Semantic search with no API key, no cost and no connection

When people read “AI search” they assume three things: that there is an API key, that there is a per-query cost, and that the text leaves the company. All three are true of most products. None of them is necessary.

What is actually required

Semantic search needs two pieces: a model that turns text into vectors, and an index you can search by proximity. Neither requires a remote server.

Small embedding models fit in a few hundred megabytes, run on CPU in milliseconds, and for search within your own corpus they perform more than well enough. The vector index is a file on your disk.

The practical consequence: three modes

Because there is no API dependency, we can offer something a cloud product cannot: a choice.

• Instant (<5ms). Lexical search over the local index. No model, no connection. Exact, predictable results: it matches the words as written.
• With AI (~400ms). Search by concept. Finds “the proposal we sent in spring” without the word “spring” appearing anywhere.
• Auto. Both in parallel. The most relevant result wins, wherever it came from. This is the default because it gets it right without anyone configuring anything.

The latency difference matters more than it looks. Three hundred milliseconds is long enough that you stop typing and wait; five is not. Instant search is not a degraded tier — for about 80% of queries it is the better one.

What you gain by sending nothing

The obvious argument is compliance: if content never leaves, there is no international transfer to document, no sub-processor to audit, and the conversation with the DPO takes five minutes instead of five weeks.

The less obvious one is internal trust. When you tell a team you are going to index their work email, the first question is never about features. It is “who will be able to read this?”. Being able to answer “nobody, including us: the index is on your laptop, and you choose which domains are excluded” is the difference between a tool being adopted and being quietly sabotaged.

And what it costs

It is only fair to state the price. The index takes up disk space on every person’s machine. The first pass over years of history takes a while. And local semantic search quality sits below that of a large hosted model — close, but below.

In return, search is free and unlimited forever, on every plan, because there is no marginal cost to pass on. That is the trade, and for a company’s operational memory it strikes us as obviously the right one.