プロンプト
An RAG architecture diagram: a user question embedding, vector search over a knowledge base, retrieved passages joining the original question in the prompt, and the model answering with citations back to the sources. Label "QUERY", "VECTOR SEARCH", "RETRIEVED CONTEXT", "ANSWER", "CITATIONS". Title: "GROUNDING THE ANSWER".. Workflow direction: Create one publication-quality science-communication illustration in a single generation. It must demonstrate both visual craft and explanatory power: a compelling central subject, a clear reading order, multiple meaningful layers of information, and a polished editorial composition that still reads at thumbnail size.
Follow the user's requested entities, relationships, sequence, viewpoint, audience, and quoted text exactly. Pick the figure grammar that best serves the topic and state it through the rendering itself: anatomy or specimen plate, cutaway or cross-section, molecular or signaling mechanism, step-by-step process or workflow, side-by-side comparison, system or network architecture, chart or plot, instrument diagram, or narrative scene. Match the rendering technique to that grammar — flat clean vector for flows and architecture, dimensional molecular rendering for micro structures, semi-realistic tissue or material rendering for cutaways, cinematic depth only for scenes — instead of forcing every topic into one house style.
Default canvas is a clean white or near-white field like a journal figure: no decorative frame, plate border, aged paper texture, or faux page furniture unless the user explicitly asks for them. Label the figure the way a journal would: direct horizontal text joined to its structure by a thin non-crossing leader line. Switch to numbered markers with a compact key only when the subject is too dense for direct labels, and include a legend only when color or symbols encode meaning. One short title, 4–8 essential labels, purposeful arrows for real movement, sequence, reaction, or causal flow with consistent arrow grammar, and at most one magnified inset that adds genuine understanding.
Render requested English or Chinese text as sharp, horizontal, high-contrast, correctly spelled text; never paraphrase quoted labels or leak prompt instructions into the artwork. Reserve a generous 10–15% outer safe area on all four sides; every title, label, leader line, legend and inset must be fully visible inside that safe area, never touching or crossing the canvas edge. Prefer fewer complete labels over cropped or microscopic copy.
Scientific accuracy outranks decoration. Preserve topology, orientation, relative position, sequence, direction and scale cues; never invent anatomy, structures, measurements, provenance, causal relationships or pseudo-data. When literal scale cannot fit, visibly signal a schematic or not-to-scale treatment. The result must be self-contained and credible as a figure a researcher could paste into a manuscript, thesis, grant, poster, textbook, classroom explainer, or public-science article.. Avoid: invented facts, misleading anatomy, false precision, generic SaaS infographic, empty decorative icons, decorative clutter, overlapping arrows, crossed leaders, duplicated labels, garbled text, misspelled labels, tiny unreadable text, cropped text, clipped legend, edge-touching callout, prompt text, logo, citation, watermark
