{"passport":{"unfragile":{"@version":"1.0","version":"2026-05","artifact":{"id":"awesome-gpt-4","slug":"gpt-4","name":"GPT-4","type":"model","url":"https://openai.com/research/gpt-4","page_url":"https://unfragile.ai/gpt-4","categories":["model-training"],"tags":[],"pricing":{"model":"unknown","free":false,"starting_price":null},"status":"active","verified":false},"capabilities":[{"id":"awesome-gpt-4__cap_0","uri":"capability://image.visual.multimodal.text.and.image.understanding.with.unified.transformer.architecture","name":"multimodal text and image understanding with unified transformer architecture","description":"GPT-4 processes both text and image inputs through a single transformer-based architecture that encodes visual information into the same token space as language tokens, enabling joint reasoning across modalities. The model uses vision encoders to convert images into embeddings that integrate seamlessly with the language model's attention mechanisms, allowing it to answer questions about images, read text within images, and reason about visual content in context with textual prompts.","intents":["analyze charts, diagrams, and screenshots to extract insights or answer questions about visual content","perform OCR and text extraction from images while maintaining semantic understanding of context","generate descriptions, captions, or detailed analysis of images based on visual features","solve visual reasoning tasks like counting objects, identifying spatial relationships, or interpreting infographics"],"best_for":["document processing teams handling mixed text-image workflows","accessibility tool builders creating alt-text and image description systems","data extraction pipelines requiring visual document understanding"],"limitations":["image resolution and aspect ratio constraints limit fine-grained visual detail extraction","performance degrades on images with dense text or complex layouts compared to specialized OCR systems","no video frame-by-frame analysis — requires static image inputs only","latency for image processing is higher than text-only inference due to vision encoder overhead"],"requires":["API access to OpenAI's GPT-4 with vision capability enabled","images in JPEG, PNG, GIF, or WebP format","image size under 20MB per OpenAI API limits"],"input_types":["text prompts","images (JPEG, PNG, GIF, WebP)","mixed text and image in single request"],"output_types":["text descriptions","structured analysis","extracted text from images","reasoning chains"],"categories":["image-visual","text-generation-language"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-4__cap_1","uri":"capability://text.generation.language.long.context.reasoning.with.extended.token.window","name":"long-context reasoning with extended token window","description":"GPT-4 supports an 8K token context window (later extended to 32K and 128K in variants), enabling the model to maintain coherence and reasoning across significantly longer documents, codebases, or conversation histories than GPT-3.5. The implementation uses standard transformer attention with optimizations to manage computational complexity at scale, allowing developers to pass entire files, specifications, or multi-turn conversations without truncation.","intents":["analyze entire source files or multi-file codebases for refactoring or security review without splitting into chunks","maintain conversation context across 50+ turns without losing semantic understanding of earlier exchanges","process long-form documents (research papers, legal contracts, technical specifications) end-to-end","perform comparative analysis across multiple documents by holding all content in context simultaneously"],"best_for":["code review and refactoring teams working with large monolithic files","legal and compliance teams analyzing lengthy contracts or regulatory documents","research and analysis teams comparing multiple sources in a single reasoning pass"],"limitations":["token counting is approximate — actual token usage may vary by 5-10% due to tokenizer behavior","latency increases linearly with context length; 128K token requests are 8-10x slower than 8K requests","attention mechanism still has O(n²) complexity, making extremely long contexts (>200K tokens) computationally expensive","model may lose focus on early context when processing very long sequences, requiring explicit re-emphasis"],"requires":["API access to GPT-4 with appropriate context window variant (8K, 32K, or 128K)","token counting library or manual estimation to stay within limits","sufficient API rate limits to handle longer processing times"],"input_types":["text documents","source code","conversation histories","concatenated multi-file content"],"output_types":["analysis and summaries","refactoring suggestions","structured insights","comparative findings"],"categories":["text-generation-language","planning-reasoning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-4__cap_10","uri":"capability://data.processing.analysis.structured.data.extraction.and.schema.based.output.generation","name":"structured data extraction and schema-based output generation","description":"GPT-4 extracts structured data from unstructured text and generates outputs conforming to specified schemas (JSON, XML, CSV) through instruction-following and constraint adherence. The model parses natural language, documents, or semi-structured data and maps it to defined schemas, enabling developers to build data extraction pipelines without custom parsing logic, though output validation is still required.","intents":["extract entities, relationships, and attributes from documents or text and output as structured JSON or XML","parse semi-structured data (tables, lists, forms) and convert to standardized formats","generate API responses or database records conforming to specified schemas","build data pipelines that convert unstructured content into structured formats for downstream processing"],"best_for":["data extraction and ETL teams processing documents at scale","API and integration teams generating structured responses","teams building knowledge graphs or structured databases from unstructured sources"],"limitations":["extraction accuracy varies by domain and data complexity; specialized NER or information extraction models may outperform on specific tasks","schema violations still occur — output validation and error handling are required","performance degrades on complex nested schemas or ambiguous data","no built-in data validation or type checking — downstream systems must validate extracted data"],"requires":["API access to GPT-4","clear schema definition (JSON schema, XML DTD, or natural language specification)","output validation logic to catch schema violations"],"input_types":["unstructured text","documents","semi-structured data","natural language descriptions"],"output_types":["JSON","XML","CSV","structured data conforming to schema"],"categories":["data-processing-analysis","text-generation-language"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-4__cap_11","uri":"capability://text.generation.language.conversational.dialogue.with.multi.turn.context.management","name":"conversational dialogue with multi-turn context management","description":"GPT-4 maintains coherent multi-turn conversations by tracking context across exchanges, using transformer attention to weight relevant prior messages and maintain consistency in responses. The model can engage in extended dialogues, remember user preferences and context from earlier turns, and adapt responses based on conversation history, enabling developers to build conversational AI systems without explicit state management.","intents":["build chatbots and conversational interfaces that maintain context across 20+ turns without losing coherence","create customer support systems that remember user issues and context from previous interactions","implement interactive tutoring systems that adapt explanations based on student understanding and prior exchanges","develop dialogue-based applications (games, interactive fiction) with consistent character and narrative context"],"best_for":["customer support and help desk teams building conversational interfaces","educational technology platforms creating interactive tutoring systems","game and entertainment companies building dialogue-driven experiences"],"limitations":["context window limits conversation length — very long conversations (>50 turns) may exceed token limits","model may lose focus on early context or contradict earlier statements in very long conversations","no persistent memory across sessions — each conversation starts fresh without access to prior interactions","context management is implicit — no explicit state tracking or memory management, making debugging difficult"],"requires":["API access to GPT-4","conversation history management (storing and passing prior messages)","token counting to ensure conversation history fits within context window"],"input_types":["user messages","conversation history","system prompts defining conversation behavior"],"output_types":["conversational responses","contextually appropriate replies","dialogue continuations"],"categories":["text-generation-language","planning-reasoning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-4__cap_12","uri":"capability://planning.reasoning.reasoning.based.problem.decomposition.and.planning","name":"reasoning-based problem decomposition and planning","description":"GPT-4 decomposes complex problems into sub-tasks and generates step-by-step plans through chain-of-thought reasoning patterns, using transformer attention to identify dependencies and logical structure. The model can break down multi-step problems, generate execution plans, and reason about intermediate steps, enabling developers to build planning and reasoning systems without explicit planning algorithms.","intents":["decompose complex software engineering tasks (system design, architecture planning) into actionable steps","generate project plans and task breakdowns for project management systems","solve multi-step reasoning problems by explicitly working through intermediate steps","create debugging workflows that systematically identify and isolate issues"],"best_for":["project management and planning teams automating task decomposition","software engineering teams using AI for architecture and design planning","educational systems teaching problem-solving and reasoning skills"],"limitations":["reasoning quality degrades on problems requiring more than 15-20 steps; specialized planning algorithms are more reliable for complex planning","no formal verification of reasoning correctness — intermediate steps may contain logical errors","planning is not optimal — generated plans may be suboptimal compared to specialized planning algorithms","reasoning is opaque — difficult to debug or understand why specific reasoning paths were chosen"],"requires":["API access to GPT-4","clear problem specification","human review of generated plans before execution"],"input_types":["problem descriptions","constraints and requirements","context and background information"],"output_types":["step-by-step plans","task decompositions","reasoning chains","intermediate steps with explanations"],"categories":["planning-reasoning","text-generation-language"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-4__cap_2","uri":"capability://text.generation.language.few.shot.and.zero.shot.task.adaptation.via.prompt.engineering","name":"few-shot and zero-shot task adaptation via prompt engineering","description":"GPT-4 demonstrates strong in-context learning capabilities, allowing developers to specify task behavior through natural language instructions and examples without fine-tuning. The model uses transformer attention to recognize patterns in provided examples and apply them to new inputs, enabling rapid task adaptation by simply modifying the prompt structure, example selection, and instruction clarity.","intents":["adapt the model to domain-specific tasks (medical coding, legal analysis, technical support) by providing 2-5 examples in the prompt","implement custom classification, extraction, or generation tasks without training or fine-tuning infrastructure","create specialized personas or writing styles by demonstrating the desired behavior in examples","build task-specific workflows that can be modified by non-technical users through prompt templates"],"best_for":["rapid prototyping teams building proof-of-concepts for new use cases","non-technical product managers creating task variations without engineering overhead","teams with limited ML infrastructure who need quick task customization"],"limitations":["performance plateaus with more than 5-10 examples; additional examples may introduce noise rather than improve accuracy","prompt sensitivity is high — small wording changes can significantly alter output quality and consistency","no persistent learning — each request requires re-specification of task context and examples","few-shot learning is less reliable than fine-tuning for tasks requiring precise, consistent formatting or domain-specific terminology"],"requires":["clear, well-structured natural language instructions","representative examples that cover edge cases and desired output formats","understanding of prompt engineering best practices (chain-of-thought, role-playing, explicit constraints)"],"input_types":["natural language instructions","example input-output pairs","task descriptions","constraint specifications"],"output_types":["task-specific text","structured data matching example format","classifications","extractions"],"categories":["text-generation-language","planning-reasoning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-4__cap_3","uri":"capability://code.generation.editing.code.generation.and.understanding.across.40.programming.languages","name":"code generation and understanding across 40+ programming languages","description":"GPT-4 generates syntactically correct, idiomatic code across Python, JavaScript, TypeScript, Java, C++, Go, Rust, SQL, and 30+ other languages through training on diverse code repositories and documentation. The model understands language-specific idioms, standard libraries, and common patterns, enabling it to generate production-quality code snippets, complete functions, and suggest refactorings with language-aware context awareness.","intents":["generate complete functions or classes from natural language specifications with correct syntax and idioms","debug code by analyzing error messages and suggesting fixes with explanations of root causes","refactor code to improve readability, performance, or adherence to language-specific best practices","translate code between languages while preserving logic and adapting to target language idioms"],"best_for":["individual developers and small teams using AI-assisted coding for productivity gains","teams building code generation tools or IDE plugins that need a capable backend","educational contexts where students learn programming with AI assistance"],"limitations":["generated code may contain subtle bugs or security vulnerabilities — requires human review before production use","performance optimization suggestions are generic; specialized profiling tools outperform on performance-critical code","understanding of complex legacy codebases is limited without explicit context and file references","no real-time compilation or execution feedback — cannot validate generated code against actual runtime behavior"],"requires":["API access to GPT-4","clear code context or specifications for generation tasks","human code review and testing before deployment"],"input_types":["natural language specifications","existing code snippets","error messages and stack traces","refactoring requests"],"output_types":["code snippets","complete functions or classes","refactoring suggestions","bug fixes with explanations"],"categories":["code-generation-editing","text-generation-language"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-4__cap_4","uri":"capability://planning.reasoning.mathematical.reasoning.and.symbolic.problem.solving","name":"mathematical reasoning and symbolic problem-solving","description":"GPT-4 demonstrates improved mathematical reasoning capabilities compared to GPT-3.5, solving algebra, calculus, geometry, and logic problems through step-by-step symbolic manipulation and reasoning. The model uses chain-of-thought patterns to break complex problems into intermediate steps, enabling it to work through multi-step proofs, equation solving, and formal logic problems with higher accuracy than previous versions.","intents":["solve mathematical problems (algebra, calculus, statistics) with step-by-step explanations suitable for educational contexts","verify mathematical proofs or identify logical errors in formal reasoning","generate mathematical content for educational materials, textbooks, or assessment systems","assist with symbolic computation tasks like equation simplification or formula derivation"],"best_for":["educational technology platforms building AI tutoring systems","mathematics educators creating problem sets and solution explanations","researchers and students needing symbolic reasoning assistance"],"limitations":["performance degrades on problems requiring more than 10-15 reasoning steps; specialized symbolic math engines (Mathematica, SymPy) are more reliable for complex problems","cannot perform numerical computation with arbitrary precision — floating-point arithmetic limitations apply","no integration with computer algebra systems — cannot verify solutions through symbolic computation","struggles with novel problem types not well-represented in training data; standard textbook problems are more reliable"],"requires":["API access to GPT-4","clear problem statement with mathematical notation or natural language description","for verification: human review or comparison against known solutions"],"input_types":["mathematical problems in natural language or notation","equations and formulas","proofs or logical statements","multi-step problem specifications"],"output_types":["step-by-step solutions","mathematical explanations","derived formulas","proof verification or corrections"],"categories":["planning-reasoning","text-generation-language"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-4__cap_5","uri":"capability://text.generation.language.instruction.following.and.constraint.adherence.with.high.consistency","name":"instruction-following and constraint adherence with high consistency","description":"GPT-4 demonstrates superior instruction-following capabilities, reliably adhering to complex constraints, output format specifications, and multi-part instructions through improved training on instruction-following datasets. The model maintains consistency across multiple requests with the same system prompt, enabling developers to build reliable, deterministic workflows by specifying precise constraints and expected output formats.","intents":["enforce strict output formatting (JSON, CSV, XML, markdown) without post-processing or parsing errors","implement content moderation or filtering by specifying rejection criteria and response templates","build deterministic workflows where the same input reliably produces the same output format and structure","create role-based systems where the model maintains a consistent persona and behavior across multiple turns"],"best_for":["production systems requiring high consistency and reliability in model outputs","teams building structured data extraction pipelines that depend on consistent formatting","applications where output format violations would break downstream processing"],"limitations":["consistency is high but not guaranteed — edge cases and adversarial inputs can still cause format violations","complex nested constraints may conflict, requiring careful prompt engineering to resolve","instruction-following degrades when constraints are contradictory or when instructions exceed ~500 tokens","no formal verification — cannot guarantee constraint satisfaction without output validation"],"requires":["API access to GPT-4","clear, unambiguous constraint specifications","output validation logic to catch edge cases where constraints are violated"],"input_types":["system prompts with constraints","multi-part instructions","format specifications","role definitions"],"output_types":["formatted text matching specifications","structured data (JSON, XML, CSV)","constrained responses"],"categories":["text-generation-language","safety-moderation"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-4__cap_6","uri":"capability://text.generation.language.knowledge.based.question.answering.with.factual.grounding","name":"knowledge-based question answering with factual grounding","description":"GPT-4 answers factual questions across diverse domains (history, science, current events, technical topics) by retrieving and synthesizing knowledge from its training data, which includes text up to April 2023. The model uses transformer attention to identify relevant knowledge patterns and synthesize coherent answers, though it operates without real-time information access or explicit knowledge base retrieval, making it subject to knowledge cutoff and hallucination risks.","intents":["answer factual questions about historical events, scientific concepts, technical topics, and general knowledge","provide explanations and context for complex topics suitable for educational or informational purposes","generate summaries of topics by synthesizing knowledge across multiple domains","fact-check claims by comparing them against known information in the model's training data"],"best_for":["educational platforms building AI tutoring or question-answering systems","content creation teams needing background research and topic explanations","customer support systems answering common factual questions"],"limitations":["knowledge cutoff at April 2023 — cannot answer questions about events or information after that date","hallucination risk is significant for obscure topics or specific factual claims — model may generate plausible-sounding but false information","no explicit knowledge base retrieval — cannot cite sources or provide confidence scores for factual claims","performance degrades on very recent events, emerging technologies, or niche domain knowledge"],"requires":["API access to GPT-4","understanding that responses may contain hallucinations and require fact-checking","for production systems: integration with external knowledge bases or fact-checking systems"],"input_types":["factual questions","topic requests","claims for verification"],"output_types":["factual answers","explanations and context","summaries","verification results"],"categories":["text-generation-language","memory-knowledge"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-4__cap_7","uri":"capability://text.generation.language.creative.writing.and.content.generation.with.stylistic.control","name":"creative writing and content generation with stylistic control","description":"GPT-4 generates creative content (stories, poetry, marketing copy, dialogue) with improved stylistic control and coherence compared to GPT-3.5, using transformer attention to maintain narrative consistency across long-form content. The model can adapt to specified writing styles, tones, and genres through prompt engineering, enabling developers to generate diverse content types without fine-tuning.","intents":["generate marketing copy, product descriptions, and promotional content tailored to specific audiences and tones","create fictional narratives, stories, or dialogue with consistent characters and plot coherence","write poetry or creative content in specified styles (haiku, sonnet, free verse) with appropriate structure","generate multiple content variations for A/B testing or creative exploration"],"best_for":["marketing and content teams generating copy at scale","creative writing platforms and tools","game developers and interactive fiction creators","educational content creation for engaging learning materials"],"limitations":["generated content may lack originality or contain subtle plagiarism from training data","stylistic consistency degrades in very long-form content (>5000 words) without explicit reinforcement","creative quality is subjective and highly dependent on prompt quality; poor prompts yield mediocre results","no built-in plagiarism detection or originality verification"],"requires":["API access to GPT-4","clear style, tone, and content specifications","human review and editing for quality assurance"],"input_types":["style and tone specifications","content briefs or outlines","genre or format requirements","example content for style matching"],"output_types":["creative text","marketing copy","narrative content","poetry or structured creative content"],"categories":["text-generation-language"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-4__cap_8","uri":"capability://text.generation.language.multilingual.translation.and.cross.lingual.understanding","name":"multilingual translation and cross-lingual understanding","description":"GPT-4 translates between 100+ languages with improved fluency and cultural adaptation compared to GPT-3.5, using transformer attention to understand semantic meaning and idiomatic expressions across languages. The model can translate technical documents, creative content, and conversational text while preserving tone and meaning, though it lacks the specialized optimization of dedicated translation services.","intents":["translate technical documentation, code comments, and API documentation across languages","localize marketing content and user-facing text while preserving brand voice and cultural appropriateness","translate creative content (stories, poetry, dialogue) while maintaining stylistic and emotional intent","provide cross-lingual understanding for multilingual customer support or content moderation"],"best_for":["global product teams localizing content for multiple markets","technical documentation teams translating specifications and guides","content platforms serving multilingual audiences"],"limitations":["translation quality varies significantly by language pair; high-resource languages (Spanish, French, German) are more reliable than low-resource languages","cultural adaptation is limited — model may miss context-specific idioms or cultural references","specialized terminology in technical or domain-specific content may be mistranslated","no real-time feedback or iterative refinement — requires human review for quality assurance"],"requires":["API access to GPT-4","source text in supported language","target language specification","for quality assurance: human review by native speakers"],"input_types":["text in any supported language","technical documentation","creative content","conversational text"],"output_types":["translated text","localized content","cross-lingual summaries"],"categories":["text-generation-language"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"awesome-gpt-4__cap_9","uri":"capability://safety.moderation.safety.aware.content.generation.with.reduced.harmful.outputs","name":"safety-aware content generation with reduced harmful outputs","description":"GPT-4 incorporates safety training through RLHF and constitutional AI techniques to reduce generation of harmful, biased, or inappropriate content compared to GPT-3.5. The model uses learned safety patterns to refuse unsafe requests, provide warnings for sensitive topics, and generate more balanced perspectives on controversial subjects, though it is not immune to jailbreaks or adversarial prompts.","intents":["generate content for public-facing applications with reduced risk of harmful outputs or offensive language","implement content moderation workflows that leverage the model's safety training to filter or flag problematic content","build systems that handle sensitive topics (mental health, violence, discrimination) with appropriate warnings and balanced perspectives","create educational content on controversial topics with nuanced, balanced treatment"],"best_for":["public-facing applications and platforms requiring content safety","educational and healthcare systems handling sensitive topics","teams building content moderation or safety systems"],"limitations":["safety training is probabilistic — adversarial prompts and jailbreaks can still elicit harmful outputs","safety guardrails may be overly conservative, refusing legitimate requests or generating unnecessary warnings","bias reduction is improved but not eliminated — model may still reflect biases present in training data","safety training is not transparent — difficult to predict which requests will be refused or flagged"],"requires":["API access to GPT-4","understanding that safety is probabilistic and not guaranteed","additional safety layers (content filtering, human review) for high-stakes applications"],"input_types":["any user input","requests on sensitive topics","content moderation queries"],"output_types":["safety-aware responses","refusals with explanations","balanced perspectives on controversial topics","content flags and warnings"],"categories":["safety-moderation","text-generation-language"],"confidence":0.5,"matches":0,"success_rate":0}],"trust":{"score":46,"verified":false,"data_access_risk":"high","permissions":["API access to OpenAI's GPT-4 with vision capability enabled","images in JPEG, PNG, GIF, or WebP format","image size under 20MB per OpenAI API limits","API access to GPT-4 with appropriate context window variant (8K, 32K, or 128K)","token counting library or manual estimation to stay within limits","sufficient API rate limits to handle longer processing times","API access to GPT-4","clear schema definition (JSON schema, XML DTD, or natural language specification)","output validation logic to catch schema violations","conversation history management (storing and passing prior messages)"],"failure_modes":["image resolution and aspect ratio constraints limit fine-grained visual detail extraction","performance degrades on images with dense text or complex layouts compared to specialized OCR systems","no video frame-by-frame analysis — requires static image inputs only","latency for image processing is higher than text-only inference due to vision encoder overhead","token counting is approximate — actual token usage may vary by 5-10% due to tokenizer behavior","latency increases linearly with context length; 128K token requests are 8-10x slower than 8K requests","attention mechanism still has O(n²) complexity, making extremely long contexts (>200K tokens) computationally expensive","model may lose focus on early context when processing very long sequences, requiring explicit re-emphasis","extraction accuracy varies by domain and data complexity; specialized NER or information extraction models may outperform on specific tasks","schema violations still occur — output validation and error handling are required","builder identity is not verified yet","no observed match outcomes yet"],"rank_breakdown":{"adoption":0.05,"quality":0.35,"ecosystem":0.25,"match_graph":0.25,"freshness":0.75,"weights":{"adoption":0.35,"quality":0.2,"ecosystem":0.1,"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":"2026-06-17T09:51:03.041Z","last_scraped_at":"2026-05-03T14:00:20.516Z","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=gpt-4","compare_url":"https://unfragile.ai/compare?artifact=gpt-4"}},"signature":"sPn1ryXR7/Ny9coLvehbiTOGzGJmk1QLMngT1QQz65V180ss6EY1BpE91utOLlGrqXgS8CP/E8TVgl78r0HDCg==","signedAt":"2026-06-21T18:47:58.615Z","signedBy":"unfragile.ai","version":1},"_links":{"self":"https://unfragile.ai/api/v1/passport/gpt-4","artifact":"https://unfragile.ai/gpt-4","verify":"https://unfragile.ai/api/v1/verify?slug=gpt-4","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"}}