c7c457ae2a
Reconcile 6-week divergence: Gitea main (injection-defense, output-defense, prompt-guard-client, admin-auth, start-with-env, dashboard-v2, savings-calculator, race-mode, gamification + 13 more modules) merged with Erik's deployed features (usage-report endpoint, per-device entries, CEST timezone, cost-panel, bridge routing). ecosystem.config.cjs excluded (live token, never commit).
97 lines
2.7 KiB
YAML
97 lines
2.7 KiB
YAML
# LLM Gateway Model Configuration
|
|
# Ollama base URL: http://192.168.178.169:11434
|
|
|
|
ollama_base_url: "http://127.0.0.1:11434"
|
|
|
|
tiers:
|
|
fast:
|
|
timeout_ms: 10000
|
|
error_threshold_percent: 50
|
|
circuit_breaker_reset_ms: 15000
|
|
medium:
|
|
timeout_ms: 30000
|
|
error_threshold_percent: 50
|
|
circuit_breaker_reset_ms: 20000
|
|
large:
|
|
timeout_ms: 120000
|
|
error_threshold_percent: 30
|
|
circuit_breaker_reset_ms: 45000
|
|
code_generation:
|
|
timeout_ms: 180000
|
|
error_threshold_percent: 20
|
|
circuit_breaker_reset_ms: 60000
|
|
|
|
models:
|
|
# Fast tier
|
|
qwen2.5:3b:
|
|
tier: fast
|
|
context_length: 32768
|
|
strengths: [classification, summarization, routing]
|
|
max_tokens_default: 512
|
|
|
|
qwen2.5:7b:
|
|
tier: fast
|
|
context_length: 32768
|
|
strengths: [classification, summarization, short_analysis]
|
|
max_tokens_default: 1024
|
|
|
|
qwen2.5:7b-instruct:
|
|
tier: fast
|
|
context_length: 32768
|
|
strengths: [classification, summarization, short_analysis]
|
|
max_tokens_default: 1024
|
|
|
|
qwen2.5-coder:7b-instruct:
|
|
tier: fast
|
|
context_length: 32768
|
|
strengths: [code_generation, technical_analysis, routing]
|
|
max_tokens_default: 512
|
|
|
|
# ─── MAGATAMA — Fine-tuned Security Intelligence (Context X) ─────────────────
|
|
magatama:32b:
|
|
tier: large
|
|
context_length: 131072
|
|
strengths: [security_analysis, threat_intelligence, compliance, bgp_security, incident_response, nis2, ciso_reporting, complex_writing, deep_analysis, technical]
|
|
max_tokens_default: 4096
|
|
description: "MAGATAMA まがたま — TEPPEKI 7-pillar security AI, fine-tuned on Qwen2.5-32B"
|
|
|
|
# Medium tier
|
|
qwen2.5:14b:
|
|
tier: medium
|
|
context_length: 131072
|
|
strengths: [general, writing, analysis, coding, dialogue]
|
|
max_tokens_default: 2048
|
|
|
|
magatama-llm-v2-0:latest:
|
|
tier: medium
|
|
context_length: 131072
|
|
strengths: [general, writing, analysis, coding, dialogue]
|
|
max_tokens_default: 2048
|
|
|
|
magatama-coder:latest:
|
|
tier: medium
|
|
context_length: 65536
|
|
strengths: [code_generation, technical_analysis, debugging]
|
|
max_tokens_default: 2048
|
|
|
|
# Large tier
|
|
qwen2.5:32b:
|
|
tier: large
|
|
context_length: 131072
|
|
strengths: [complex_writing, deep_analysis, technical, security_analysis]
|
|
max_tokens_default: 4096
|
|
|
|
# Fallback chains per tier
|
|
fallback_chains:
|
|
fast: [qwen2.5:7b-instruct, qwen2.5-coder:7b-instruct]
|
|
medium: [magatama-llm-v2-0:latest, magatama-coder:latest, qwen2.5:7b-instruct]
|
|
large: [magatama:32b, magatama-llm-v2-0:latest]
|
|
code_generation: [magatama-coder:latest, qwen2.5-coder:7b-instruct]
|
|
|
|
# Cross-tier fallback when primary tier fails
|
|
tier_fallback:
|
|
code_generation: large
|
|
large: medium
|
|
medium: fast
|
|
fast: null
|