Repository navigation
Decide what moving Kudos into the knowledge database commits us to #175
Description
Part of digdir/kunnskapsassistenten#151.
ADR 0042 (the knowledge database, #125) sets out what any dataset gets when it moves in. It does not decide moving Kudos. These four decisions are taken here, each with its cost.
Prices are Azure list prices in USD for Norway East, before any agreement discount, at 730 hours a month. They come from the Azure Retail Prices API on 2026-10-29. Each link below returns the one price used. To browse instead of query: the pricing pages for Azure Database for PostgreSQL flexible server and Azure OpenAI, with Norway East chosen.
These are list prices for a size estimated in a container. #132 measures a Kudos-sized build on a real Azure server and prices the size it supports, standby and backups included, for digdir/kunnskapsassistenten#30. Its numbers replace the ones in section 2 when they arrive.
1. Move the full Kudos corpus
45,961 documents and about 3.1 million pages and vectors (KUDOS listing, 2026-10-27).
- Embeddings: the pilot used 13,504,576
text-embedding-3-largetokens for 299 documents (timings.md), which cost $1.80 (costs.csv). Scaled to 45,961 documents, that is about 2.1 billion tokens. Azure charges $0.13 per million tokens on a global deployment and $0.158 on a regional one in Norway East. So the cost is about $270 one-off on a global deployment, or about $330 on a regional one. Which applies depends on decision 4. A later model change costs the same again. - Conversion: Marker GPU time is not measured (ADR 0042, Open). Until a host is decided, the text layer is used, at no extra cost.
2. Server size and standby
| Size | Compute per month | Storage per month | Total |
|---|---|---|---|
Standard_D2ds_v5, 2 vCores, 8 GiB: the size ADR 0042 starts at |
$174 ($0.239/h) | $10 (64 GiB at $0.151) | about $185 |
Standard_E8ds_v5, 8 vCores, 64 GiB: what Kudos would need |
$969 ($1.328/h) | $19 (128 GiB at $0.151) | about $990 |
| The same, with a zone-redundant standby | about $1,980, if Azure bills the standby like the primary (not checked; #132 prices it) |
Backup storage is $0.131 per GB a month. How much of it Azure includes for free is not checked.
3. Remove the shared Typesense (#150)
Saves the shared cluster's cost, which is not known here. It only follows once Kudos runs on the knowledge database and an evaluation run has checked it.
4. Where Kudos may be processed (#131)
No direct cost of its own. #131 declares Kudos global for now, because every embedding model in the catalog is declared global and a dataset declared norway cannot be embedded at all. The decision here is whether that stays: it decides whether Kudos files may go to a Marker host outside Norway, and whether embeddings may use a global deployment (decision 1).
Activity