{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreia7uszgyuvagkryba5mow6db7z77pwndcb5fud2tllyd52h65npvu",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mebamcdjlmv2"
  },
  "path": "/t/wfgy-ai-clinic-a-small-er-for-rag-and-llm-failures/173172#post_1",
  "publishedAt": "2026-02-07T08:33:33.000Z",
  "site": "https://discuss.huggingface.co",
  "tags": [
    "https://chatgpt.com/share/68b9b7ad-51e4-8000-90ee-a25522da01d7"
  ],
  "textContent": "I want to share one small thing today.\n\nThis is not an ad, not a product launch.\nIt is just a tool I built for myself to debug RAG / LLM pipelines, and it helped me so many times that it feels wrong to keep it only for me.\n\nWhen we build RAG, many bugs look the same on the surface.\nModel answers feel “kind of wrong”, and we guess randomly: maybe vector DB problem, maybe prompt, maybe top-k, maybe need bigger model. We change many things, but still do not really know what is actually broken.\n\nBecause of this, I wrote down the common failure patterns and turned them into a small “AI clinic” inside a ChatGPT shared conversation. It is not a new model. It is just a fixed way of thinking about sixteen types of RAG / LLM failures, with some math / system view behind it.\n\nLink here:\nhttps://chatgpt.com/share/68b9b7ad-51e4-8000-90ee-a25522da01d7\n\nHow to use is very simple:\n\n  * copy-paste your real problem (question, model answer, expected answer)\n  * add any logs, screenshots, top-k results, vector DB name (FAISS, Qdrant, Weaviate, Milvus, pgvector, etc)\n  * write in normal language what you already tried\n\n\n\nThe “clinic” will try to:\n\n  * restate your problem in plain English\n  * guess which kind of failure you are hitting\n  * point to the likely broken layer (retrieval, embedding, reasoning, routing, deployment)\n  * propose a few small experiments to confirm or reject the guess\n\n\n\nFor me this changed the workflow from “try 10 random fixes” to “run 2–3 targeted checks”.\n\nNo signup, no extra website, just that ChatGPT share link.\n\nIf you are building RAG, document QA, internal copilots or agent workflows, and you have one of those bugs that feels wrong but you cannot name it, you can just copy-paste your case into this clinic and see if the diagnosis is useful. Take what helps, ignore the rest.\n\n",
  "title": "WFGY AI Clinic: a small “ER” for RAG and LLM failures"
}