Embeddings: the hidden costs of an AI knowledge base
For the AI to answer from your knowledge base, every section needs a vector — and so does every visitor question. These embeddings cost almost nothing per call but occur constantly. SilentChat computes them at IONOS in Germany, books them separately and skips unchanged content.
Where embeddings arise
- On save: when an article is published or changed, SilentChat splits it into sections of about 512 tokens with 100 tokens of overlap, prefixes each with the article title and computes the vectors in the background. Only published articles and active snippets go into the index.
- When reading your website: new or changed pages are embedded.
- On every visitor question: the question itself is embedded to find matching sections. If a widget isn't connected to the knowledge base, this step is skipped.
- On reindexing: everything at once, in batches of 64 sections.
What it costs
We use the BAAI/bge-m3 model at IONOS with 1024 dimensions. According to the IONOS price list (checked on 11 August 2026) it costs 2 US cents per million input tokens. A 512-token section therefore costs about a thousandth of a cent.
Amounts this small misled us twice: first, embeddings were booked at zero; then every call was rounded to whole cents — a single query cost zero again. Since 14 September 2026, SilentChat calculates in millionths of a cent and carries remainders per tenant. Bookings are split into index, visitor question and duplicate detection.
How we avoid needless recomputation
- Only what changed: when re-reading a site, the crawler asks whether a page has changed. If the server answers “not modified”, nothing is loaded; otherwise SilentChat compares a checksum of the content and skips unchanged pages.
- Interval by plan: automatic re-reading runs every 30 days (Growth), every 7 days (Professional) or daily (Enterprise).
- No search without purpose: widgets without a knowledge base create no search embeddings.
Changing models means rebuilding
Until 13 August 2026, embeddings ran through OpenAI in the US, with 1536 dimensions. Vectors from two models aren't comparable. For the switch to IONOS, the column was therefore recreated empty and every knowledge base was reindexed.
To keep the model from starting cold, SilentChat sends it a small request every minute. These requests are booked without a tenant and don't count against any customer budget.