through our backend.

3

Retention

The AI provider does not use your data for model training. Vecteur retains chat history for 12 months, then anonymizes it. See Privacy Policy.

4

Your Control

You can export or delete your data at any time (GDPR Art. 15-22) via Settings or by contacting privacy@vecteur.space.

Data Minimization

Vecteur applies data minimization before sending context to the AI provider.

Your query and design parameters
Workspace context relevant to your request
Email addresses (stripped automatically)
Other users' projects or data
IP addresses or raw browser data

Best Practices for AI-Assisted Engineering

Question Numerical Results

The AI reasons well but can make errors in interpretation. Cross-check key numbers against hand calculations or reference tools.

Provide Context

Include mission phase, constraints, and assumptions. The more specific your question, the better the result.

Iterate and Refine

Use follow-up questions to refine analyses. Multi-turn conversations preserve context and improve results.

Explore Before Committing

Use the AI for trade-space exploration and concept generation. Validate final designs through formal review.

Cross-Check Against Standards

For critical decisions, reference ECSS, NASA-STD, or your company standards. The AI is a starting point, not the final word.

Audit the Reasoning Chain

Before committing to a design, trace the full reasoning. Can you independently reproduce the key results?

Technical Details

AI Provider

Vecteur can use Anthropic or Mistral as AI provider. Model tier is selected automatically by workflow stage.

All model details (ID, version, provider, tokens) are logged with each interaction for auditing.

Physics Engine

Physics computations run on Vecteur's own validated engine, cross-checked against analytical solutions and reference databases (SPICE, Orekit).

The AI orchestrates the analysis but does not perform physics calculations itself.

Provenance

Every AI-generated result carries metadata: model ID, provider, timestamp, and an ai_generated: true flag.

This satisfies EU AI Act Art. 50 transparency requirements.

Execution Control

You can stop a running analysis at any time. Long-running computations maintain a heartbeat connection.

Compliance

This guide fulfills EU AI Act Art. 4 (AI Literacy). See our Compliance page for full GDPR and AI Act documentation.