Improve Dream QA Modal feedback latency

app.py

@@ -15,7 +15,7 @@ from dream_customs.ui.app import build_demo
 
 LOCAL_GRADIO_PORT = 7862
 HF_SPACE_GRADIO_PORT = 7860
-BROWSER_ASR_TIMEOUT_SECONDS = 150.0
+BROWSER_ASR_TIMEOUT_SECONDS = 20.0
 
 
 def _default_server_port() -> int:
@@ -32,7 +32,7 @@ def _browser_asr_timeout_seconds() -> float:
     if not value:
         return BROWSER_ASR_TIMEOUT_SECONDS
     try:
-        return max(10.0, float(value))
+        return max(5.0, float(value))
     except ValueError:
         return BROWSER_ASR_TIMEOUT_SECONDS
 

dream_customs/app_logic.py

@@ -39,7 +39,7 @@ DEFAULT_HOSTED_TIMEOUT_SECONDS = 60.0
 DEFAULT_ASR_TIMEOUT_SECONDS = 45.0
 DEFAULT_TEXT_TEMPERATURE = 0.0
 DEFAULT_VISION_TEMPERATURE = 0.1
-DEFAULT_TEXT_MAX_TOKENS = 780
+DEFAULT_TEXT_MAX_TOKENS = 560
 DEFAULT_VISION_MAX_TOKENS = 320
 DEFAULT_TEXT_LATENCY_BUDGET_MS = 3500
 DEFAULT_VISION_LATENCY_BUDGET_MS = 6500

dream_customs/models.py

@@ -362,7 +362,7 @@ class OllamaTextClient:
         parsed = self._generate_json(
             prompt,
             '{"visitor_name":"string","questions":["string","string"],"tone_note":"string"}',
-            num_predict=320,
+            num_predict=240,
         )
         if not parsed:
             return self.fallback.generate_negotiation(prompt)
@@ -383,7 +383,7 @@ class OllamaTextClient:
                 '"risk_level":"string","alliance_reading":"string","practical_suggestion":"string",'
                 '"weird_task":"string","bedtime_release":"string","safety_note":"string"}'
             ),
-            num_predict=700,
+            num_predict=560,
         )
         if not parsed:
             return self.fallback.generate_pact(prompt)
@@ -410,7 +410,7 @@ class OllamaTextClient:
                 '"followup_questions":["string"],"user_answers":["string"],"interpretation":"string",'
                 '"today_tip":"string","tiny_action":"string","caring_note":"string","safety_note":"string"}'
             ),
-            num_predict=780,
+            num_predict=560,
         )
         if not parsed:
             return self.fallback.generate_today_tip(prompt)
@@ -426,7 +426,7 @@ class OllamaTextClient:
                 '"today_bridge":"string","visual_evidence":["string"],'
                 '"safety_flags":["string"],"language":"en"}'
             ),
-            num_predict=520,
+            num_predict=360,
         )
         if not parsed:
             return self.fallback.generate_brief(prompt)
@@ -627,6 +627,7 @@ class HostedMiniCPMTextClient:
         timeout: float = 60.0,
         temperature: float = 0.0,
         max_tokens: int = 780,
+        latency_budget_ms: int = 0,
         fallback: Optional[FakeTextClient] = None,
     ):
         self.endpoint = endpoint.strip()
@@ -634,8 +635,15 @@ class HostedMiniCPMTextClient:
         self.timeout = timeout
         self.temperature = max(0.0, min(float(temperature), 0.7))
         self.max_tokens = max(64, min(int(max_tokens), 1200))
+        self.latency_budget_ms = max(0, int(latency_budget_ms or 0))
         self.fallback = fallback or FakeTextClient()
 
+    @property
+    def request_timeout(self) -> float:
+        if self.latency_budget_ms <= 0:
+            return self.timeout
+        return max(1.0, min(float(self.timeout), self.latency_budget_ms / 1000.0))
+
     def _post_json(self, prompt: str, max_tokens: int = 700) -> Optional[Dict[str, Any]]:
         if not self.endpoint:
             return None
@@ -656,7 +664,7 @@ class HostedMiniCPMTextClient:
             method="POST",
         )
         try:
-            with urllib.request.urlopen(request, timeout=self.timeout) as response:
+            with urllib.request.urlopen(request, timeout=self.request_timeout) as response:
                 return json.loads(response.read().decode("utf-8"))
         except (OSError, TimeoutError, urllib.error.URLError, json.JSONDecodeError):
             return None
@@ -676,7 +684,7 @@ class HostedMiniCPMTextClient:
         parsed = self._generate_json(
             prompt,
             '{"visitor_name":"string","questions":["string"],"tone_note":"string"}',
-            max_tokens=360,
+            max_tokens=240,
         )
         if not parsed:
             return self.fallback.generate_negotiation(prompt)
@@ -697,7 +705,7 @@ class HostedMiniCPMTextClient:
                 '"risk_level":"string","alliance_reading":"string","practical_suggestion":"string",'
                 '"weird_task":"string","bedtime_release":"string","safety_note":"string"}'
             ),
-            max_tokens=780,
+            max_tokens=560,
         )
         if not parsed:
             return self.fallback.generate_pact(prompt)
@@ -724,7 +732,7 @@ class HostedMiniCPMTextClient:
                 '"followup_questions":["string"],"user_answers":["string"],"interpretation":"string",'
                 '"today_tip":"string","tiny_action":"string","caring_note":"string","safety_note":"string"}'
             ),
-            max_tokens=780,
+            max_tokens=560,
         )
         if not parsed:
             return self.fallback.generate_today_tip(prompt)
@@ -740,7 +748,7 @@ class HostedMiniCPMTextClient:
                 '"today_bridge":"string","visual_evidence":["string"],'
                 '"safety_flags":["string"],"language":"en"}'
             ),
-            max_tokens=520,
+            max_tokens=360,
         )
         if not parsed:
             return self.fallback.generate_brief(prompt)
@@ -777,6 +785,7 @@ class HostedMiniCPMVisionClient:
         timeout: float = 60.0,
         temperature: float = 0.1,
         max_tokens: int = 320,
+        latency_budget_ms: int = 0,
         fallback: Optional[FakeVisionClient] = None,
     ):
         self.endpoint = endpoint.strip()
@@ -784,8 +793,15 @@ class HostedMiniCPMVisionClient:
         self.timeout = timeout
         self.temperature = max(0.0, min(float(temperature), 0.7))
         self.max_tokens = max(64, min(int(max_tokens), 800))
+        self.latency_budget_ms = max(0, int(latency_budget_ms or 0))
         self.fallback = fallback or FakeVisionClient()
 
+    @property
+    def request_timeout(self) -> float:
+        if self.latency_budget_ms <= 0:
+            return self.timeout
+        return max(1.0, min(float(self.timeout), self.latency_budget_ms / 1000.0))
+
     def _post_image(self, image_path: str, prompt: Optional[str] = None) -> Optional[Dict[str, Any]]:
         if not self.endpoint:
             return None
@@ -811,7 +827,7 @@ class HostedMiniCPMVisionClient:
             method="POST",
         )
         try:
-            with urllib.request.urlopen(request, timeout=self.timeout) as response:
+            with urllib.request.urlopen(request, timeout=self.request_timeout) as response:
                 return json.loads(response.read().decode("utf-8"))
         except (OSError, TimeoutError, urllib.error.URLError, json.JSONDecodeError):
             return None
@@ -876,16 +892,24 @@ class HostedASRClient:
         endpoint: str = "",
         token: str = "",
         timeout: float = 45.0,
+        latency_budget_ms: int = 0,
         fallback: Optional[FakeASRClient] = None,
         fallback_enabled: bool = True,
     ):
         self.endpoint = endpoint.strip()
         self.token = token.strip()
         self.timeout = timeout
+        self.latency_budget_ms = max(0, int(latency_budget_ms or 0))
         self.fallback = fallback or FakeASRClient()
         self.fallback_enabled = fallback_enabled
         self.last_error = ""
 
+    @property
+    def request_timeout(self) -> float:
+        if self.latency_budget_ms <= 0:
+            return self.timeout
+        return max(1.0, min(float(self.timeout), self.latency_budget_ms / 1000.0))
+
     def _fallback_transcript(self, audio_path: Optional[str]) -> str:
         if not self.fallback_enabled:
             return ""
@@ -918,7 +942,7 @@ class HostedASRClient:
             method="POST",
         )
         try:
-            with urllib.request.urlopen(request, timeout=self.timeout) as response:
+            with urllib.request.urlopen(request, timeout=self.request_timeout) as response:
                 payload = json.loads(response.read().decode("utf-8"))
         except (OSError, TimeoutError, urllib.error.URLError, json.JSONDecodeError) as exc:
             self.last_error = f"{exc.__class__.__name__}: {exc}"

dream_customs/prompts.py

@@ -58,7 +58,7 @@ Write a non-diagnostic interpretation draft, one waking-life Today Tip / 今日
 and one weird little thing / 古怪的小事.
 First answer the user's stated question directly. If the user sounds scared, sad,
 overwhelmed, guilty, lonely, or asks for comfort, follow that emotion before giving any action.
-The interpretation must be step-by-step: use 2 to 4 short layers that move from
+The interpretation must be compact: use 1 to 2 short layers that move from
 the user's feeling, to concrete dream anchors, to the follow-up answers, to one gentle way to care for today.
 Do not collapse every dream into productivity advice such as opening a task,
 writing a first line, or making the first step smaller.
@@ -79,7 +79,8 @@ or a reason to seek professional help unless the user explicitly reports severe
 self-harm, harm to others, or inability to function.
 Do not infer work stress, fear, loneliness, anxiety, guilt, or self-blame unless the user actually says that feeling.
 Avoid generic wellness filler such as warm water, relaxing music, positive mindset, or "tomorrow will be better".
-Keep the whole result short, warm, emotionally responsive, and specific to the user's answer.
+Keep the whole JSON compact: each user-facing field should be one short sentence when possible.
+Keep the whole result warm, emotionally responsive, and specific to the user's answer.
 The weird little thing must be harmless, legal, low-cost, non-embarrassing, and not a command to solve the whole problem.
 Avoid demanding phrases such as "immediately", "must", "fix it", or "solve it".
 If the user asks for comfort, caring_note should be warm, specific, and validating.

dream_customs/ui/app.py

@@ -800,8 +800,8 @@ def _agent_dream_qa(dream_text: str, mood: str = "", answer: str = "", language:
         image_value=None,
         audio_value=None,
         mood=mood,
-        text_backend=DEFAULT_TEXT_BACKEND,
-        vision_backend=DEFAULT_VISION_BACKEND,
+        text_backend="demo",
+        vision_backend="demo",
         language=language,
     )
     view = json.loads(view_json)
@@ -919,7 +919,7 @@ def _mic_html(language: str = DEFAULT_LANGUAGE) -> str:
     data-empty="{escape(copy['mic_empty'])}"
     data-error="{escape(copy['mic_error'])}"
     data-timeout="{escape(copy['mic_timeout'])}"
-    data-timeout-ms="140000"
+    data-timeout-ms="20000"
   >
     <span class="dc-mic-glyph" aria-hidden="true"></span>
   </button>

modal_backend/contracts.py

@@ -24,11 +24,11 @@ def normalize_text_payload(payload: Dict[str, Any]) -> Dict[str, Any]:
                     if content:
                         parts.append(content)
             prompt = "\n".join(parts).strip()
-    max_tokens = payload.get("max_tokens", 700)
+    max_tokens = payload.get("max_tokens", 560)
     try:
         max_tokens = int(max_tokens)
     except (TypeError, ValueError):
-        max_tokens = 700
+        max_tokens = 560
     max_tokens = max(64, min(max_tokens, 1200))
     temperature = payload.get("temperature", 0.0)
     try:

tests/test_modal_contract.py

@@ -22,7 +22,7 @@ def test_normalize_text_payload_accepts_openai_style_messages():
         {"messages": [{"role": "user", "content": "Dream of an elevator."}]}
     )
     assert payload["prompt"] == "Dream of an elevator."
-    assert payload["max_tokens"] == 700
+    assert payload["max_tokens"] == 560
 
 
 def test_normalize_text_payload_clamps_max_tokens():

tests/test_ui_actions.py

@@ -71,10 +71,20 @@ def test_browser_voice_uses_same_origin_asr_proxy():
     assert "_browser_asr_client()" in source
     assert "fallback_enabled=False" in source
     assert "asyncio.to_thread(asr_client.transcribe, temp_path)" in source
-    assert "BROWSER_ASR_TIMEOUT_SECONDS = 150.0" in source
+    assert "BROWSER_ASR_TIMEOUT_SECONDS = 20.0" in source
+    assert 'data-timeout-ms="20000"' in ui_app._mic_html("en")
     assert "DREAM_CUSTOMS_HOSTED_TOKEN" not in source
 
 
+def test_agent_api_uses_fast_local_question_before_modal_tip():
+    source = inspect.getsource(ui_app._agent_dream_qa)
+
+    assert 'text_backend="demo"' in source
+    assert 'vision_backend="demo"' in source
+    assert "answer_to_card_action(" in source
+    assert "skip_to_card_action(" in source
+
+
 def test_image_upload_is_composer_plus_drawer():
     source = inspect.getsource(ui_app.build_demo)
 
@@ -322,7 +332,7 @@ def test_mobile_reset_restores_calm_mood():
         60,
         0.2,
         0.1,
-        780,
+        560,
         320,
         "demo",
         "",