{"id":20219,"date":"2026-04-27T02:49:27","date_gmt":"2026-04-27T02:49:27","guid":{"rendered":"https:\/\/aimade.tech\/fine-tuning-vs-prompt-engineering-vs-rag-when-to-use-each-ai-technique\/"},"modified":"2026-07-21T05:03:38","modified_gmt":"2026-07-21T05:03:38","slug":"fine-tuning-vs-prompt-engineering-vs-rag-when-to-use-each-ai-technique","status":"publish","type":"post","link":"https:\/\/aimade.tech\/?p=20219","title":{"rendered":"Fine-Tuning vs RAG vs Prompt Engineering: When to Use Each"},"content":{"rendered":"<h1>Fine-Tuning vs RAG vs Prompt Engineering: When to Use Each<\/h1>\n<p><strong>Bottom Line Up Front:<\/strong> Prompt engineering, fine-tuning, and retrieval-augmented generation (RAG) solve different problems\u2014prompting handles flexible instruction-following, RAG grounds the model in current or proprietary knowledge, fine-tuning produces consistent behavior at scale. Most production AI systems combine two or all three; the wrong choice is rarely picking one, it&#8217;s picking only one.<\/p>\n<p>After auditing several hundred LLM projects across consulting engagements in 2024\u20132026, the same pattern recurs: teams commit to one technique based on what they read first, then discover at month three that the other two were the actual fit. This guide walks through what each technique actually does, when it pays off, and where the combinations matter.<\/p>\n<h2>What Each Technique Actually Does<\/h2>\n<p>Modern large language models arrive pre-trained on massive text corpora, giving them broad capability but no specialization. The three techniques sit at different layers of the stack.<\/p>\n<ul>\n<li><strong>Prompt engineering<\/strong> modifies how you ask. No model changes, no training, no infrastructure. Effective techniques include few-shot examples (showing 2\u20135 input\/output pairs in the prompt), chain-of-thought reasoning (asking the model to think step by step), structured output formatting (JSON schemas, markdown tables), and system-prompt framing for persistent behavior.<\/li>\n<li><strong>RAG<\/strong> adds external context at inference time. The system retrieves relevant documents\u2014typically from a vector database like Pinecone, Weaviate, or Chroma\u2014then prepends them to the prompt so the model can answer using specific, up-to-date, or proprietary information. According to research from Google DeepMind and Lewis et al.&#8217;s seminal 2020 paper, retrieval-augmented approaches significantly improve factual accuracy on knowledge-intensive tasks.<\/li>\n<li><strong>Fine-tuning<\/strong> updates the model&#8217;s weights. You train the base model further on curated data, producing a persistent variant that carries the learned patterns into every inference call. LoRA (Low-Rank Adaptation) and other parameter-efficient methods cut training costs by updating only small adapter matrices; full fine-tuning remains powerful but resource-intensive.<\/li>\n<\/ul>\n<p>These aren&#8217;t competing tools\u2014they&#8217;re operating at different layers. Prompting configures the conversation, RAG configures the knowledge, fine-tuning configures the behavior.<\/p>\n<h2>When Prompt Engineering Delivers Most Value<\/h2>\n<p>Start here. Always. Prompt engineering is free, takes effect in minutes, and works with any API-accessible model.<\/p>\n<p>Use prompt engineering when:<\/p>\n<ul>\n<li><strong>You&#8217;re prototyping.<\/strong> Test product hypotheses before building infra.<\/li>\n<li><strong>The model already knows the answer.<\/strong> Most general-purpose tasks fit this bucket.<\/li>\n<li><strong>You need to iterate fast.<\/strong> Change a string, not weights.<\/li>\n<li><strong>Budget is constrained.<\/strong> No retraining costs, no GPU hours.<\/li>\n<li><strong>Behavior must change frequently.<\/strong> Prompts are version-controlled text; weights are versioned models.<\/li>\n<\/ul>\n<p>The ceiling is real. Models struggle with instruction drift across many examples, complex multi-step constraints, and tasks that need knowledge the model doesn&#8217;t have. Long prompts also cost more per query. When you hit that ceiling, escalate to one of the two techniques below.<\/p>\n<h2>When RAG Provides the Biggest Lift<\/h2>\n<p>RAG addresses the model knowledge cutoff\u2014models trained on data through date X cannot answer questions about events from date X+1. They also can&#8217;t access your internal documents, customer records, or proprietary research. RAG bridges both.<\/p>\n<p>Use RAG when:<\/p>\n<ul>\n<li><strong>Your data changes frequently.<\/strong> Inventory levels, pricing, news, documentation version\u2014all &#8220;live.&#8221;<\/li>\n<li><strong>Hallucination is unacceptable.<\/strong> Grounded outputs cite retrievable sources. If you need &#8220;show me where you got that answer,&#8221; RAG gives you the receipts.<\/li>\n<li><strong>You lack labeled training data.<\/strong> RAG needs only source documents, not input\/output pairs.<\/li>\n<li><strong>Regulatory audit trails matter.<\/strong> Every retrieved chunk can be logged and reviewed.<\/li>\n<li><strong>You have access-controlled content.<\/strong> RAG can enforce document-level permissions before the model sees anything.<\/li>\n<\/ul>\n<p>The trade-off is system complexity. RAG requires document ingestion pipelines, embedding model choice (text-embedding-3-small, Voyage, Cohere), vector storage, retrieval orchestration (LangChain, LlamaIndex, Haystack), and chunking strategy decisions. Latency also rises by 50\u2013300ms depending on retrieval depth. Caching helps, but the operational surface is real.<\/p>\n<p>The most common production RAG win: fine-tune for behavioral consistency, RAG for knowledge currency. The two techniques compose.<\/p>\n<h2>When Fine-Tuning Pays Off<\/h2>\n<p>Fine-tuning is expensive and slow, but it&#8217;s the only technique that produces a persistent model variant. The base model you start with and the fine-tuned version that comes out are different artifacts with different inference characteristics.<\/p>\n<p>Use fine-tuning when:<\/p>\n<ul>\n<li><strong>Consistency matters across thousands of queries.<\/strong> Prompt engineering will drift; a fine-tuned model won&#8217;t.<\/li>\n<li><strong>Domain jargon is essential.<\/strong> Legal, medical, scientific, or industry-specific vocabulary that the base model mishandles.<\/li>\n<li><strong>Latency is critical.<\/strong> A fine-tuned small model often beats a prompted large model on both speed and cost.<\/li>\n<li><strong>You&#8217;re optimizing for evaluable metrics.<\/strong> If you can score outputs and the patterns are learnable, fine-tuning will land them in the model weights.<\/li>\n<li><strong>You have abundant high-quality data.<\/strong> Hundreds to thousands of labeled examples, ideally more.<\/li>\n<\/ul>\n<p>Fine-tuning has two real ceilings. First, knowledge is frozen at training time\u2014fine-tuned models know what they knew when trained. Pair with RAG if your data updates. Second, fine-tuning inherits everything in the base model; it amplifies base behavior, good and bad. Don&#8217;t fine-tune to fix a hallucination problem the base model has; you&#8217;ll just get a fine-tuned hallucinator.<\/p>\n<h2>Comparison Table<\/h2>\n<table>\n<thead>\n<tr>\n<th>Factor<\/th>\n<th>Prompt Engineering<\/th>\n<th>RAG<\/th>\n<th>Fine-Tuning<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Setup cost<\/td>\n<td>None<\/td>\n<td>Medium (vector DB, ingestion)<\/td>\n<td>High (compute, data prep)<\/td>\n<\/tr>\n<tr>\n<td>Knowledge updates<\/td>\n<td>Edit prompt<\/td>\n<td>Update document store<\/td>\n<td>Retrain<\/td>\n<\/tr>\n<tr>\n<td>Latency overhead<\/td>\n<td>Minimal<\/td>\n<td>50\u2013300ms retrieval<\/td>\n<td>None (model-side)<\/td>\n<\/tr>\n<tr>\n<td>Per-query cost<\/td>\n<td>Prompt token cost<\/td>\n<td>Prompt + retrieved context cost<\/td>\n<td>Often lower (smaller fine-tuned model)<\/td>\n<\/tr>\n<tr>\n<td>Hallucination mitigation<\/td>\n<td>Marginal<\/td>\n<td>Strong (grounded)<\/td>\n<td>Model-dependent<\/td>\n<\/tr>\n<tr>\n<td>Behavioral consistency<\/td>\n<td>Varies with prompt<\/td>\n<td>Same<\/td>\n<td>High (learned into weights)<\/td>\n<\/tr>\n<tr>\n<td>Best for<\/td>\n<td>Prototyping, general use, fast iteration<\/td>\n<td>Live data, audit trails, zero-training scenarios<\/td>\n<td>Domain specialization, high-volume consistency<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h2>How to Choose<\/h2>\n<p>Three variables matter most, in this order: (1) do you need new knowledge, (2) do you need behavioral consistency, (3) what&#8217;s your budget and timeline.<\/p>\n<p><strong>Need new knowledge, no consistency requirement, low budget:<\/strong> RAG. The knowledge is what you actually need; behavior is incidental.<\/p>\n<p><strong>No new knowledge, need consistency, mid budget:<\/strong> Fine-tuning. Behavior is the deliverable; you can use any general base model.<\/p>\n<p><strong>Need both:<\/strong> Both. Fine-tune for tone\/format\/voice. RAG for facts. This is the architecture behind most production chatbots that don&#8217;t hallucinate and don&#8217;t sound robotic.<\/p>\n<p><strong>Neither:<\/strong> Just prompt. Most use cases fit here. Don&#8217;t over-engineer.<\/p>\n<h2>What Most Projects Get Wrong<\/h2>\n<p>Three recurring failure modes show up across consulting work.<\/p>\n<p><strong>Fine-tuning first.<\/strong> Teams invest weeks in fine-tuning a model to answer questions from documents, when RAG would have worked immediately and updated dynamically. Fine-tuning is not a knowledge-injection mechanism\u2014it&#8217;s a behavior-shaping mechanism. Once trained, the model knows nothing it didn&#8217;t know at training.<\/p>\n<p><strong>RAG without evaluation.<\/strong> Retrieval quality varies wildly. A model that retrieves the wrong chunks will hallucinate confidently. Add retrieval accuracy eval before shipping, not after.<\/p>\n<p><strong>Prompt-only at scale.<\/strong> A prompt that works for one user will hit instruction-following ceilings at 10,000 concurrent varied users. If your downstream metrics drop at volume, escalate to fine-tuning or hybrid.<\/p>\n<h2>Production Architecture Recommendations<\/h2>\n<p>Three recipes that work for most teams in 2026.<\/p>\n<p><strong>Customer support.<\/strong> RAG over knowledge base + ticket history, fine-tuned tone for brand voice. The RAG handles facts; the fine-tuning handles &#8220;your company sounds like this.&#8221;<\/p>\n<p><strong>Document Q&amp;A.<\/strong> Pure RAG. Citation requirements make fine-tuning the wrong tool\u2014fine-tuned answers can&#8217;t always cite sources the way retrieval-driven answers can.<\/p>\n<p><strong>Code assistants.<\/strong> Fine-tuned on the team&#8217;s codebase for style and conventions; prompt the base model with retrieved documentation snippets for library specifics. The Composer-style tools all use this layered approach.<\/p>\n<h2>Choosing the Stack<\/h2>\n<p>Practical defaults:<\/p>\n<ul>\n<li>Start with prompt engineering on a strong base model (Claude Sonnet 4.5, GPT-5, Gemini 2.5 Pro).<\/li>\n<li>Add RAG the moment factual accuracy or currency matters.<\/li>\n<li>Fine-tune only when you&#8217;ve outgrown prompting and have evaluation data proving the gap.<\/li>\n<li>Plan the architecture to support all three; you&#8217;ll need at least two by year two.<\/li>\n<\/ul>\n<p>The teams that get this right treat prompt engineering, RAG, and fine-tuning as layers in a stack, not competitors on a menu. Each technique exists to do something the others can&#8217;t. The error isn&#8217;t picking one\u2014it&#8217;s refusing to combine them.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Fine-Tuning vs RAG vs Prompt Engineering: When to Use Each Bottom Line Up Front: Prompt engineering, fine-tuning, and retrieval-augmented generation (RAG) solve different problems\u2014prompting handles flexible instruction-following, RAG grounds the model in current or proprietary knowledge, fine-tuning produces consistent behavior at scale. Most production AI systems combine two or all three; the wrong choice is [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":20467,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_feature_clip_id":0,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_publicize_message":"","jetpack_publicize_feature_enabled":true,"jetpack_social_post_already_shared":false,"jetpack_social_options":{"image_generator_settings":{"template":"highway","default_image_id":0,"font":"","enabled":false},"version":2},"jetpack_post_was_ever_published":false},"categories":[8],"tags":[],"class_list":["post-20219","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tools-resources"],"jetpack_publicize_connections":[],"jetpack_sharing_enabled":true,"jetpack-related-posts":[{"id":20671,"url":"https:\/\/aimade.tech\/?p=20671","url_meta":{"origin":20219,"position":0},"title":"Small language models in 2026: when 7B beats 70B","author":"Mr. Technology","date":"July 30, 2026","format":false,"excerpt":"Small language models in 2026 \u2014 when 7B beats 70B, with the cost-adjusted benchmark of Llama-3.1-8B vs GPT-4o across 11 enterprise tasks. The 2026 cutoff.","rel":"","context":"In &quot;AI Models&quot;","block_context":{"text":"AI Models","link":"https:\/\/aimade.tech\/?cat=297"},"img":{"alt_text":"","src":"","width":0,"height":0},"classes":[]},{"id":20645,"url":"https:\/\/aimade.tech\/?p=20645","url_meta":{"origin":20219,"position":1},"title":"RAG isn&#8217;t dead: retrieval-augmented production in 2026","author":"","date":"July 24, 2026","format":false,"excerpt":"RAG production 2026 is winning \u2014 hybrid retrieval, reranking, and eval gates are now default. We surveyed 47 teams and broke down 3 production case studies.","rel":"","context":"In &quot;AI Deep Dives&quot;","block_context":{"text":"AI Deep Dives","link":"https:\/\/aimade.tech\/?cat=304"},"img":{"alt_text":"Dark editorial research desk with two monitors showing a RAG pipeline: vector database on the left, retriever-reranker-LLM flow on the right","src":"https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/07\/rag-production-2026-hero-scaled.jpg?fit=1200%2C670&ssl=1&resize=350%2C200","width":350,"height":200,"srcset":"https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/07\/rag-production-2026-hero-scaled.jpg?fit=1200%2C670&ssl=1&resize=350%2C200 1x, https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/07\/rag-production-2026-hero-scaled.jpg?fit=1200%2C670&ssl=1&resize=525%2C300 1.5x, https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/07\/rag-production-2026-hero-scaled.jpg?fit=1200%2C670&ssl=1&resize=700%2C400 2x, https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/07\/rag-production-2026-hero-scaled.jpg?fit=1200%2C670&ssl=1&resize=1050%2C600 3x"},"classes":[]},{"id":20639,"url":"https:\/\/aimade.tech\/?p=20639","url_meta":{"origin":20219,"position":2},"title":"LLM Context Windows: Why Your 1M-Token Model Only Uses 32K","author":"Lucy Monday","date":"July 22, 2026","format":false,"excerpt":"LLM context window limits explained: RULER and LongBench v2 show frontier models lose 50%+ accuracy past 64K. A 1M-token window is the ceiling, not the deliverable.","rel":"","context":"In &quot;AI Deep Dives&quot;","block_context":{"text":"AI Deep Dives","link":"https:\/\/aimade.tech\/?cat=304"},"img":{"alt_text":"Long printed document roll spilling off an editorial research desk with a cyan accent - representing how LLM context windows are advertised long but used short.","src":"https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/07\/aimade-context-window-hero.png?fit=1200%2C686&ssl=1&resize=350%2C200","width":350,"height":200,"srcset":"https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/07\/aimade-context-window-hero.png?fit=1200%2C686&ssl=1&resize=350%2C200 1x, https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/07\/aimade-context-window-hero.png?fit=1200%2C686&ssl=1&resize=525%2C300 1.5x, https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/07\/aimade-context-window-hero.png?fit=1200%2C686&ssl=1&resize=700%2C400 2x, https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/07\/aimade-context-window-hero.png?fit=1200%2C686&ssl=1&resize=1050%2C600 3x"},"classes":[]},{"id":1561,"url":"https:\/\/aimade.tech\/?p=1561","url_meta":{"origin":20219,"position":3},"title":"Midjourney v7 Review: Is It Worth the Upgrade?","author":"Mr. Technology","date":"April 9, 2026","format":false,"excerpt":"Hey guys, Monday here. Midjourney v7 dropped a few weeks ago, and I finally had a chance to put it through its paces properly. Short answer: yes, it's worth the upgrade if you work with AI image generation. Long answer: below. What You Need to Know:Midjourney v7 delivers significantly improved\u2026","rel":"","context":"In &quot;Midjourney &amp; Creative AI&quot;","block_context":{"text":"Midjourney &amp; Creative AI","link":"https:\/\/aimade.tech\/?cat=314"},"img":{"alt_text":"","src":"https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/04\/midjourney-v7-cover.jpg?fit=1024%2C1024&ssl=1&resize=350%2C200","width":350,"height":200,"srcset":"https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/04\/midjourney-v7-cover.jpg?fit=1024%2C1024&ssl=1&resize=350%2C200 1x, https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/04\/midjourney-v7-cover.jpg?fit=1024%2C1024&ssl=1&resize=525%2C300 1.5x, https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/04\/midjourney-v7-cover.jpg?fit=1024%2C1024&ssl=1&resize=700%2C400 2x"},"classes":[]},{"id":20651,"url":"https:\/\/aimade.tech\/?p=20651","url_meta":{"origin":20219,"position":4},"title":"AI Energy Consumption 2026: Cost per Query and Grid Demand","author":"","date":"July 25, 2026","format":false,"excerpt":"AI energy consumption in 2026, measured per Gemini prompt and grid scale, with the engineering variables that change every estimate\u2014start auditing now.","rel":"","context":"In &quot;Tools &amp; Resources&quot;","block_context":{"text":"Tools &amp; Resources","link":"https:\/\/aimade.tech\/?cat=8"},"img":{"alt_text":"AI data center power monitoring console with server racks and energy graphs","src":"https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/07\/ai-energy-consumption-2026-scaled.jpg?fit=1200%2C670&ssl=1&resize=350%2C200","width":350,"height":200,"srcset":"https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/07\/ai-energy-consumption-2026-scaled.jpg?fit=1200%2C670&ssl=1&resize=350%2C200 1x, https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/07\/ai-energy-consumption-2026-scaled.jpg?fit=1200%2C670&ssl=1&resize=525%2C300 1.5x, https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/07\/ai-energy-consumption-2026-scaled.jpg?fit=1200%2C670&ssl=1&resize=700%2C400 2x, https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/07\/ai-energy-consumption-2026-scaled.jpg?fit=1200%2C670&ssl=1&resize=1050%2C600 3x"},"classes":[]},{"id":1473,"url":"https:\/\/aimade.tech\/?p=1473","url_meta":{"origin":20219,"position":5},"title":"Google&#8217;s Gemma 4 Now Runs on a Raspberry Pi \u2014 And It Is Actually Useful","author":"Mr. Technology","date":"April 7, 2026","format":false,"excerpt":"Hey guys, Mr. Technology here. I have been waiting YEARS for this. Open-source AI models that you can actually run locally \u2014 not some sad demo that barely fits in memory, but something genuinely useful. Google just made a big leap with Gemma 4, and it runs on my Raspberry\u2026","rel":"","context":"In &quot;Open Source AI&quot;","block_context":{"text":"Open Source AI","link":"https:\/\/aimade.tech\/?cat=307"},"img":{"alt_text":"","src":"https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/04\/gemma4-cover.jpg?fit=1024%2C1024&ssl=1&resize=350%2C200","width":350,"height":200,"srcset":"https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/04\/gemma4-cover.jpg?fit=1024%2C1024&ssl=1&resize=350%2C200 1x, https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/04\/gemma4-cover.jpg?fit=1024%2C1024&ssl=1&resize=525%2C300 1.5x, https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/04\/gemma4-cover.jpg?fit=1024%2C1024&ssl=1&resize=700%2C400 2x"},"classes":[]}],"jetpack_featured_media_url":"https:\/\/i0.wp.com\/aimade.tech\/wp-content\/uploads\/2026\/05\/img-03-agents-sdk.png?fit=1376%2C768&ssl=1","_links":{"self":[{"href":"https:\/\/aimade.tech\/index.php?rest_route=\/wp\/v2\/posts\/20219","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aimade.tech\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aimade.tech\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aimade.tech\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/aimade.tech\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=20219"}],"version-history":[{"count":2,"href":"https:\/\/aimade.tech\/index.php?rest_route=\/wp\/v2\/posts\/20219\/revisions"}],"predecessor-version":[{"id":20629,"href":"https:\/\/aimade.tech\/index.php?rest_route=\/wp\/v2\/posts\/20219\/revisions\/20629"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aimade.tech\/index.php?rest_route=\/wp\/v2\/media\/20467"}],"wp:attachment":[{"href":"https:\/\/aimade.tech\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=20219"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aimade.tech\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=20219"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aimade.tech\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=20219"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}