When your internet connection is censored or disabled, you will want to learn things like how to effectively grow your own food.
EoWKit downloads a copy of Wikipedia and an offline LLM to your computer so you can still get answers to hard problems, even when you've been cut off from the internet.
Tested Feature
✅ macOS, wikipedia_en_all_mini_2025-06.zim, phi3:mini, reranker=NO
❌ Windows
❌ Linux
Download the latest release for your OS under Assets.
eowkit install
Unsigned builds (macOS and Windows):
macOS (Gatekeeper)
eowkit → Open → Openxattr -dr com.apple.quarantine /path/to/eowkit
Windows (SmartScreen/Defender)
eowkit.exe → Properties → check "Unblock" → Apply → OKPrompts for directories: downloads, ZIM storage, models (OLLAMA_MODELS)
Lets you choose the Wikipedia snapshot and LLM model
Warns if RAM is too low or target disk is nearly full (90% rule)
Offers to enable the reranker and auto-download ONNX + vocab
Edit configs/eowkit.toml
[paths]
downloads_dir = "/mnt/downloads" # where large downloads land
zim_dir = "/mnt/wiki" # where .zim files live
models_dir = "/mnt/ollama/models" # where Ollama stores models
library_xml = "/mnt/wiki/library.xml" # optional: serve many ZIMs
[wiki]
zim = "wikipedia_en_all_nopic_2025-08.zim" # resolved under paths.zim_dir
kiwix_port = 8080
bind = "127.0.0.1"
To serve many ZIMs:
kiwix-manage /mnt/wiki/library.xml add /mnt/wiki/*.zim
Set paths.library_xml and leave wiki.zim as-is (it’ll be ignored when a library is present).
eowkit run
eowkit install-reranker
This downloads a cross-encoder ONNX model + vocab into models_dir and flips TOML:
[reranker]
enabled = true
onnx_model = "models/ce-minilm-l6.onnx"
tokenizer_vocab = "models/vocab.txt"
max_seq_len = 256
[llm_runtime]
num_threads = 8
downloads_dir and then are copied to zim_dir.paths.models_dir via OLLAMA_MODELS when the daemon is spawned.What happens: • kiwix-serve exposes the offline Wikipedia. • Ollama is started (honors paths.models_dir via OLLAMA_MODELS). • The runner retrieves top articles and answers with citations (titles). If the reranker is enabled and its files exist, results are re-ranked with the cross-encoder.
When your internet connection is censored or disabled, you will want to learn things like how to effectively grow your own food.
EoWKit downloads a copy of Wikipedia and an offline LLM to your computer so you can still get answers to hard problems, even when you've been cut off from the internet.
Tested Feature
✅ macOS, wikipedia_en_all_mini_2025-06.zim, phi3:mini, reranker=NO
❌ Windows
❌ Linux
Download the latest release for your OS under Assets.
eowkit install
Unsigned builds (macOS and Windows):
macOS (Gatekeeper)
eowkit → Open → Openxattr -dr com.apple.quarantine /path/to/eowkit
Windows (SmartScreen/Defender)
eowkit.exe → Properties → check "Unblock" → Apply → OKPrompts for directories: downloads, ZIM storage, models (OLLAMA_MODELS)
Lets you choose the Wikipedia snapshot and LLM model
Warns if RAM is too low or target disk is nearly full (90% rule)
Offers to enable the reranker and auto-download ONNX + vocab
Edit configs/eowkit.toml
[paths]
downloads_dir = "/mnt/downloads" # where large downloads land
zim_dir = "/mnt/wiki" # where .zim files live
models_dir = "/mnt/ollama/models" # where Ollama stores models
library_xml = "/mnt/wiki/library.xml" # optional: serve many ZIMs
[wiki]
zim = "wikipedia_en_all_nopic_2025-08.zim" # resolved under paths.zim_dir
kiwix_port = 8080
bind = "127.0.0.1"
To serve many ZIMs:
kiwix-manage /mnt/wiki/library.xml add /mnt/wiki/*.zim
Set paths.library_xml and leave wiki.zim as-is (it’ll be ignored when a library is present).
eowkit run
eowkit install-reranker
This downloads a cross-encoder ONNX model + vocab into models_dir and flips TOML:
[reranker]
enabled = true
onnx_model = "models/ce-minilm-l6.onnx"
tokenizer_vocab = "models/vocab.txt"
max_seq_len = 256
[llm_runtime]
num_threads = 8
downloads_dir and then are copied to zim_dir.paths.models_dir via OLLAMA_MODELS when the daemon is spawned.What happens: • kiwix-serve exposes the offline Wikipedia. • Ollama is started (honors paths.models_dir via OLLAMA_MODELS). • The runner retrieves top articles and answers with citations (titles). If the reranker is enabled and its files exist, results are re-ranked with the cross-encoder.