Private AI for internal wiki search
Semantic search finds the wiki page that answers a question even when it uses different words than the query, which is where keyword search fails and people give up and ask a coworker. It also means every page, including the candid ones, goes through an embedding model somewhere. Run that model yourself on a dedicated machine and the wiki never passes through a vendor's indexing service.
What's in a company wiki
Engineering wikis collect the most unguarded writing in a company: postmortems that name what broke and who was paged, architecture decision records that explain why a shortcut was taken, runbooks with internal hostnames and access procedures, and onboarding notes that describe how things really work. That candor makes the wiki useful. It also makes it something nobody wants in a third party's embedding service, especially the infrastructure details that would help an attacker as much as a new hire.
A cloud AI search product embeds and stores every one of those pages outside the company's systems, often through an integration whose data-handling terms few people on the team have read.
What changes when the search pipeline is self-hosted
Running the embedding model and vector index on a dedicated DGX Spark means pages are chunked, embedded, indexed and searched by models you run. No page goes to an AI vendor's embedding API. Engineers still ask "how do we roll back a bad deploy" in plain language and land on the right runbook.
A rented Spark is a GPUwerk-run machine (its own in Czechia, or a partner's in the country the console shows), operated under its published DPA, not part of your network. Check that your security policy allows a hosting provider before runbooks with internal hostnames go on it.
Open WebUI is a practical way to stand this up, using the retrieval approach in the guide to on-premise RAG. Point it at a wiki export or API connection, let it handle the chunk-embed-retrieve loop, and swap in a stronger multilingual embedding model if the default underperforms on technical or non-English content. If you already run embeddings for another workload (see self-hosted semantic search), reuse that infrastructure instead of standing up a second pipeline.
Chunking and staleness matter more than the model
Search quality over a wiki depends more on how pages are chunked and how stale content is handled than on which model answers the query. A wiki with years of unmaintained pages will surface outdated runbooks as readily as current ones. The pipeline needs a way to weight or flag recency, and the wiki needs an owner willing to prune pages that are no longer accurate.
What this doesn't solve
Self-hosted search doesn't fix a wiki that's poorly organized or full of stale pages, and it doesn't replace access controls that were already supposed to restrict who can read which pages. What it removes is the exposure of sending an entire internal knowledge base through a third party's indexing service to get search that should have worked in the first place. See pricing for what a dedicated Spark costs for an engineering team's wiki.