Upload app.py with huggingface_hub

app.py

@@ -14,8 +14,10 @@ GROQ_KEY = os.environ.get("GROQ_API_KEY", "")
 HF_TOKEN = os.environ.get("HF_TOKEN", "")
 HISTORY_REPO = "Saicharan21/cardiolab-chat-history"
 PAPERS_DB_REPO = "Saicharan21/cardiolab-papers-db"
+CARDIOLAB_MODEL = "Saicharan21/CardioLab-AI-Model"
 
 CHAT_MODELS = {
+    "CardioLab Fine-tuned (SJSU)": "cardiolab",
     "Llama 3.3 70B (Best)": "llama-3.3-70b-versatile",
     "Llama 3.1 8B (Fast)": "llama-3.1-8b-instant",
     "Mixtral 8x7B": "mixtral-8x7b-32768",
@@ -28,18 +30,20 @@ KNOWHOW = ("MCL: Sylgard 184 PDMS 10:1 ratio 48hr cure green laser PIV 70bpm 5L/
     "MHV: 27mm SJM Regent bileaflet trileaflet monoleaflet pediatric. "
     "Equipment: Heska HT5 analyzer PIV green laser Tygon tubing Arduino Uno.")
 
-# ── LOAD PAPERS ON STARTUP ─────────────────────────────────────────
+# ── LOAD PAPERS + FINE-TUNED MODEL ON STARTUP ─────────────────────
 CHUNKS = []
 METADATA = []
 EMBEDDINGS = None
 PAPERS_LOADED = False
 EMBEDDER = None
+CARDIOLAB_TOKENIZER = None
+CARDIOLAB_LLM = None
+CARDIOLAB_MODEL_LOADED = False
 
 def load_papers():
     global CHUNKS, METADATA, EMBEDDINGS, PAPERS_LOADED, EMBEDDER
     try:
         from sentence_transformers import SentenceTransformer
-        print("Loading paper database from HuggingFace...")
         chunks_path = hf_hub_download(repo_id=PAPERS_DB_REPO, filename="chunks.json", repo_type="dataset", token=HF_TOKEN)
         meta_path = hf_hub_download(repo_id=PAPERS_DB_REPO, filename="metadata.json", repo_type="dataset", token=HF_TOKEN)
         emb_path = hf_hub_download(repo_id=PAPERS_DB_REPO, filename="embeddings.npy", repo_type="dataset", token=HF_TOKEN)
@@ -48,18 +52,41 @@ def load_papers():
         EMBEDDINGS = np.load(emb_path)
         EMBEDDER = SentenceTransformer("all-MiniLM-L6-v2")
         PAPERS_LOADED = True
-        papers_count = len(set(m["paper"] for m in METADATA))
-        print(f"Loaded {len(CHUNKS)} chunks from {papers_count} SJSU papers!")
+        print(f"Papers loaded: {len(CHUNKS)} chunks from {len(set(m['paper'] for m in METADATA))} papers")
         return True
     except Exception as e:
         print(f"Paper load error: {e}")
         return False
 
+def load_cardiolab_model():
+    global CARDIOLAB_TOKENIZER, CARDIOLAB_LLM, CARDIOLAB_MODEL_LOADED
+    try:
+        import torch
+        from transformers import AutoModelForCausalLM, AutoTokenizer
+        from peft import PeftModel
+        print("Loading CardioLab fine-tuned model...")
+        base_model = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
+        CARDIOLAB_TOKENIZER = AutoTokenizer.from_pretrained(CARDIOLAB_MODEL, token=HF_TOKEN)
+        CARDIOLAB_TOKENIZER.pad_token = CARDIOLAB_TOKENIZER.eos_token
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+        CARDIOLAB_LLM = AutoModelForCausalLM.from_pretrained(
+            CARDIOLAB_MODEL, token=HF_TOKEN,
+            torch_dtype=torch.float16 if device=="cuda" else torch.float32,
+            device_map="auto" if device=="cuda" else None,
+            low_cpu_mem_usage=True
+        )
+        CARDIOLAB_MODEL_LOADED = True
+        print(f"CardioLab model loaded on {device}!")
+        return True
+    except Exception as e:
+        print(f"CardioLab model load error: {e}")
+        return False
+
 load_papers()
+load_cardiolab_model()
 
 # ── SEMANTIC SEARCH ────────────────────────────────────────────────
 def search_papers(query, n=4):
-    global CHUNKS, METADATA, EMBEDDINGS, EMBEDDER, PAPERS_LOADED
     if not PAPERS_LOADED or EMBEDDINGS is None or EMBEDDER is None:
         return "", []
     try:
@@ -77,16 +104,42 @@ def search_papers(query, n=4):
             meta = METADATA[idx]
             score = float(scores[idx])
             if score > 0.25:
-                results.append({"chunk": chunk, "paper": meta["paper"], "pillar": meta.get("pillar",""), "score": score})
+                results.append({"chunk":chunk,"paper":meta["paper"],"pillar":meta.get("pillar",""),"score":score})
                 if meta["paper"] not in seen:
-                    context += chr(10)+"=== FROM: "+meta["paper"]+" ["+meta.get("pillar","")+"] ==="+chr(10)
+                    context += chr(10)+"=== FROM: "+meta["paper"]+" ==="+chr(10)
                     seen.add(meta["paper"])
                 context += chunk[:500]+chr(10)
         return context, results
     except Exception as e:
-        print(f"Search error: {e}")
         return "", []
 
+def answer_with_cardiolab_model(question, paper_context=""):
+    if not CARDIOLAB_MODEL_LOADED:
+        return None
+    try:
+        import torch
+        system = "You are CardioLab AI for SJSU Biomedical Engineering."
+        if paper_context:
+            system += " Use these SJSU research papers: "+paper_context[:500]
+        prompt = f"<|system|>{system}</s><|user|>{question}</s><|assistant|>"
+        inputs = CARDIOLAB_TOKENIZER(prompt, return_tensors="pt", truncation=True, max_length=512)
+        device = next(CARDIOLAB_LLM.parameters()).device
+        inputs = {k:v.to(device) for k,v in inputs.items()}
+        with torch.no_grad():
+            outputs = CARDIOLAB_LLM.generate(
+                **inputs, max_new_tokens=200, do_sample=True,
+                temperature=0.3, pad_token_id=CARDIOLAB_TOKENIZER.eos_token_id
+            )
+        response = CARDIOLAB_TOKENIZER.decode(outputs[0], skip_special_tokens=True)
+        if "<|assistant|>" in response:
+            answer = response.split("<|assistant|>")[-1].strip()
+        else:
+            answer = response[len(prompt):].strip() if len(response) > len(prompt) else response
+        return answer if len(answer) > 20 else None
+    except Exception as e:
+        print(f"CardioLab model error: {e}")
+        return None
+
 CSS = """
 body, .gradio-container { background: #f7f7f8 !important; font-family: -apple-system, BlinkMacSystemFont, Segoe UI, sans-serif !important; }
 .tab-nav { background: #ffffff !important; border-bottom: 1px solid #e5e7eb !important; padding: 0 16px !important; display: flex !important; flex-wrap: wrap !important; }
@@ -120,7 +173,7 @@ HEADER = """<div style="background:linear-gradient(135deg,#0a0f2e 0%,#1a0a0a 100
 <svg width="100" height="28" viewBox="0 0 120 32"><polyline points="0,16 20,16 26,4 30,28 34,2 38,26 44,16 120,16" fill="none" stroke="#c1121f" stroke-width="2.5" stroke-linecap="round"/></svg>
 <div style="font-size:2em;font-weight:900;letter-spacing:2px;"><span style="color:#ffffff;">Cardio</span><span style="color:#c1121f;">Lab</span><span style="color:#ffffff;"> AI</span></div>
 <svg width="100" height="28" viewBox="0 0 120 32" style="transform:scaleX(-1);"><polyline points="0,16 20,16 26,4 30,28 34,2 38,26 44,16 120,16" fill="none" stroke="#c1121f" stroke-width="2.5" stroke-linecap="round"/></svg></div>
-<div style="color:#9ca3af;font-size:0.68em;letter-spacing:2px;text-transform:uppercase;">RAG Agent | 16 SJSU Papers | Llama 3.3 70B | 5 Search Sources</div></div>
+<div style="color:#9ca3af;font-size:0.68em;letter-spacing:2px;text-transform:uppercase;">Fine-tuned on 16 SJSU Papers | RAG | Llama 3.3 70B | 5 AI Models</div></div>
 <div style="display:flex;align-items:center;gap:14px;">
 <div style="text-align:right;"><div style="color:#9ca3af;font-size:0.68em;text-transform:uppercase;">Research Pillars</div>
 <div style="color:#ffffff;font-size:0.72em;margin-top:3px;">MHV  CKD  FSI</div>
@@ -130,7 +183,6 @@ HEADER = """<div style="background:linear-gradient(135deg,#0a0f2e 0%,#1a0a0a 100
 <polyline points="25,45 32,45 35,35 38,55 41,30 44,50 50,45 75,45" fill="none" stroke="white" stroke-width="2.5" stroke-linecap="round" opacity="0.9"/></svg></div></div>
 <div style="height:3px;background:linear-gradient(90deg,#0057a8,#c1121f,#e8a020,#c1121f,#0057a8);"></div></div>"""
 
-# ── SESSION MANAGEMENT ─────────────────────────────────────────────
 def load_all_sessions():
     if not HF_TOKEN: return {}
     try:
@@ -150,8 +202,7 @@ def save_all_sessions(sessions):
 
 def get_session_list():
     s = load_all_sessions()
-    if not s: return ["No saved sessions"]
-    return list(reversed(list(s.keys())))
+    return list(reversed(list(s.keys()))) if s else ["No saved sessions"]
 
 def save_session(history, name):
     if not history: return "Nothing to save", gr.update()
@@ -165,8 +216,7 @@ def save_session(history, name):
 def load_session(name):
     if not name or "No saved" in name: return [], "Select a session"
     sessions = load_all_sessions()
-    if name in sessions: return sessions[name]["messages"], "Loaded: "+name
-    return [], "Not found"
+    return (sessions[name]["messages"], "Loaded: "+name) if name in sessions else ([], "Not found")
 
 def delete_session(name):
     if not name or "No saved" in name: return "Select a session", gr.update()
@@ -179,115 +229,34 @@ def delete_session(name):
 
 def new_chat(): return [], "", "New chat started"
 
-# ── SEARCH FUNCTIONS ───────────────────────────────────────────────
-def expand_query_ai(query, model_id="llama-3.3-70b-versatile"):
-    if not GROQ_KEY: return query
-    try:
-        client = Groq(api_key=GROQ_KEY)
-        resp = client.chat.completions.create(model=model_id,
-            messages=[{"role":"system","content":"Biomedical PubMed expert. Convert to optimized MeSH terms for heart valves hemodynamics PIV thrombogenicity FSI microfluidics CKD creatinine. Return ONLY terms."},
-            {"role":"user","content":"Optimize: "+query}],max_tokens=80)
-        return resp.choices[0].message.content.strip() or query
-    except: return query
-
-def fetch_pubmed(query, n=6):
-    try:
-        forced = query+" AND (heart valve OR hemodynamics OR microfluidic OR thrombogen OR creatinine OR PIV OR CFD OR CKD OR fluid structure)"
-        r = requests.get("https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi",
-            params={"db":"pubmed","term":forced,"retmax":n,"retmode":"json","sort":"date","field":"tiab"},timeout=12)
-        ids = r.json()["esearchresult"]["idlist"]
-        if not ids: return []
-        r2 = requests.get("https://eutils.ncbi.nlm.nih.gov/entrez/eutils/efetch.fcgi",
-            params={"db":"pubmed","id":",".join(ids),"retmode":"xml","rettype":"abstract"},timeout=12)
-        import xml.etree.ElementTree as ET
-        root = ET.fromstring(r2.content)
-        results = []
-        for article in root.findall(".//PubmedArticle"):
-            try:
-                title = article.find(".//ArticleTitle").text or "No title"
-                pmid = article.find(".//PMID").text or ""
-                year_el = article.find(".//PubDate/Year")
-                year = year_el.text if year_el is not None else ""
-                results.append({"source":"PubMed","title":str(title),"year":year,
-                    "url":"https://pubmed.ncbi.nlm.nih.gov/"+pmid,"citations":"N/A"})
-            except: continue
-        return results
-    except: return []
-
-def fetch_scholar(query, n=6):
-    try:
-        r = requests.get("https://api.semanticscholar.org/graph/v1/paper/search",
-            params={"query":query,"limit":n,"fields":"title,year,url,citationCount"},timeout=12)
-        papers = r.json().get("data",[])
-        results = []
-        for p in papers:
-            year = p.get("year",0) or 0
-            if int(year) < 2015: continue
-            results.append({"source":"Scholar","title":p.get("title",""),"year":str(year),
-                "url":p.get("url",""),"citations":str(p.get("citationCount",0))})
-        results.sort(key=lambda x:(x["year"],int(x["citations"]) if x["citations"].isdigit() else 0),reverse=True)
-        return results
-    except: return []
-
-def fetch_europe_pmc(query, n=5):
-    try:
-        r = requests.get("https://www.ebi.ac.uk/europepmc/webservices/rest/search",
-            params={"query":query,"format":"json","pageSize":n,"sort":"P_PDATE_D desc"},timeout=12)
-        articles = r.json().get("resultList",{}).get("result",[])
-        results = []
-        for a in articles:
-            year = str(a.get("pubYear",""))
-            if year and int(year) < 2015: continue
-            pmid = a.get("pmid",""); doi = a.get("doi","")
-            url = ("https://pubmed.ncbi.nlm.nih.gov/"+pmid if pmid else "https://doi.org/"+doi if doi else "")
-            if not url: continue
-            results.append({"source":"Europe PMC","title":a.get("title",""),"year":year,
-                "url":url,"citations":str(a.get("citedByCount",0))})
-        return results
-    except: return []
-
-def quick_search(query, search_model="Llama 3.3 70B (Best)"):
-    if not query.strip(): return "Please enter a research topic."
-    model_id = CHAT_MODELS.get(search_model, "llama-3.3-70b-versatile")
-    expanded = expand_query_ai(query, model_id)
-    r1 = fetch_pubmed(expanded, n=6)
-    r2 = fetch_scholar(expanded, n=6)
-    r3 = fetch_europe_pmc(expanded, n=5)
-    sjsu_url = "https://scholarworks.sjsu.edu/do/search/?q="+requests.utils.quote(query)+"&context=6781027"
-    all_results = r1+r2+r3
-    seen = set()
-    unique = []
-    for r in all_results:
-        key = r["title"][:50].lower().strip()
-        if key not in seen and r["url"]:
-            seen.add(key); unique.append(r)
-    out = "QUERY: "+query+chr(10)+"AI EXPANDED: "+expanded+chr(10)
-    out += "="*45+chr(10)+chr(10)
-    groups = {"PubMed":[],"Scholar":[],"Europe PMC":[]}
-    for r in unique[:20]:
-        if r["source"] in groups: groups[r["source"]].append(r)
-    for source, papers in groups.items():
-        if not papers: continue
-        out += "--- "+source+" ---"+chr(10)
-        for p in papers:
-            out += p["title"][:85]+" ("+p["year"]+")"
-            if p["citations"] not in ("N/A","",): out += " | "+p["citations"]+" citations"
-            out += chr(10)+"  "+p["url"]+chr(10)+chr(10)
-    out += "--- SJSU ScholarWorks ---"+chr(10)
-    out += "Search SJSU papers: "+sjsu_url+chr(10)
-    return out
-
 def get_pubmed_chat(query, n=3):
     try:
         r = requests.get("https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi",
             params={"db":"pubmed","term":query+" AND (heart valve OR hemodynamics OR microfluidic OR thrombogen OR creatinine OR CKD)","retmax":n,"retmode":"json","sort":"date","field":"tiab"},timeout=10)
         ids = r.json()["esearchresult"]["idlist"]
-        if not ids: return ""
-        return chr(10).join(["https://pubmed.ncbi.nlm.nih.gov/"+i for i in ids])
+        return chr(10).join(["https://pubmed.ncbi.nlm.nih.gov/"+i for i in ids]) if ids else ""
     except: return ""
 
-# ── CHAT WITH RAG ──────────────────────────────────────────────────
 def research_chat(message, history, chat_model="Llama 3.3 70B (Best)"):
+    if not message.strip(): return "", history
+    paper_context, paper_results = search_papers(message, n=4)
+
+    # Use fine-tuned CardioLab model if selected
+    if chat_model == "CardioLab Fine-tuned (SJSU)" and CARDIOLAB_MODEL_LOADED:
+        answer = answer_with_cardiolab_model(message, paper_context)
+        if answer:
+            if paper_results:
+                unique_papers = list(dict.fromkeys([r["paper"] for r in paper_results]))
+                answer += chr(10)+chr(10)+"Sources from SJSU CardioLab papers:"
+                for p in unique_papers[:3]:
+                    answer += chr(10)+"  - "+p.replace('.pdf','').replace('_',' ')
+            pubmed = get_pubmed_chat(message, n=2)
+            if pubmed: answer += chr(10)+"PubMed: "+pubmed
+            history.append({"role":"user","content":message})
+            history.append({"role":"assistant","content":"[CardioLab Fine-tuned Model] "+answer})
+            return "", history
+
+    # Fall back to Groq models
     if not GROQ_KEY:
         history.append({"role":"user","content":message})
         history.append({"role":"assistant","content":"Error: Add GROQ_API_KEY to Space Settings."})
@@ -295,11 +264,10 @@ def research_chat(message, history, chat_model="Llama 3.3 70B (Best)"):
     try:
         model_id = CHAT_MODELS.get(chat_model, "llama-3.3-70b-versatile")
         client = Groq(api_key=GROQ_KEY)
-        paper_context, paper_results = search_papers(message, n=4)
         if paper_context:
             system_prompt = ("You are CardioLab AI for SJSU Biomedical Engineering. "
                 "Answer using SJSU CardioLab research papers below. "
-                "Always cite the paper name when using specific data. Be precise with numbers and protocols."+chr(10)+chr(10)+
+                "Always cite the paper name when using specific data."+chr(10)+chr(10)+
                 "SJSU CARDIOLAB PAPERS:"+chr(10)+paper_context+chr(10)+chr(10)+
                 "ADDITIONAL KNOWLEDGE: "+KNOWHOW)
         else:
@@ -348,6 +316,65 @@ def voice_chat(audio, history):
         history.append({"role":"assistant","content":"Voice error: "+str(e)})
         return history
 
+def expand_query_ai(query):
+    if not GROQ_KEY: return query
+    try:
+        client = Groq(api_key=GROQ_KEY)
+        resp = client.chat.completions.create(model="llama-3.1-8b-instant",
+            messages=[{"role":"system","content":"Biomedical PubMed expert. Convert to MeSH terms for heart valves hemodynamics PIV thrombogenicity FSI microfluidics CKD. Return ONLY terms."},
+            {"role":"user","content":"Optimize: "+query}],max_tokens=80)
+        return resp.choices[0].message.content.strip() or query
+    except: return query
+
+def quick_search(query, search_model="Llama 3.3 70B (Best)"):
+    if not query.strip(): return "Please enter a topic."
+    expanded = expand_query_ai(query)
+    results = []
+    try:
+        forced = expanded+" AND (heart valve OR hemodynamics OR microfluidic OR thrombogen OR creatinine OR PIV OR CFD OR CKD)"
+        r = requests.get("https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi",
+            params={"db":"pubmed","term":forced,"retmax":8,"retmode":"json","sort":"date","field":"tiab"},timeout=12)
+        ids = r.json()["esearchresult"]["idlist"]
+        if ids:
+            r2 = requests.get("https://eutils.ncbi.nlm.nih.gov/entrez/eutils/efetch.fcgi",
+                params={"db":"pubmed","id":",".join(ids),"retmode":"xml","rettype":"abstract"},timeout=12)
+            import xml.etree.ElementTree as ET
+            root = ET.fromstring(r2.content)
+            for article in root.findall(".//PubmedArticle"):
+                try:
+                    title = article.find(".//ArticleTitle").text or "No title"
+                    pmid = article.find(".//PMID").text or ""
+                    year_el = article.find(".//PubDate/Year")
+                    year = year_el.text if year_el is not None else ""
+                    results.append({"source":"PubMed","title":str(title),"year":year,"url":"https://pubmed.ncbi.nlm.nih.gov/"+pmid,"citations":"N/A"})
+                except: continue
+    except: pass
+    try:
+        r = requests.get("https://api.semanticscholar.org/graph/v1/paper/search",
+            params={"query":expanded,"limit":6,"fields":"title,year,url,citationCount"},timeout=12)
+        for p in r.json().get("data",[]):
+            year = p.get("year",0) or 0
+            if int(year) >= 2015:
+                results.append({"source":"Scholar","title":p.get("title",""),"year":str(year),"url":p.get("url",""),"citations":str(p.get("citationCount",0))})
+    except: pass
+    out = "QUERY: "+query+chr(10)+"AI EXPANDED: "+expanded+chr(10)+"="*45+chr(10)+chr(10)
+    groups = {"PubMed":[],"Scholar":[]}
+    seen = set()
+    for r in results:
+        key = r["title"][:50].lower()
+        if key not in seen and r["url"]:
+            seen.add(key); groups[r["source"]].append(r)
+    for source, papers in groups.items():
+        if not papers: continue
+        out += "--- "+source+" ---"+chr(10)
+        for p in papers[:8]:
+            out += p["title"][:85]+" ("+p["year"]+")"
+            if p["citations"] not in ("N/A","","0"): out += " | "+p["citations"]+" citations"
+            out += chr(10)+"  "+p["url"]+chr(10)+chr(10)
+    out += "--- SJSU ScholarWorks ---"+chr(10)
+    out += "https://scholarworks.sjsu.edu/do/search/?q="+requests.utils.quote(query)+"&context=6781027"
+    return out
+
 def analyze_upad_photo(image):
     if image is None: return None, "Upload a uPAD photo first."
     try:
@@ -394,7 +421,7 @@ def analyze_piv_csv(file,theme="White"):
                 ax.plot(xv,df[vc],color="#c1121f",linewidth=2.5,marker="o",markersize=5)
                 ax.fill_between(xv,df[vc],alpha=0.15,color="#c1121f")
                 ax.axhline(y=2.0,color="#f59e0b",linestyle="--",linewidth=2,label="Risk: 2.0 m/s")
-                ax.set_ylabel("Velocity (m/s)",color=ac,fontsize=11); ax.legend(fontsize=9,labelcolor=fg,facecolor=pb)
+                ax.set_ylabel("Velocity (m/s)",color=ac); ax.legend(fontsize=9,labelcolor=fg,facecolor=pb)
         def ps(ax):
             if sc2:
                 xp=xv.values if tc else x
@@ -402,13 +429,13 @@ def analyze_piv_csv(file,theme="White"):
                 ax.fill_between(xp,df[sc2],alpha=0.15,color="#0057a8")
                 ax.axhline(y=5,color="#f59e0b",linestyle="--",linewidth=2,label="Caution 5 Pa")
                 ax.axhline(y=10,color="#c1121f",linestyle="--",linewidth=2,label="High risk 10 Pa")
-                ax.set_ylabel("Shear (Pa)",color=ac,fontsize=11); ax.legend(fontsize=9,labelcolor=fg,facecolor=pb)
+                ax.set_ylabel("Shear (Pa)",color=ac); ax.legend(fontsize=9,labelcolor=fg,facecolor=pb)
         def psc(ax):
             if vc and sc2:
                 s3=ax.scatter(df[vc],df[sc2],c=x,cmap="RdYlGn_r",s=90,edgecolors=fg,linewidth=0.5,zorder=5)
                 cb=plt.colorbar(s3,ax=ax,label="Time"); cb.ax.yaxis.label.set_color(fg); cb.ax.tick_params(colors=ac)
                 ax.axvline(x=2.0,color="#f59e0b",linestyle="--",linewidth=2); ax.axhline(y=10,color="#c1121f",linestyle="--",linewidth=2)
-                ax.set_xlabel("Velocity (m/s)",color=ac,fontsize=11); ax.set_ylabel("Shear (Pa)",color=ac,fontsize=11)
+                ax.set_xlabel("Velocity (m/s)",color=ac); ax.set_ylabel("Shear (Pa)",color=ac)
         def psum(ax):
             ax.axis("off"); risk=[]
             st="CLINICAL SUMMARY"+chr(10)+"="*20+chr(10)+chr(10)
@@ -460,22 +487,19 @@ def analyze_tgt_csv(file,theme="White"):
                         ax.fill_between(xp,yp,alpha=0.15,color=color)
                         for xi,yi in zip(xp,yp): ax.annotate(str(round(yi,1)),(xi,yi),textcoords="offset points",xytext=(0,10),ha="center",color=fg,fontsize=10,fontweight="bold")
                     ax.axhline(y=lim,color="#f59e0b",linestyle="--",linewidth=2.5,label=ll)
-                    ax.legend(fontsize=10,labelcolor=fg,facecolor=pb)
-                    ax.set_ylabel(yl,color=ac,fontsize=11)
+                    ax.legend(fontsize=10,labelcolor=fg,facecolor=pb); ax.set_ylabel(yl,color=ac)
                     mv=round(float(np.max(yp)),2)
                     ax.set_title(title+chr(10)+"Max: "+str(mv)+" - "+("HIGH" if mv>lim else "NORMAL"),color=fg,fontweight="bold",fontsize=12)
             return mk_chart(fn,title,bg,fg,gc,ac,pb)
-        i1=mk2(tatc,"#c1121f","TAT (ng/mL)",8,"Normal: 8","TAT Thrombin-Antithrombin")
-        i2=mk2(pfc,"#0057a8","PF1.2 (nmol/L)",2.0,"Normal: 2.0","PF1.2 Prothrombin Fragment")
-        i3=mk2(hc,"#2ecc71","Free Hemoglobin (mg/L)",20,"Normal: 20","Free Hemoglobin",bar=True)
-        i4=mk2(plc,"#e8a020","Platelet Count",150,"Normal min: 150","Platelet Count")
+        i1=mk2(tatc,"#c1121f","TAT (ng/mL)",8,"Normal: 8","TAT"); i2=mk2(pfc,"#0057a8","PF1.2",2.0,"Normal: 2.0","PF1.2")
+        i3=mk2(hc,"#2ecc71","Free Hgb (mg/L)",20,"Normal: 20","Free Hemoglobin",bar=True); i4=mk2(plc,"#e8a020","Platelets",150,"Normal>150","Platelets")
         ai=""
         if GROQ_KEY:
             try:
                 client=Groq(api_key=GROQ_KEY)
                 resp=client.chat.completions.create(model="llama-3.3-70b-versatile",
-                    messages=[{"role":"system","content":"Hematology expert SJSU CardioLab. Give thrombogenicity risk."},
-                    {"role":"user","content":"TGT from 27mm SJM Regent:"+chr(10)+df.describe().to_string()[:500]}],max_tokens=250)
+                    messages=[{"role":"system","content":"Hematology expert. Give thrombogenicity risk."},
+                    {"role":"user","content":"TGT:"+chr(10)+df.describe().to_string()[:500]}],max_tokens=250)
                 ai=chr(10)+"AI: "+resp.choices[0].message.content
             except: pass
         return i1,i2,i3,i4,"TGT: "+str(len(df))+" rows"+ai
@@ -504,27 +528,26 @@ def generate_image(prompt):
                 r=requests.post(url,headers=headers,json={"inputs":enhanced,"parameters":{"num_inference_steps":8}},timeout=60)
                 if r.status_code==200: return Image.open(io.BytesIO(r.content)),"Generated!",desc
             except: continue
-        return None,"Models busy. Try again.",desc
+        return None,"Models busy.",desc
     except Exception as e: return None,"Error: "+str(e),""
 
 def piv_manual(v,s,h):
-    vr="HIGH - stenosis" if float(v)>2.0 else "NORMAL"
-    sr="HIGH - thrombosis" if float(s)>10 else "ELEVATED" if float(s)>5 else "NORMAL"
-    return "Velocity: "+str(v)+" m/s - "+vr+chr(10)+"Shear: "+str(s)+" Pa - "+sr+chr(10)+"HR: "+str(h)+" bpm"
+    vr="HIGH-stenosis" if float(v)>2.0 else "NORMAL"
+    sr="HIGH-thrombosis" if float(s)>10 else "ELEVATED" if float(s)>5 else "NORMAL"
+    return "Velocity: "+str(v)+" m/s — "+vr+chr(10)+"Shear: "+str(s)+" Pa — "+sr+chr(10)+"HR: "+str(h)+" bpm"
 
 def tgt_manual(t,p,h,pl,tm):
     risk=sum([float(t)>15,float(p)>2.0,float(h)>50,float(pl)<150])
     return "TAT:"+str(t)+" PF1.2:"+str(p)+chr(10)+"Hemo:"+str(h)+" Plt:"+str(pl)+chr(10)+"RESULT: "+("HIGH RISK" if risk>=3 else "MODERATE" if risk>=2 else "LOW RISK")
 
-# ── UI ─────────────────────────────────────────────────────────────
 with gr.Blocks(title="CardioLab AI - SJSU", css=CSS) as demo:
     gr.HTML(HEADER)
 
     papers_count = len(set(m["paper"] for m in METADATA)) if PAPERS_LOADED else 0
-    chunks_count = len(CHUNKS) if PAPERS_LOADED else 0
-    status_color = "#27ae60" if PAPERS_LOADED else "#e67e22"
-    status_msg = f"RAG Active: {chunks_count} chunks from {papers_count} SJSU papers | AI reads actual lab papers before every answer" if PAPERS_LOADED else "Loading paper database..."
-    gr.HTML(f'''<div style="background:{status_color};color:white;text-align:center;padding:6px;font-size:0.8em;font-weight:700;">{status_msg}</div>''')
+    model_status = "CardioLab Fine-tuned Model LOADED" if CARDIOLAB_MODEL_LOADED else "Fine-tuned model loading..."
+    rag_status = f"RAG: {len(CHUNKS)} chunks from {papers_count} SJSU papers" if PAPERS_LOADED else "RAG: loading..."
+    gr.HTML(f'''<div style="background:#1a7340;color:white;text-align:center;padding:7px;font-size:0.82em;font-weight:700;">
+    {rag_status} | {model_status} | Select "CardioLab Fine-tuned (SJSU)" in Model dropdown to use your custom model!</div>''')
 
     with gr.Tabs():
         with gr.Tab("Chat"):
@@ -542,11 +565,11 @@ with gr.Blocks(title="CardioLab AI - SJSU", css=CSS) as demo:
                         delete_btn = gr.Button("Del", variant="secondary", scale=1)
                     session_status = gr.Textbox(label="", lines=1, interactive=False, container=False)
                 with gr.Column(scale=4):
-                    chatbot = gr.Chatbot(label="", height=480, show_label=False, container=False)
+                    chatbot = gr.Chatbot(label="", height=460, show_label=False, container=False)
                     with gr.Row():
-                        msg_box = gr.Textbox(placeholder="Ask anything about CardioLab — AI searches 16 SJSU papers + PubMed live...", label="", lines=2, scale=4, container=False)
-                        with gr.Column(scale=1, min_width=120):
-                            chat_model_dd = gr.Dropdown(choices=list(CHAT_MODELS.keys()), value="Llama 3.3 70B (Best)", label="Model")
+                        msg_box = gr.Textbox(placeholder="Ask anything — AI searches 16 SJSU papers + PubMed...", label="", lines=2, scale=4, container=False)
+                        with gr.Column(scale=1, min_width=160):
+                            chat_model_dd = gr.Dropdown(choices=list(CHAT_MODELS.keys()), value="Llama 3.3 70B (Best)", label="AI Model")
                             send_btn = gr.Button("Send", variant="primary")
                             clear_btn = gr.Button("Clear", variant="secondary")
             send_btn.click(research_chat, inputs=[msg_box, chatbot, chat_model_dd], outputs=[msg_box, chatbot])
@@ -567,12 +590,12 @@ with gr.Blocks(title="CardioLab AI - SJSU", css=CSS) as demo:
             voice_clear.click(lambda: [], outputs=voice_chatbot)
 
         with gr.Tab("Papers"):
-            gr.Markdown("### Search PubMed + Semantic Scholar + Europe PMC + SJSU ScholarWorks")
+            gr.Markdown("### Search PubMed + Semantic Scholar + SJSU ScholarWorks")
             with gr.Row():
-                search_input = gr.Textbox(placeholder="e.g. bileaflet mechanical heart valve hemodynamics thrombogenicity", label="Research Topic", scale=3)
+                search_input = gr.Textbox(placeholder="e.g. bileaflet mechanical heart valve thrombogenicity hemodynamics", label="Research Topic", scale=3)
                 search_model_dd = gr.Dropdown(choices=list(CHAT_MODELS.keys()), value="Llama 3.3 70B (Best)", label="AI Model", scale=1)
-                search_btn = gr.Button("Search All Sources", variant="primary", scale=1)
-            search_output = gr.Textbox(label="AI Ranked Results", lines=25)
+                search_btn = gr.Button("Search", variant="primary", scale=1)
+            search_output = gr.Textbox(label="Results", lines=22)
             search_btn.click(quick_search, inputs=[search_input, search_model_dd], outputs=search_output)
             search_input.submit(quick_search, inputs=[search_input, search_model_dd], outputs=search_output)
 
@@ -583,9 +606,9 @@ with gr.Blocks(title="CardioLab AI - SJSU", css=CSS) as demo:
             piv_btn = gr.Button("Analyze PIV Data", variant="primary")
             piv_result = gr.Textbox(label="AI Analysis", lines=4)
             with gr.Row():
-                piv_c1=gr.Image(label="Velocity Profile",type="pil"); piv_c2=gr.Image(label="Shear Stress",type="pil")
+                piv_c1=gr.Image(label="Velocity",type="pil"); piv_c2=gr.Image(label="Shear Stress",type="pil")
             with gr.Row():
-                piv_c3=gr.Image(label="Velocity vs Shear",type="pil"); piv_c4=gr.Image(label="Clinical Summary",type="pil")
+                piv_c3=gr.Image(label="Vel vs Shear",type="pil"); piv_c4=gr.Image(label="Clinical Summary",type="pil")
             piv_btn.click(analyze_piv_csv, inputs=[piv_file,piv_theme], outputs=[piv_c1,piv_c2,piv_c3,piv_c4,piv_result])
 
         with gr.Tab("TGT CSV"):
@@ -604,14 +627,14 @@ with gr.Blocks(title="CardioLab AI - SJSU", css=CSS) as demo:
             with gr.Row():
                 with gr.Column():
                     photo_input = gr.Image(label="Upload uPAD Photo", type="numpy", height=260)
-                    analyze_btn = gr.Button("Analyze uPAD Photo", variant="primary")
+                    analyze_btn = gr.Button("Analyze uPAD", variant="primary")
                 with gr.Column():
                     photo_img = gr.Image(label="Detection Zone", type="pil", height=260)
                     photo_text = gr.Textbox(label="CKD Result", lines=8)
             analyze_btn.click(analyze_upad_photo, inputs=photo_input, outputs=[photo_img, photo_text])
             with gr.Row():
                 r=gr.Number(label="R",value=210); g=gr.Number(label="G",value=140); b=gr.Number(label="B",value=80)
-                out3=gr.Textbox(label="Manual Result",lines=3)
+                out3=gr.Textbox(label="Result",lines=3)
             gr.Button("Analyze RGB",variant="secondary").click(
                 lambda r,g,b:"Creatinine: "+str(max(0,round(0.02*(r-b)-0.5,2)))+" mg/dL"+chr(10)+("Normal" if max(0,round(0.02*(r-b)-0.5,2))<1.2 else "Borderline" if max(0,round(0.02*(r-b)-0.5,2))<1.5 else "CKD"),
                 inputs=[r,g,b],outputs=out3)
@@ -620,7 +643,7 @@ with gr.Blocks(title="CardioLab AI - SJSU", css=CSS) as demo:
             with gr.Row():
                 img_prompt = gr.Textbox(placeholder="e.g. 27mm bileaflet mechanical heart valve cross section", label="Describe image", lines=2, scale=4)
                 with gr.Column(scale=1):
-                    img_btn = gr.Button("Generate Image", variant="primary")
+                    img_btn = gr.Button("Generate", variant="primary")
                     img_status = gr.Textbox(label="Status", lines=1)
             img_desc = gr.Textbox(label="AI Description", lines=2, interactive=False)
             img_output = gr.Image(label="Generated Image", type="pil", height=400)
@@ -638,10 +661,10 @@ with gr.Blocks(title="CardioLab AI - SJSU", css=CSS) as demo:
                 with gr.Column():
                     t1=gr.Number(label="TAT ng/mL",value=18); t2=gr.Number(label="PF1.2",value=2.5)
                     t3=gr.Number(label="Hemoglobin mg/L",value=60); t4=gr.Number(label="Platelets",value=140)
-                    t5=gr.Number(label="Time minutes",value=40); out2=gr.Textbox(label="Result",lines=6)
+                    t5=gr.Number(label="Time min",value=40); out2=gr.Textbox(label="Result",lines=6)
                     gr.Button("Analyze TGT",variant="primary").click(tgt_manual,inputs=[t1,t2,t3,t4,t5],outputs=out2)
 
     gr.HTML("""<div style="text-align:center;padding:10px;border-top:1px solid #e5e7eb;background:#f9fafb;">
-    <span style="color:#9ca3af;font-size:0.75em;">CardioLab AI v35 | SJSU Biomedical Engineering | RAG + 16 Papers Embedded | Inspired by <a href="https://github.com/snap-stanford/Biomni" style="color:#c1121f;">Biomni Stanford</a> | <a href="https://github.com/pranatechsol/Cardio-Lab-Ai" style="color:#0057a8;">GitHub</a> | Apache 2.0 | $0 Cost</span></div>""")
+    <span style="color:#9ca3af;font-size:0.75em;">CardioLab AI v36 | SJSU Biomedical Engineering | Fine-tuned on 16 SJSU Papers | RAG + Custom Model | Inspired by <a href="https://github.com/snap-stanford/Biomni" style="color:#c1121f;">Biomni Stanford</a> | <a href="https://github.com/pranatechsol/Cardio-Lab-Ai" style="color:#0057a8;">GitHub</a> | Apache 2.0 | $0 Cost</span></div>""")
 
 demo.launch()