[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-nv-tech-assistant":3,"mdc--yuhnae-key":34,"related-org-nvidia-nv-tech-assistant":1357,"related-repo-nvidia-nv-tech-assistant":1516},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":11,"stars":23,"repoUrl":24,"updatedAt":25,"license":26,"forks":27,"topics":28,"repo":29,"sourceUrl":32,"mdContent":33},"nv-tech-assistant","answer NVIDIA technical questions","Answer NVIDIA technical questions by searching authorized sources and citing retrieved evidence rather than relying on memory. Use for NVIDIA SDK or library usage and sample code, technology recommendations (including models and NGC\u002FGitHub assets), adoption stories, troubleshooting, and product integration planning. Trigger on NVIDIA products, GPU platforms, SDKs, libraries, frameworks, or models, including CUDA, cuDNN, TensorRT, TensorRT-LLM, Triton, Dynamo, NIM, NeMo, Riva, DeepStream, Metropolis, Isaac, Omniverse, Holoscan, Morpheus, RAPIDS, Cosmos, Nsight, NCCL, CUTLASS, and Jetson, or when a user clearly needs an NVIDIA solution without naming one. Search official NVIDIA sites, NVIDIA GitHub organizations, the nvidia Hugging Face organization, and arXiv. Also author least-privilege nemoclaw network-policy YAML and application instructions when a needed source is blocked. Prefer this skill over answering NVIDIA specifics directly.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},"nvidia","NVIDIA","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fnvidia.png",[12,16,19,20],{"name":13,"slug":14,"type":15},"Documentation","documentation","tag",{"name":17,"slug":18,"type":15},"Reference","reference",{"name":9,"slug":8,"type":15},{"name":21,"slug":22,"type":15},"SDK","sdk",72,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fnemoclaw-community","2026-08-05T05:58:19.93533",null,32,[],{"repoUrl":24,"stars":23,"forks":27,"topics":30,"description":31},[],"Our repository of community driven examples, showcases, and integrations","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fnemoclaw-community\u002Ftree\u002FHEAD\u002Fexamples\u002Frecipes\u002Fnvidia\u002Fnv-tech-assistant\u002Fskills\u002Fnv-tech-assistant","---\nname: nv-tech-assistant\ndescription: >-\n  Answer NVIDIA technical questions by searching authorized sources and citing retrieved evidence\n  rather than relying on memory. Use for NVIDIA SDK or library usage and sample code, technology\n  recommendations (including models and NGC\u002FGitHub assets), adoption stories, troubleshooting, and\n  product integration planning. Trigger on NVIDIA products, GPU platforms, SDKs, libraries,\n  frameworks, or models, including CUDA, cuDNN, TensorRT, TensorRT-LLM, Triton, Dynamo, NIM, NeMo,\n  Riva, DeepStream, Metropolis, Isaac, Omniverse, Holoscan, Morpheus, RAPIDS, Cosmos, Nsight, NCCL,\n  CUTLASS, and Jetson, or when a user clearly needs an NVIDIA solution without naming one. Search\n  official NVIDIA sites, NVIDIA GitHub organizations, the nvidia Hugging Face organization, and\n  arXiv. Also author least-privilege nemoclaw network-policy YAML and application instructions\n  when a needed source is blocked. Prefer this skill over answering NVIDIA specifics directly.\n---\n\n# NVIDIA Technical Helper\n\n## What this is for\n\nNVIDIA's technology surface is enormous, overlapping, and fast-moving: dozens of SDKs with\nsimilar names, models that get renamed and re-released, repos that move between GitHub orgs, and\nversion numbers that turn last quarter's answer into misinformation. A confident answer from\nmemory is the single biggest failure mode here — it produces plausible SDK names, invented API\nsignatures, dead repos, and models that never existed.\n\nSo the whole job of this skill is **grounding**: turn every NVIDIA technical question into\nsearches against authorized sources, retrieve the actual pages, and answer only from what you\njust read — with a link for every claim. If you can't find it, say so. A short honest answer with\nreal links beats a long confident one that might be fiction.\n\n## Core principles\n\n1. **Ground everything you retrieved this session; don't answer NVIDIA specifics from memory.**\n   Product names, versions, APIs, model IDs, repo paths, and \"does X support Y\" must come from a\n   page you actually fetched during this task. Your prior knowledge is fine for *deciding where to\n   search*, not for stating facts.\n\n2. **Cite every claim with the exact URL** you got it from. If a statement has no source you\n   retrieved, either go find one or drop the statement.\n\n3. **Include code only when the user asks for it or it is essentially necessary.** Most answers\n   don't need a code block — a link to the official example\u002Fdoc file (\"the runnable client is at\n   \u003CURL>\") lets the user check the details themselves and keeps the answer fast. Reserve inline\n   code for when the user explicitly wants sample code or the answer is meaningless without it\n   (e.g. the fix *is* a specific flag or call).\n\n4. **When you do include code, quote it verbatim from its source.** Retrieve the real\n   file (from the official repo, docs, or model card) and quote it exactly, with the source URL\n   and — for repos — the branch\u002Ftag or \"as of \u003Cdate>\". Do not synthesize, \"clean up\", or fill in\n   APIs from memory. If you must trim for length, show the real excerpt and link the full file. If\n   the user explicitly asks you to adapt it, keep changes minimal and mark clearly what you\n   changed versus the original.\n\n5. **Disambiguate the product first.** NVIDIA names collide and get confused constantly (TensorRT\n   vs. TensorRT-LLM, Triton Inference Server vs. other \"Triton\"s, NeMo the framework vs. unrelated\n   \"nemo\", NIM vs. NGC). If the target is ambiguous, resolve it before searching — see\n   `references\u002Fnvidia-landscape.md` for a disambiguation map and canonical entry points.\n\n6. **Prefer primary sources; label everything else.** Official docs and NVIDIA-owned repos\u002Fmodel\n   cards are ground truth. The developer blog is official but time-stamped (features change).\n   Forum posts and community repos are useful — especially for troubleshooting — but flag them as\n   unofficial and note whether an NVIDIA staff member or an accepted\u002F\"solved\" answer backs them.\n\n7. **Pin versions and dates.** State the SDK\u002Flibrary version, model revision, or article date you\n   read, because NVIDIA guidance changes fast. For a page's publish date, check its metadata (the\n   `datePublished` \u002F `article:published_time` field or the JSON-LD block in the page source), not\n   just the visible text — many NVIDIA pages show no visible date but carry one in metadata, so\n   \"no date\" is usually wrong. If the user is on a different version, say your answer may not\n   transfer.\n\n8. **Degrade gracefully when a source is unavailable.** If a page won't load, returns nothing\n   useful, or a source is unreachable, don't give up silently — use the other sources and be\n   explicit in your answer about what you couldn't check. Never fill the gap with a guess.\n\n9. **Be honest about gaps.** \"I couldn't find an official source for X\" and \"the only result is a\n   community post, unverified\" are good, expected answers. Never paper over a gap with a guess.\n\n## Workflow\n\n**Step 1 — Classify and disambiguate.** Identify which of the five question types this is (see\nplaybooks below) and pin down the exact NVIDIA product(s) involved. If unsure which product the\nuser means, resolve it first (`references\u002Fnvidia-landscape.md`).\n\n**Step 2 — Pick sources.** Route to the right sources for the question type (playbooks below).\nAlmost always search at least two independent sources so you can cross-check.\n\n**Step 3 — Search.** Prefer the Brave `WebSearch` tool if it is available — scope queries with\n`site:` filters (e.g. `site:docs.nvidia.com`, `site:github.com NVIDIA`). If `WebSearch` is\nunavailable or thin, fall back to the structured JSON search endpoints and `WebFetch`. All of the\nexact URL patterns (GitHub API, Hugging Face API, forum `search.json`, blog `wp-json`, arXiv API,\nNGC, docs) are in `references\u002Fsources-and-search.md` — read it when you need a recipe.\n\n**Step 4 — Retrieve and verify.** Open the actual pages\u002Ffiles. Confirm the thing exists and is\ncurrent before you rely on it. For a recommendation, open the repo\u002Fmodel\u002FSDK page and check it's\nNVIDIA-owned (or clearly note if it's third-party), maintained, and matches the ask.\n\n**Step 5 — Answer with citations.** Use the output format below. Lead with the direct answer, back\nevery claim with a link, quote code verbatim, and close with caveats (version\u002Fdate, unofficial\nsources, anything unavailable or unverified).\n\n## Question-type playbooks\n\n### 1. \"How do I use \u003CSDK>? Show me sample code.\"\n- **Sources:** official docs (`docs.nvidia.com`) for the how-to and API; the SDK's NVIDIA GitHub\n  repo for runnable examples (`examples\u002F`, `samples\u002F`, `tutorials\u002F`, README quickstart); the\n  developer blog for walkthroughs; a Hugging Face model card if it's a model.\n- **Do:** find the current install\u002Fsetup steps and the smallest complete example. Quote the code\n  **verbatim** with its source URL and version\u002Fbranch. Note prerequisites (CUDA version, GPU\n  requirement, OS) exactly as the source states them.\n- **Don't:** hand-write example code from memory or merge fragments from different versions.\n\n### 2. \"Recommend NVIDIA technology for \u003Ctask>.\"\n- **Sources:** the NGC catalog (`catalog.ngc.nvidia.com`) and `build.nvidia.com` for models\u002FNIMs;\n  NVIDIA GitHub orgs (repo search, scoped to NVIDIA — see rules below); the `nvidia` Hugging Face\n  org for models\u002Fdatasets; docs and the blog to confirm the intended use case.\n- **Do:** verify each candidate actually exists and is current by opening its page. Give 1–3\n  concrete options with what each is for, maturity\u002Frecency signals (last update, stars, downloads),\n  and links. Distinguish NVIDIA-owned from community\u002Fthird-party.\n- **Don't:** recommend from memory, or list deprecated\u002Frenamed products. If nothing fits, say so.\n\n### 3. \"Show me a success story \u002F real adoption of \u003Cproduct>.\"\n- **Sources:** the developer blog (`developer.nvidia.com\u002Fblog`) and `blogs.nvidia.com` for\n  customer stories and case studies; `www.nvidia.com` customer-story\u002Findustry pages; the SDK's\n  docs\u002Frepo \"used by\" or partner mentions.\n- **Do:** cite the specific case study (company, what they built, result) with its URL and date.\n  For the date, check the page metadata (`datePublished`\u002FJSON-LD), not just the visible text —\n  these pages often carry the date only in metadata. Attribute metrics as vendor\u002Fself-reported.\n  Prefer named, published cases over vague claims.\n- **Don't:** invent customers or metrics. If you only find NVIDIA's own marketing framing, say so.\n\n### 4. \"I hit \u003Cerror\u002Fissue> with \u003CSDK>. How do I fix it?\"\n- **Sources:** the developer **forums** (`forums.developer.nvidia.com` via `search.json`) — this\n  is the primary source for real errors; GitHub **Issues** on the SDK's repo; docs\n  troubleshooting\u002Frelease-notes sections for known issues.\n- **Do:** search the exact error text (and a generalized version). Prefer threads with an accepted\n  answer, a `status:solved` marker, or an NVIDIA staff reply. Summarize the fix and link the\n  thread; quote the working commands\u002Fcode verbatim. Note the versions the fix applied to.\n- **Don't:** present a community workaround as official, or guess at a fix. If unresolved\n  upstream, say that and link the open issue.\n\n### 5. \"How could my product use NVIDIA SDKs?\" (integration\u002Farchitecture)\n- **Sources:** docs (architecture\u002Fdeployment guides), the blog (reference architectures,\n  blueprints), NGC\u002F`build.nvidia.com` for deployable components, relevant repos for integration\n  examples.\n- **Do:** map the user's product\u002Fgoal to specific NVIDIA components, citing each. Sketch a\n  concrete path (which SDK does what, how they connect) grounded in real docs\u002Fexamples. Call out\n  hardware\u002Flicensing prerequisites you find.\n- **Don't:** produce a generic \"NVIDIA is great for AI\" pitch. Every recommended piece needs a\n  link and a reason.\n\n## Search source rules\n\n- **GitHub:** focus on NVIDIA-owned repositories for anything you recommend or quote as\n  authoritative. Scope searches to NVIDIA's organizations (e.g. `org:NVIDIA`) and confirm the repo\n  owner before treating it as official. Community repos can be mentioned but must be labeled as\n  such, with a note on activity\u002Fmaintenance. Known NVIDIA orgs and how to search the API are in\n  `references\u002Fsources-and-search.md`.\n- **Hugging Face:** prioritize the official org at `https:\u002F\u002Fhuggingface.co\u002Fnvidia` — query with\n  `?author=nvidia` first. Only broaden beyond it when nvidia has nothing suitable, and then clearly\n  mark non-NVIDIA models as community\u002Fthird-party and check their license and provenance.\n- **Cross-check:** for recommendations and factual claims, corroborate across at least two sources\n  (e.g., docs + repo, or NGC + blog) before stating something as settled.\n\n## Data sources\n\nSearch and fetch from these authorized sources — this is where NVIDIA's ground truth lives. Match\nthe source to the question (see the playbooks) and cross-check across at least two before treating\nsomething as settled.\n\n| Source | Best for | Where to look |\n|---|---|---|\n| Official docs | how-to, API reference, supported versions, known issues | `docs.nvidia.com` |\n| Product hubs | product overview, downloads, getting started | `developer.nvidia.com\u002F\u003Cproduct>` |\n| Developer blog | walkthroughs, reference architectures, blueprints, success stories | `developer.nvidia.com\u002Fblog` |\n| Corporate blog & customer stories | named adoption and case studies | `blogs.nvidia.com`, `www.nvidia.com` |\n| NGC catalog | pretrained models, containers, Helm charts | `catalog.ngc.nvidia.com` |\n| API catalog \u002F NIM | deployable inference microservices, hosted model endpoints | `build.nvidia.com` |\n| Developer forums | real-world errors and fixes (troubleshooting) | `forums.developer.nvidia.com` |\n| GitHub (NVIDIA orgs) | official example code, repos, issues | `github.com` \u002F `api.github.com` (NVIDIA, `rapidsai`, `NVIDIA-AI-IOT`, …) |\n| Hugging Face (nvidia org) | open models and datasets | `huggingface.co\u002Fnvidia` |\n| arXiv | the research behind a technique | `arxiv.org` |\n\n`references\u002Fsources-and-search.md` has the exact search recipes and URL patterns (forum\n`search.json`, blog `wp-json`, GitHub\u002FHugging Face APIs, arXiv API) for each of these — read it\nwhen you need a recipe.\n\n## Unblocking sources: nemoclaw network policies\n\nIf a source you need is blocked by the sandbox network policy — or the user asks to open access to\na new site — don't just report the gap: author a nemoclaw network-policy YAML for the blocked\nhost(s) and walk the user through applying it. Read `references\u002Fnemoclaw-network-policy.md` for\nthe full guide; `references\u002Fnemoclaw-policy-template.yaml` is an annotated starting point.\n\nUse least privilege for every endpoint. For sources that the assistant only searches, fetches, or\ndownloads from, set `access: read-only`, `protocol: rest`, and `enforcement: enforce`. Do not set\n`tls: skip` for ordinary HTTPS endpoints. **DO NOT set `access: full` for an API endpoint unless\nit is absolutely required by the user's requested workflow.** An endpoint being an API or\nsupporting search does not justify full access when GET\u002FHEAD is sufficient. If broader access or\nskipped TLS inspection is unavoidable, explain the exact requirement and security tradeoff.\n\nYou **cannot** apply the policy yourself — you are running inside the sandbox and are not allowed\nto change its network policy from within. Tell the user this explicitly. In short: give them the\nYAML (one file per site, listing every hostname it needs — subdomains are not implied), tell them\nto save it **on their host machine** as `\u003Cdomain-name>.yaml`, then run from the host:\n\n```bash\nnemoclaw \u003Csbx-name> policy-add --from-file \u003Cdomain-name>.yaml\n```\n\nThis applies the policy immediately to the running sandbox — no sandbox rebuild needed. Once the\nuser confirms, retry the blocked fetch to verify.\n\n## Output format\n\nAdapt to the question, but keep this backbone:\n\n- **Answer** — the direct response in 1–3 sentences up front.\n- **Details \u002F evidence** — the specifics, each with an inline link. Note versions and dates.\n- **Code** — only if the user asked for code or the answer needs it (core principle 3); otherwise\n  link the official example\u002Fdoc file instead. When included: fenced and quoted **verbatim**, with\n  the source URL and branch\u002Ftag above it. Mark any modification explicitly.\n- **Sources** — a short list of every URL you actually retrieved, so the user can verify.\n- **Caveats** — versions the answer assumes, anything unofficial\u002Fcommunity, and anything you\n  couldn't confirm or reach.\n\nKeep it as long as it needs to be and no longer. Prefer a tight, verifiable answer over a\ncomprehensive-sounding one.\n",{"data":35,"body":36},{"name":4,"description":6},{"type":37,"children":38},"root",[39,48,55,61,74,80,227,233,250,260,341,351,361,367,379,450,462,515,526,586,597,664,670,707,713,777,783,788,1065,1088,1094,1115,1166,1193,1271,1276,1282,1287,1346,1351],{"type":40,"tag":41,"props":42,"children":44},"element","h1",{"id":43},"nvidia-technical-helper",[45],{"type":46,"value":47},"text","NVIDIA Technical Helper",{"type":40,"tag":49,"props":50,"children":52},"h2",{"id":51},"what-this-is-for",[53],{"type":46,"value":54},"What this is for",{"type":40,"tag":56,"props":57,"children":58},"p",{},[59],{"type":46,"value":60},"NVIDIA's technology surface is enormous, overlapping, and fast-moving: dozens of SDKs with\nsimilar names, models that get renamed and re-released, repos that move between GitHub orgs, and\nversion numbers that turn last quarter's answer into misinformation. A confident answer from\nmemory is the single biggest failure mode here — it produces plausible SDK names, invented API\nsignatures, dead repos, and models that never existed.",{"type":40,"tag":56,"props":62,"children":63},{},[64,66,72],{"type":46,"value":65},"So the whole job of this skill is ",{"type":40,"tag":67,"props":68,"children":69},"strong",{},[70],{"type":46,"value":71},"grounding",{"type":46,"value":73},": turn every NVIDIA technical question into\nsearches against authorized sources, retrieve the actual pages, and answer only from what you\njust read — with a link for every claim. If you can't find it, say so. A short honest answer with\nreal links beats a long confident one that might be fiction.",{"type":40,"tag":49,"props":75,"children":77},{"id":76},"core-principles",[78],{"type":46,"value":79},"Core principles",{"type":40,"tag":81,"props":82,"children":83},"ol",{},[84,103,113,136,152,171,181,207,217],{"type":40,"tag":85,"props":86,"children":87},"li",{},[88,93,95,101],{"type":40,"tag":67,"props":89,"children":90},{},[91],{"type":46,"value":92},"Ground everything you retrieved this session; don't answer NVIDIA specifics from memory.",{"type":46,"value":94},"\nProduct names, versions, APIs, model IDs, repo paths, and \"does X support Y\" must come from a\npage you actually fetched during this task. Your prior knowledge is fine for ",{"type":40,"tag":96,"props":97,"children":98},"em",{},[99],{"type":46,"value":100},"deciding where to\nsearch",{"type":46,"value":102},", not for stating facts.",{"type":40,"tag":85,"props":104,"children":105},{},[106,111],{"type":40,"tag":67,"props":107,"children":108},{},[109],{"type":46,"value":110},"Cite every claim with the exact URL",{"type":46,"value":112}," you got it from. If a statement has no source you\nretrieved, either go find one or drop the statement.",{"type":40,"tag":85,"props":114,"children":115},{},[116,121,123],{"type":40,"tag":67,"props":117,"children":118},{},[119],{"type":46,"value":120},"Include code only when the user asks for it or it is essentially necessary.",{"type":46,"value":122}," Most answers\ndon't need a code block — a link to the official example\u002Fdoc file (\"the runnable client is at\n",{"type":40,"tag":124,"props":125,"children":126},"url",{},[127,129,134],{"type":46,"value":128},"\") lets the user check the details themselves and keeps the answer fast. Reserve inline\ncode for when the user explicitly wants sample code or the answer is meaningless without it\n(e.g. the fix ",{"type":40,"tag":96,"props":130,"children":131},{},[132],{"type":46,"value":133},"is",{"type":46,"value":135}," a specific flag or call).",{"type":40,"tag":85,"props":137,"children":138},{},[139,144,146],{"type":40,"tag":67,"props":140,"children":141},{},[142],{"type":46,"value":143},"When you do include code, quote it verbatim from its source.",{"type":46,"value":145}," Retrieve the real\nfile (from the official repo, docs, or model card) and quote it exactly, with the source URL\nand — for repos — the branch\u002Ftag or \"as of ",{"type":40,"tag":147,"props":148,"children":149},"date",{},[150],{"type":46,"value":151},"\". Do not synthesize, \"clean up\", or fill in\nAPIs from memory. If you must trim for length, show the real excerpt and link the full file. If\nthe user explicitly asks you to adapt it, keep changes minimal and mark clearly what you\nchanged versus the original.",{"type":40,"tag":85,"props":153,"children":154},{},[155,160,162,169],{"type":40,"tag":67,"props":156,"children":157},{},[158],{"type":46,"value":159},"Disambiguate the product first.",{"type":46,"value":161}," NVIDIA names collide and get confused constantly (TensorRT\nvs. TensorRT-LLM, Triton Inference Server vs. other \"Triton\"s, NeMo the framework vs. unrelated\n\"nemo\", NIM vs. NGC). If the target is ambiguous, resolve it before searching — see\n",{"type":40,"tag":163,"props":164,"children":166},"code",{"className":165},[],[167],{"type":46,"value":168},"references\u002Fnvidia-landscape.md",{"type":46,"value":170}," for a disambiguation map and canonical entry points.",{"type":40,"tag":85,"props":172,"children":173},{},[174,179],{"type":40,"tag":67,"props":175,"children":176},{},[177],{"type":46,"value":178},"Prefer primary sources; label everything else.",{"type":46,"value":180}," Official docs and NVIDIA-owned repos\u002Fmodel\ncards are ground truth. The developer blog is official but time-stamped (features change).\nForum posts and community repos are useful — especially for troubleshooting — but flag them as\nunofficial and note whether an NVIDIA staff member or an accepted\u002F\"solved\" answer backs them.",{"type":40,"tag":85,"props":182,"children":183},{},[184,189,191,197,199,205],{"type":40,"tag":67,"props":185,"children":186},{},[187],{"type":46,"value":188},"Pin versions and dates.",{"type":46,"value":190}," State the SDK\u002Flibrary version, model revision, or article date you\nread, because NVIDIA guidance changes fast. For a page's publish date, check its metadata (the\n",{"type":40,"tag":163,"props":192,"children":194},{"className":193},[],[195],{"type":46,"value":196},"datePublished",{"type":46,"value":198}," \u002F ",{"type":40,"tag":163,"props":200,"children":202},{"className":201},[],[203],{"type":46,"value":204},"article:published_time",{"type":46,"value":206}," field or the JSON-LD block in the page source), not\njust the visible text — many NVIDIA pages show no visible date but carry one in metadata, so\n\"no date\" is usually wrong. If the user is on a different version, say your answer may not\ntransfer.",{"type":40,"tag":85,"props":208,"children":209},{},[210,215],{"type":40,"tag":67,"props":211,"children":212},{},[213],{"type":46,"value":214},"Degrade gracefully when a source is unavailable.",{"type":46,"value":216}," If a page won't load, returns nothing\nuseful, or a source is unreachable, don't give up silently — use the other sources and be\nexplicit in your answer about what you couldn't check. Never fill the gap with a guess.",{"type":40,"tag":85,"props":218,"children":219},{},[220,225],{"type":40,"tag":67,"props":221,"children":222},{},[223],{"type":46,"value":224},"Be honest about gaps.",{"type":46,"value":226}," \"I couldn't find an official source for X\" and \"the only result is a\ncommunity post, unverified\" are good, expected answers. Never paper over a gap with a guess.",{"type":40,"tag":49,"props":228,"children":230},{"id":229},"workflow",[231],{"type":46,"value":232},"Workflow",{"type":40,"tag":56,"props":234,"children":235},{},[236,241,243,248],{"type":40,"tag":67,"props":237,"children":238},{},[239],{"type":46,"value":240},"Step 1 — Classify and disambiguate.",{"type":46,"value":242}," Identify which of the five question types this is (see\nplaybooks below) and pin down the exact NVIDIA product(s) involved. If unsure which product the\nuser means, resolve it first (",{"type":40,"tag":163,"props":244,"children":246},{"className":245},[],[247],{"type":46,"value":168},{"type":46,"value":249},").",{"type":40,"tag":56,"props":251,"children":252},{},[253,258],{"type":40,"tag":67,"props":254,"children":255},{},[256],{"type":46,"value":257},"Step 2 — Pick sources.",{"type":46,"value":259}," Route to the right sources for the question type (playbooks below).\nAlmost always search at least two independent sources so you can cross-check.",{"type":40,"tag":56,"props":261,"children":262},{},[263,268,270,276,278,284,286,292,294,300,302,307,309,315,317,323,325,331,333,339],{"type":40,"tag":67,"props":264,"children":265},{},[266],{"type":46,"value":267},"Step 3 — Search.",{"type":46,"value":269}," Prefer the Brave ",{"type":40,"tag":163,"props":271,"children":273},{"className":272},[],[274],{"type":46,"value":275},"WebSearch",{"type":46,"value":277}," tool if it is available — scope queries with\n",{"type":40,"tag":163,"props":279,"children":281},{"className":280},[],[282],{"type":46,"value":283},"site:",{"type":46,"value":285}," filters (e.g. ",{"type":40,"tag":163,"props":287,"children":289},{"className":288},[],[290],{"type":46,"value":291},"site:docs.nvidia.com",{"type":46,"value":293},", ",{"type":40,"tag":163,"props":295,"children":297},{"className":296},[],[298],{"type":46,"value":299},"site:github.com NVIDIA",{"type":46,"value":301},"). If ",{"type":40,"tag":163,"props":303,"children":305},{"className":304},[],[306],{"type":46,"value":275},{"type":46,"value":308}," is\nunavailable or thin, fall back to the structured JSON search endpoints and ",{"type":40,"tag":163,"props":310,"children":312},{"className":311},[],[313],{"type":46,"value":314},"WebFetch",{"type":46,"value":316},". All of the\nexact URL patterns (GitHub API, Hugging Face API, forum ",{"type":40,"tag":163,"props":318,"children":320},{"className":319},[],[321],{"type":46,"value":322},"search.json",{"type":46,"value":324},", blog ",{"type":40,"tag":163,"props":326,"children":328},{"className":327},[],[329],{"type":46,"value":330},"wp-json",{"type":46,"value":332},", arXiv API,\nNGC, docs) are in ",{"type":40,"tag":163,"props":334,"children":336},{"className":335},[],[337],{"type":46,"value":338},"references\u002Fsources-and-search.md",{"type":46,"value":340}," — read it when you need a recipe.",{"type":40,"tag":56,"props":342,"children":343},{},[344,349],{"type":40,"tag":67,"props":345,"children":346},{},[347],{"type":46,"value":348},"Step 4 — Retrieve and verify.",{"type":46,"value":350}," Open the actual pages\u002Ffiles. Confirm the thing exists and is\ncurrent before you rely on it. For a recommendation, open the repo\u002Fmodel\u002FSDK page and check it's\nNVIDIA-owned (or clearly note if it's third-party), maintained, and matches the ask.",{"type":40,"tag":56,"props":352,"children":353},{},[354,359],{"type":40,"tag":67,"props":355,"children":356},{},[357],{"type":46,"value":358},"Step 5 — Answer with citations.",{"type":46,"value":360}," Use the output format below. Lead with the direct answer, back\nevery claim with a link, quote code verbatim, and close with caveats (version\u002Fdate, unofficial\nsources, anything unavailable or unverified).",{"type":40,"tag":49,"props":362,"children":364},{"id":363},"question-type-playbooks",[365],{"type":46,"value":366},"Question-type playbooks",{"type":40,"tag":368,"props":369,"children":371},"h3",{"id":370},"_1-how-do-i-use-show-me-sample-code",[372,374],{"type":46,"value":373},"1. \"How do I use ",{"type":40,"tag":22,"props":375,"children":376},{},[377],{"type":46,"value":378},"? Show me sample code.\"",{"type":40,"tag":380,"props":381,"children":382},"ul",{},[383,423,440],{"type":40,"tag":85,"props":384,"children":385},{},[386,391,393,399,401,407,408,414,415,421],{"type":40,"tag":67,"props":387,"children":388},{},[389],{"type":46,"value":390},"Sources:",{"type":46,"value":392}," official docs (",{"type":40,"tag":163,"props":394,"children":396},{"className":395},[],[397],{"type":46,"value":398},"docs.nvidia.com",{"type":46,"value":400},") for the how-to and API; the SDK's NVIDIA GitHub\nrepo for runnable examples (",{"type":40,"tag":163,"props":402,"children":404},{"className":403},[],[405],{"type":46,"value":406},"examples\u002F",{"type":46,"value":293},{"type":40,"tag":163,"props":409,"children":411},{"className":410},[],[412],{"type":46,"value":413},"samples\u002F",{"type":46,"value":293},{"type":40,"tag":163,"props":416,"children":418},{"className":417},[],[419],{"type":46,"value":420},"tutorials\u002F",{"type":46,"value":422},", README quickstart); the\ndeveloper blog for walkthroughs; a Hugging Face model card if it's a model.",{"type":40,"tag":85,"props":424,"children":425},{},[426,431,433,438],{"type":40,"tag":67,"props":427,"children":428},{},[429],{"type":46,"value":430},"Do:",{"type":46,"value":432}," find the current install\u002Fsetup steps and the smallest complete example. Quote the code\n",{"type":40,"tag":67,"props":434,"children":435},{},[436],{"type":46,"value":437},"verbatim",{"type":46,"value":439}," with its source URL and version\u002Fbranch. Note prerequisites (CUDA version, GPU\nrequirement, OS) exactly as the source states them.",{"type":40,"tag":85,"props":441,"children":442},{},[443,448],{"type":40,"tag":67,"props":444,"children":445},{},[446],{"type":46,"value":447},"Don't:",{"type":46,"value":449}," hand-write example code from memory or merge fragments from different versions.",{"type":40,"tag":368,"props":451,"children":453},{"id":452},"_2-recommend-nvidia-technology-for",[454,456],{"type":46,"value":455},"2. \"Recommend NVIDIA technology for ",{"type":40,"tag":457,"props":458,"children":459},"task",{},[460],{"type":46,"value":461},".\"",{"type":40,"tag":380,"props":463,"children":464},{},[465,497,506],{"type":40,"tag":85,"props":466,"children":467},{},[468,472,474,480,482,488,490,495],{"type":40,"tag":67,"props":469,"children":470},{},[471],{"type":46,"value":390},{"type":46,"value":473}," the NGC catalog (",{"type":40,"tag":163,"props":475,"children":477},{"className":476},[],[478],{"type":46,"value":479},"catalog.ngc.nvidia.com",{"type":46,"value":481},") and ",{"type":40,"tag":163,"props":483,"children":485},{"className":484},[],[486],{"type":46,"value":487},"build.nvidia.com",{"type":46,"value":489}," for models\u002FNIMs;\nNVIDIA GitHub orgs (repo search, scoped to NVIDIA — see rules below); the ",{"type":40,"tag":163,"props":491,"children":493},{"className":492},[],[494],{"type":46,"value":8},{"type":46,"value":496}," Hugging Face\norg for models\u002Fdatasets; docs and the blog to confirm the intended use case.",{"type":40,"tag":85,"props":498,"children":499},{},[500,504],{"type":40,"tag":67,"props":501,"children":502},{},[503],{"type":46,"value":430},{"type":46,"value":505}," verify each candidate actually exists and is current by opening its page. Give 1–3\nconcrete options with what each is for, maturity\u002Frecency signals (last update, stars, downloads),\nand links. Distinguish NVIDIA-owned from community\u002Fthird-party.",{"type":40,"tag":85,"props":507,"children":508},{},[509,513],{"type":40,"tag":67,"props":510,"children":511},{},[512],{"type":46,"value":447},{"type":46,"value":514}," recommend from memory, or list deprecated\u002Frenamed products. If nothing fits, say so.",{"type":40,"tag":368,"props":516,"children":518},{"id":517},"_3-show-me-a-success-story-real-adoption-of",[519,521],{"type":46,"value":520},"3. \"Show me a success story \u002F real adoption of ",{"type":40,"tag":522,"props":523,"children":524},"product",{},[525],{"type":46,"value":461},{"type":40,"tag":380,"props":527,"children":528},{},[529,561,577],{"type":40,"tag":85,"props":530,"children":531},{},[532,536,538,544,545,551,553,559],{"type":40,"tag":67,"props":533,"children":534},{},[535],{"type":46,"value":390},{"type":46,"value":537}," the developer blog (",{"type":40,"tag":163,"props":539,"children":541},{"className":540},[],[542],{"type":46,"value":543},"developer.nvidia.com\u002Fblog",{"type":46,"value":481},{"type":40,"tag":163,"props":546,"children":548},{"className":547},[],[549],{"type":46,"value":550},"blogs.nvidia.com",{"type":46,"value":552}," for\ncustomer stories and case studies; ",{"type":40,"tag":163,"props":554,"children":556},{"className":555},[],[557],{"type":46,"value":558},"www.nvidia.com",{"type":46,"value":560}," customer-story\u002Findustry pages; the SDK's\ndocs\u002Frepo \"used by\" or partner mentions.",{"type":40,"tag":85,"props":562,"children":563},{},[564,568,570,575],{"type":40,"tag":67,"props":565,"children":566},{},[567],{"type":46,"value":430},{"type":46,"value":569}," cite the specific case study (company, what they built, result) with its URL and date.\nFor the date, check the page metadata (",{"type":40,"tag":163,"props":571,"children":573},{"className":572},[],[574],{"type":46,"value":196},{"type":46,"value":576},"\u002FJSON-LD), not just the visible text —\nthese pages often carry the date only in metadata. Attribute metrics as vendor\u002Fself-reported.\nPrefer named, published cases over vague claims.",{"type":40,"tag":85,"props":578,"children":579},{},[580,584],{"type":40,"tag":67,"props":581,"children":582},{},[583],{"type":46,"value":447},{"type":46,"value":585}," invent customers or metrics. If you only find NVIDIA's own marketing framing, say so.",{"type":40,"tag":368,"props":587,"children":589},{"id":588},"_4-i-hit-errorissue-with-how-do-i-fix-it",[590,592],{"type":46,"value":591},"4. \"I hit \u003Cerror\u002Fissue> with ",{"type":40,"tag":22,"props":593,"children":594},{},[595],{"type":46,"value":596},". How do I fix it?\"",{"type":40,"tag":380,"props":598,"children":599},{},[600,638,655],{"type":40,"tag":85,"props":601,"children":602},{},[603,607,609,614,616,622,624,629,631,636],{"type":40,"tag":67,"props":604,"children":605},{},[606],{"type":46,"value":390},{"type":46,"value":608}," the developer ",{"type":40,"tag":67,"props":610,"children":611},{},[612],{"type":46,"value":613},"forums",{"type":46,"value":615}," (",{"type":40,"tag":163,"props":617,"children":619},{"className":618},[],[620],{"type":46,"value":621},"forums.developer.nvidia.com",{"type":46,"value":623}," via ",{"type":40,"tag":163,"props":625,"children":627},{"className":626},[],[628],{"type":46,"value":322},{"type":46,"value":630},") — this\nis the primary source for real errors; GitHub ",{"type":40,"tag":67,"props":632,"children":633},{},[634],{"type":46,"value":635},"Issues",{"type":46,"value":637}," on the SDK's repo; docs\ntroubleshooting\u002Frelease-notes sections for known issues.",{"type":40,"tag":85,"props":639,"children":640},{},[641,645,647,653],{"type":40,"tag":67,"props":642,"children":643},{},[644],{"type":46,"value":430},{"type":46,"value":646}," search the exact error text (and a generalized version). Prefer threads with an accepted\nanswer, a ",{"type":40,"tag":163,"props":648,"children":650},{"className":649},[],[651],{"type":46,"value":652},"status:solved",{"type":46,"value":654}," marker, or an NVIDIA staff reply. Summarize the fix and link the\nthread; quote the working commands\u002Fcode verbatim. Note the versions the fix applied to.",{"type":40,"tag":85,"props":656,"children":657},{},[658,662],{"type":40,"tag":67,"props":659,"children":660},{},[661],{"type":46,"value":447},{"type":46,"value":663}," present a community workaround as official, or guess at a fix. If unresolved\nupstream, say that and link the open issue.",{"type":40,"tag":368,"props":665,"children":667},{"id":666},"_5-how-could-my-product-use-nvidia-sdks-integrationarchitecture",[668],{"type":46,"value":669},"5. \"How could my product use NVIDIA SDKs?\" (integration\u002Farchitecture)",{"type":40,"tag":380,"props":671,"children":672},{},[673,689,698],{"type":40,"tag":85,"props":674,"children":675},{},[676,680,682,687],{"type":40,"tag":67,"props":677,"children":678},{},[679],{"type":46,"value":390},{"type":46,"value":681}," docs (architecture\u002Fdeployment guides), the blog (reference architectures,\nblueprints), NGC\u002F",{"type":40,"tag":163,"props":683,"children":685},{"className":684},[],[686],{"type":46,"value":487},{"type":46,"value":688}," for deployable components, relevant repos for integration\nexamples.",{"type":40,"tag":85,"props":690,"children":691},{},[692,696],{"type":40,"tag":67,"props":693,"children":694},{},[695],{"type":46,"value":430},{"type":46,"value":697}," map the user's product\u002Fgoal to specific NVIDIA components, citing each. Sketch a\nconcrete path (which SDK does what, how they connect) grounded in real docs\u002Fexamples. Call out\nhardware\u002Flicensing prerequisites you find.",{"type":40,"tag":85,"props":699,"children":700},{},[701,705],{"type":40,"tag":67,"props":702,"children":703},{},[704],{"type":46,"value":447},{"type":46,"value":706}," produce a generic \"NVIDIA is great for AI\" pitch. Every recommended piece needs a\nlink and a reason.",{"type":40,"tag":49,"props":708,"children":710},{"id":709},"search-source-rules",[711],{"type":46,"value":712},"Search source rules",{"type":40,"tag":380,"props":714,"children":715},{},[716,741,767],{"type":40,"tag":85,"props":717,"children":718},{},[719,724,726,732,734,739],{"type":40,"tag":67,"props":720,"children":721},{},[722],{"type":46,"value":723},"GitHub:",{"type":46,"value":725}," focus on NVIDIA-owned repositories for anything you recommend or quote as\nauthoritative. Scope searches to NVIDIA's organizations (e.g. ",{"type":40,"tag":163,"props":727,"children":729},{"className":728},[],[730],{"type":46,"value":731},"org:NVIDIA",{"type":46,"value":733},") and confirm the repo\nowner before treating it as official. Community repos can be mentioned but must be labeled as\nsuch, with a note on activity\u002Fmaintenance. Known NVIDIA orgs and how to search the API are in\n",{"type":40,"tag":163,"props":735,"children":737},{"className":736},[],[738],{"type":46,"value":338},{"type":46,"value":740},".",{"type":40,"tag":85,"props":742,"children":743},{},[744,749,751,757,759,765],{"type":40,"tag":67,"props":745,"children":746},{},[747],{"type":46,"value":748},"Hugging Face:",{"type":46,"value":750}," prioritize the official org at ",{"type":40,"tag":163,"props":752,"children":754},{"className":753},[],[755],{"type":46,"value":756},"https:\u002F\u002Fhuggingface.co\u002Fnvidia",{"type":46,"value":758}," — query with\n",{"type":40,"tag":163,"props":760,"children":762},{"className":761},[],[763],{"type":46,"value":764},"?author=nvidia",{"type":46,"value":766}," first. Only broaden beyond it when nvidia has nothing suitable, and then clearly\nmark non-NVIDIA models as community\u002Fthird-party and check their license and provenance.",{"type":40,"tag":85,"props":768,"children":769},{},[770,775],{"type":40,"tag":67,"props":771,"children":772},{},[773],{"type":46,"value":774},"Cross-check:",{"type":46,"value":776}," for recommendations and factual claims, corroborate across at least two sources\n(e.g., docs + repo, or NGC + blog) before stating something as settled.",{"type":40,"tag":49,"props":778,"children":780},{"id":779},"data-sources",[781],{"type":46,"value":782},"Data sources",{"type":40,"tag":56,"props":784,"children":785},{},[786],{"type":46,"value":787},"Search and fetch from these authorized sources — this is where NVIDIA's ground truth lives. Match\nthe source to the question (see the playbooks) and cross-check across at least two before treating\nsomething as settled.",{"type":40,"tag":789,"props":790,"children":791},"table",{},[792,816],{"type":40,"tag":793,"props":794,"children":795},"thead",{},[796],{"type":40,"tag":797,"props":798,"children":799},"tr",{},[800,806,811],{"type":40,"tag":801,"props":802,"children":803},"th",{},[804],{"type":46,"value":805},"Source",{"type":40,"tag":801,"props":807,"children":808},{},[809],{"type":46,"value":810},"Best for",{"type":40,"tag":801,"props":812,"children":813},{},[814],{"type":46,"value":815},"Where to look",{"type":40,"tag":817,"props":818,"children":819},"tbody",{},[820,842,864,885,912,933,954,975,1021,1043],{"type":40,"tag":797,"props":821,"children":822},{},[823,829,834],{"type":40,"tag":824,"props":825,"children":826},"td",{},[827],{"type":46,"value":828},"Official docs",{"type":40,"tag":824,"props":830,"children":831},{},[832],{"type":46,"value":833},"how-to, API reference, supported versions, known issues",{"type":40,"tag":824,"props":835,"children":836},{},[837],{"type":40,"tag":163,"props":838,"children":840},{"className":839},[],[841],{"type":46,"value":398},{"type":40,"tag":797,"props":843,"children":844},{},[845,850,855],{"type":40,"tag":824,"props":846,"children":847},{},[848],{"type":46,"value":849},"Product hubs",{"type":40,"tag":824,"props":851,"children":852},{},[853],{"type":46,"value":854},"product overview, downloads, getting started",{"type":40,"tag":824,"props":856,"children":857},{},[858],{"type":40,"tag":163,"props":859,"children":861},{"className":860},[],[862],{"type":46,"value":863},"developer.nvidia.com\u002F\u003Cproduct>",{"type":40,"tag":797,"props":865,"children":866},{},[867,872,877],{"type":40,"tag":824,"props":868,"children":869},{},[870],{"type":46,"value":871},"Developer blog",{"type":40,"tag":824,"props":873,"children":874},{},[875],{"type":46,"value":876},"walkthroughs, reference architectures, blueprints, success stories",{"type":40,"tag":824,"props":878,"children":879},{},[880],{"type":40,"tag":163,"props":881,"children":883},{"className":882},[],[884],{"type":46,"value":543},{"type":40,"tag":797,"props":886,"children":887},{},[888,893,898],{"type":40,"tag":824,"props":889,"children":890},{},[891],{"type":46,"value":892},"Corporate blog & customer stories",{"type":40,"tag":824,"props":894,"children":895},{},[896],{"type":46,"value":897},"named adoption and case studies",{"type":40,"tag":824,"props":899,"children":900},{},[901,906,907],{"type":40,"tag":163,"props":902,"children":904},{"className":903},[],[905],{"type":46,"value":550},{"type":46,"value":293},{"type":40,"tag":163,"props":908,"children":910},{"className":909},[],[911],{"type":46,"value":558},{"type":40,"tag":797,"props":913,"children":914},{},[915,920,925],{"type":40,"tag":824,"props":916,"children":917},{},[918],{"type":46,"value":919},"NGC catalog",{"type":40,"tag":824,"props":921,"children":922},{},[923],{"type":46,"value":924},"pretrained models, containers, Helm charts",{"type":40,"tag":824,"props":926,"children":927},{},[928],{"type":40,"tag":163,"props":929,"children":931},{"className":930},[],[932],{"type":46,"value":479},{"type":40,"tag":797,"props":934,"children":935},{},[936,941,946],{"type":40,"tag":824,"props":937,"children":938},{},[939],{"type":46,"value":940},"API catalog \u002F NIM",{"type":40,"tag":824,"props":942,"children":943},{},[944],{"type":46,"value":945},"deployable inference microservices, hosted model endpoints",{"type":40,"tag":824,"props":947,"children":948},{},[949],{"type":40,"tag":163,"props":950,"children":952},{"className":951},[],[953],{"type":46,"value":487},{"type":40,"tag":797,"props":955,"children":956},{},[957,962,967],{"type":40,"tag":824,"props":958,"children":959},{},[960],{"type":46,"value":961},"Developer forums",{"type":40,"tag":824,"props":963,"children":964},{},[965],{"type":46,"value":966},"real-world errors and fixes (troubleshooting)",{"type":40,"tag":824,"props":968,"children":969},{},[970],{"type":40,"tag":163,"props":971,"children":973},{"className":972},[],[974],{"type":46,"value":621},{"type":40,"tag":797,"props":976,"children":977},{},[978,983,988],{"type":40,"tag":824,"props":979,"children":980},{},[981],{"type":46,"value":982},"GitHub (NVIDIA orgs)",{"type":40,"tag":824,"props":984,"children":985},{},[986],{"type":46,"value":987},"official example code, repos, issues",{"type":40,"tag":824,"props":989,"children":990},{},[991,997,998,1004,1006,1012,1013,1019],{"type":40,"tag":163,"props":992,"children":994},{"className":993},[],[995],{"type":46,"value":996},"github.com",{"type":46,"value":198},{"type":40,"tag":163,"props":999,"children":1001},{"className":1000},[],[1002],{"type":46,"value":1003},"api.github.com",{"type":46,"value":1005}," (NVIDIA, ",{"type":40,"tag":163,"props":1007,"children":1009},{"className":1008},[],[1010],{"type":46,"value":1011},"rapidsai",{"type":46,"value":293},{"type":40,"tag":163,"props":1014,"children":1016},{"className":1015},[],[1017],{"type":46,"value":1018},"NVIDIA-AI-IOT",{"type":46,"value":1020},", …)",{"type":40,"tag":797,"props":1022,"children":1023},{},[1024,1029,1034],{"type":40,"tag":824,"props":1025,"children":1026},{},[1027],{"type":46,"value":1028},"Hugging Face (nvidia org)",{"type":40,"tag":824,"props":1030,"children":1031},{},[1032],{"type":46,"value":1033},"open models and datasets",{"type":40,"tag":824,"props":1035,"children":1036},{},[1037],{"type":40,"tag":163,"props":1038,"children":1040},{"className":1039},[],[1041],{"type":46,"value":1042},"huggingface.co\u002Fnvidia",{"type":40,"tag":797,"props":1044,"children":1045},{},[1046,1051,1056],{"type":40,"tag":824,"props":1047,"children":1048},{},[1049],{"type":46,"value":1050},"arXiv",{"type":40,"tag":824,"props":1052,"children":1053},{},[1054],{"type":46,"value":1055},"the research behind a technique",{"type":40,"tag":824,"props":1057,"children":1058},{},[1059],{"type":40,"tag":163,"props":1060,"children":1062},{"className":1061},[],[1063],{"type":46,"value":1064},"arxiv.org",{"type":40,"tag":56,"props":1066,"children":1067},{},[1068,1073,1075,1080,1081,1086],{"type":40,"tag":163,"props":1069,"children":1071},{"className":1070},[],[1072],{"type":46,"value":338},{"type":46,"value":1074}," has the exact search recipes and URL patterns (forum\n",{"type":40,"tag":163,"props":1076,"children":1078},{"className":1077},[],[1079],{"type":46,"value":322},{"type":46,"value":324},{"type":40,"tag":163,"props":1082,"children":1084},{"className":1083},[],[1085],{"type":46,"value":330},{"type":46,"value":1087},", GitHub\u002FHugging Face APIs, arXiv API) for each of these — read it\nwhen you need a recipe.",{"type":40,"tag":49,"props":1089,"children":1091},{"id":1090},"unblocking-sources-nemoclaw-network-policies",[1092],{"type":46,"value":1093},"Unblocking sources: nemoclaw network policies",{"type":40,"tag":56,"props":1095,"children":1096},{},[1097,1099,1105,1107,1113],{"type":46,"value":1098},"If a source you need is blocked by the sandbox network policy — or the user asks to open access to\na new site — don't just report the gap: author a nemoclaw network-policy YAML for the blocked\nhost(s) and walk the user through applying it. Read ",{"type":40,"tag":163,"props":1100,"children":1102},{"className":1101},[],[1103],{"type":46,"value":1104},"references\u002Fnemoclaw-network-policy.md",{"type":46,"value":1106}," for\nthe full guide; ",{"type":40,"tag":163,"props":1108,"children":1110},{"className":1109},[],[1111],{"type":46,"value":1112},"references\u002Fnemoclaw-policy-template.yaml",{"type":46,"value":1114}," is an annotated starting point.",{"type":40,"tag":56,"props":1116,"children":1117},{},[1118,1120,1126,1127,1133,1135,1141,1143,1149,1151,1164],{"type":46,"value":1119},"Use least privilege for every endpoint. For sources that the assistant only searches, fetches, or\ndownloads from, set ",{"type":40,"tag":163,"props":1121,"children":1123},{"className":1122},[],[1124],{"type":46,"value":1125},"access: read-only",{"type":46,"value":293},{"type":40,"tag":163,"props":1128,"children":1130},{"className":1129},[],[1131],{"type":46,"value":1132},"protocol: rest",{"type":46,"value":1134},", and ",{"type":40,"tag":163,"props":1136,"children":1138},{"className":1137},[],[1139],{"type":46,"value":1140},"enforcement: enforce",{"type":46,"value":1142},". Do not set\n",{"type":40,"tag":163,"props":1144,"children":1146},{"className":1145},[],[1147],{"type":46,"value":1148},"tls: skip",{"type":46,"value":1150}," for ordinary HTTPS endpoints. ",{"type":40,"tag":67,"props":1152,"children":1153},{},[1154,1156,1162],{"type":46,"value":1155},"DO NOT set ",{"type":40,"tag":163,"props":1157,"children":1159},{"className":1158},[],[1160],{"type":46,"value":1161},"access: full",{"type":46,"value":1163}," for an API endpoint unless\nit is absolutely required by the user's requested workflow.",{"type":46,"value":1165}," An endpoint being an API or\nsupporting search does not justify full access when GET\u002FHEAD is sufficient. If broader access or\nskipped TLS inspection is unavoidable, explain the exact requirement and security tradeoff.",{"type":40,"tag":56,"props":1167,"children":1168},{},[1169,1171,1176,1178,1183,1185,1191],{"type":46,"value":1170},"You ",{"type":40,"tag":67,"props":1172,"children":1173},{},[1174],{"type":46,"value":1175},"cannot",{"type":46,"value":1177}," apply the policy yourself — you are running inside the sandbox and are not allowed\nto change its network policy from within. Tell the user this explicitly. In short: give them the\nYAML (one file per site, listing every hostname it needs — subdomains are not implied), tell them\nto save it ",{"type":40,"tag":67,"props":1179,"children":1180},{},[1181],{"type":46,"value":1182},"on their host machine",{"type":46,"value":1184}," as ",{"type":40,"tag":163,"props":1186,"children":1188},{"className":1187},[],[1189],{"type":46,"value":1190},"\u003Cdomain-name>.yaml",{"type":46,"value":1192},", then run from the host:",{"type":40,"tag":1194,"props":1195,"children":1200},"pre",{"className":1196,"code":1197,"language":1198,"meta":1199,"style":1199},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","nemoclaw \u003Csbx-name> policy-add --from-file \u003Cdomain-name>.yaml\n","bash","",[1201],{"type":40,"tag":163,"props":1202,"children":1203},{"__ignoreMap":1199},[1204],{"type":40,"tag":1205,"props":1206,"children":1209},"span",{"class":1207,"line":1208},"line",1,[1210,1216,1222,1228,1234,1239,1244,1249,1253,1258,1262,1266],{"type":40,"tag":1205,"props":1211,"children":1213},{"style":1212},"--shiki-light:#E2931D;--shiki-default:#FFCB6B;--shiki-dark:#FFCB6B",[1214],{"type":46,"value":1215},"nemoclaw",{"type":40,"tag":1205,"props":1217,"children":1219},{"style":1218},"--shiki-light:#39ADB5;--shiki-default:#89DDFF;--shiki-dark:#89DDFF",[1220],{"type":46,"value":1221}," \u003C",{"type":40,"tag":1205,"props":1223,"children":1225},{"style":1224},"--shiki-light:#91B859;--shiki-default:#C3E88D;--shiki-dark:#C3E88D",[1226],{"type":46,"value":1227},"sbx-nam",{"type":40,"tag":1205,"props":1229,"children":1231},{"style":1230},"--shiki-light:#90A4AE;--shiki-default:#EEFFFF;--shiki-dark:#BABED8",[1232],{"type":46,"value":1233},"e",{"type":40,"tag":1205,"props":1235,"children":1236},{"style":1218},[1237],{"type":46,"value":1238},">",{"type":40,"tag":1205,"props":1240,"children":1241},{"style":1224},[1242],{"type":46,"value":1243}," policy-add",{"type":40,"tag":1205,"props":1245,"children":1246},{"style":1224},[1247],{"type":46,"value":1248}," --from-file",{"type":40,"tag":1205,"props":1250,"children":1251},{"style":1218},[1252],{"type":46,"value":1221},{"type":40,"tag":1205,"props":1254,"children":1255},{"style":1224},[1256],{"type":46,"value":1257},"domain-nam",{"type":40,"tag":1205,"props":1259,"children":1260},{"style":1230},[1261],{"type":46,"value":1233},{"type":40,"tag":1205,"props":1263,"children":1264},{"style":1218},[1265],{"type":46,"value":1238},{"type":40,"tag":1205,"props":1267,"children":1268},{"style":1224},[1269],{"type":46,"value":1270},".yaml\n",{"type":40,"tag":56,"props":1272,"children":1273},{},[1274],{"type":46,"value":1275},"This applies the policy immediately to the running sandbox — no sandbox rebuild needed. Once the\nuser confirms, retry the blocked fetch to verify.",{"type":40,"tag":49,"props":1277,"children":1279},{"id":1278},"output-format",[1280],{"type":46,"value":1281},"Output format",{"type":40,"tag":56,"props":1283,"children":1284},{},[1285],{"type":46,"value":1286},"Adapt to the question, but keep this backbone:",{"type":40,"tag":380,"props":1288,"children":1289},{},[1290,1300,1310,1326,1336],{"type":40,"tag":85,"props":1291,"children":1292},{},[1293,1298],{"type":40,"tag":67,"props":1294,"children":1295},{},[1296],{"type":46,"value":1297},"Answer",{"type":46,"value":1299}," — the direct response in 1–3 sentences up front.",{"type":40,"tag":85,"props":1301,"children":1302},{},[1303,1308],{"type":40,"tag":67,"props":1304,"children":1305},{},[1306],{"type":46,"value":1307},"Details \u002F evidence",{"type":46,"value":1309}," — the specifics, each with an inline link. Note versions and dates.",{"type":40,"tag":85,"props":1311,"children":1312},{},[1313,1318,1320,1324],{"type":40,"tag":67,"props":1314,"children":1315},{},[1316],{"type":46,"value":1317},"Code",{"type":46,"value":1319}," — only if the user asked for code or the answer needs it (core principle 3); otherwise\nlink the official example\u002Fdoc file instead. When included: fenced and quoted ",{"type":40,"tag":67,"props":1321,"children":1322},{},[1323],{"type":46,"value":437},{"type":46,"value":1325},", with\nthe source URL and branch\u002Ftag above it. Mark any modification explicitly.",{"type":40,"tag":85,"props":1327,"children":1328},{},[1329,1334],{"type":40,"tag":67,"props":1330,"children":1331},{},[1332],{"type":46,"value":1333},"Sources",{"type":46,"value":1335}," — a short list of every URL you actually retrieved, so the user can verify.",{"type":40,"tag":85,"props":1337,"children":1338},{},[1339,1344],{"type":40,"tag":67,"props":1340,"children":1341},{},[1342],{"type":46,"value":1343},"Caveats",{"type":46,"value":1345}," — versions the answer assumes, anything unofficial\u002Fcommunity, and anything you\ncouldn't confirm or reach.",{"type":40,"tag":56,"props":1347,"children":1348},{},[1349],{"type":46,"value":1350},"Keep it as long as it needs to be and no longer. Prefer a tight, verifiable answer over a\ncomprehensive-sounding one.",{"type":40,"tag":1352,"props":1353,"children":1354},"style",{},[1355],{"type":46,"value":1356},"html .light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html.light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"items":1358,"total":1515},[1359,1375,1393,1404,1416,1430,1443,1457,1470,1481,1495,1504],{"slug":1360,"name":1360,"fn":1361,"description":1362,"org":1363,"tags":1364,"stars":1372,"repoUrl":1373,"updatedAt":1374},"nemoclaw-user-guide","retrieve NemoClaw documentation and configuration","Guides human users' AI agents to the NemoClaw docs MCP server and canonical Fern documentation in Markdown form. Use when users ask how to install, configure, operate, troubleshoot, secure, or learn NemoClaw with an AI coding assistant. Trigger keywords - nemoclaw docs, use nemoclaw with ai agent, nemoclaw mcp docs, nemoclaw install help, nemoclaw quickstart, nemoclaw markdown docs, llms.txt, agent skills.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1365,1366,1369],{"name":13,"slug":14,"type":15},{"name":1367,"slug":1368,"type":15},"MCP","mcp",{"name":1370,"slug":1371,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":1376,"name":1376,"fn":1377,"description":1378,"org":1379,"tags":1380,"stars":1390,"repoUrl":1391,"updatedAt":1392},"mcore-build-and-dependency","manage Megatron-LM development environments","Container-based dev environment setup and dependency management for Megatron-LM. Covers acquiring and launching the CI container, uv package management, and updating uv.lock.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1381,1384,1387],{"name":1382,"slug":1383,"type":15},"Containers","containers",{"name":1385,"slug":1386,"type":15},"Deployment","deployment",{"name":1388,"slug":1389,"type":15},"Python","python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":1394,"name":1394,"fn":1395,"description":1396,"org":1397,"tags":1398,"stars":1390,"repoUrl":1391,"updatedAt":1403},"mcore-bump-base-image","update NVIDIA PyTorch base images","Bump the NVIDIA PyTorch base image (`nvcr.io\u002Fnvidia\u002Fpytorch:YY.MM-py3`) used by Megatron-LM CI. Covers the two pin sites (GitHub CI in `docker\u002F.ngc_version.dev` and GitLab CI in `.gitlab\u002Fstages\u002F01.build.yml`), the post-bump CI loop (re-run functional tests, refresh golden values, mark broken tests), and the gotchas that bit PRs",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1399,1402],{"name":1400,"slug":1401,"type":15},"CI\u002FCD","ci-cd",{"name":1385,"slug":1386,"type":15},"2026-07-14T05:25:59.97109",{"slug":1405,"name":1405,"fn":1406,"description":1407,"org":1408,"tags":1409,"stars":1390,"repoUrl":1391,"updatedAt":1415},"mcore-cicd","manage CI\u002FCD pipelines for Megatron-LM","CI\u002FCD reference for Megatron-LM. Covers CI pipeline structure, PR scope labels, triggering internal GitLab CI (which force-pushes the current branch to a pull-request\u002FBRANCH ref — always dry-run and verify the destination first; never run against shared or protected branches), and CI failure investigation.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1410,1411,1412],{"name":1400,"slug":1401,"type":15},{"name":1385,"slug":1386,"type":15},{"name":1413,"slug":1414,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":1417,"name":1417,"fn":1418,"description":1419,"org":1420,"tags":1421,"stars":1390,"repoUrl":1391,"updatedAt":1429},"mcore-create-issue","investigate CI failures and create issues","Investigate a failing GitHub Actions run or job and create a GitHub issue for the failure.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1422,1425,1426],{"name":1423,"slug":1424,"type":15},"Debugging","debugging",{"name":1413,"slug":1414,"type":15},{"name":1427,"slug":1428,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":1431,"name":1431,"fn":1432,"description":1433,"org":1434,"tags":1435,"stars":1390,"repoUrl":1391,"updatedAt":1442},"mcore-linting-and-formatting","lint and format Megatron-LM code","Linting and formatting for Megatron-LM. Covers running autoformat.sh, tools (ruff, black, isort, pylint, mypy), and code style rules.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1436,1439],{"name":1437,"slug":1438,"type":15},"Best Practices","best-practices",{"name":1440,"slug":1441,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":1444,"name":1444,"fn":1445,"description":1446,"org":1447,"tags":1448,"stars":1390,"repoUrl":1391,"updatedAt":1456},"mcore-migrate-gpt-to-hybrid","migrate Megatron-LM models to HybridModel","Migration guide for moving Megatron Core GPTModel checkpoints, model providers, training commands, and layer mappings to HybridModel.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1449,1452,1455],{"name":1450,"slug":1451,"type":15},"Machine Learning","machine-learning",{"name":1453,"slug":1454,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":1458,"name":1458,"fn":1459,"description":1460,"org":1461,"tags":1462,"stars":1390,"repoUrl":1391,"updatedAt":1469},"mcore-onboard-gb200-1node-tests","onboard functional tests for GB200","Onboard 1-node GitHub MR functional tests for GB200 from existing mr-scoped 2-node tests.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1463,1466],{"name":1464,"slug":1465,"type":15},"QA","qa",{"name":1467,"slug":1468,"type":15},"Testing","testing","2026-07-14T05:25:53.673039",{"slug":1471,"name":1471,"fn":1472,"description":1473,"org":1474,"tags":1475,"stars":1390,"repoUrl":1391,"updatedAt":1480},"mcore-run-on-slurm","launch distributed training jobs on SLURM","How to launch distributed Megatron-LM training jobs on a SLURM cluster. Covers a minimal sbatch skeleton, environment-variable setup for torch.distributed.run, CUDA_DEVICE_MAX_CONNECTIONS rules across hardware and parallelism modes, container conventions, monitoring, and per-rank failure diagnosis.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1476,1477],{"name":1385,"slug":1386,"type":15},{"name":1478,"slug":1479,"type":15},"Infrastructure","infrastructure","2026-07-14T05:25:49.362534",{"slug":1482,"name":1482,"fn":1483,"description":1484,"org":1485,"tags":1486,"stars":1390,"repoUrl":1391,"updatedAt":1494},"mcore-split-pr","split pull requests to reduce review load","Split a PR into multiple PRs to reduce the number of required CODEOWNERS reviewer groups.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1487,1490,1491],{"name":1488,"slug":1489,"type":15},"Code Review","code-review",{"name":1413,"slug":1414,"type":15},{"name":1492,"slug":1493,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":1496,"name":1496,"fn":1497,"description":1498,"org":1499,"tags":1500,"stars":1390,"repoUrl":1391,"updatedAt":1503},"mcore-testing","run and manage Megatron-LM tests","Test system for Megatron-LM. Covers test layout, recipe YAML structure, adding and running unit and functional tests, golden values, marker filters, and CI parity.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1501,1502],{"name":1464,"slug":1465,"type":15},{"name":1467,"slug":1468,"type":15},"2026-07-14T05:25:54.928983",{"slug":1505,"name":1505,"fn":1506,"description":1507,"org":1508,"tags":1509,"stars":1390,"repoUrl":1391,"updatedAt":1514},"nightly-sync","manage nightly main-to-dev sync workflows","Domain knowledge for the nightly main-to-dev sync workflow. Covers merge strategy, CI architecture, failure investigation, and known issues.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1510,1513],{"name":1511,"slug":1512,"type":15},"Automation","automation",{"name":1400,"slug":1401,"type":15},"2026-07-30T05:29:03.275638",525,{"items":1517,"total":1610},[1518,1535,1547,1559,1571,1578,1593],{"slug":1519,"name":1519,"fn":1520,"description":1521,"org":1522,"tags":1523,"stars":23,"repoUrl":24,"updatedAt":1534},"cross-source-gap-analysis","identify alignment gaps across communication sources","Compare findings across Slack, GitHub, NVIDIA forums, and Outlook to identify alignment gaps, missing coverage, and follow-ups.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1524,1527,1528,1531],{"name":1525,"slug":1526,"type":15},"Analysis","analysis",{"name":1413,"slug":1414,"type":15},{"name":1529,"slug":1530,"type":15},"Outlook Email","outlook-email",{"name":1532,"slug":1533,"type":15},"Slack","slack","2026-08-05T05:58:20.663418",{"slug":1536,"name":1536,"fn":1537,"description":1538,"org":1539,"tags":1540,"stars":23,"repoUrl":24,"updatedAt":1546},"github-readonly-live","read live GitHub repository data","Read the configured live GitHub repository through authenticated, policy-scoped GitHub REST GET requests.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1541,1544,1545],{"name":1542,"slug":1543,"type":15},"API Development","api-development",{"name":1413,"slug":1414,"type":15},{"name":9,"slug":8,"type":15},"2026-08-05T05:58:19.175067",{"slug":1548,"name":1548,"fn":1549,"description":1550,"org":1551,"tags":1552,"stars":23,"repoUrl":24,"updatedAt":1558},"nemoclaw-autoheal","guide Hermes availability and auto-heal setup","Guide users through Hermes availability checks and the optional host-side auto-heal setup without crossing the sandbox-to-host boundary.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1553,1554,1555],{"name":1423,"slug":1424,"type":15},{"name":9,"slug":8,"type":15},{"name":1556,"slug":1557,"type":15},"Operations","operations","2026-08-05T05:58:22.804006",{"slug":1560,"name":1560,"fn":1561,"description":1562,"org":1563,"tags":1564,"stars":23,"repoUrl":24,"updatedAt":1570},"nemoclaw-nvteam","route engineering and community work","Route product, program, engineering, data and ML, quality, SRE, security, and developer-community work through the eight local role lenses packaged with the developer-community-chief-of-staff recipe in nemoclaw-community. Use for explicit NVTeam or persona activation, cross-functional readiness, developer relations, community enablement, technical-enablement work, or automatic specialist routing within this recipe. Do not use for a standalone question about core NemoClaw product capabilities unless the user explicitly requests NVTeam. This skill is Community-recipe behavior, not a built-in NemoClaw capability.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1565,1568,1569],{"name":1566,"slug":1567,"type":15},"Management","management",{"name":9,"slug":8,"type":15},{"name":1556,"slug":1557,"type":15},"2026-08-05T05:58:31.181122",{"slug":4,"name":4,"fn":5,"description":6,"org":1572,"tags":1573,"stars":23,"repoUrl":24,"updatedAt":25},{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1574,1575,1576,1577],{"name":13,"slug":14,"type":15},{"name":9,"slug":8,"type":15},{"name":17,"slug":18,"type":15},{"name":21,"slug":22,"type":15},{"slug":1579,"name":1579,"fn":1580,"description":1581,"org":1582,"tags":1583,"stars":23,"repoUrl":24,"updatedAt":1592},"outlook-email-search","search Outlook emails via Microsoft Graph","Search the Outlook mailbox via Microsoft Graph to find and read emails that help answer user questions.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1584,1587,1590,1591],{"name":1585,"slug":1586,"type":15},"Communications","communications",{"name":1588,"slug":1589,"type":15},"Microsoft 365","microsoft-365",{"name":1529,"slug":1530,"type":15},{"name":1370,"slug":1371,"type":15},"2026-08-05T05:58:19.55877",{"slug":1594,"name":1594,"fn":1595,"description":1596,"org":1597,"tags":1598,"stars":23,"repoUrl":24,"updatedAt":1609},"payment-ops-playbook","manage payment operations workflows","The standard FinGuard payment-operations workflow — screen, explain, prepare, hand off to a human. Never release.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1599,1602,1605,1606],{"name":1600,"slug":1601,"type":15},"Compliance","compliance",{"name":1603,"slug":1604,"type":15},"Finance","finance",{"name":1556,"slug":1557,"type":15},{"name":1607,"slug":1608,"type":15},"Payments","payments","2026-08-05T05:58:31.551123",15]