defaultline/settings.yaml at main · teetangh/defaultline · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
# Single source of truth for swappable architectural choices.
# See docs/07-decision-agnosticism.md for the swap-cost matrix.

# ---------------------------------------------------------------------------
# LLM tiering — currently configured for OpenRouter free tier per Riverline's
# email recommending free providers (NVIDIA NIM / Cloudflare Workers AI /
# OpenRouter). To swap to Anthropic, replace each role with the commented-out
# block beneath it.
#
# Free-tier rate limits are tight (~20 req/min on OpenRouter free models). If
# you hit them, either (a) reduce eval cycle size, (b) swap to NVIDIA NIM (more
# generous limits), or (c) swap to a paid provider.
llm:
  agent:
    provider: openai_compatible
    model: openai/gpt-oss-120b:free
    base_url: https://openrouter.ai/api/v1
    api_key_env: OPENROUTER_API_KEY
    max_tokens: 1500
    temperature: 0.0
  sim:
    provider: openai_compatible
    model: openai/gpt-oss-20b:free
    base_url: https://openrouter.ai/api/v1
    api_key_env: OPENROUTER_API_KEY
    max_tokens: 1500
    temperature: 0.7
  judge:
    provider: openai_compatible
    model: minimax/minimax-m2.5:free
    base_url: https://openrouter.ai/api/v1
    api_key_env: OPENROUTER_API_KEY
    max_tokens: 2000
    temperature: 0.0
    # MiniMax is on a separate rate-limit pool from the gpt-oss family used
    # by agent + sim, so judge calls don't compete with simulation calls.
    # Cross-family also reduces same-family-judge bias.
  mutator:
    provider: openai_compatible
    model: nvidia/nemotron-3-super-120b-a12b:free
    base_url: https://openrouter.ai/api/v1
    api_key_env: OPENROUTER_API_KEY
    max_tokens: 4000
    temperature: 0.5
    # Nemotron is reasoning-tuned and works well for reflective prompt
    # mutation. Different rate-limit pool from minimax (judge).

  # ---------------------------------------------------------------------------
  # Voice-specialized agent. Voice cannot tolerate free-tier OpenRouter
  # throttle backoff (8-28s per call) — Twilio webhook timeouts and borrower
  # patience both fail. Groq runs Llama 3.3 70B on custom LPU silicon at
  # 250-400 tokens/sec; per-call latency is typically 0.5-1.5s including
  # network. Free tier: 30 RPM, 6,000 TPM — comfortable for voice.
  #
  # Get a key at console.groq.com (Google login works), paste as
  # GROQ_API_KEY in .env. If the key is missing, voice sessions fall back
  # to the default `agent` role automatically (see chat_session.py).
  agent_voice:
    provider: openai_compatible
    model: llama-3.3-70b-versatile
    base_url: https://api.groq.com/openai/v1
    api_key_env: GROQ_API_KEY
    max_tokens: 250
    temperature: 0.0

# To switch to Anthropic Claude tiered (requires ANTHROPIC_API_KEY in .env):
# llm:
#   agent:    {provider: anthropic, model: claude-sonnet-4-6, max_tokens: 1500, temperature: 0.0}
#   sim:      {provider: anthropic, model: claude-haiku-4-5,  max_tokens: 1500, temperature: 0.7}
#   judge:    {provider: anthropic, model: claude-haiku-4-5,  max_tokens: 2000, temperature: 0.0}
#   mutator:  {provider: anthropic, model: claude-opus-4-7,   max_tokens: 4000, temperature: 0.5}

# To switch to NVIDIA NIM (more generous free rate limits):
#   provider: openai_compatible
#   model: meta/llama-3.3-70b-instruct
#   base_url: https://integrate.api.nvidia.com/v1
#   api_key_env: NVIDIA_NIM_API_KEY

context_budget:
  total_per_agent: 2000
  max_handoff: 500

voice:
  provider: pipecat       # pipecat | mock | vapi | bland | retell
  pipecat:
    stt: deepgram         # deepgram | assemblyai | google
    tts: rime             # rime | cartesia | elevenlabs | inworld
    telephony: twilio

learning:
  optimizer: gepa         # gepa | custom
  promotion_method: hierarchical_bootstrap   # hierarchical_bootstrap | flat_bootstrap | paired_t | bayesian_sequential
  eval:
    personas: [cooperative, combative, evasive, confused, distressed]
    conversations_per_persona: 6
    seeds_per_conversation: 3
    bootstrap_resamples: 10000
    ci_level: 0.95
    min_per_persona_delta: -0.05
  composite_weights:
    resolution: 0.4
    compliance: 0.4
    continuity: 0.1
    naturalness: 0.1
  rubrics: [strict, lenient, continuity]
  inter_rubric_correlation_threshold: 0.7

compliance:
  mode: hybrid            # hybrid | llm_only | rules_only
  rule_set_version: v1
  hard_veto: true

policy:
  active_version: v1