Update app.py

app.py

@@ -4,11 +4,11 @@ import csv
 import re
 import base64
 import textwrap
+
 import requests
 import streamlit as st
 from PIL import Image
 from audio_recorder_streamlit import audio_recorder
-from huggingface_hub import InferenceClient
 
 # =========================
 # BASIC CONFIG
@@ -23,33 +23,23 @@ st.set_page_config(
 # --- Keys / endpoints ---
 
 OPENROUTER_API_KEY = os.getenv("OPENROUTER_API_KEY")
-
-# HF_TOKEN is the new standard name; fall back to HF_API_TOKEN if needed
-HF_TOKEN = os.getenv("HF_TOKEN") or os.getenv("HF_API_TOKEN")
-
 if not OPENROUTER_API_KEY:
     st.warning("Set OPENROUTER_API_KEY in your Space secrets (OpenRouter) to enable AI features.")
     st.stop()
 
+HF_TOKEN = os.getenv("HF_TOKEN") or os.getenv("HF_API_TOKEN")  # for Whisper via HF router
+
 OPENROUTER_URL = "https://openrouter.ai/api/v1/chat/completions"
 
 # IMPORTANT: set these to valid OpenRouter model slugs from your OpenRouter dashboard.
-VISION_MODEL = os.getenv("VISION_MODEL", "nvidia/nemotron-nano-12b-v2-vl:free")        # example; update to real slug if different
-REASONING_MODEL = os.getenv("REASONING_MODEL", "nvidia/nemotron-nano-12b-v2-vl:free")   # or another instruct-capable model
-
-# Whisper model to use on Hugging Face Inference
-HF_WHISPER_MODEL = os.getenv("HF_WHISPER_MODEL", "openai/whisper-small")
+VISION_MODEL = os.getenv("VISION_MODEL", "nvidia/nemotron-nano-12b-v2-vl:free")        # example; update to actual slug if needed
+REASONING_MODEL = os.getenv("REASONING_MODEL", "nvidia/nemotron-nano-12b-v2-vl:free")   # or another instruct model
 
-# Create HF Inference client if token available
-hf_asr_client = None
-if HF_TOKEN:
-    try:
-        hf_asr_client = InferenceClient(
-            provider="hf-inference",
-            api_key=HF_TOKEN,
-        )
-    except Exception:
-        hf_asr_client = None  # fail soft; app still runs without ASR
+# Whisper via HF Inference router (remote, no local load)
+HF_WHISPER_URL = os.getenv(
+    "HF_WHISPER_URL",
+    "https://router.huggingface.co/hf-inference/models/openai/whisper-large-v3",
+)
 
 DPD_BASE_URL = "https://health-products.canada.ca/api/drug/drugproduct"
 RECALLS_FEED_URL = os.getenv(
@@ -250,6 +240,11 @@ def load_city_options(path: str = "canadacities.csv"):
 
 
 CITY_OPTIONS = load_city_options()
+if not CITY_OPTIONS:
+    st.warning(
+        "Could not load cities from canadacities.csv. "
+        "Autocomplete list will be empty until that file is available."
+    )
 
 
 def split_city_label(label: str):
@@ -419,7 +414,7 @@ def tool_hqontario_context_city(city_label: str) -> str:
     return ""
 
 # =========================
-# DPD & RECALLS HELPERS
+# DPD & RECALLS
 # =========================
 
 def tool_lookup_drug_products(text: str) -> str:
@@ -505,16 +500,11 @@ def tool_get_wait_times_awareness() -> str:
 # =========================
 
 def call_openrouter_chat(model: str, messages, temperature: float = 0.3):
-    """
-    Helper for OpenRouter's /chat/completions.
-    Returns a readable error message if the response is not valid JSON.
-    """
     headers = {
         "Authorization": f"Bearer {OPENROUTER_API_KEY}",
         "Content-Type": "application/json",
     }
 
-    # Optional attribution if you configure APP_URL + origins in OpenRouter
     app_url = os.getenv("APP_URL", "").strip()
     if app_url:
         headers["HTTP-Referer"] = app_url
@@ -527,35 +517,25 @@ def call_openrouter_chat(model: str, messages, temperature: float = 0.3):
     }
 
     try:
-        r = requests.post(OPENROUTER_URL, headers=headers, json=payload, timeout=60)
-
+        r = requests.post(OPENROUTER_URL, headers=headers, json=payload, timeout=90)
         if r.status_code != 200:
             text_snippet = r.text[:300].replace("\n", " ")
             return f"(Model call error: {r.status_code} — {text_snippet})"
-
-        try:
-            data = r.json()
-        except ValueError:
-            text_snippet = r.text[:300].replace("\n", " ")
-            return f"(Model call error: Non-JSON response — {text_snippet})"
-
+        data = r.json()
         choices = data.get("choices")
-        if not choices or "message" not in choices[0] or "content" not in choices[0]["message"]:
-            return f"(Model call error: Unexpected response format — {data})"
-
-        return choices[0]["message"]["content"].strip()
-
+        if not choices:
+            return "(Model call error: empty response.)"
+        msg = choices[0].get("message", {})
+        content = msg.get("content", "")
+        return content.strip() if isinstance(content, str) else str(content)
     except Exception as e:
         return f"(Model call unavailable: {e})"
 
 # =========================
-# VISION (Nemotron via OpenRouter)
+# VISION SUMMARIZER (Nemotron via OpenRouter)
 # =========================
 
 def call_vision_summarizer(image_bytes: bytes) -> str:
-    """
-    Structured, cautious but detailed visual assessment via vision-capable model on OpenRouter.
-    """
     if not image_bytes:
         return ""
     b64 = base64.b64encode(image_bytes).decode("utf-8")
@@ -598,47 +578,84 @@ Write as if supporting a separate triage system, not directly reassuring or diag
     return call_openrouter_chat(VISION_MODEL, messages, temperature=0.2)
 
 # =========================
-# ASR (Whisper via HF InferenceClient)
+# ASR VIA HF ROUTER (Whisper Large v3)
 # =========================
 
 def call_asr(audio_source) -> str:
     """
-    Uses Hugging Face InferenceClient for Whisper.
-    Returns transcript text on success, or "" on any failure.
-    Never surfaces raw HTTP errors to the user.
+    Use Hugging Face router + openai/whisper-large-v3 for transcription.
+    - No local model load.
+    - Returns transcript or "".
+    - Shows friendly warnings, never raw HTTP/JSON in the text area.
     """
-    if not audio_source or hf_asr_client is None:
+    if not audio_source:
         return ""
 
-    # Normalize to bytes
+    if not HF_TOKEN:
+        st.warning("Voice transcription is not configured. Please type your description.")
+        return ""
+
+    # Normalize to bytes from audio_recorder_streamlit
     if isinstance(audio_source, bytes):
         audio_bytes = audio_source
     else:
         audio_bytes = audio_source.read()
 
+    headers = {
+        "Authorization": f"Bearer {HF_TOKEN}",
+        # audio_recorder_streamlit sends wav; this is fine for Whisper
+        "Content-Type": "audio/wav",
+    }
+
     try:
-        # automatic_speech_recognition accepts bytes directly
-        result = hf_asr_client.automatic_speech_recognition(
-            audio_bytes,
-            model=HF_WHISPER_MODEL,
+        resp = requests.post(
+            HF_WHISPER_URL,
+            headers=headers,
+            data=audio_bytes,
+            timeout=120,
         )
+    except Exception as e:
+        print(f"Whisper HF router request error: {e}")
+        st.warning("Voice transcription is temporarily unavailable. Please type your description.")
+        return ""
 
-        # result can be:
-        # - dict with "text"
-        # - string
-        if isinstance(result, dict):
-            text_val = result.get("text") or result.get("generated_text") or ""
-            return (text_val or "").strip()
-        elif isinstance(result, str):
-            return result.strip()
-        else:
-            return ""
-    except Exception:
-        # Fail silently; caller will prompt user to type instead
+    if resp.status_code != 200:
+        # Log server-side, keep UI clean
+        print(f"Whisper HF router HTTP {resp.status_code}: {resp.text[:300]}")
+        st.warning("We couldn't transcribe that recording. Please type or edit your description.")
+        return ""
+
+    try:
+        data = resp.json()
+    except Exception as e:
+        print(f"Whisper HF router JSON error: {e}, body: {resp.text[:300]}")
+        st.warning("We couldn't transcribe that recording. Please type or edit your description.")
         return ""
 
+    # HF router responses for Whisper typically include "text" or similar.
+    text_val = ""
+
+    if isinstance(data, dict):
+        # Common schema: {"text": "..."} or {"generated_text": "..."} or error
+        if "error" in data:
+            print(f"Whisper HF router error field: {data['error']}")
+        text_val = data.get("text") or data.get("generated_text") or ""
+    elif isinstance(data, list) and data:
+        # Sometimes list of segments or outputs
+        first = data[0]
+        if isinstance(first, dict):
+            text_val = first.get("text") or first.get("generated_text") or ""
+        elif isinstance(first, str):
+            text_val = first
+
+    text_val = (text_val or "").strip()
+
+    if not text_val:
+        st.warning("We couldn't clearly understand that recording. Please check or type your description.")
+    return text_val
+
 # =========================
-# REASONING AGENT (Nemotron via OpenRouter)
+# REASONING AGENT
 # =========================
 
 def call_reasoning_agent(
@@ -674,8 +691,11 @@ HARD SAFETY RULES
 3. Do NOT claim to provide real-time wait times.
 4. Do NOT contradict emergency advice from standard red-flag criteria.
 5. DO NOT invent or guess facility names.
-   - You may ONLY list specific facilities that appear literally in facilities_context or hqontario_context.
-   - If facilities_context reports none found, direct users to Google Maps / official tools without naming specific clinics.
+   - You may ONLY list specific hospitals/clinics/ERs that appear LITERALLY in:
+     - facilities_context, or
+     - hqontario_context.
+   - If facilities_context says no specific local facilities were found, you MUST NOT make up examples.
+     Instead, direct users to Google Maps or official provincial "find a clinic/ER" tools.
 6. Keep the whole answer under about 350 words.
 7. Use clear, neutral, plain language that a non-clinician can understand.
 
@@ -683,90 +703,110 @@ IMAGE USE (VERY IMPORTANT)
 --------------------------
 - If a vision_summary is provided, you MUST treat it as a primary source of information.
 - Cross-check the user narrative against the vision_summary.
-- If the image description suggests more serious concern (e.g., black tissue, deep open wound, marked swelling,
-  spreading redness, thick pus, obvious deformity), escalate the triage tier even if the user's text sounds mild.
-- If there is a conflict between what is described and what is seen, favour caution and explain that visually
-  it appears safer to seek a higher level of assessment.
+- If the image description suggests more serious concern (e.g., black tissue, deep open wound,
+  marked swelling, spreading redness, thick pus, obvious deformity), escalate the triage tier
+  even if the user's text sounds mild.
+- If there is a conflict between what is described and what is seen, favour caution and explain that
+  visually it appears safer to seek a higher level of assessment.
 - You still MUST NOT assign a diagnosis or name specific diseases based on the image.
 
 TRIAGE LOGIC (INTERNAL)
 -----------------------
 You must internally choose EXACTLY ONE primary pathway, but DO NOT show letters or labels to the user.
 
-(1) "No current action needed (you appear well, continue normal care)"
-- Use when the user clearly reports feeling well or "perfect", with no symptoms or concerns.
-- Provide light wellness reminders.
-- Phrase as "based on what you've shared, you appear well." This is NOT a diagnosis.
-
-(2) "Home care reasonable (self-care + monitor)"
-- For mild, short-lived, non-progressive symptoms that do not limit normal activities.
-- Provide simple, safe self-care steps.
-- Clearly state when to escalate to pharmacist/clinic/Telehealth.
-
-(3) "Pharmacist / Walk-in clinic / Family doctor or Telehealth recommended"
-- For persistent, recurrent, or moderate symptoms, or diagnostic uncertainty without clear emergency.
-- Explain why an in-person or phone assessment is appropriate.
-
-(4) "Go to Emergency / Call 911 now"
-- For clear red-flag features:
-  - severe or rapidly worsening pain,
-  - major trauma,
-  - uncontrolled bleeding,
-  - chest pain or trouble breathing,
-  - stroke-like symptoms,
-  - signs of sepsis or systemic illness,
-  - significant numbness/weakness,
-  - concerning post-surgical changes,
-  - or visually severe findings from the image.
-- Be direct and unambiguous.
+(1) "No current action needed (you appear well, continue normal care)":
+    - Use when the user clearly reports feeling well or "perfect", with no symptoms or concerns.
+    - You may gently remind them of general wellness habits.
+    - Phrase as "based on what you've shared, you appear well"; this is NOT a diagnosis.
+
+(2) "Home care reasonable (self-care + monitor)":
+    - Use when symptoms are:
+        - mild,
+        - short-lived,
+        - non-progressive,
+        - and NOT limiting normal activities.
+    - Provide clear, safe self-care suggestions and "watch for these changes" rules.
+
+(3) "Pharmacist / Walk-in clinic / Family doctor or Telehealth recommended":
+    - Use when:
+        - symptoms are persistent or recurrent,
+        - moderate in intensity,
+        - affecting function,
+        - or there is diagnostic uncertainty but NOT clear emergency.
+    - Explain why an in-person/phone assessment is appropriate.
+
+(4) "Go to Emergency / Call 911 now":
+    - Use when there are red-flag features:
+        - severe or rapidly worsening pain,
+        - major trauma,
+        - uncontrolled bleeding,
+        - chest pain, trouble breathing,
+        - stroke signs,
+        - signs of sepsis or systemic illness,
+        - significant numbness/weakness,
+        - concerning changes after surgery or known serious conditions,
+        - or visually very concerning findings on the image.
+    - Be direct and clear.
 
 OUTPUT FORMAT (WHAT USER SEES)
 ------------------------------
 1. **Primary Recommended Pathway**
    - One brief sentence stating the recommendation WITHOUT internal labels.
 
-2. Then include sections as appropriate:
+2. Conditional sections depending on your chosen pathway:
 
 If you chose "No current action needed":
-- **Home Care / Wellness**
-- **When to seek medical advice**
-- **When this is an Emergency (ER/911)**
-- **Local & Official Resources**
-- **Important Disclaimer**
+    - **Home Care / Wellness**
+    - **When to seek medical advice**
+    - **When this is an Emergency (ER/911)**
+    - **Local & Official Resources**
+    - **Important Disclaimer**
 
 If you chose "Home care reasonable (self-care + monitor)":
-- **Home Care**
-- **When to see a Pharmacist / Clinic / Telehealth**
-- **When this is an Emergency (ER/911)**
-- **Local & Official Resources**
-- **Important Disclaimer**
+    - **Home Care**
+    - **When to see a Pharmacist / Clinic / Telehealth**
+    - **When this is an Emergency (ER/911)**
+    - **Local & Official Resources**
+    - **Important Disclaimer**
 
 If you chose "Pharmacist / Walk-in clinic / Family doctor or Telehealth recommended":
-- **When to see a Pharmacist / Clinic / Telehealth**
-- **When this is an Emergency (ER/911)**
-- **Local & Official Resources**
-- **Important Disclaimer**
+    - **When to see a Pharmacist / Clinic / Telehealth**
+    - **When this is an Emergency (ER/911)**
+    - **Local & Official Resources**
+    - **Important Disclaimer**
 
 If you chose "Go to Emergency / Call 911 now":
-- **When this is an Emergency (ER/911)**
-- **Local & Official Resources**
-- **Important Disclaimer**
-- Do NOT include pure-home-care reassurance in this case.
+    - **When this is an Emergency (ER/911)**
+    - **Local & Official Resources**
+    - **Important Disclaimer**
+    - Do NOT include home-care reassurance.
 
 LOCAL & OFFICIAL RESOURCES
 --------------------------
-- If facilities_context lists facilities, summarize them as bullet points:
-  - name, city/address, phone (if present), hours (if present), URL.
-- If none are listed, say so and direct user to Google Maps / provincial tools.
-- Optionally remind of:
-  - Health Canada Drug Product Database / Recalls site,
-  - CIHI/provincial dashboards for general (non-real-time) wait information.
+- If facilities_context lists facilities:
+    - Summarize them as bullet points:
+        - name,
+        - city or address,
+        - phone (if present),
+        - hours/“open now” (if present),
+        - link/URL.
+    - Do NOT alter names or add new ones.
+- If facilities_context indicates that no specific facilities were found:
+    - Say that directly.
+    - Instruct users to:
+        - use Google Maps with their city,
+        - or use official provincial "find a clinic/ER" / Telehealth tools.
+    - Do NOT fabricate facility names.
+- Include hqontario_context when relevant (for Ontario users).
+- Optionally remind users:
+    - Health Canada Drug Product Database and Recalls site for medication/product safety.
+    - CIHI/provincial dashboards for general (non-real-time) wait-time information.
 
 TONE
 ----
-- Calm, supportive, neutral.
-- Never alarmist, never dismissive.
-- Always remind users this is informational only and not a diagnosis.
+- Calm, supportive, non-alarming.
+- No legalese walls of text; keep it tight and readable.
+- Always remind: this is informational, not a diagnosis; if worried, they should seek real care.
 """
 
     user_message = f"""
@@ -806,7 +846,7 @@ HQ Ontario info:
     return call_openrouter_chat(REASONING_MODEL, messages, temperature=0.3)
 
 # =========================
-# STATE & NAV HELPERS
+# STATE & NAV
 # =========================
 
 if "step" not in st.session_state:
@@ -840,7 +880,7 @@ def render_steps():
     )
 
 # =========================
-# APP HEADER
+# HEADER
 # =========================
 
 st.markdown(
@@ -857,7 +897,7 @@ st.markdown(
 render_steps()
 
 # =========================
-# STEP 1: CITY + CAMERA
+# STEP 1
 # =========================
 
 if st.session_state.step == 1:
@@ -909,7 +949,7 @@ if st.session_state.step == 1:
     st.markdown("</div>", unsafe_allow_html=True)
 
 # =========================
-# STEP 2: VOICE OR TEXT (WITH AUTO-FILL)
+# STEP 2
 # =========================
 
 elif st.session_state.step == 2:
@@ -936,14 +976,11 @@ elif st.session_state.step == 2:
 
     if audio_bytes:
         st.session_state.audio_bytes = audio_bytes
-        # Only re-transcribe if this is a new recording
         if st.session_state.last_audio != audio_bytes:
             st.success("Voice note captured. Transcribing...")
             transcript = call_asr(audio_bytes)
             if transcript:
-                st.session_state.user_text = transcript.strip()
-            else:
-                st.warning("Couldn't transcribe that recording. Please type a brief description.")
+                st.session_state.user_text = transcript
             st.session_state.last_audio = audio_bytes
 
     user_text = st.text_area(
@@ -1014,7 +1051,7 @@ elif st.session_state.step == 2:
     st.markdown("</div>", unsafe_allow_html=True)
 
 # =========================
-# STEP 3: RESULT SCREEN
+# STEP 3
 # =========================
 
 elif st.session_state.step == 3: