A knowledge base where not a single character leaves your servers
Install the same Wisdombase inside your own servers. Conversion, embedding, search and answers all finish on-site, and outbound traffic is zero by default.
What stays inside, what goes out
On-prem runs the same code as the SaaS. The only difference: every step happens on your servers.
Stays inside
Uploaded files are converted by the on-site parser; originals are deleted right after.
Chunking and embeddings come from the on-site embedding server (Qwen3-Embedding). Vectors live only in your DB.
Nothing exchanged through MCP or the chatbot leaves the company.
Who read and edited what, and who uses which access key — all recorded on-site.
Goes out (only if an admin enables it)
If an admin connects an answer model, traffic goes to that address only. Internal vLLM or Korean pack: zero. External commercial LLM (GPT, Claude, Gemini): per answer only the question and the retrieved excerpts (default up to 4,800 characters) — never source documents, embeddings, the index or accounts.
Mail, payments, external login and usage reporting are all off in the on-prem edition. Licenses are verified offline.
Domestic, internal, external commercial — or no LLM at all
Four ways to deploy, whether the rule is "no confidential uploads to foreign LLMs" or "answers must be GPT/Claude grade". Whichever you choose, documents, embeddings and the index never leave your servers.
Any model served by vLLM, llama.cpp or an internal AI gateway — enter address, key and model name. Test the connection from the admin screen.
Kakao Kanana 1.5 8B and SKT A.X 4.0 Light (Apache-2.0) as ready-to-run packs on your CPU/GPU server. No extra contract or key.
Use GPT, Claude, Gemini or Solar for writing answers only. Conversion, embedding and search finish on-site; what leaves is the question plus the retrieved excerpts (default up to 4,800 characters). The admin screen shows monthly what went out.
Hybrid search, MCP and the document site all work without an LLM. Answers are returned as document excerpts.
Search quality comes from the index, not the LLM
"Does search get worse on-prem?" "What changes with an external LLM?" — Which chunks are retrieved is decided by the parser, chunking, embeddings and hybrid search, and on-prem these four always run on-site with the same code. The LLM only writes the final sentences from the retrieved excerpts. So retrieval is identical with an on-prem or an external LLM; only answer wording and speed differ.
| Step | Where | On-prem LLM (Korean pack · internal vLLM) | External commercial LLM (hybrid) |
|---|---|---|---|
| Conversion (HWP, PDF, Office → Markdown) | On-site parser | Same | Same |
| Chunking and embeddings (Qwen3-Embedding) | On-site embedding server | Same | Same |
| Retrieval (semantic + full-text · RRF) | On-site DB | Same — same index, same ranking | Same — same index, same ranking |
| Writing the answer | LLM | Bound by your server. CPU-only: tens of seconds or more per answer, GPU recommended | Provider API speed. GPT/Claude-grade wording |
| Document analysis (optional · chunking strategy, summaries) | LLM | Minutes per page on CPU — off by default | If enabled, a document sample (up to 3,500 characters per page) leaves at import — off by default |
| What leaves | — | 0 | Question + excerpts (default ≤4,800 chars) · monthly traffic on the admin screen |
The comparison page contains retrieval latency measured on the same documents and questions, answer latency and sample wording for on-prem vs. external commercial LLMs, and an example of exactly what an LLM request contains.
See the measured comparison →What gets installed
Document site, editor, MCP, chatbot, teams, audit logs. Same build as the SaaS.
HWP, HWPX, DOC, DOCX, PPTX, XLS, XLSX, PDF, HTML, EPUB and more to Markdown. The on-prem Corepin parser ships alongside.
Qwen3-Embedding served on CPU. Chunking and embedding finish on-site with no external API.
Load a Korean LLM pack for on-site answers and analysis, connect an LLM you already run, or attach an external commercial LLM (GPT, Claude, Gemini) for answers only.
From license to install in 10 minutes
Sign up and create a team. On-prem licenses are issued per team.
In Team › On-prem, click "Trial license" for a 30-day, 50-seat license (once per team).
Download the install bundle (about 2.5GB) from the same screen. The LLM pack only if you need it.
On a Linux x86_64 server with Docker, unpack and run install.sh. Drop the two license files into license/.
Sign up with an email listed in WB_SITE_ADMINS in .env to become a server admin. Under Admin › On-prem, pick an internal LLM, a Korean pack or an external commercial LLM.
Before the trial ends, request a quote with seats and term on the same screen. After the contract, the commercial license is issued and you just swap in the new license file.
What you need
Questions
Does everything from the SaaS work?
Document site, editor, import (files and URLs), MCP, chatbot, GitHub sync (internal GitHub Enterprise), team policies and audit logs all work as-is. Not included on-prem: image OCR (scanned documents), full-site crawl, Google login and outgoing email.
Can we use a Korean LLM?
Yes. We provide Kakao Kanana 1.5 8B and SKT A.X 4.0 Light packs, and any Korean model you serve internally (HyperCLOVA X, EXAONE, etc.) can be connected through an OpenAI-compatible address.
Can we attach an external commercial LLM such as GPT, Claude or Gemini?
Yes. Under Admin › On-prem › LLM connection, pick OpenAI, Anthropic Claude, Google Gemini, Azure OpenAI, Upstage Solar or BizRouter and enter the API key. Embeddings and search still finish on-site; per answer only the question and the retrieved excerpts go to that provider. Admins set the excerpt cap (default 4,800 characters) and see monthly traffic.
Does an external LLM change search quality?
Retrieval is identical. Which chunks are found is decided by the on-site parser, chunking, embeddings and hybrid search; the LLM only writes sentences from those excerpts. What changes is answer wording and speed — measured on the same documents and questions in the comparison page.
How is the license verified?
The server verifies an Ed25519-signed file offline. No phone-home, no usage reporting. After expiry, reading and search continue; only new sign-ups, imports and AI answers stop (14-day grace).
Pricing?
Quoted by seats × term. Enter seats and term in Team › On-prem to request a quote and we reply within one business day, including procurement and contract paperwork. Trial licenses are issued instantly.
Updates?
Download the new bundle from the same screen and run upgrade.sh. Your data stays.
Kubernetes and SSO?
Single-node Docker Compose today. Contact us if you need Kubernetes or SSO (OIDC/SAML).
Start with a trial license
No card. 30 days, 50 seats. Create a team and issue it from the On-prem menu.