Pain map
What breaks in document AI
290 production incidents from public threads. Every quote links to its source. Pick a category and check it yourself.
Offline install and dependencies
20 of 20 quotes
“With `HF_HUB_OFFLINE=1`, a Hub-ID speculative draft model logs `ERROR ... Error retrieving safetensors: ... offline mode is enabled` twice (plus a 2 s retry sleep), although startup succeeds”
GitHub, vllm · 5 November 2026 · open the thread “For offline boxes the workaround that saved me was pointing `RAG_EMBEDDING_ENGINE` at a local ollama (nomic-embed-text or bge-m3)”
GitHub, open-webui · 26 August 2026 · open the thread “OFFLINE_MODE does not prevent the Whisper model from downloading”
GitHub, open-webui · 31 July 2026 · open the thread “Currently, there is no straightforward way to export or copy a specific model from one machine to another.”
GitHub, ollama · 10 July 2026 · open the thread “Web fetch no longer works in offline mode with external loader”
GitHub, open-webui · 3 July 2026 · open the thread Hit this one too? We are building for it. Join the waitlist.
“The installer of AnythingLLM downloads Ollama Cuda v12 and v13 libraries, as well as other stuff such as mmpeg codecs etc.”
GitHub, anything-llm · 13 June 2026 · open the thread “As far as I understand it, Ollama 0.30 requires a separate llama-server for the GGUF models. At least with Homebrew, this isn't currently installed automatically.”
GitHub, ollama · 2 June 2026 · open the thread “privacy-restricted machines I run hit the same wall on first build.”
GitHub, ollama · 23 May 2026 · open the thread “Building Ollama in fully offline environments is currently difficult because the go dependencies must be fetched at build time.”
GitHub, ollama · 22 May 2026 · in production · open the thread “Installing npm on my offline server is difficult, and I do not need to use the Web UI.”
GitHub, llama.cpp · 17 May 2026 · open the thread “this bug definitely increases the pain level for me a bit on upgrading llama.cpp regularly on an airgapped system.”
GitHub, llama.cpp · 17 May 2026 · in production · open the thread “I'm building `llama-server` on an offline server and build fails because `llama.cpp` now has some parts of it on huggingface.co and downloads it automatically.”
GitHub, llama.cpp · 16 May 2026 · in production · open the thread “I'm trying to run Gemma 4 (google/gemma-4-31B-it) with vLLM in a strictly air-gapped environment where downloads and inference must be completely separated.”
GitHub, vllm · 5 April 2026 · in production · open the thread “In a fully offline/air-gapped environment, we use `OFFLINE_MODE=true` to isolate the OpenWebUI backend. However, this does not isolate the client browser.”
GitHub, open-webui · 10 March 2026 · in production · open the thread “so I just quickly unplugged ethernet cable, and voila. in the chat its just loading.”
GitHub, anything-llm · 5 March 2026 · open the thread “However for offline environment (without Internet access), it just mentioned the upgrade of the Docker image, but not the code.”
GitHub, ragflow · 10 February 2026 · open the thread “'punkt_tab' not bundled with OWUI Container, results in Offline Mode / Airgapped system issues”
GitHub, open-webui · 4 February 2026 · in production · open the thread “v0.7.2: --data-dir flag removed, breaks bare metal installs”
GitHub, open-webui · 12 January 2026 · open the thread “after a Windows installation it requires Internet connections and it calls home and furthermore it will not start without internet connection.”
GitHub, anything-llm · 9 November 2025 · open the thread “I'm deploying Open WebUI in a fully air-gapped (offline) environment, where no external network access is available. Currently, every embedding or reranker model has to be manually downloaded from Hugging Face and then manually added into the Docker image.”
GitHub, open-webui · 7 November 2025 · in production · open the thread Hit this one too? We are building for it. Join the waitlist.
Method
We collected 448 public posts published after 1 November 2025 in which a builder describes document AI failing. Each link was re-opened on 8 November 2026. We kept the 290 where the quote is on the page, the author describes an incident and not an opinion, and has no tie to a vendor. Quotes are verbatim and trimmed after 420 characters. The grouping is ours. A complaint is not a measurement.