Skip to content

PROMPT.md

Latest commit

 

History

History
370 lines (289 loc) · 17.2 KB

PROMPT.md

File metadata and controls

370 lines (289 loc) · 17.2 KB
Raw

Spring Vision MCP — Sample Prompts

Sample prompts to exercise every tool exposed by the spring-vision MCP server, plus advice on what kind of media to use for each. You source the media yourself — the notes describe what makes a good test image.

Every capability has two variants:

  • *_b — raw image bytes (drop a local file path; the client uploads bytes)
  • *_u — image URL (any publicly reachable https:// URL)

When prompting, just give a URL or file path. The model will pick the right tool.

Faces & people

count_faces

Prompt: "How many faces are in this image? <url-or-path>"

Good test media:

  • Single front-facing portrait (sanity check: should return 1)
  • Group photo (5–20 faces, mixed angles)
  • Crowd photo (50+ faces) to see the upper limit
  • Photo with no people (should return 0)
  • Side-profile / partial-face shot to test detector robustness

Verified URLs:

extract_face_embeddings

Prompt: "Extract face embeddings from <url> and tell me the vector dimensions."

Good test media:

  • Two clear portraits of the same person from different angles — embeddings should be similar
  • Two portraits of different people — embeddings should diverge
  • A photo with multiple faces — should return one embedding per face

Verified URLs:

detect_emotions

Prompt: "What emotions are visible in <url>?"

Good test media:

  • Stock photos labeled "happy/sad/angry/surprised" face — easy ground truth
  • Movie still with strong emotional expression
  • Neutral-expression portrait as a control

Verified URLs:

detect_demographics

Prompt: "Estimate age and gender for the people in <path>."

Good test media:

  • Mixed-age group photo (child, adult, elderly)
  • Single portraits across age groups
  • Note: these models are statistically biased — don't use for anything consequential

Verified URLs:

detect_deepfake

Prompt: "Is this image a deepfake? <url>"

Good test media:

  • Real photo of a public figure (control — should classify as real)
  • Known deepfake samples from FaceForensics++ or the DFDC dataset
  • AI-generated portrait from thispersondoesnotexist.com (interesting edge case — synthetic but not a face-swap)

Verified URLs:

authenticate_access

Prompt: "Run access authentication on <path>."

Good test media:

  • Clear front-facing portrait, eyes open, neutral light
  • Same person in poor lighting / partial occlusion (test rejection)
  • Photo with no face (should deny)

Verified URLs:

Detection & recognition

detect_objects

Prompt: "What objects are in <url>? Give me bounding boxes."

Good test media:

  • Street scene (cars, pedestrians, bicycles, traffic lights) — COCO classes
  • Kitchen photo (bowl, cup, fork, microwave)
  • Living room (sofa, tv, laptop, person, cat/dog)
  • Sports photo (ball, racket, person)
  • Indoor photo with no recognizable COCO objects (should return empty/low confidence)

Verified URLs:

detect_hands

Prompt: "Find hands in <path> and return their positions."

Good test media:

  • Sign-language photo
  • Hands holding objects (close-up)
  • Two-person handshake
  • Photo with no hands visible (control)

Verified URLs:

detect_poses

Prompt: "Detect human poses in <url> and list the joint positions."

Good test media:

  • Yoga pose, single person, plain background
  • Athletic action (running, jumping, kicking)
  • Dance still
  • Multiple people in frame
  • Partially occluded body (test joint inference)

Verified URLs:

recognize_actions

Prompt: "What action is happening in <url>?"

Good test media:

  • Sports action stills (basketball dunk, soccer kick, swimming)
  • Cooking / typing / writing — everyday actions
  • Static portrait (control — should be low confidence or "standing")

Verified URLs:

detect_threats

Prompt: "Scan <path> for security threats."

Good test media:

  • Stock photos of weapons (knife, firearm) — clearly visible
  • Airport-style luggage X-ray images (if you can find labeled samples)
  • Crowd photo with no threats (control)
  • ⚠️ Use only legally-sourced/stock imagery

Verified URLs:

detect_vehicle

Prompt: "Detect vehicles in <url> and tell me the type and bounding box."

Good test media:

  • Street scene with multiple vehicle types (car, bus, truck, motorcycle)
  • Single sports car / sedan close-up (sanity check)
  • Bicycle or motorcycle (two-wheeler categories)
  • Aerial / parking lot shot (many cars, varied scale)
  • Landscape with no vehicles (control — should return empty)

Verified URLs:

detect_vehicle_damages

Prompt: "Find damage on any vehicle in <url> — classify type and severity."

Good test media:

  • Insurance-claim-style photos: visible dent, scratch, crack, broken glass, flat tire
  • Close-up of a single damaged panel (one clear defect)
  • Multi-damage shot (front-end collision)
  • Pristine vehicle (control — should return no damage or low-confidence)
  • ⚠️ Requires the bundled yolov11n-car-damage ONNX model; without it the tool fails loudly

Verified URLs:

classify_image

Prompt: "Classify <url> and give me the top 5 predictions with confidence."

Good test media:

  • Single dominant subject (one animal, one object) — ImageNet class
  • Iconic ImageNet examples: golden retriever, espresso cup, sports car, banana
  • Ambiguous/empty scene (sky, wall) to see low-confidence behavior

Verified URLs:

Text & metadata

extract_text (OCR)

Prompt: "Read all text in <url>."

Good test media:

  • Screenshot of a webpage or document
  • Photo of a printed page (book, receipt, business card)
  • Street sign / shop signage
  • Handwritten note (test handwriting limits)
  • Multi-language sample (English + non-Latin script)
  • Low-contrast / rotated text to test robustness

Verified URLs:

scan_barcode

Prompt: "Scan the barcode in <path> and decode it."

Good test media:

  • QR code (generate one at qr-code-generator.com pointing to any URL)
  • Product barcode (EAN-13/UPC) from a packaged item
  • Boarding pass PDF417
  • Multiple barcodes in one image
  • Blurry / partially occluded barcode (test ZXing's tolerance)

Verified URLs:

extract_image_metadata

Prompt: "Extract EXIF and GPS metadata from <path>."

Good test media:

  • Photo taken with a smartphone with location services enabled (rich EXIF + GPS)
  • DSLR shot (camera make/model, lens, exposure data)
  • Screenshot or downloaded JPEG from a CDN (typically stripped — should return minimal metadata)
  • PNG file (limited EXIF support — useful negative test)

Verified URLs:

Health & safety

detect_fall

Prompt: "Assess fall risk in <url>."

Good test media:

  • Standing person, upright posture (control — low risk)
  • Person lying on the floor (high-risk classification)
  • Person sitting on the ground (ambiguous case)
  • Sports action photo where body is horizontal mid-air (false-positive test)

Verified URLs:

analyze_stress

Prompt: "Analyze stress level in <path>."

Good test media:

  • Portrait with relaxed, smiling expression (low)
  • Portrait with furrowed brow, tense expression (high)
  • Stock "stressed at desk" photo
  • Note: this is heuristic — treat output as a vibes-check, not a clinical signal

Verified URLs:

detect_nsfw

Prompt: "Is <url> safe for work?"

Good test media:

  • Family-friendly stock photo (control — should be normal)
  • Beach/swimsuit photo (edge case for the model's threshold)
  • ⚠️ Don't source actual NSFW for testing in shared environments — trust the normal classification on benign inputs as your primary signal

Verified URLs:

estimate_heart_rate_f

Prompt: "Estimate heart rate from this sequence of frames: [<path1>, <path2>, ..., <pathN>]"

Good test media:

  • 30–300 consecutive frames extracted from a short video (1–10 seconds at 30fps) of a person's face holding still
  • Use ffmpeg -i input.mp4 -vf fps=30 frame_%04d.png to extract frames
  • Subject should be well-lit, face visible, minimal motion
  • Test with a recording where you know the reference HR (smartwatch reading)
  • ⚠️ rPPG is sensitive — expect noisy results without a proper recording setup

Quick sources for test media

  • Generic photos / stock: Unsplash, Pexels, Pixabay
  • Faces (synthetic, ethics-safe): thispersondoesnotexist.com
  • Object detection benchmarks: COCO dataset sample images (cocodataset.org)
  • Pose / action: MPII Human Pose, AVA, Kinetics sample frames
  • OCR: Take a phone photo of any printed page; or use ICDAR sample images
  • Barcodes: Generate at qr-code-generator.com / barcode.tec-it.com
  • EXIF-rich photos: Anything straight off your phone (avoid CDN downloads — they strip metadata)
  • Deepfakes: FaceForensics++ samples (research access required)

Caveat — current platform state

The local jar runs on macOS arm64, but the bundled PyTorch native library is linux-x86_64. Tools that rely on PyTorch models (object detection, segmentation, pose, RetinaFace) currently fall back to synthetic results or generic detectors. The following are not affected and work fully:

  • OCR (extract_text)
  • Barcode (scan_barcode)
  • EXIF/GPS metadata (extract_image_metadata)
  • ONNX-based YOLO detectors where bundled models are used

If a tool consistently returns suspiciously generic output, suspect the PyTorch native fallback rather than the input image.