docs: warn against setting OLLAMA_INTENT_MODEL larger than the chat model

A deployment had OLLAMA_INTENT_MODEL=qwen2.5:7b while the chat model was a 4b,
so every auxiliary call (intent judge, tool router, place extraction, query
decomposition) ran on the bigger, slower model and added latency to each
command. Make the .env.example comment state the invariant explicitly.

.env.example

@@ -59,11 +59,15 @@ OLLAMA_BASE_URL=http://127.0.0.1:11434
 # free-form chit-chat. Swap back to qwen3:8b for the strongest tool-calling.
 OLLAMA_CHAT_MODEL=qwen2.5:3b
 # Model for the auxiliary small-model calls: intent judge, tool router, weather
-# place extraction, query decomposition. BLANK (default) reuses OLLAMA_CHAT_MODEL
-# so the stack runs on one already-warm model. The code's built-in default
-# (gemma4:e2b) is NOT pulled by this stack, so leaving this unset previously made
-# every router/extractor call silently fail. Only set this if you also pull the
-# model into Ollama.
+# place extraction, query decomposition. These are classification/JSON calls,
+# NOT the spoken answer, so a small fast model here cuts 2-3 big-model round
+# trips per command without touching answer quality. BLANK uses the stack
+# default qwen2.5:3b, which ollama-init pulls automatically. Set it equal to
+# OLLAMA_CHAT_MODEL to run everything on one resident model instead (saves VRAM
+# at the cost of slower routing when the chat model is large).
+# NEVER set this LARGER than OLLAMA_CHAT_MODEL: the auxiliary calls would then
+# run on the bigger, slower model and add latency to every command (the exact
+# opposite of the split's purpose). Keep it <= the chat model, blank, or equal.
 OLLAMA_INTENT_MODEL=
 OLLAMA_EMBED_MODEL=nomic-embed-text
 WHISPER_MODEL=small
@@ -226,6 +230,7 @@ COMPOSE_FILE=docker-compose.yml:docker-compose.gpu-linux.yml
 # OLLAMA_CHAT_MODEL=qwen2.5:7b   # quality (needs ~5GB VRAM + whisper small)
 # OLLAMA_CHAT_MODEL=qwen2.5:3b   # speed (fits easily, faster on 8GB GPUs)
 # WHISPER_MODEL=small            # small frees VRAM for a bigger LLM; medium=more accurate
+# STT_BEAM_SIZE=5                # beam search (5) > greedy (1) for accuracy; lower for speed
 # MELO_DEVICE=cuda               # cpu if no GPU on the bot host
 
 # --- Settings web UI (http://localhost:8765/settings on the bot host) ---

README.md

@@ -87,7 +87,7 @@ COMPOSE_FILE=docker-compose.yml;docker-compose.gpu-windows.yml
 > Linux와 Windows는 GPU를 컨테이너에 넣는 방식이 달라서 override 파일이 갈립니다. Linux는 CDI(`devices: nvidia.com/gpu=all`), Windows(Docker Desktop)는 Compose의 `deploy.resources.reservations.devices`(`driver: nvidia`)를 씁니다. 호스트 사전 준비는 아래 "GPU 가속" 절 참고.
 
 `docker compose up` 한 번이면 자동으로:
-- Ollama 서버가 뜨고, `ollama-init`이 채팅/임베딩 모델을 **자동 pull**
+- Ollama 서버가 뜨고, `ollama-init`이 채팅/보조(의도·라우팅)/임베딩 모델을 **자동 pull** (보조 모델 `OLLAMA_INTENT_MODEL`은 기본 `qwen2.5:3b`로, 큰 채팅 모델은 답변에만 쓰고 내부 분류 호출은 이 작은 모델이 처리)
 - VNC+XFCE 데스크톱 + 크롬 + Python 브릿지가 기동
 - Whisper STT 모델 / Piper TTS 음성 자동 다운로드(볼륨에 캐시)
 

agents/llm.md

@@ -1,7 +1,7 @@
 # 자비스 운영자 지시
 
 - 너의 이름은 자비스다.
-- 모든 답변은 음성(TTS)으로 읽혀 나간다. 그러니 최대한 간결하게, 한두 문장으로 답한다. 목록, 마크다운, 이모지, 그리고 소리 내어 읽기 어려운 특수문자는 쓰지 않는다.
+- 모든 답변은 음성(TTS)으로 읽혀 나간다. 그러니 무조건 한 문장으로만 답한다. 두 문장 이상 쓰지 않는다. 목록, 마크다운, 이모지, 그리고 소리 내어 읽기 어려운 특수문자는 쓰지 않는다.
 - 정해진 문구에만 반응하지 말고, 실제 사람처럼 말의 뉘앙스와 맥락으로 의도를 알아듣고 처리한다.
 
 화면 속 크롬(방송 화면)에서 유튜브를 다룰 때 (화면에 보여야 하므로 반드시 on-screen 브라우저 제어 도구로 수행한다):

bridge/server.py

@@ -87,6 +87,17 @@ VAD_MIN_SPEECH_MS = int(os.environ.get("VAD_MIN_SPEECH_MS", "200"))
 # Korean phrase decoded as Chinese) and shaves a little latency. Empty = auto.
 STT_LANGUAGE = os.environ.get("STT_LANGUAGE", "ko").strip() or None
 
+# Whisper decoding accuracy knobs. beam_size=1 is greedy decoding — fast but the
+# least accurate; beam search (5 is the Whisper default) explores alternatives
+# and noticeably improves recognition on short, accented, or noisy Discord-mic
+# speech. condition_on_previous_text=False stops Whisper from feeding a previous
+# clip's transcript back in as a prompt, which on isolated short utterances
+# causes repetition loops and drift rather than helping. Both are env-tunable so
+# accuracy/latency can be traded without a code change (lower STT_BEAM_SIZE for
+# speed, raise it for accuracy).
+STT_BEAM_SIZE = max(1, int(os.environ.get("STT_BEAM_SIZE", "5")))
+STT_CONDITION_ON_PREV = os.environ.get("STT_CONDITION_ON_PREV", "0") in ("1", "true", "True", "yes", "on")
+
 # TTS engine: "edge" (Microsoft Edge TTS, natural Korean neural voice) is the
 # primary voice. "melo" (a warm MeloTTS worker) and "piper" remain selectable.
 def _tts_engine_setting() -> str:
@@ -243,7 +254,12 @@ def transcribe(wav_bytes: bytes) -> dict:
         print("[bridge] no speech detected (VAD) — skipping STT", flush=True)
         return {"text": "", "language": None, "note": "음성 아님(VAD 차단)"}
 
-    segments, info = _whisper.transcribe(audio, beam_size=1, language=STT_LANGUAGE)
+    segments, info = _whisper.transcribe(
+        audio,
+        beam_size=STT_BEAM_SIZE,
+        language=STT_LANGUAGE,
+        condition_on_previous_text=STT_CONDITION_ON_PREV,
+    )
     # Second line of defence: drop non-speech / hallucinated segments by
     # Whisper's own no_speech_prob. The no_speech_prob hard cutoff (plus the VAD
     # pre-gate above) is what rejects noise/hallucinations. The avg_logprob

docker-compose.yml

@@ -40,6 +40,9 @@ services:
     environment:
       OLLAMA_HOST: http://ollama:11434
       CHAT_MODEL: ${OLLAMA_CHAT_MODEL:-qwen2.5:3b}
+      # Small auxiliary model for intent/router/extraction calls (see javis
+      # service). Pulled here so the split is ready out of the box.
+      INTENT_MODEL: ${OLLAMA_INTENT_MODEL:-qwen2.5:3b}
       EMBED_MODEL: ${OLLAMA_EMBED_MODEL:-nomic-embed-text}
     entrypoint: ["/bin/sh", "-c"]
     command:
@@ -48,6 +51,10 @@ services:
         until ollama list >/dev/null 2>&1; do sleep 2; done;
         echo "[ollama-init] pulling $$CHAT_MODEL";
         ollama pull "$$CHAT_MODEL";
+        if [ -n "$$INTENT_MODEL" ] && [ "$$INTENT_MODEL" != "$$CHAT_MODEL" ]; then
+          echo "[ollama-init] pulling $$INTENT_MODEL (auxiliary intent/router model)";
+          ollama pull "$$INTENT_MODEL";
+        fi;
         echo "[ollama-init] pulling $$EMBED_MODEL";
         ollama pull "$$EMBED_MODEL";
         echo "[ollama-init] models ready.";
@@ -62,6 +69,14 @@ services:
       # Point the brain at the ollama service and the bot at the in-container bridge.
       OLLAMA_BASE_URL: http://ollama:11434
       OLLAMA_CHAT_MODEL: ${OLLAMA_CHAT_MODEL:-qwen2.5:3b}
+      # Auxiliary small-model calls (intent judge, tool router, arg extraction,
+      # query decomposition) run on this fast model so the big chat model only
+      # runs for the actual spoken answer. With the GPU on, this is the main
+      # per-turn latency win: a command no longer pays the big model's cost 2-3
+      # times for routing/extraction. Defaults to qwen2.5:3b (the project's
+      # reference small model, clean Korean on classification); set it equal to
+      # OLLAMA_CHAT_MODEL to fold everything back onto one resident model.
+      OLLAMA_INTENT_MODEL: ${OLLAMA_INTENT_MODEL:-qwen2.5:3b}
       OLLAMA_EMBED_MODEL: ${OLLAMA_EMBED_MODEL:-nomic-embed-text}
       WHISPER_MODEL: ${WHISPER_MODEL:-medium}
       WHISPER_DEVICE: ${WHISPER_DEVICE:-cuda}
@@ -82,6 +97,9 @@ services:
       PLANNER_ENABLED: ${PLANNER_ENABLED:-0}
       # Lock STT to Korean (skip Whisper auto-detect).
       STT_LANGUAGE: ${STT_LANGUAGE:-ko}
+      # Whisper decode accuracy: beam search (5) over greedy (1) lifts recognition
+      # on short/noisy Discord speech. Lower to 1 for minimum latency.
+      STT_BEAM_SIZE: ${STT_BEAM_SIZE:-5}
       VOICE_SILENCE_MS: ${VOICE_SILENCE_MS:-600}
       BRIDGE_URL: http://127.0.0.1:8765
       # Split-deployment role: full (default, all-in-one), browser (only the

docker/entrypoint.sh

@@ -10,12 +10,12 @@ set -euo pipefail
 : "${OLLAMA_BASE_URL:=http://ollama:11434}"
 : "${OLLAMA_CHAT_MODEL:=qwen3:8b}"
 # Auxiliary small-model calls (intent judge, tool router, weather place
-# extraction, query decomposition). The code default is gemma4:e2b, which this
-# stack does not pull, so those calls would silently fail and fall open —
-# crippling tool routing and arg extraction. Reuse the (already warm) chat model
-# by default so everything runs on one resident model; override if you pull a
-# dedicated small model.
-: "${OLLAMA_INTENT_MODEL:=${OLLAMA_CHAT_MODEL}}"
+# extraction, query decomposition). Default to a small fast model so the big
+# chat model only runs for the actual spoken answer — the main per-turn latency
+# win once the GPU is in use, since the 2-3 routing/extraction calls per command
+# no longer pay the big model's cost. ollama-init pulls this model. Set it equal
+# to OLLAMA_CHAT_MODEL to fold everything back onto one resident model.
+: "${OLLAMA_INTENT_MODEL:=qwen2.5:3b}"
 # Cap chat-model output tokens per turn (worst-case latency guard). Spoken
 # answers are 1-2 sentences; 512 is safe headroom above tool-call JSON. 0 = off.
 : "${OLLAMA_NUM_PREDICT:=512}"

docs/llm_contexts.md

@@ -19,14 +19,14 @@ Every distinct LLM call in Jarvis, what feeds it, what consumes it, and how it i
   - Time + location context (re-injected each turn)
   - Tool schema: native via `generate_tools_json_schema()` ([src/jarvis/tools/registry.py](src/jarvis/tools/registry.py)) or text fallback via `_text_tool_call_guidance()` ([engine.py:68](src/jarvis/reply/engine.py:68))
   - Tool results from prior turns (raw or digested — see #5)
-- **Output**: OpenAI-style `{content, tool_calls, thinking}`. Consumed by the tool orchestrator and TTS pipeline. Natural-language content is delivered immediately; no post-turn evaluator runs.
+- **Output**: OpenAI-style `{content, tool_calls, thinking}`. Consumed by the tool orchestrator and TTS pipeline. Natural-language content is delivered immediately; no post-turn evaluator runs. Spoken-answer length: the persona (`system_prompt.py`) and `voice_style` (`prompts/system.py`) both constrain the reply to a SINGLE sentence — any dry aside must fold into that one sentence as a trailing clause, never a second sentence. This keeps TTS latency down (synth time scales with text length) and matches the `agents/llm.md` operator instruction.
 - **Limits**: `num_ctx: 8192` (explicit). Output `num_predict: cfg.ollama_num_predict` (default 512, `0`/negative disables) caps generated tokens per turn — a worst-case latency guard for short spoken answers; the headroom stays above tool-call JSON so it does not truncate tool calls (both native and text tool-call paths). Timeout `llm_chat_timeout_sec` (45s). Auto-fallback from native to text tool-calls on HTTP 400 (`ToolsNotSupportedError`), sticky for the session. Risk: `fetch_web_page` truncates at 50,000 chars (~37k tokens) — mitigated for SMALL models by tool-result digest (#5) which compresses the payload before it enters the messages history. LARGE models receive the raw payload and may silently see a truncated context.
 
 ## 2. Intent Judge
 
 - **File**: [src/jarvis/listening/intent_judge.py](src/jarvis/listening/intent_judge.py) — `IntentJudge.evaluate()`.
 - **Trigger**: on a speech segment *only if* there is an engagement signal (wake word detected, hot-window active, or TTS playing). Pure ambient speech skips it.
-- **Model / gating**: `cfg.intent_judge_model` (default `gemma4:e2b`, ~2B). Falls back to text-based wake detection if Ollama is unavailable.
+- **Model / gating**: `cfg.intent_judge_model`. Code-level default `gemma4:e2b` (~2B); the **Docker stack** renders it from `OLLAMA_INTENT_MODEL` (default `qwen2.5:3b`, pulled by `ollama-init`), kept deliberately **separate from `ollama_chat_model`** so this judge and the tool router (#3, #7) run on a small fast model while the big chat model is reserved for the spoken answer. Setting `OLLAMA_INTENT_MODEL = OLLAMA_CHAT_MODEL` folds them back onto one resident model. Falls back to text-based wake detection if Ollama is unavailable.
 - **Inputs**:
   - Rolling transcript buffer (last 120s, with timestamps)
   - Wake-word timestamp (if any), normalised aliases
@@ -246,7 +246,7 @@ user input
 3. Pre-warm the intent-judge model before TTS finishes.
 4. Cache tool-router (#7) output by query hash.
 5. Give each digest its own timeout budget rather than sharing `llm_digest_timeout_sec` (today a slow memory digest can starve the max-turn digest).
-6. Consider single-model deployments: router+planner prefer `intent_judge_model`; loading a second model hurts cold-start latency on small hardware.
+6. Two-model vs single-model tradeoff: the Docker default keeps a **separate** small `intent_judge_model` (`OLLAMA_INTENT_MODEL=qwen2.5:3b`) so routing/judging/extraction don't pay the big chat model's per-call cost — the main win once the GPU holds both models resident. On VRAM-constrained hardware, fold them onto one model by setting `OLLAMA_INTENT_MODEL = OLLAMA_CHAT_MODEL` (saves a resident model at the cost of slower routing when the chat model is large).
 7. Narrow `llm_thinking_enabled` to router/planner only, not every context.
 8. Reduce `intent_judge_timeout_sec` (15s) or race it against text-based wake detection to avoid blocking the audio loop.
 

src/jarvis/reply/prompts/prompts.spec.md

@@ -43,7 +43,7 @@ from jarvis.reply.prompts import (
 Both model sizes share these base components:
 - `asr_note`: Voice transcription error handling
 - `inference_guidance`: Prefer inference over clarification
-- `voice_style`: Concise, conversational responses
+- `voice_style`: Single-sentence, conversational responses (spoken aloud, so one sentence only — never more)
 
 Model-size-specific components:
 - `tool_incentives`: When/how aggressively to use tools

src/jarvis/reply/prompts/system.py

@@ -26,8 +26,8 @@ INFERENCE_GUIDANCE = (
 # Voice assistant communication style - concise, conversational
 VOICE_STYLE = (
     "Keep responses concise and conversational since this is a voice assistant. "
-    "Two to three sentences maximum. Prioritize clarity and brevity - users are listening, not reading. "
-    "Avoid unnecessary elaboration unless specifically requested. "
+    "Reply in a SINGLE sentence - never more than one sentence. Prioritize clarity and brevity - users are listening, not reading. "
+    "Avoid unnecessary elaboration. "
     "Do NOT offer follow-up suggestions or ask if the user wants more info - just respond directly. "
     "IMPORTANT: Always respond in natural language - never output JSON, code, or structured data as your response. "
     "NEVER use markdown formatting in your replies: no asterisks for emphasis (**bold**, *italic*), "

src/jarvis/system_prompt.py

@@ -62,10 +62,11 @@ _SYSTEM_PROMPT_TEMPLATE: str = (
     "Tone rails (hard): never mean, never condescending, never passive-aggressive, never "
     "sulking, never preachy, never sycophantic ('great question', 'I'd be happy to'). "
     "Sarcasm points at the situation, the topic, or mildly at yourself — never at the user. "
-    "Shape for casual, factual, or small-talk replies: state the answer in a sentence, then add "
-    "one short dry observation about it (an understated aside, a raised-eyebrow remark, a gentle "
-    "noticing of the irony). One aside — not two, not a joke opener, not a joke-shaped sentence "
-    "replacing the answer. The aside is a tail, not the head. "
+    "Shape for casual, factual, or small-talk replies: give the answer in a SINGLE sentence. If a "
+    "dry aside fits, fold it into that same sentence as a short trailing clause — never add it as "
+    "a second sentence, never a joke opener, never a joke-shaped sentence replacing the answer. "
+    "Whenever the wit would require a second sentence, drop the wit and keep the one-sentence "
+    "answer. The aside is a tail inside the sentence, not a head and not a new sentence. "
     "Examples of the MOVE (shape, not wording — never copy these): stating a fact and then noting "
     "its mild absurdity; giving the weather and then commenting on what it implies for the day; "
     "answering a trivia question and then offering a wry footnote about the subject; admitting "
@@ -79,8 +80,8 @@ _SYSTEM_PROMPT_TEMPLATE: str = (
     "butler clichés, and never address the user as 'sir', 'madam', 'my liege', or similar. "
     "Never stack multiple jokes in one reply. "
     "Be concise, conversational, and actionable. "
-    "This is a spoken voice assistant: answer in ONE short sentence whenever possible "
-    "(two at the very most). No lists, no preamble, no 'is there anything else' offers. "
+    "This is a spoken voice assistant: your ENTIRE reply must be a single short sentence. "
+    "Never write a second sentence. No lists, no preamble, no 'is there anything else' offers. "
     "When a controlBrowser tool is available, use IT (never webSearch) for anything that "
     "should happen in the on-screen browser — opening a site, searching on a site "
     "(controlBrowser action 'search' with the right site), clicking, typing — because only "

tests/test_intent_model_split.py

@@ -0,0 +1,62 @@
+"""The docker deployment must run auxiliary calls on a small model.
+
+Latency win: intent judging, tool routing and arg extraction are
+classification/JSON calls, not the spoken answer. Running them on a small fast
+model means the big chat model only runs once per command (for the answer),
+instead of 2-3 times per command for routing/extraction too.
+
+The wiring is: docker/jarvis-config.template.json renders `intent_judge_model`
+from `${OLLAMA_INTENT_MODEL}`, and the reply engine's resolver falls through
+`tool_router_model -> intent_judge_model -> ollama_chat_model`. The template
+sets no `tool_router_model`, so the auxiliary model is whatever
+`OLLAMA_INTENT_MODEL` renders to. These tests pin that behaviour end to end.
+"""
+
+import json
+import string
+from pathlib import Path
+
+import pytest
+
+from jarvis.reply.engine import resolve_tool_router_model
+
+TEMPLATE = Path(__file__).resolve().parent.parent / "docker" / "jarvis-config.template.json"
+
+
+def _render(**env) -> dict:
+    raw = TEMPLATE.read_text(encoding="utf-8")
+    return json.loads(string.Template(raw).safe_substitute(**env))
+
+
+class _Cfg:
+    """cfg stand-in carrying only the fields the resolver reads. The template
+    does not render `tool_router_model`, so it stays empty here too."""
+
+    def __init__(self, rendered: dict):
+        self.tool_router_model = rendered.get("tool_router_model", "") or ""
+        self.intent_judge_model = rendered.get("intent_judge_model", "") or ""
+        self.ollama_chat_model = rendered.get("ollama_chat_model", "") or ""
+
+
+def test_template_renders_separate_intent_model():
+    cfg = _render(OLLAMA_CHAT_MODEL="qwen3:8b", OLLAMA_INTENT_MODEL="qwen2.5:3b")
+    assert cfg["ollama_chat_model"] == "qwen3:8b"
+    assert cfg["intent_judge_model"] == "qwen2.5:3b"
+    assert cfg["intent_judge_model"] != cfg["ollama_chat_model"]
+
+
+@pytest.mark.unit
+def test_aux_calls_route_to_small_model_not_chat_model():
+    """The whole point: with a big chat model and a small intent model, tool
+    routing must resolve to the small model, leaving the big model for answers."""
+    cfg = _Cfg(_render(OLLAMA_CHAT_MODEL="qwen3:8b", OLLAMA_INTENT_MODEL="qwen2.5:3b"))
+    assert resolve_tool_router_model(cfg) == "qwen2.5:3b"
+
+
+@pytest.mark.unit
+def test_folding_intent_onto_chat_model_keeps_one_model():
+    """Setting OLLAMA_INTENT_MODEL == OLLAMA_CHAT_MODEL folds everything back
+    onto a single resident model (the documented VRAM-saving opt-out)."""
+    cfg = _Cfg(_render(OLLAMA_CHAT_MODEL="qwen2.5:3b", OLLAMA_INTENT_MODEL="qwen2.5:3b"))
+    assert resolve_tool_router_model(cfg) == "qwen2.5:3b"
+    assert cfg.intent_judge_model == cfg.ollama_chat_model

tests/test_prompts.py

@@ -121,6 +121,18 @@ class TestPromptComponents:
             assert prompts.voice_style, f"{size.value} missing voice_style"
             assert prompts.tool_guidance, f"{size.value} missing tool_guidance"
 
+    def test_voice_style_enforces_single_sentence(self):
+        """voice_style must cap replies at one sentence (spoken aloud). The old
+        'Two to three sentences maximum' wording let the model run long, which
+        also slowed TTS since synth time scales with text length."""
+        from jarvis.reply.prompts import get_system_prompts, ModelSize
+
+        for size in [ModelSize.SMALL, ModelSize.LARGE]:
+            voice_style = get_system_prompts(size).voice_style
+            assert "SINGLE sentence" in voice_style, f"{size.value} voice_style not single-sentence"
+            assert "never more than one sentence" in voice_style
+            assert "Two to three" not in voice_style
+
     def test_to_list_returns_non_empty_strings(self):
         """to_list() returns only non-empty prompt strings."""
         from jarvis.reply.prompts import get_system_prompts, ModelSize

tests/test_system_prompt.py

@@ -44,6 +44,22 @@ class TestBuildSystemPrompt:
         assert "in the user's language" not in prompt
         assert "in Korean" in prompt
 
+    def test_persona_enforces_single_sentence(self):
+        # Spoken replies must be one sentence (TTS latency scales with text
+        # length, and the user asked for one-sentence answers). The persona must
+        # state the single-sentence rule and must NOT carry the old "two at the
+        # very most" allowance that let the model run long.
+        prompt = build_system_prompt("Jarvis")
+        assert "single short sentence" in prompt
+        assert "Never write a second sentence" in prompt
+        assert "two at the very most" not in prompt
+
+    def test_persona_aside_does_not_authorise_a_second_sentence(self):
+        # The dry aside must fold into the one sentence, not become a 2nd one.
+        prompt = build_system_prompt("Jarvis")
+        assert "SINGLE sentence" in prompt
+        assert "never add it as " in prompt
+
 
 class TestOutputLanguageDirective:
     """A deployment may lock replies to a single language via OUTPUT_LANGUAGE.