docs: LLM-rutingsstrategi — lett front, tung bakkant per modul

Updated the status and added LLM routing strategy details, including models, pricing, and configuration. Improved cost estimates and region availability information.
This commit is contained in:
chrischristiansen-glitch 2026-05-24 11:37:55 +02:00 committed by GitHub
parent 761c32e10a
commit 3f5a559600
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194

View File

@ -1,7 +1,7 @@
# VAUCO MASTERPLAN — Chris Christiansen # VAUCO MASTERPLAN — Chris Christiansen
**Sist oppdatert: 2026-05-24 CEST** **Sist oppdatert: 2026-05-24 CEST**
**Status: Repo ferdig — GCP gjenstår (A4A9 neste)** **Status: Repo ferdig — GCP gjenstaar (A4A9 neste) | LLM-strategi: Lett front / tung bakkant — se LLM-RUTINGSSTRATEGI**
--- ---
@ -240,5 +240,54 @@ Terraform-ressurser:
--- ---
*OSVauco-NMTMD-GCOS | propane-will-491900-m5 | us-central1*
---
## LLM-RUTINGSSTRATEGI
**Prinsipp: Lett modell i front, tung modell ved behov i bakkant — vurdert per modul og kost.**
| Modell | Pris input/output (per 1M) | EU-tilgang | Bruksomrade |
|---|---|---|---|
| `gemini-2.0-flash` | $0.10 / $0.40 | Ja (europe-west4) | Standard — alle enkle kall |
| `gemini-2.5-pro` | $1.25 / $10.00 | Ja (europe-west4) | Kompleks reasoning, compliance |
| `gemini-3.5-flash` | $1.50 / $9.00 | Nei (kun US/global) | Tung orkestrering, Fase C |
**Rutetabell per modul:**
| Modul / Agent | Standard modell | Eskalering til | Trigger |
|---|---|---|---|
| OPAX billing-agent | `gemini-2.0-flash` | — | Aldri — strukturert data |
| RAG retrieval + svar | `gemini-2.0-flash` | `gemini-2.5-pro` | Komplekse flertrinns-sporsmal |
| Memory Bank oppslag | `gemini-2.0-flash` | — | Aldri — faktahenting |
| TUI-cockpit (operator) | `gemini-2.0-flash` | `gemini-3.5-flash` | Bruker ber om dyp analyse |
| vauco-os (kundeflate) | `gemini-2.0-flash` | `gemini-2.5-pro` | Compliance/kliniske sporsmal |
| Fase C: Terraform-gen | `gemini-3.5-flash` | — | Standard for denne modulen |
| Fase C: Kunde-onboarding | `gemini-2.5-pro` | `gemini-3.5-flash` | Multi-step orkestrering |
**Konfigurasjon i `.env`:**
```env
ORCHESTRATOR_MODEL=gemini-2.0-flash
SUBAGENT_MODEL=gemini-2.0-flash
REASONING_MODEL=gemini-3.5-flash
```
**Regionsstatus (mai 2026):**
- `gemini-2.0-flash` + `gemini-2.5-pro` tilgjengelig `europe-west4` ok
- `gemini-3.5-flash` kun `us-central1` / global endpoint
- Navarende prosjektregion: `us-central1` — dekker alle modeller ok
**Kostnadsestimat med ruting:**
| Scenario | Uten ruting (3.5 Flash alt) | Med ruting | Besparelse |
|---|---|---|---|
| Dev daglig | ~$15-30/mnd | ~$1-5/mnd | ~85% |
| 1 kunde prod | ~$60-120/mnd | ~$8-20/mnd | ~85% |
---
*OSVauco-NMTMD-GCOS | propane-will-491900-m5 | us-central1* *OSVauco-NMTMD-GCOS | propane-will-491900-m5 | us-central1*
*GitHub: chrischristiansen-glitch/OSVauco* *GitHub: chrischristiansen-glitch/OSVauco*
*GitHub: chrischristiansen-glitch/OSVauco*