{"passport":{"unfragile":{"@version":"1.0","version":"2026-05","artifact":{"id":"pypi_pypi-segment-anything","slug":"pypi-segment-anything","name":"segment-anything","type":"repo","url":"https://pypi.org/project/segment-anything/","page_url":"https://unfragile.ai/pypi-segment-anything","categories":["model-training"],"tags":[],"pricing":{"model":"open_source","free":true,"starting_price":null},"status":"active","verified":false},"capabilities":[{"id":"pypi_pypi-segment-anything__cap_0","uri":"capability://image.visual.zero.shot.image.segmentation.with.prompt.based.masks","name":"zero-shot image segmentation with prompt-based masks","description":"Generates precise object segmentation masks from images using a vision transformer encoder-decoder architecture that accepts flexible prompts (points, bounding boxes, text descriptions, or mask hints). The model uses a two-stage process: an image encoder processes the full image into embeddings, then a lightweight mask decoder generates segmentation masks conditioned on prompt embeddings, enabling real-time inference without task-specific fine-tuning.","intents":["segment arbitrary objects in images without training on task-specific datasets","generate multiple mask hypotheses for ambiguous objects using different prompt strategies","integrate segmentation into downstream vision pipelines without retraining","build interactive segmentation tools where users click or draw to refine masks"],"best_for":["computer vision engineers building general-purpose segmentation systems","researchers prototyping vision applications without labeled training data","teams building interactive annotation or image editing tools","developers integrating segmentation into multi-modal AI systems"],"limitations":["requires high-resolution images (1024x1024 recommended) for optimal accuracy; performance degrades on small objects or cluttered scenes","prompt quality directly impacts output quality — ambiguous prompts may generate multiple competing masks requiring disambiguation logic","inference latency ~500ms per image on CPU, ~50-100ms on GPU; batch processing not optimized for real-time video","model weights are ~375MB (ViT-B) to ~1.2GB (ViT-L); requires significant memory for edge deployment","struggles with transparent objects, reflections, and fine-grained boundaries; post-processing often needed for production use"],"requires":["Python 3.8+","PyTorch 1.9+ with CUDA 11.0+ (for GPU acceleration)","NumPy, Pillow for image I/O","4GB+ RAM minimum; 8GB+ recommended for batch processing","GPU with 6GB+ VRAM for real-time inference (NVIDIA/AMD/Intel Arc supported via PyTorch)"],"input_types":["image (PNG, JPEG, BMP, TIFF)","point prompts (x, y coordinates as tuples)","bounding box prompts (x_min, y_min, x_max, y_max)","mask hints (binary or soft masks as numpy arrays)","negative prompts (points/boxes to exclude from segmentation)"],"output_types":["binary segmentation masks (H x W boolean arrays)","soft masks with confidence scores (H x W float [0, 1])","multiple mask hypotheses ranked by model confidence","IoU predictions per mask (intersection-over-union quality estimates)"],"categories":["image-visual","computer-vision"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-segment-anything__cap_1","uri":"capability://image.visual.multi.prompt.mask.disambiguation.and.refinement","name":"multi-prompt mask disambiguation and refinement","description":"Generates multiple candidate segmentation masks for a single image and ranks them by model confidence, allowing users or downstream systems to select the most appropriate mask or iteratively refine masks by adding positive/negative prompts. The decoder outputs IoU predictions alongside masks, enabling confidence-based filtering and automatic selection of high-quality masks without manual review.","intents":["handle ambiguous segmentation scenarios where object boundaries are unclear or multiple valid interpretations exist","enable interactive refinement workflows where users iteratively add prompts to improve mask quality","automatically select the highest-confidence mask for downstream tasks without manual intervention","build confidence-aware pipelines that flag low-confidence segmentations for human review"],"best_for":["interactive annotation platforms requiring user feedback loops","quality assurance systems that need confidence metrics for mask validation","autonomous systems that must handle ambiguous inputs gracefully","research teams studying segmentation robustness and failure modes"],"limitations":["IoU predictions are model estimates, not ground-truth accuracy; can be overconfident on out-of-distribution images","iterative refinement adds latency — each prompt requires a full forward pass through the decoder (~50-100ms per iteration)","no built-in mechanism to detect when additional prompts won't improve masks; users must implement stopping criteria","mask quality still depends on prompt quality; poor prompts generate poor masks regardless of disambiguation logic"],"requires":["Python 3.8+","PyTorch 1.9+","NumPy for mask manipulation","GPU recommended for interactive workflows (sub-100ms latency requirement)"],"input_types":["image (PNG, JPEG, etc.)","initial prompt (point, box, or mask)","refinement prompts (additional points/boxes to add or subtract)","confidence threshold (float [0, 1] for filtering masks)"],"output_types":["ranked list of candidate masks with IoU scores","selected mask (highest confidence or user-chosen)","mask quality metrics (IoU estimate, stability across prompts)","refinement suggestions (recommended next prompts to improve mask)"],"categories":["image-visual","planning-reasoning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-segment-anything__cap_10","uri":"capability://image.visual.semantic.and.instance.segmentation.with.class.agnostic.masks","name":"semantic and instance segmentation with class-agnostic masks","description":"Generates class-agnostic segmentation masks (no class labels) that can be post-processed to produce semantic or instance segmentation by applying clustering, connected-component analysis, or external classifiers. The model outputs masks without semantic information, enabling flexible downstream classification and enabling use cases where class information is not available at inference time.","intents":["segment all objects in an image without knowing object classes in advance","generate instance masks that can be classified by external models","enable semantic segmentation by applying classifiers to instance masks","support open-vocabulary segmentation by combining with CLIP or other classifiers"],"best_for":["open-vocabulary or zero-shot segmentation pipelines","systems combining segmentation with external classifiers","instance segmentation without class labels","research on class-agnostic segmentation"],"limitations":["class-agnostic masks lack semantic information; downstream classification required for semantic segmentation","no built-in instance ID assignment; requires external clustering or connected-component analysis","cannot leverage class-specific priors; may segment background or non-objects","post-processing (clustering, classification) adds latency and complexity"],"requires":["Python 3.8+","PyTorch 1.9+","SciPy for connected-component analysis","Optional: external classifier (CLIP, ResNet, etc.) for semantic information"],"input_types":["image (PNG, JPEG, etc.)","prompt (point, box, or mask)","optional: external classifier for semantic labels"],"output_types":["class-agnostic mask (H x W boolean array)","instance masks (list of H x W arrays after clustering)","semantic labels (if external classifier provided)","instance metadata (area, centroid, confidence)"],"categories":["image-visual","data-processing-analysis"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-segment-anything__cap_2","uri":"capability://image.visual.efficient.image.encoding.with.frozen.vision.transformer.backbone","name":"efficient image encoding with frozen vision transformer backbone","description":"Pre-computes and caches image embeddings using a frozen ViT encoder (ViT-B, ViT-L, or ViT-H variants), enabling fast mask decoding for multiple prompts on the same image without re-encoding. The encoder processes images at 1024x1024 resolution and outputs 64x64 feature maps; embeddings are cached in memory or disk, reducing per-prompt latency from ~500ms to ~50-100ms.","intents":["segment multiple objects in the same image with minimal latency overhead","build interactive tools where users click multiple times on one image without waiting for re-encoding","batch-process prompts on the same image efficiently by reusing cached embeddings","reduce computational cost for multi-prompt workflows by amortizing encoding cost"],"best_for":["interactive annotation and image editing applications","batch processing systems with multiple prompts per image","resource-constrained environments where encoding is the bottleneck","real-time segmentation pipelines requiring sub-100ms latency per prompt"],"limitations":["cached embeddings consume ~50-100MB per image (depending on ViT variant); limits in-memory caching for large datasets","frozen encoder cannot adapt to domain-specific image characteristics; fine-tuning requires retraining the entire model","embedding cache must be invalidated if image is modified; no incremental update mechanism","ViT-H variant requires 24GB+ VRAM for encoding; ViT-B/L more practical for deployment"],"requires":["Python 3.8+","PyTorch 1.9+","GPU with 6GB+ VRAM (ViT-B) or 12GB+ (ViT-L/H)","Storage for cached embeddings (50-100MB per image)"],"input_types":["image (PNG, JPEG, BMP, TIFF)","image resolution (automatically resized to 1024x1024)"],"output_types":["image embeddings (64x64x256 for ViT-B, 64x64x768 for ViT-L)","cached embedding tensors (PyTorch or NumPy format)","embedding metadata (resolution, encoder variant, timestamp)"],"categories":["image-visual","data-processing-analysis"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-segment-anything__cap_3","uri":"capability://image.visual.batch.segmentation.with.heterogeneous.prompts","name":"batch segmentation with heterogeneous prompts","description":"Processes multiple images and prompts in batches, supporting mixed prompt types (some images with point prompts, others with boxes or masks) in a single forward pass. The implementation pads prompts to a fixed size and uses attention masking to ignore padding tokens, enabling efficient GPU utilization without requiring homogeneous prompt types across the batch.","intents":["segment large image collections with varying prompt types in a single batch job","maximize GPU utilization by batching heterogeneous prompts without padding overhead","reduce total inference time for multi-image, multi-prompt workflows","process datasets where different images require different prompt modalities"],"best_for":["data annotation pipelines processing thousands of images","batch segmentation jobs on cloud infrastructure (AWS, GCP, Azure)","research teams benchmarking segmentation across large datasets","production systems requiring throughput optimization"],"limitations":["batch size is limited by GPU memory; typical batch size 4-16 on 24GB VRAM","heterogeneous prompts require padding and masking, adding ~10-20% overhead vs homogeneous batches","no built-in distributed batching across multiple GPUs; requires external orchestration (Ray, Horovod)","batch processing loses interactive refinement capability; designed for static, pre-defined prompts"],"requires":["Python 3.8+","PyTorch 1.9+ with CUDA 11.0+","GPU with 12GB+ VRAM for batch size 8+","NumPy, Pillow for image I/O"],"input_types":["list of images (PNG, JPEG, etc.)","list of prompts (points, boxes, masks, or mixed)","batch size (integer, 1-64 depending on GPU memory)","prompt padding strategy (fixed size or dynamic)"],"output_types":["batch of segmentation masks (B x H x W boolean arrays)","batch of IoU predictions (B x num_masks float arrays)","batch metadata (image IDs, prompt types, processing time per image)"],"categories":["image-visual","automation-workflow"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-segment-anything__cap_4","uri":"capability://image.visual.automatic.mask.post.processing.and.refinement","name":"automatic mask post-processing and refinement","description":"Applies morphological operations (erosion, dilation, opening, closing) and contour-based filtering to refine raw model outputs, removing noise, filling holes, and smoothing boundaries. Post-processing is configurable and can be applied selectively based on mask quality estimates (IoU predictions), enabling automatic quality improvement without manual tuning.","intents":["clean up noisy segmentation masks from model output without manual editing","remove small spurious regions and fill holes in masks automatically","smooth jagged boundaries for better visual quality in downstream applications","apply quality-dependent post-processing (aggressive for low-confidence masks, minimal for high-confidence)"],"best_for":["production segmentation pipelines requiring high-quality outputs","image editing and annotation tools with automatic cleanup","quality assurance systems that need consistent mask quality","downstream tasks (instance tracking, 3D reconstruction) sensitive to mask quality"],"limitations":["morphological operations can distort fine details and thin structures; aggressive post-processing may remove valid small objects","contour smoothing adds ~50-100ms per mask; not suitable for real-time interactive workflows","post-processing parameters (kernel size, iterations) require tuning per domain; no automatic parameter selection","cannot recover information lost in model prediction; post-processing cannot fix fundamental segmentation errors"],"requires":["Python 3.8+","OpenCV (cv2) for morphological operations","SciPy for advanced contour processing","NumPy for mask manipulation"],"input_types":["binary segmentation mask (H x W boolean array)","soft mask with confidence scores (H x W float [0, 1])","post-processing configuration (kernel size, operation type, iterations)","quality threshold (IoU estimate for conditional post-processing)"],"output_types":["refined binary mask (H x W boolean array)","smoothed contours (list of (x, y) coordinates)","mask statistics (area, perimeter, compactness)","processing metadata (operations applied, time elapsed)"],"categories":["image-visual","data-processing-analysis"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-segment-anything__cap_5","uri":"capability://image.visual.multi.scale.segmentation.with.image.pyramid.processing","name":"multi-scale segmentation with image pyramid processing","description":"Processes images at multiple scales (0.5x, 1.0x, 2.0x original resolution) and combines predictions using ensemble voting or confidence-weighted averaging, improving robustness to scale variations and small object detection. The implementation reuses cached embeddings at the base scale and computes additional embeddings for upsampled/downsampled variants, trading memory for improved accuracy.","intents":["segment objects at varying scales (small objects in large images, large objects in small images)","improve robustness by combining predictions across scales using ensemble methods","detect small objects that may be missed at single-scale inference","reduce false positives by filtering predictions that don't align across scales"],"best_for":["datasets with objects at highly variable scales (aerial imagery, medical imaging)","applications requiring high recall for small objects","robustness-critical systems where ensemble voting improves reliability","research on scale-invariant segmentation"],"limitations":["multi-scale processing increases latency by 2-4x (encoding at 3 scales); not suitable for real-time applications","memory consumption scales with number of scales; 3 scales require ~3x embedding storage","ensemble voting requires careful weighting; equal weighting can amplify errors if all scales fail similarly","upsampling/downsampling introduces interpolation artifacts; may degrade accuracy on already-small objects"],"requires":["Python 3.8+","PyTorch 1.9+","GPU with 12GB+ VRAM for multi-scale processing","OpenCV or PIL for image resizing"],"input_types":["image (PNG, JPEG, etc.)","scale factors (list of floats, e.g., [0.5, 1.0, 2.0])","prompt (point, box, or mask)","ensemble strategy (voting, confidence-weighted averaging, max-confidence)"],"output_types":["ensemble mask (H x W boolean array)","per-scale masks (list of H x W arrays)","confidence map (H x W float [0, 1] indicating agreement across scales)","scale-specific IoU predictions"],"categories":["image-visual","planning-reasoning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-segment-anything__cap_6","uri":"capability://image.visual.point.based.interactive.segmentation.with.click.refinement","name":"point-based interactive segmentation with click refinement","description":"Enables interactive segmentation where users click on image regions to provide positive/negative point prompts, with real-time mask updates after each click. The implementation maintains a prompt history and iteratively refines masks by accumulating prompts, using the previous mask as a hint for the next iteration to improve consistency and reduce flicker.","intents":["build interactive annotation tools where users click to segment objects","enable real-time mask refinement with visual feedback after each click","support undo/redo by maintaining prompt history","create user-friendly segmentation workflows without requiring bounding boxes or complex input"],"best_for":["interactive annotation platforms (web, desktop, mobile)","image editing tools requiring object selection","user studies on segmentation interaction patterns","accessibility-focused applications (click-based interaction)"],"limitations":["latency must be <100ms per click for smooth interaction; requires GPU and optimized inference","user click accuracy directly impacts mask quality; misclicks require undo/redo","no automatic stopping criterion; users must decide when mask is good enough","prompt history grows with each click, potentially causing memory issues in long sessions (>100 clicks)"],"requires":["Python 3.8+","PyTorch 1.9+ with GPU support","Web framework (Flask, FastAPI) for interactive UI","WebSocket or similar for real-time client-server communication","GPU with 6GB+ VRAM for <100ms latency"],"input_types":["image (PNG, JPEG, etc.)","click coordinates (x, y as integers)","click type (positive or negative)","previous mask (optional, for consistency)"],"output_types":["updated segmentation mask (H x W boolean array)","mask confidence (IoU prediction)","visualization (mask overlay on image)","interaction metadata (click history, refinement steps)"],"categories":["image-visual","tool-use-integration"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-segment-anything__cap_7","uri":"capability://image.visual.bounding.box.based.segmentation.with.automatic.refinement","name":"bounding-box-based segmentation with automatic refinement","description":"Accepts bounding box prompts (x_min, y_min, x_max, y_max) and generates segmentation masks for objects within the box. The implementation can automatically refine boxes by detecting object boundaries within the box region, or generate multiple masks for ambiguous boxes, enabling coarse-to-fine segmentation workflows.","intents":["segment objects from object detection outputs (bounding boxes)","enable segmentation workflows that start with coarse bounding boxes","integrate with object detection pipelines (YOLO, Faster R-CNN) for end-to-end instance segmentation","support bounding-box-based annotation tools"],"best_for":["instance segmentation pipelines combining detection and segmentation","annotation tools with bounding box input","integration with existing object detection systems","datasets where bounding boxes are already available"],"limitations":["bounding box quality directly impacts segmentation; loose boxes may include background, tight boxes may exclude object parts","cannot segment objects partially visible in the box; requires box to fully contain object","no automatic box refinement; users must provide accurate boxes","ambiguous boxes (multiple objects) generate multiple masks requiring disambiguation"],"requires":["Python 3.8+","PyTorch 1.9+","NumPy for box manipulation"],"input_types":["image (PNG, JPEG, etc.)","bounding box (x_min, y_min, x_max, y_max as integers)","optional: box confidence or source (detection model)"],"output_types":["segmentation mask (H x W boolean array)","multiple masks for ambiguous boxes (list of masks with confidence scores)","refined box (optional, based on mask boundaries)","mask statistics (area, IoU with original box)"],"categories":["image-visual","data-processing-analysis"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-segment-anything__cap_8","uri":"capability://image.visual.mask.based.iterative.segmentation.with.hint.propagation","name":"mask-based iterative segmentation with hint propagation","description":"Accepts a previous segmentation mask as input and uses it as a hint to refine or extend segmentation in subsequent iterations. The mask is encoded alongside point/box prompts and passed to the decoder, enabling iterative refinement where each iteration builds on the previous mask, useful for correcting errors or extending segmentation to new regions.","intents":["iteratively refine masks by providing previous mask as context","correct segmentation errors by masking out incorrect regions and re-segmenting","extend segmentation to new regions by providing partial masks as hints","build multi-step segmentation workflows where each step refines the previous"],"best_for":["iterative annotation workflows requiring multiple refinement steps","error correction pipelines where masks are reviewed and refined","segmentation of complex objects requiring multi-step decomposition","research on mask propagation and refinement"],"limitations":["mask hints can bias the model; incorrect hints may propagate errors","no automatic detection of when mask is good enough; requires external stopping criterion","iterative refinement adds latency; each iteration requires a full decoder forward pass","mask encoding adds complexity; requires careful handling of mask resolution and alignment"],"requires":["Python 3.8+","PyTorch 1.9+","NumPy for mask manipulation"],"input_types":["image (PNG, JPEG, etc.)","previous mask (H x W boolean or soft array)","refinement prompt (point, box, or additional mask)","refinement mode (extend, correct, or refine)"],"output_types":["refined mask (H x W boolean array)","mask difference (regions changed from previous iteration)","confidence in refinement (IoU prediction)","iteration metadata (step number, changes made)"],"categories":["image-visual","planning-reasoning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-segment-anything__cap_9","uri":"capability://automation.workflow.efficient.model.variant.selection.and.deployment","name":"efficient model variant selection and deployment","description":"Provides three pre-trained model variants (ViT-B, ViT-L, ViT-H) with different speed-accuracy tradeoffs, enabling users to select the appropriate model for their hardware and latency constraints. The implementation includes model loading, quantization support (int8, fp16), and export to ONNX/TorchScript for deployment on edge devices and cloud infrastructure.","intents":["choose model variant based on available hardware (mobile, edge, cloud)","optimize inference latency for real-time applications","reduce model size for edge deployment via quantization","export models for deployment on non-Python runtimes (C++, JavaScript, mobile)"],"best_for":["teams deploying segmentation across heterogeneous hardware (mobile, edge, cloud)","resource-constrained environments (mobile, IoT, embedded systems)","production systems requiring latency/accuracy tradeoffs","cross-platform deployment (web, mobile, desktop)"],"limitations":["ViT-B is smallest but may have lower accuracy on complex scenes; ViT-H requires 24GB+ VRAM","quantization (int8, fp16) reduces accuracy by 1-5%; requires validation per use case","ONNX export requires careful handling of dynamic shapes; some operations may not be supported","TorchScript export may not support all Python features; requires model refactoring for complex logic"],"requires":["Python 3.8+","PyTorch 1.9+","ONNX Runtime (optional, for ONNX inference)","TensorRT (optional, for NVIDIA GPU optimization)","CoreML tools (optional, for iOS deployment)"],"input_types":["model variant selection (ViT-B, ViT-L, or ViT-H)","quantization config (none, int8, fp16)","export format (PyTorch, ONNX, TorchScript)","target hardware (CPU, GPU, mobile)"],"output_types":["loaded model (PyTorch module)","quantized model (int8 or fp16 weights)","exported model (ONNX, TorchScript, or CoreML format)","model metadata (variant, size, latency estimates)"],"categories":["automation-workflow","tool-use-integration"],"confidence":0.5,"matches":0,"success_rate":0}],"trust":{"score":22,"verified":false,"data_access_risk":"high","permissions":["Python 3.8+","PyTorch 1.9+ with CUDA 11.0+ (for GPU acceleration)","NumPy, Pillow for image I/O","4GB+ RAM minimum; 8GB+ recommended for batch processing","GPU with 6GB+ VRAM for real-time inference (NVIDIA/AMD/Intel Arc supported via PyTorch)","PyTorch 1.9+","NumPy for mask manipulation","GPU recommended for interactive workflows (sub-100ms latency requirement)","SciPy for connected-component analysis","Optional: external classifier (CLIP, ResNet, etc.) for semantic information"],"failure_modes":["requires high-resolution images (1024x1024 recommended) for optimal accuracy; performance degrades on small objects or cluttered scenes","prompt quality directly impacts output quality — ambiguous prompts may generate multiple competing masks requiring disambiguation logic","inference latency ~500ms per image on CPU, ~50-100ms on GPU; batch processing not optimized for real-time video","model weights are ~375MB (ViT-B) to ~1.2GB (ViT-L); requires significant memory for edge deployment","struggles with transparent objects, reflections, and fine-grained boundaries; post-processing often needed for production use","IoU predictions are model estimates, not ground-truth accuracy; can be overconfident on out-of-distribution images","iterative refinement adds latency — each prompt requires a full forward pass through the decoder (~50-100ms per iteration)","no built-in mechanism to detect when additional prompts won't improve masks; users must implement stopping criteria","mask quality still depends on prompt quality; poor prompts generate poor masks regardless of disambiguation logic","class-agnostic masks lack semantic information; downstream classification required for semantic segmentation","builder identity is not verified yet","no observed match outcomes yet"],"rank_breakdown":{"adoption":0.05,"quality":0.32,"ecosystem":0.3,"match_graph":0.25,"freshness":0.5,"weights":{"adoption":0.3,"quality":0.2,"ecosystem":0.15,"match_graph":0.3,"freshness":0.05}},"observed_outcomes":{"matches":0,"success_rate":0,"avg_confidence":0,"top_intents":[],"last_matched_at":null},"maintenance":{"status":"active","updated_at":"2026-05-24T12:16:25.060Z","last_scraped_at":"2026-05-03T15:20:20.420Z","last_commit":null},"community":{"stars":null,"forks":null,"weekly_downloads":null,"model_downloads":null,"model_likes":null}},"distribution":{"claim_url":"https://unfragile.ai/submit?claim=pypi-segment-anything","compare_url":"https://unfragile.ai/compare?artifact=pypi-segment-anything"}},"signature":"Bevnfaonmh2vbnlkgn+nhbJyO9LuBjiwsPOuQpUb+4SxgkS6UliWI1KSU86hFeSuDyqcL+xJWYJ9G49gxQ+FBg==","signedAt":"2026-06-20T18:53:28.450Z","signedBy":"unfragile.ai","version":1},"_links":{"self":"https://unfragile.ai/api/v1/passport/pypi-segment-anything","artifact":"https://unfragile.ai/pypi-segment-anything","verify":"https://unfragile.ai/api/v1/verify?slug=pypi-segment-anything","publicKey":"https://unfragile.ai/api/v1/trust-passport-public-key","spec":"https://unfragile.ai/trust","schema":"https://unfragile.ai/schema.json","docs":"https://unfragile.ai/docs"}}