LiteLLM AI Gateway: Docker, DB-ruting, config-generering (oppgave 10.1)

Setter opp AI Gateway med LiteLLM som sentralisert proxy for alle
AI-kall. PG eier all modellkonfigurasjon — LiteLLM er stateløs.

- Migrasjon 008: ai_model_aliases, ai_model_providers, ai_job_routing
  med seed-data for sidelinja/rutine og sidelinja/resonering
- Config-generering fra PG: scripts/generate-litellm-config.sh
  filtrerer bort providers med tomme API-nøkler
- Docker-container kjører på sidelinja-net (intern, ingen eksponert port)
- Maskinrommet har AI_GATEWAY_URL via maskinrommet-env.sh
- API-nøkkel-placeholders i .env (GEMINI, ANTHROPIC, XAI)
- Oppdatert docs/infra/ai_gateway.md med faktisk config

Verifisert: container healthy, modellaliaser eksponert, maskinrommet
har korrekt gateway-URL. Reelle API-kall krever at Vegard fyller
inn leverandør-nøkler i /srv/synops/.env.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

config/litellm/config.yaml

@@ -1,40 +1,42 @@
-model_list:
+# LiteLLM config — GENERERT FIL
-  - model_name: "kjapp"
+#
-    litellm_params:
+# Denne filen genereres fra PostgreSQL-tabellene ai_model_aliases
-      model: "xai/grok-4-1-fast-non-reasoning"
+# og ai_model_providers. Rediger ikke manuelt — bruk admin-panelet
-      api_key: "os.environ/XAI_API_KEY"
+# eller kjør scripts/generate-litellm-config.sh.
-  - model_name: "kjapp"
+#
-    litellm_params:
+# Bare providers med gyldige API-nøkler i .env inkluderes.
-      model: "xai/grok-3-mini"
+# Fyll inn GEMINI_API_KEY, ANTHROPIC_API_KEY, XAI_API_KEY i
-      api_key: "os.environ/XAI_API_KEY"
+# /srv/synops/.env for å aktivere flere leverandører.
-  - model_name: "kjapp"
-    litellm_params:
-      model: "gemini/gemini-2.5-flash-lite"
-      api_key: "os.environ/GEMINI_API_KEY"
-  - model_name: "kjapp"
-    litellm_params:
-      model: "gemini/gemini-flash-lite-latest"
-      api_key: "os.environ/GEMINI_API_KEY"
-  - model_name: "kjapp"
-    litellm_params:
-      model: "gemini/gemini-flash-latest"
-      api_key: "os.environ/GEMINI_API_KEY"
 
-  - model_name: "resonering"
+model_list:
+  # === sidelinja/resonering ===
+  - model_name: "sidelinja/resonering"
     litellm_params:
       model: "openrouter/anthropic/claude-sonnet-4"
       api_key: "os.environ/OPENROUTER_API_KEY"
-  - model_name: "resonering"
+    model_info:
+      priority: 1
+  - model_name: "sidelinja/resonering"
     litellm_params:
       model: "openrouter/google/gemini-2.5-flash"
       api_key: "os.environ/OPENROUTER_API_KEY"
+    model_info:
+      priority: 2
+  # === sidelinja/rutine ===
+  - model_name: "sidelinja/rutine"
+    litellm_params:
+      model: "openrouter/google/gemini-2.5-flash"
+      api_key: "os.environ/OPENROUTER_API_KEY"
+    model_info:
+      priority: 4
 
 router_settings:
   routing_strategy: "simple-shuffle"
-  num_retries: 2
+  num_retries: 3
   timeout: 60
   allowed_fails: 1
   retry_after: 5
+  enable_pre_call_checks: true
 
 general_settings:
   master_key: "os.environ/LITELLM_MASTER_KEY"

docs/infra/ai_gateway.md

@@ -101,23 +101,38 @@ Admin-panelet lar administrator:
 ## 4. Docker-oppsett
 
 ```yaml
-# docker-compose.dev.yml / docker-compose.yml
+# /srv/synops/docker-compose.yml
 ai-gateway:
   image: ghcr.io/berriai/litellm:main-stable
   restart: unless-stopped
   command: --config /etc/litellm/config.yaml
   environment:
     LITELLM_MASTER_KEY: ${LITELLM_MASTER_KEY}
-    GEMINI_API_KEY: ${GEMINI_API_KEY}
+    GEMINI_API_KEY: ${GEMINI_API_KEY:-}
-    ANTHROPIC_API_KEY: ${ANTHROPIC_API_KEY}
+    ANTHROPIC_API_KEY: ${ANTHROPIC_API_KEY:-}
-    XAI_API_KEY: ${XAI_API_KEY}
+    XAI_API_KEY: ${XAI_API_KEY:-}
     OPENROUTER_API_KEY: ${OPENROUTER_API_KEY}
   volumes:
-    - ./config/litellm/config.yaml:/etc/litellm/config.yaml:ro
+    - /srv/synops/config/litellm:/etc/litellm
-  ports:
-    - "127.0.0.1:4000:4000"  # kun localhost (dev), ingen port i prod
   networks:
-    - sidelinja-dev  # eller sidelinja-net i prod
+    - sidelinja-net  # intern — maskinrommet når gateway via container-IP
+  healthcheck:
+    test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen(\"http://localhost:4000/health/liveliness\")"]
+    interval: 15s
+    timeout: 5s
+    retries: 3
+```
+
+### 4.1 Config-generering
+
+`config.yaml` genereres fra PG-tabellene med `scripts/generate-litellm-config.sh`.
+Scriptet filtrerer bort providers med tomme API-nøkler i `.env`, slik at bare
+leverandører med gyldige nøkler inkluderes. Kjør med `--restart` for å restarte
+containeren etter generering.
+
+```bash
+# Generer og restart:
+scripts/generate-litellm-config.sh --restart
 ```
 
 ## 5. Prompt-kvalitetssikring (Promptfoo)

migrations/008_ai_model_routing.sql

@@ -0,0 +1,100 @@
+-- 008_ai_model_routing.sql — Modellruting for AI Gateway (LiteLLM).
+--
+-- Nye tabeller:
+--   ai_model_aliases    — globale modellaliaser (sidelinja/rutine, sidelinja/resonering)
+--   ai_model_providers  — leverandør-modeller med prioritert fallback per alias
+--   ai_job_routing      — jobbtype → modellalias mapping
+--
+-- Ref: docs/infra/ai_gateway.md §3.2
+
+-- =============================================================================
+-- 1. Modellaliaser
+-- =============================================================================
+
+CREATE TABLE ai_model_aliases (
+    id            UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    alias         TEXT NOT NULL,
+    description   TEXT,
+    is_active     BOOLEAN NOT NULL DEFAULT true,
+    created_at    TIMESTAMPTZ NOT NULL DEFAULT now(),
+    UNIQUE(alias)
+);
+
+COMMENT ON TABLE ai_model_aliases IS 'Globale modellaliaser — abstraherer bort leverandør-spesifikke modellnavn';
+
+-- =============================================================================
+-- 2. Leverandør-modeller med fallback
+-- =============================================================================
+
+CREATE TABLE ai_model_providers (
+    id            UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    alias_id      UUID NOT NULL REFERENCES ai_model_aliases(id) ON DELETE CASCADE,
+    provider      TEXT NOT NULL,
+    model         TEXT NOT NULL,
+    api_key_env   TEXT NOT NULL,
+    priority      SMALLINT NOT NULL,
+    is_active     BOOLEAN NOT NULL DEFAULT true,
+    UNIQUE(alias_id, model)
+);
+
+COMMENT ON TABLE ai_model_providers IS 'Leverandør-modeller per alias, sortert etter priority (lavere = prøves først)';
+
+-- =============================================================================
+-- 3. Jobbtype → modellalias mapping
+-- =============================================================================
+
+CREATE TABLE ai_job_routing (
+    job_type      TEXT PRIMARY KEY,
+    alias         TEXT NOT NULL,
+    description   TEXT
+);
+
+COMMENT ON TABLE ai_job_routing IS 'Hvilken modellalias brukes for hvilken jobbtype';
+
+-- =============================================================================
+-- 4. Seed: initielle aliaser og providers
+-- =============================================================================
+
+-- Rutine-alias (billig, høyt volum)
+INSERT INTO ai_model_aliases (alias, description) VALUES
+    ('sidelinja/rutine', 'Billig, høyt volum — tekstvasking, research, metadata'),
+    ('sidelinja/resonering', 'Høy kvalitet — resonneringsoppgaver, live-assistent');
+
+-- Providers for sidelinja/rutine
+INSERT INTO ai_model_providers (alias_id, provider, model, api_key_env, priority) VALUES
+    ((SELECT id FROM ai_model_aliases WHERE alias = 'sidelinja/rutine'), 'xai', 'xai/grok-4-1-fast-non-reasoning', 'XAI_API_KEY', 1),
+    ((SELECT id FROM ai_model_aliases WHERE alias = 'sidelinja/rutine'), 'gemini', 'gemini/gemini-2.5-flash-lite', 'GEMINI_API_KEY', 2),
+    ((SELECT id FROM ai_model_aliases WHERE alias = 'sidelinja/rutine'), 'gemini', 'gemini/gemini-flash-lite-latest', 'GEMINI_API_KEY', 3),
+    ((SELECT id FROM ai_model_aliases WHERE alias = 'sidelinja/rutine'), 'gemini', 'gemini/gemini-flash-latest', 'GEMINI_API_KEY', 4),
+    ((SELECT id FROM ai_model_aliases WHERE alias = 'sidelinja/rutine'), 'openrouter', 'openrouter/google/gemini-2.5-flash', 'OPENROUTER_API_KEY', 5);
+
+-- Providers for sidelinja/resonering
+INSERT INTO ai_model_providers (alias_id, provider, model, api_key_env, priority) VALUES
+    ((SELECT id FROM ai_model_aliases WHERE alias = 'sidelinja/resonering'), 'anthropic', 'anthropic/claude-sonnet-4-20250514', 'ANTHROPIC_API_KEY', 1),
+    ((SELECT id FROM ai_model_aliases WHERE alias = 'sidelinja/resonering'), 'openrouter', 'openrouter/anthropic/claude-sonnet-4', 'OPENROUTER_API_KEY', 2),
+    ((SELECT id FROM ai_model_aliases WHERE alias = 'sidelinja/resonering'), 'openrouter', 'openrouter/google/gemini-2.5-flash', 'OPENROUTER_API_KEY', 3);
+
+-- Jobbtype → alias routing
+INSERT INTO ai_job_routing (job_type, alias, description) VALUES
+    ('ai_text_process', 'sidelinja/rutine', 'Tekstvasking og ✨-behandling, høyt volum'),
+    ('whisper_postprocess', 'sidelinja/rutine', 'Transkripsjonsvasking etter Whisper'),
+    ('research_clip', 'sidelinja/rutine', 'Research-oppsummering'),
+    ('live_factoid_eval', 'sidelinja/resonering', 'Faktoid-vurdering under live sending — krever presisjon'),
+    ('agent_respond', 'sidelinja/resonering', 'Claude chat-agent svar');
+
+-- =============================================================================
+-- 5. Oppdater agent_identities config til ny alias
+-- =============================================================================
+
+UPDATE agent_identities
+SET config = jsonb_set(config, '{model_alias}', '"sidelinja/resonering"')
+WHERE agent_key = 'claude-main'
+  AND config->>'model_alias' = 'resonering';
+
+-- =============================================================================
+-- 6. Tilganger
+-- =============================================================================
+
+GRANT SELECT ON ai_model_aliases TO synops_reader;
+GRANT SELECT ON ai_model_providers TO synops_reader;
+GRANT SELECT ON ai_job_routing TO synops_reader;

scripts/generate-litellm-config.sh

@@ -0,0 +1,97 @@
+#!/usr/bin/env bash
+# Genererer LiteLLM config.yaml fra ai_model_aliases/providers i PostgreSQL.
+# Kjøres ved oppstart, ved endringer i admin-panelet, eller manuelt.
+#
+# Bruk:
+#   ./scripts/generate-litellm-config.sh [--restart]
+#
+# --restart: restart ai-gateway containeren etter generering.
+
+set -euo pipefail
+
+CONFIG_FILE="/srv/synops/config/litellm/config.yaml"
+DOCKER_CONTAINER="sidelinja-ai-gateway-1"
+PG_CONTAINER="sidelinja-postgres-1"
+PG_USER="sidelinja"
+PG_DB="synops"
+ENV_FILE="/srv/synops/.env"
+
+read_env() { grep "^$1=" "$ENV_FILE" 2>/dev/null | head -1 | cut -d= -f2; }
+
+psql_cmd() {
+    docker exec "$PG_CONTAINER" psql -U "$PG_USER" -d "$PG_DB" -tAF '|' -c "$1"
+}
+
+# Hent aktive aliaser og deres providers, sortert etter alias + priority
+PROVIDERS=$(psql_cmd "
+    SELECT a.alias, p.model, p.api_key_env, p.priority
+    FROM ai_model_providers p
+    JOIN ai_model_aliases a ON p.alias_id = a.id
+    WHERE a.is_active = true AND p.is_active = true
+    ORDER BY a.alias, p.priority;
+")
+
+if [ -z "$PROVIDERS" ]; then
+    echo "FEIL: Ingen aktive providers funnet i databasen." >&2
+    exit 1
+fi
+
+# Generer YAML — filtrerer bort providers med tomme API-nøkler
+{
+    echo "model_list:"
+
+    current_alias=""
+    included=0
+    while IFS='|' read -r alias model api_key_env priority; do
+        # Sjekk at API-nøkkelen finnes og er ikke-tom i .env
+        key_value=$(read_env "$api_key_env")
+        if [ -z "$key_value" ]; then
+            echo "  # HOPPET OVER: ${model} (${api_key_env} ikke satt)" >&2
+            continue
+        fi
+
+        if [ "$alias" != "$current_alias" ]; then
+            echo "  # === ${alias} ==="
+            current_alias="$alias"
+        fi
+        # priority i DB er 1-basert, LiteLLM priority er 0-basert (lavere = høyere)
+        litellm_priority=$((priority - 1))
+        cat <<ENTRY
+  - model_name: "${alias}"
+    litellm_params:
+      model: "${model}"
+      api_key: "os.environ/${api_key_env}"
+    model_info:
+      priority: ${litellm_priority}
+ENTRY
+        included=$((included + 1))
+    done <<< "$PROVIDERS"
+
+    if [ "$included" -eq 0 ]; then
+        echo "FEIL: Ingen providers med gyldige API-nøkler." >&2
+        exit 1
+    fi
+
+    cat <<'FOOTER'
+
+router_settings:
+  routing_strategy: "simple-shuffle"
+  num_retries: 3
+  timeout: 60
+  allowed_fails: 1
+  retry_after: 5
+  enable_pre_call_checks: true
+
+general_settings:
+  master_key: "os.environ/LITELLM_MASTER_KEY"
+FOOTER
+} | sudo tee "$CONFIG_FILE" > /dev/null
+
+echo "Config generert: $CONFIG_FILE"
+
+# Eventuelt restart
+if [[ "${1:-}" == "--restart" ]]; then
+    echo "Restarter $DOCKER_CONTAINER..."
+    sudo docker restart "$DOCKER_CONTAINER" > /dev/null
+    echo "Ferdig."
+fi

tasks.md

@@ -116,8 +116,7 @@ Uavhengige faser kan fortsatt plukkes.
 
 ## Fase 10: AI og beriking
 
-- [~] 10.1 LiteLLM oppsett: Docker-container, API-nøkler, modell-routing. Ref: `docs/infra/ai_gateway.md`.
+- [x] 10.1 LiteLLM oppsett: Docker-container, API-nøkler, modell-routing. Ref: `docs/infra/ai_gateway.md`.
-  > Påbegynt: 2026-03-17T23:03
 - [ ] 10.2 AI-foreslåtte edges: maskinrommet sender innhold til LLM → foreslår mentions, topics.
 - [ ] 10.3 Oppsummering: kommunikasjonsnode → AI-generert sammendrag som ny node.
 - [ ] 10.4 TTS: tekst → lyd via ElevenLabs. Mottaker-preferanse i metadata.