{"id":127,"date":"2026-02-12T19:42:41","date_gmt":"2026-02-12T19:42:41","guid":{"rendered":"https:\/\/brgt.com.br\/blog-admin\/?p=127"},"modified":"2026-02-12T19:42:43","modified_gmt":"2026-02-12T19:42:43","slug":"testei-6-combinacoes-de-llms-para-um-bot-corporativo-de-busca-o-resultado-me-surpreendeu","status":"publish","type":"post","link":"https:\/\/brgt.com.br\/blog-admin\/testei-6-combinacoes-de-llms-para-um-bot-corporativo-de-busca-o-resultado-me-surpreendeu\/","title":{"rendered":"Testei 6 combina\u00e7\u00f5es de LLMs para um bot corporativo de busca. O resultado me surpreendeu."},"content":{"rendered":"\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">O problema<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Aqui na BRGT Engenharia lidamos todos os dias com uma grande quantidade  de arquivos: propostas comerciais, projetos estruturais, planilhas de or\u00e7amento, plantas DWG, relat\u00f3rios de sondagem, contratos e at\u00e9 scripts de automa\u00e7\u00e3o.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">O problema era simples (e comum em qualquer empresa):<br>a informa\u00e7\u00e3o existia, mas encontrar r\u00e1pido era dif\u00edcil.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">A solu\u00e7\u00e3o que criamos foi um bot no Telegram com intelig\u00eancia artificial que indexa os arquivos do servidor e responde perguntas em linguagem natural.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Exemplo real:<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u201cQual foi o valor do or\u00e7amento da \u00faltima proposta?\u201d<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">O bot busca nos arquivos, interpreta e responde.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Como o sistema funciona (arquitetura RAG)<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">O sistema usa uma arquitetura RAG (Retrieval-Augmented Generation) com duas etapas de LLM:<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>1. Extra\u00e7\u00e3o de termos<\/strong><br>O modelo converte a pergunta do usu\u00e1rio em palavras-chave otimizadas para busca.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>2. Gera\u00e7\u00e3o de resposta<\/strong><br>Outro modelo recebe os arquivos encontrados como contexto e gera a resposta final.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">A grande d\u00favida era:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Qual combina\u00e7\u00e3o de modelos usar?<\/li>\n\n\n\n<li>Um modelo caro?<\/li>\n\n\n\n<li>Um modelo local gratuito?<\/li>\n\n\n\n<li>Um modelo barato na nuvem?<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Para evitar achismo, rodamos <strong>6 benchmarks sistem\u00e1ticos<\/strong>.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Metodologia<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">Avalia\u00e7\u00e3o automatizada<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Todos os testes seguiram o mesmo processo:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>10 consultas reais simulando o uso di\u00e1rio<\/li>\n\n\n\n<li>Avalia\u00e7\u00e3o autom\u00e1tica em 5 crit\u00e9rios:\n<ul class=\"wp-block-list\">\n<li>Precis\u00e3o<\/li>\n\n\n\n<li>Completude<\/li>\n\n\n\n<li>Relev\u00e2ncia<\/li>\n\n\n\n<li>Clareza<\/li>\n\n\n\n<li>Utilidade<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>Uma resposta de refer\u00eancia considerada nota 10<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Tudo comparado de forma objetiva.<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\"><strong>Nota metodol\u00f3gica importante<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Todas as notas apresentadas s\u00e3o relativas a uma resposta de refer\u00eancia avaliada automaticamente pelo modelo mais avan\u00e7ado dispon\u00edvel no dia do teste: <strong>OpenAI GPT-5.2<\/strong>.<br>Isso garante padroniza\u00e7\u00e3o, consist\u00eancia e compara\u00e7\u00e3o objetiva entre os benchmarks.<\/p>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Modelos testados<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>Modelo<\/th><th>Tipo<\/th><th>Custo input (1M tokens)<\/th><th>Custo output (1M tokens)<\/th><\/tr><\/thead><tbody><tr><td>GPT-5.2<\/td><td>API<\/td><td>$1,25<\/td><td>$10,00<\/td><\/tr><tr><td>GPT-5-mini<\/td><td>API<\/td><td>$0,25<\/td><td>$2,00<\/td><\/tr><tr><td>GPT-5-nano<\/td><td>API<\/td><td>$0,05<\/td><td>$0,40<\/td><\/tr><tr><td>GPT-4o-mini<\/td><td>API<\/td><td>$0,15<\/td><td>$0,60<\/td><\/tr><tr><td>Ollama qwen2.5:7b<\/td><td>Local<\/td><td>Gr\u00e1tis<\/td><td>Gr\u00e1tis<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">O \u00edndice continha cerca de <strong>1.100 arquivos<\/strong> com busca full-text.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Benchmark 1 \u2014 Local vs Nuvem<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Pergunta:<\/strong> modelo local compete com API?<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>Modelo<\/th><th>Nota<\/th><th>Tempo m\u00e9dio<\/th><\/tr><\/thead><tbody><tr><td>GPT-4o-mini<\/td><td>5,4<\/td><td>2,0 s<\/td><\/tr><tr><td>Ollama qwen2.5:7b<\/td><td>4,4<\/td><td>6,1 s<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">Insight importante:<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">O problema n\u00e3o era s\u00f3 o modelo \u2014 era usar o mesmo modelo fraco para extrair termos e responder.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Benchmark 2 \u2014 Usando modelos diferentes para cada etapa<\/h2>\n\n\n\n<figure class=\"wp-block-table is-style-regular\"><table class=\"has-fixed-layout\"><thead><tr><th>Combina\u00e7\u00e3o (extra\u00e7\u00e3o + resposta)<\/th><th>Nota<\/th><th>Tempo<\/th><\/tr><\/thead><tbody><tr><td>4o-mini + 5.2<\/td><td>7,3<\/td><td>7,2 s<\/td><\/tr><tr><td>Ollama + 5.2<\/td><td>6,5<\/td><td>8,0 s<\/td><\/tr><tr><td>5.2 + 4o-mini<\/td><td>6,1<\/td><td>2,4 s<\/td><\/tr><tr><td>5.2 + Ollama<\/td><td>5,9<\/td><td>7,1 s<\/td><\/tr><tr><td>4o-mini + 4o-mini<\/td><td>5,2<\/td><td>2,5 s<\/td><\/tr><tr><td>4o-mini + Ollama<\/td><td>4,6<\/td><td>7,0 s<\/td><\/tr><tr><td>Ollama + 4o-mini<\/td><td>4,3<\/td><td>4,1 s<\/td><\/tr><tr><td>Ollama + Ollama<\/td><td>3,9<\/td><td>8,1 s<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">Insight:<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">O modelo de resposta impacta muito mais na qualidade final do que o de extra\u00e7\u00e3o.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Benchmark 3 \u2014 Fam\u00edlia GPT-5<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>Combina\u00e7\u00e3o<\/th><th>Nota<\/th><th>Custo (10 consultas)<\/th><\/tr><\/thead><tbody><tr><td>5.2 + 5-nano<\/td><td>8,0<\/td><td>$0,03<\/td><\/tr><tr><td>5.2 + 5-mini<\/td><td>7,5<\/td><td>$0,08<\/td><\/tr><tr><td>5-nano + 5.2<\/td><td>6,6<\/td><td>$0,48<\/td><\/tr><tr><td>5-mini + 5.2<\/td><td>6,3<\/td><td>$0,48<\/td><\/tr><tr><td>5-nano + 5-mini<\/td><td>6,2<\/td><td>$0,01<\/td><\/tr><tr><td>5-mini + 5-nano<\/td><td>6,2<\/td><td>$0,01<\/td><\/tr><tr><td>5-mini + 5-mini<\/td><td>6,1<\/td><td>$0,08<\/td><\/tr><tr><td>5-nano + 5-nano<\/td><td>5,9<\/td><td>$0,01<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">Conclus\u00e3o:<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Modelo forte para extrair + modelo barato para responder manteve alta qualidade com custo baix\u00edssimo.<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u26a0\ufe0f <strong>Descoberta<\/strong> Cr\u00edtica:<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Modelos menores e mais baratos retornaram <strong>string vazia na etapa de extra\u00e7\u00e3o de termos<\/strong> em m\u00faltiplos testes.<br>A tarefa parece simples, mas exige capacidade de racioc\u00ednio que modelos compactos ainda n\u00e3o entregam com confiabilidade.<\/p>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Benchmark 4 \u2014 Varia\u00e7\u00e3o do prompt do sistema<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>Prompt<\/th><th>Nota<\/th><th>Descri\u00e7\u00e3o<\/th><\/tr><\/thead><tbody><tr><td>Especialista t\u00e9cnico<\/td><td>7,9<\/td><td>Persona de engenheiro civil s\u00eanior<\/td><\/tr><tr><td>Original<\/td><td>7,3<\/td><td>Gen\u00e9rico<\/td><\/tr><tr><td>Detalhado<\/td><td>7,2<\/td><td>Contexto da empresa<\/td><\/tr><tr><td>Passo-a-passo<\/td><td>7,1<\/td><td>Racioc\u00ednio guiado<\/td><\/tr><tr><td>Anti-alucina\u00e7\u00e3o<\/td><td>6,0<\/td><td>Regras r\u00edgidas<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">Prompts muito restritivos pioraram o resultado.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Benchmark 5 \u2014 Prompt de extra\u00e7\u00e3o<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>Tipo<\/th><th>Nota<\/th><th>Observa\u00e7\u00e3o<\/th><\/tr><\/thead><tbody><tr><td>Categorizado<\/td><td>7,6<\/td><td>Separa\u00e7\u00e3o por tipo<\/td><\/tr><tr><td>Original<\/td><td>7,5<\/td><td>Simples<\/td><\/tr><tr><td>Expans\u00e3o sem\u00e2ntica<\/td><td>6,7<\/td><td>Sin\u00f4nimos t\u00e9cnicos<\/td><\/tr><tr><td>Contextualizado<\/td><td>6,4<\/td><td>Regras anti-gen\u00e9rico<\/td><\/tr><tr><td>Few-shot<\/td><td>6,4<\/td><td>Exemplos<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">Conclus\u00e3o:<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Prompt simples funciona.<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u26a0\ufe0f <strong>Descoberta <\/strong>Cr\u00edtica:<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Prompts excessivamente restritivos (anti-alucina\u00e7\u00e3o) reduziram a utilidade das respostas.<br>Ao tentar impedir qualquer risco de erro, o modelo passou a responder de forma curta, vaga e pouco pr\u00e1tica.<\/p>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Benchmark 6 \u2014 Compara\u00e7\u00e3o final de modelos de resposta<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>Modelo<\/th><th>Nota<\/th><th>Custo (10 consultas)<\/th><th>Tempo<\/th><\/tr><\/thead><tbody><tr><td>GPT-5.2<\/td><td>8,6<\/td><td>$0,318<\/td><td>5,7 s<\/td><\/tr><tr><td>GPT-5-nano<\/td><td>7,7<\/td><td>$0,0108<\/td><td>17,2 s<\/td><\/tr><tr><td>GPT-4o-mini<\/td><td>6,6<\/td><td>$0,0079<\/td><td>1,6 s<\/td><\/tr><tr><td>Ollama<\/td><td>5,9<\/td><td>Gr\u00e1tis<\/td><td>10 s<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">O modelo econ\u00f4mico entregou cerca de <strong>90% da qualidade por 3% do custo<\/strong>.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Configura\u00e7\u00e3o final (produ\u00e7\u00e3o)<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>Etapa<\/th><th>Escolha<\/th><\/tr><\/thead><tbody><tr><td>Extra\u00e7\u00e3o<\/td><td>Modelo robusto (5.2)<\/td><\/tr><tr><td>Resposta<\/td><td>Modelo econ\u00f4mico (5.2 nano)<\/td><\/tr><tr><td>Prompt<\/td><td>Persona t\u00e9cnica<\/td><\/tr><tr><td>Extra\u00e7\u00e3o<\/td><td>Simples e direto<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Proje\u00e7\u00e3o de custos<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>Uso<\/th><th>Consultas\/dia<\/th><th>Custo mensal<\/th><\/tr><\/thead><tbody><tr><td>Leve<\/td><td>5<\/td><td>~R$ 0,70<\/td><\/tr><tr><td>Moderado<\/td><td>10<\/td><td>~R$ 1,40<\/td><\/tr><tr><td>Intenso<\/td><td>50<\/td><td>~R$ 6,90<\/td><\/tr><tr><td>Pesado<\/td><td>100<\/td><td>~R$ 13,80<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">O que aprendemos<\/h2>\n\n\n\n<ol class=\"wp-block-list\">\n<li>Separe extra\u00e7\u00e3o e resposta<\/li>\n\n\n\n<li>O modelo de resposta pesa mais<\/li>\n\n\n\n<li>Modelos baratos n\u00e3o extraem bem termos<\/li>\n\n\n\n<li>Persona t\u00e9cnica melhora respostas<\/li>\n\n\n\n<li>Anti-alucina\u00e7\u00e3o excessiva piora<\/li>\n\n\n\n<li>Prompt complexo atrapalha<\/li>\n\n\n\n<li>Local nem sempre compensa<\/li>\n<\/ol>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Arquitetura final (resumo simples)<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Usu\u00e1rio pergunta no Telegram \u2192<br>Modelo extrai palavras-chave \u2192<br>Busca nos arquivos \u2192<br>Modelo gera resposta \u2192<br>Resposta enviada.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Simples, r\u00e1pido e barato.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">C\u00f3digo aberto?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">O sistema est\u00e1 em produ\u00e7\u00e3o na BRGT Engenharia.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Os scripts de benchmark podem ser compartilhados para quem quiser implementar.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Stack:<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Python<\/li>\n\n\n\n<li>Telegram Bot<\/li>\n\n\n\n<li>OpenAI API<\/li>\n\n\n\n<li>SQLite<\/li>\n\n\n\n<li>Leitura de PDF, DWG e planilhas<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Vers\u00e3o t\u00e9cnica completa<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A vers\u00e3o completa do benchmark, com todos os dados, tabelas detalhadas e an\u00e1lises t\u00e9cnicas aprofundadas, est\u00e1 dispon\u00edvel aqui:<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\ud83d\udc49 <strong><a href=\"https:\/\/drive.google.com\/drive\/folders\/14Q5Cez1kqJVDb9vzD78cI7az3LSulmAL?usp=drive_link\">Acessar vers\u00e3o t\u00e9cnica completa do benchmark<\/a><\/strong><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Gostou do conte\u00fado? Nos siga nas redes sociais e acompanhe nossos pr\u00f3ximos posts.<\/strong><\/p>\n","protected":false},"excerpt":{"rendered":"<p>O problema Aqui na BRGT Engenharia lidamos todos os dias com uma grande quantidade de arquivos: propostas comerciais, projetos estruturais, planilhas de or\u00e7amento, plantas DWG, relat\u00f3rios de sondagem, contratos e at\u00e9 scripts de automa\u00e7\u00e3o. O problema era simples (e comum em qualquer empresa):a informa\u00e7\u00e3o existia, mas encontrar r\u00e1pido era dif\u00edcil. A solu\u00e7\u00e3o que criamos foi &hellip; <a href=\"https:\/\/brgt.com.br\/blog-admin\/testei-6-combinacoes-de-llms-para-um-bot-corporativo-de-busca-o-resultado-me-surpreendeu\/\" class=\"more-link\">Continue lendo<span class=\"screen-reader-text\"> &#8220;Testei 6 combina\u00e7\u00f5es de LLMs para um bot corporativo de busca. O resultado me surpreendeu.&#8221;<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":142,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"pagelayer_contact_templates":[],"_pagelayer_content":"","footnotes":""},"categories":[1],"tags":[],"class_list":["post-127","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/brgt.com.br\/blog-admin\/wp-json\/wp\/v2\/posts\/127","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/brgt.com.br\/blog-admin\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/brgt.com.br\/blog-admin\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/brgt.com.br\/blog-admin\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/brgt.com.br\/blog-admin\/wp-json\/wp\/v2\/comments?post=127"}],"version-history":[{"count":5,"href":"https:\/\/brgt.com.br\/blog-admin\/wp-json\/wp\/v2\/posts\/127\/revisions"}],"predecessor-version":[{"id":143,"href":"https:\/\/brgt.com.br\/blog-admin\/wp-json\/wp\/v2\/posts\/127\/revisions\/143"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/brgt.com.br\/blog-admin\/wp-json\/wp\/v2\/media\/142"}],"wp:attachment":[{"href":"https:\/\/brgt.com.br\/blog-admin\/wp-json\/wp\/v2\/media?parent=127"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/brgt.com.br\/blog-admin\/wp-json\/wp\/v2\/categories?post=127"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/brgt.com.br\/blog-admin\/wp-json\/wp\/v2\/tags?post=127"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}