Semantic Search and Your Data

Lore has two kinds of search. Keyword search runs entirely on Lore's own servers — nothing in your knowledge base is sent anywhere. Semantic search — search by meaning rather than by exact words — cannot work that way. It needs a language model to turn text into vectors, and Lore does not run language models. So on accounts with semantic search, some knowledge base text is sent to a third party to be processed. This page explains exactly what, exactly when, and exactly to whom, so the trade is yours to make with the facts in hand.

Where Your Account Stands

As of 15 August 2026, semantic search is the standard search on Lore accounts — it was enabled for all existing accounts on that date, because measured on real knowledge bases it finds what keyword search misses (paraphrases, synonyms, questions phrased in your own words). Newly created accounts start on keyword search and are switched on as part of onboarding.

If you would rather your content never left Lore's servers, ask and your account goes back to keyword-only search — from that moment nothing in the rest of this page applies to it, and no content leaves at any point. Vectors already computed stay inside your encrypted knowledge base file like any other column; what was previously sent for processing cannot be unsent, which is why this page is explicit about what is sent and when.

What Is Sent, and When

With semantic search on, text leaves Lore's servers at two moments and no others.

When an entry is written or updated

The headings and bodies of all of that entry's sections are sent together to be turned into vectors — the model indexes an entry as one document, so changing any one section re-sends its siblings. This happens when your AI creates or edits an entry, and once in bulk when semantic search is first enabled for your account, to index what is already there.

When you or your AI runs a meaning-based search

The search query is sent, along with the text of the candidate sections that matched, so the results can be ordered by relevance. Each candidate goes with its entry's name, type and description, which the ranking step reads alongside the section. An exact-word search sends nothing — it is answered entirely from Lore's own index.

Who Receives It

VoyageAI, a United States company, over an encrypted (TLS) connection. Lore uses two of their models: voyage-context-3 to produce the vectors and rerank-2.5 to order results. VoyageAI's handling of the text it receives is governed by their own terms and privacy policy, not by Lore's — if that matters to your decision, read theirs before enabling this.

What Stays Inside Lore

Turning It Off

Switching an account back to keyword search takes effect on the very next request — no waiting period and no redeploy. From that moment nothing further is sent. Vectors already stored stay inside your encrypted database, where they are inert; they are reused if you ever switch back on, and removed with the rest of your data if you delete your account.

There is also a switch above the per-account one: Lore can be run with no VoyageAI credential configured at all, which makes these requests impossible process-wide, for every account, regardless of individual settings.

If It Fails, Search Still Works

Semantic search is an addition to keyword search, not a replacement for it. If VoyageAI is unreachable or returns an error, that search silently falls back to keyword results rather than failing — and the response tells your AI that it did, so it can say so rather than quietly presenting weaker results as though they were the best available.

The Short Version

If anything here is unclear, or you want your account switched to keyword-only search, get in touch — this is exactly the kind of trade-off worth a conversation.