Update app.py

app.py

@@ -7,8 +7,7 @@ import textwrap
 import requests
 import streamlit as st
 from PIL import Image
-from openai import OpenAI
-from audio_recorder_streamlit import audio_recorder  # mic button
+from audio_recorder_streamlit import audio_recorder
 
 # =========================
 # BASIC CONFIG
@@ -20,16 +19,28 @@ st.set_page_config(
     layout="centered"
 )
 
-OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
-if not OPENAI_API_KEY:
-    st.warning("Set OPENAI_API_KEY in your Space secrets to enable AI features.")
+# --- Keys / endpoints ---
+
+OPENROUTER_API_KEY = os.getenv("OPENROUTER_API_KEY")
+HF_API_TOKEN = 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()
 
-client = OpenAI(api_key=OPENAI_API_KEY)
+if not HF_API_TOKEN:
+    st.warning("Set HF_API_TOKEN in your Space secrets (Hugging Face Inference) to enable speech-to-text.")
+    # we don't stop; app still works without voice
+
+OPENROUTER_URL = "https://openrouter.ai/api/v1/chat/completions"
 
-VISION_MODEL = "gpt-4.1-mini"      # image-capable model
-REASONING_MODEL = "gpt-4o"         # reasoning / instructions
-ASR_MODEL = "whisper-1"            # transcription
+# IMPORTANT: put the exact OpenRouter model IDs here.
+# Check https://openrouter.ai/models for the correct slugs.
+VISION_MODEL = os.getenv("VISION_MODEL", "nvidia/nemotron-nano-12b-vl")       # <-- adjust if needed
+REASONING_MODEL = os.getenv("REASONING_MODEL", "nvidia/nemotron-nano-12b-vl") # <-- or another Nemotron/Instruct model
+
+HF_WHISPER_MODEL = os.getenv("HF_WHISPER_MODEL", "openai/whisper-large-v3")
+HF_WHISPER_URL = f"https://api-inference.huggingface.co/models/{HF_WHISPER_MODEL}"
 
 DPD_BASE_URL = "https://health-products.canada.ca/api/drug/drugproduct"
 RECALLS_FEED_URL = os.getenv(
@@ -75,7 +86,6 @@ st.markdown(
         justify-content: center;
         gap: 8px;
         margin-bottom: 0.8rem;
-        /* kill any weird pill styles */
         background: transparent !important;
         box-shadow: none !important;
         padding: 0 !important;
@@ -160,7 +170,7 @@ def shorten(text, max_chars=260):
 def extract_phone(text: str) -> str:
     if not text:
         return ""
-    m = re.search(r"(\(?\d{3}\)?[\\s\\-\\.]?\\d{3}[\\s\\-\\.]?\\d{4})", text)
+    m = re.search(r"(\(?\d{3}\)?[\s\-\.]?\d{3}[\s\-\.]?\d{4})", text)
     return m.group(1) if m else ""
 
 
@@ -486,13 +496,41 @@ def tool_get_wait_times_awareness() -> str:
 
 
 # =========================
-# OPENAI HELPERS
+# OPENROUTER HELPER
+# =========================
+
+def call_openrouter_chat(model: str, messages, temperature: float = 0.3):
+    """
+    Generic helper for OpenRouter chat/completions API with OpenAI-format messages.
+    """
+    headers = {
+        "Authorization": f"Bearer {OPENROUTER_API_KEY}",
+        "Content-Type": "application/json",
+        # Optional but recommended by OpenRouter:
+        "HTTP-Referer": "https://seth0330-save.hf.space",
+        "X-Title": "CareCall AI (Canada)",
+    }
+    payload = {
+        "model": model,
+        "messages": messages,
+        "temperature": temperature,
+    }
+    try:
+        r = requests.post(OPENROUTER_URL, headers=headers, json=payload, timeout=60)
+        r.raise_for_status()
+        data = r.json()
+        return data["choices"][0]["message"]["content"].strip()
+    except Exception as e:
+        return f"(Model call unavailable: {e})"
+
+
+# =========================
+# VISION (Nemotron via OpenRouter)
 # =========================
 
 def call_vision_summarizer(image_bytes: bytes) -> str:
     """
-    Structured, cautious but detailed visual assessment.
-    Treats the image as a primary signal, without diagnosing.
+    Structured, cautious but detailed visual assessment via vision-capable model on OpenRouter.
     """
     if not image_bytes:
         return ""
@@ -514,59 +552,74 @@ Describe clearly and systematically:
 6. Any visible discharge, bleeding, or crusting.
 7. Any objects/devices/dressings in view (e.g., ring, bandage, splint).
 8. Any visually concerning features that might suggest higher risk
-   (e.g., blackened tissue, rapidly spreading redness pattern, deep open wound, thick pus) — describe ONLY what is seen.
+   (e.g., blackened tissue, spreading redness, deep open wound, thick pus) — describe ONLY what is seen.
 
 Keep it about 120–180 words.
 Write as if supporting a separate triage system, not directly reassuring or diagnosing the patient.
 """
 
-    try:
-        resp = client.chat.completions.create(
-            model=VISION_MODEL,
-            messages=[{
-                "role": "user",
-                "content": [
-                    {"type": "text", "text": prompt},
-                    {
-                        "type": "image_url",
-                        "image_url": {"url": f"data:image/jpeg;base64,{b64}"},
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": prompt},
+                {
+                    "type": "image_url",
+                    "image_url": {
+                        "url": f"data:image/jpeg;base64,{b64}"
                     },
-                ],
-            }],
-            temperature=0.2,
-        )
-        return resp.choices[0].message.content.strip()
-    except Exception as e:
-        return f"(Vision analysis unavailable: {e})"
+                },
+            ],
+        }
+    ]
 
+    return call_openrouter_chat(VISION_MODEL, messages, temperature=0.2)
+
+
+# =========================
+# ASR (Whisper via HF Inference)
+# =========================
 
 def call_asr(audio_source) -> str:
     """
-    Accepts:
+    Uses Hugging Face Inference API for Whisper.
+    Expects:
     - bytes from audio_recorder_streamlit, or
-    - a file-like object (if extended in future).
+    - file-like object.
     """
-    if not audio_source:
+    if not audio_source or not HF_API_TOKEN:
         return ""
-    try:
-        if isinstance(audio_source, bytes):
-            bio = io.BytesIO(audio_source)
-            bio.name = "voice.wav"
-        else:
-            audio_bytes = audio_source.read()
-            bio = io.BytesIO(audio_bytes)
-            bio.name = getattr(audio_source, "name", "voice.wav")
 
-        result = client.audio.transcriptions.create(
-            model=ASR_MODEL,
-            file=bio
-        )
-        txt = getattr(result, "text", None)
-        return txt.strip() if isinstance(txt, str) else str(result)
+    # Normalize to bytes
+    if isinstance(audio_source, bytes):
+        audio_bytes = audio_source
+    else:
+        audio_bytes = audio_source.read()
+
+    headers = {
+        "Authorization": f"Bearer {HF_API_TOKEN}",
+        "Content-Type": "audio/wav",  # works for most; HF autodetects
+    }
+
+    try:
+        resp = requests.post(HF_WHISPER_URL, headers=headers, data=audio_bytes, timeout=120)
+        resp.raise_for_status()
+        data = resp.json()
+        # ASR pipeline returns {"text": "..."}
+        if isinstance(data, dict) and "text" in data:
+            return data["text"].strip()
+        # Some models may return list-style
+        if isinstance(data, list) and data and isinstance(data[0], dict):
+            return (data[0].get("text") or data[0].get("generated_text") or "").strip()
+        return ""
     except Exception as e:
         return f"(Transcription unavailable: {e})"
 
 
+# =========================
+# REASONING AGENT (Nemotron via OpenRouter)
+# =========================
+
 def call_reasoning_agent(
     narrative: str,
     vision_summary: str = "",
@@ -621,23 +674,17 @@ You must internally choose EXACTLY ONE primary pathway, but DO NOT show letters
 
 (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.
-- Example: "I feel healthy and really perfect and fresh."
 - Provide light wellness reminders.
-- This is NOT a diagnosis; phrase as "based on what you've shared, you appear well."
+- 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
-  (e.g., light headache, small bruise, mild soreness, mild cold symptoms without red flags).
+- 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.
-- Do NOT automatically push these to doctor/ER unless they persist, worsen, or match red flags.
 
 (3) "Pharmacist / Walk-in clinic / Family doctor or Telehealth recommended"
-- For persistent, recurrent, or moderate symptoms,
-  or when there is diagnostic uncertainty but no clear emergency.
+- For persistent, recurrent, or moderate symptoms, or diagnostic uncertainty without clear emergency.
 - Explain why an in-person or phone assessment is appropriate.
-- Suggest pharmacist for minor medication/symptom questions.
-- Suggest walk-in/clinic/family doctor/Telehealth when a physical exam/history is needed.
 
 (4) "Go to Emergency / Call 911 now"
 - For clear red-flag features:
@@ -648,21 +695,16 @@ You must internally choose EXACTLY ONE primary pathway, but DO NOT show letters
   - stroke-like symptoms,
   - signs of sepsis or systemic illness,
   - significant numbness/weakness,
-  - concerning changes after surgery,
-  - or visually severe findings in the image.
+  - concerning post-surgical changes,
+  - or visually severe findings from the image.
 - Be direct and unambiguous.
 
 OUTPUT FORMAT (WHAT USER SEES)
 ------------------------------
 1. **Primary Recommended Pathway**
    - One brief sentence stating the recommendation WITHOUT internal labels.
-     Examples:
-     - "No current action needed (you appear well, continue normal care)."
-     - "Home care is reasonable with self-care and monitoring."
-     - "An in-person assessment with a walk-in clinic, your family doctor, or Telehealth is recommended."
-     - "You should go to an emergency department or call 911 now."
 
-2. Then include sections as appropriate for the chosen pathway:
+2. Then include sections as appropriate:
 
 If you chose "No current action needed":
 - **Home Care / Wellness**
@@ -688,14 +730,14 @@ If you chose "Go to Emergency / Call 911 now":
 - **When this is an Emergency (ER/911)**
 - **Local & Official Resources**
 - **Important Disclaimer**
-- Do NOT include home-care-only reassurance in this case.
+- Do NOT include pure-home-care reassurance in this case.
 
 LOCAL & OFFICIAL RESOURCES
 --------------------------
 - If facilities_context lists facilities, summarize them as bullet points:
-  - name, city/address, phone (if present), hours (if present), and URL.
+  - 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:
+- Optionally remind of:
   - Health Canada Drug Product Database / Recalls site,
   - CIHI/provincial dashboards for general (non-real-time) wait information.
 
@@ -735,21 +777,12 @@ HQ Ontario info:
 {hqontario_context or "(not applicable)"}
 """
 
-    try:
-        resp = client.chat.completions.create(
-            model=REASONING_MODEL,
-            temperature=0.3,
-            messages=[
-                {"role": "system", "content": system_prompt},
-                {"role": "user", "content": user_message},
-            ],
-        )
-        return resp.choices[0].message.content.strip()
-    except Exception as e:
-        return (
-            f"(Reasoning agent unavailable: {e})\n\n"
-            "Please contact a licensed provider, Telehealth, or emergency services as appropriate."
-        )
+    messages = [
+        {"role": "system", "content": system_prompt},
+        {"role": "user", "content": user_message},
+    ]
+
+    return call_openrouter_chat(REASONING_MODEL, messages, temperature=0.3)
 
 
 # =========================
@@ -809,11 +842,11 @@ render_steps()
 if st.session_state.step == 1:
     st.markdown('<div class="card">', unsafe_allow_html=True)
     st.markdown(
-        '<div class="summary-title">Step 1 · Where are you, and what area is this?</div>',
+        '<div class="summary-title">Step 1 · Where & what are we looking at?</div>',
         unsafe_allow_html=True,
     )
     st.markdown(
-        '<div class="label-soft">Select your city and (optionally) take a clear photo of the area you are concerned about.</div>',
+        '<div class="label-soft">Select your city and (optionally) take a clear photo of the area you are worried about.</div>',
         unsafe_allow_html=True,
     )
 
@@ -865,7 +898,7 @@ elif st.session_state.step == 2:
         unsafe_allow_html=True,
     )
     st.markdown(
-        '<div class="label-soft">Use the mic to describe your concern, or type instead. A clear story helps the AI match what it sees in the photo.</div>',
+        '<div class="label-soft">Use the mic to describe your concern, or type instead. A clear story helps match what we see in the photo.</div>',
         unsafe_allow_html=True,
     )
 
@@ -891,7 +924,7 @@ elif st.session_state.step == 2:
     st.session_state.user_text = user_text
 
     st.markdown(
-        '<div class="label-soft">When ready, get your one recommended pathway.</div>',
+        '<div class="label-soft">When you are ready, get your one recommended pathway.</div>',
         unsafe_allow_html=True,
     )