Upload app.py with huggingface_hub

app.py

@@ -13,12 +13,7 @@ from huggingface_hub import HfApi, hf_hub_download
 GROQ_KEY = os.environ.get("GROQ_API_KEY", "")
 HF_TOKEN = os.environ.get("HF_TOKEN", "")
 HISTORY_REPO = "Saicharan21/cardiolab-chat-history"
-
-KNOWHOW = ("MCL: Sylgard 184 PDMS 10:1 ratio 48hr cure green laser PIV 70bpm 5L/min. "
-    "TGT: Arduino Uno Stepper Motor 150mL blood sampled at 0 20 40 60min measures TAT PF1.2 hemolysis platelets. "
-    "uPAD: Jaffe reaction creatinine plus picric acid gives orange-red color normal 0.6-1.2 mg/dL CKD above 1.5. "
-    "MHV: 27mm SJM Regent bileaflet also trileaflet monoleaflet pediatric. "
-    "Equipment: Heska HT5 hematology analyzer time-resolved PIV Tygon tubing Arduino Uno.")
+PAPERS_DB_REPO = "Saicharan21/cardiolab-papers-db"
 
 CHAT_MODELS = {
     "Llama 3.3 70B (Best)": "llama-3.3-70b-versatile",
@@ -27,6 +22,71 @@ CHAT_MODELS = {
     "Gemma 2 9B": "gemma2-9b-it",
 }
 
+KNOWHOW = ("MCL: Sylgard 184 PDMS 10:1 ratio 48hr cure green laser PIV 70bpm 5L/min. "
+    "TGT: Arduino Uno Stepper Motor 150mL blood 0 20 40 60min TAT PF1.2 hemolysis platelets. "
+    "uPAD: Jaffe reaction creatinine picric acid orange-red 0.6-1.2 mg/dL CKD above 1.5. "
+    "MHV: 27mm SJM Regent bileaflet trileaflet monoleaflet pediatric. "
+    "Equipment: Heska HT5 analyzer PIV green laser Tygon tubing Arduino Uno.")
+
+# ── LOAD PAPERS ON STARTUP ─────────────────────────────────────────
+CHUNKS = []
+METADATA = []
+EMBEDDINGS = None
+PAPERS_LOADED = False
+EMBEDDER = None
+
+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)
+        with open(chunks_path) as f: CHUNKS = json.load(f)
+        with open(meta_path) as f: METADATA = json.load(f)
+        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!")
+        return True
+    except Exception as e:
+        print(f"Paper load error: {e}")
+        return False
+
+load_papers()
+
+# ── 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:
+        q_emb = EMBEDDER.encode([query])
+        norms = np.linalg.norm(EMBEDDINGS, axis=1, keepdims=True)
+        emb_norm = EMBEDDINGS / (norms + 1e-10)
+        q_norm = q_emb / (np.linalg.norm(q_emb) + 1e-10)
+        scores = (emb_norm @ q_norm.T).flatten()
+        top_idx = np.argsort(scores)[::-1][:n]
+        context = ""
+        results = []
+        seen = set()
+        for idx in top_idx:
+            chunk = CHUNKS[idx]
+            meta = METADATA[idx]
+            score = float(scores[idx])
+            if score > 0.25:
+                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)
+                    seen.add(meta["paper"])
+                context += chunk[:500]+chr(10)
+        return context, results
+    except Exception as e:
+        print(f"Search error: {e}")
+        return "", []
+
 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; }
@@ -39,32 +99,32 @@ textarea { background: #ffffff !important; color: #1a202c !important; border: 1p
 button.primary { background: #c1121f !important; color: white !important; border: none !important; border-radius: 8px !important; font-weight: 600 !important; }
 button.secondary { background: #f3f4f6 !important; color: #374151 !important; border: 1px solid #d1d5db !important; border-radius: 8px !important; }
 input[type=number] { background: #f9fafb !important; color: #1a202c !important; border: 1px solid #d1d5db !important; border-radius: 8px !important; }
-label span { color: #374151 !important; font-weight: 500 !important; font-size: 0.85em !important; }
 """
 
-HEADER = """<div style="background:linear-gradient(135deg,#0a0f2e 0%,#1a0a0a 100%);padding:0;border-bottom:3px solid #c1121f;position:relative;overflow:hidden;">
-<svg style="position:absolute;top:0;left:0;width:100%;height:100%;opacity:0.07;" viewBox="0 0 1200 120" preserveAspectRatio="none">
+HEADER = """<div style="background:linear-gradient(135deg,#0a0f2e 0%,#1a0a0a 100%);padding:0;border-bottom:3px solid #c1121f;overflow:hidden;">
+<svg style="position:absolute;opacity:0.07;width:100%;height:100%;" viewBox="0 0 1200 120" preserveAspectRatio="none">
 <polyline points="0,60 100,60 130,20 150,100 170,10 200,90 220,60 400,60 430,20 450,100 470,10 500,90 520,60 700,60 730,20 750,100 770,10 800,90 820,60 1000,60 1030,20 1050,100 1070,10 1100,90 1120,60 1200,60" fill="none" stroke="#c1121f" stroke-width="3"/>
 </svg>
 <div style="max-width:1200px;margin:0 auto;padding:16px 24px;display:flex;align-items:center;justify-content:space-between;position:relative;z-index:1;">
 <div style="display:flex;align-items:center;gap:14px;">
 <svg width="55" height="55" viewBox="0 0 100 100"><circle cx="50" cy="35" r="28" fill="#0057a8" opacity="0.9"/><ellipse cx="50" cy="14" rx="22" ry="10" fill="#0057a8"/>
-<polygon points="30,14 33,4 36,14" fill="#e8a020"/><polygon points="36,12 39,2 42,12" fill="#e8a020"/><polygon points="42,11 45,1 48,11" fill="#e8a020"/>
-<polygon points="48,11 51,1 54,11" fill="#e8a020"/><polygon points="54,12 57,2 60,12" fill="#e8a020"/><polygon points="60,14 63,4 66,14" fill="#e8a020"/>
+<polygon points="30,14 33,4 36,14" fill="#e8a020"/><polygon points="36,12 39,2 42,12" fill="#e8a020"/>
+<polygon points="42,11 45,1 48,11" fill="#e8a020"/><polygon points="48,11 51,1 54,11" fill="#e8a020"/>
+<polygon points="54,12 57,2 60,12" fill="#e8a020"/><polygon points="60,14 63,4 66,14" fill="#e8a020"/>
 <rect x="36" y="30" width="28" height="22" rx="4" fill="#0057a8"/><rect x="40" y="35" width="8" height="12" rx="2" fill="#e8a020"/>
 <rect x="34" y="50" width="32" height="8" rx="4" fill="#0057a8"/></svg>
-<div><div style="color:#9ca3af;font-size:0.7em;font-weight:500;letter-spacing:2px;text-transform:uppercase;">San Jose State University</div>
+<div><div style="color:#9ca3af;font-size:0.7em;letter-spacing:2px;text-transform:uppercase;">San Jose State University</div>
 <div style="color:#e8a020;font-size:0.82em;font-weight:700;">Biomedical Engineering</div></div></div>
 <div style="text-align:center;flex:1;padding:0 20px;">
 <div style="display:flex;align-items:center;justify-content:center;gap:10px;margin-bottom:3px;">
-<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" stroke-linejoin="round"/></svg>
+<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" stroke-linejoin="round"/></svg></div>
-<div style="color:#9ca3af;font-size:0.7em;letter-spacing:2px;text-transform:uppercase;">AI Research Agent | Biomni Stanford | Llama 3.3 70B</div></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="display:flex;align-items:center;gap:14px;">
-<div style="text-align:right;"><div style="color:#9ca3af;font-size:0.68em;letter-spacing:1px;text-transform:uppercase;">Research Pillars</div>
+<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>
-<div style="color:#9ca3af;font-size:0.62em;margin-top:2px;">MCL · PIV · TGT · uPAD · COMSOL</div></div>
+<div style="color:#9ca3af;font-size:0.62em;margin-top:2px;">MCL  PIV  TGT  uPAD  COMSOL</div></div>
 <svg width="48" height="48" viewBox="0 0 100 90">
 <path d="M50 85 C50 85 5 55 5 30 C5 15 18 5 30 5 C38 5 45 9 50 15 C55 9 62 5 70 5 C82 5 95 15 95 30 C95 55 50 85 50 85Z" fill="#c1121f" opacity="0.9"/>
 <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>
@@ -75,7 +135,7 @@ def load_all_sessions():
     if not HF_TOKEN: return {}
     try:
         path = hf_hub_download(repo_id=HISTORY_REPO, filename="chat_history.json", repo_type="dataset", token=HF_TOKEN)
-        with open(path, "r") as f: return json.load(f)
+        with open(path) as f: return json.load(f)
     except: return {}
 
 def save_all_sessions(sessions):
@@ -83,8 +143,8 @@ def save_all_sessions(sessions):
     try:
         api2 = HfApi(token=HF_TOKEN)
         api2.upload_file(path_or_fileobj=json.dumps(sessions, indent=2).encode(),
-            path_in_repo="chat_history.json", repo_id=HISTORY_REPO, repo_type="dataset",
-            token=HF_TOKEN, commit_message="Update")
+            path_in_repo="chat_history.json", repo_id=HISTORY_REPO,
+            repo_type="dataset", token=HF_TOKEN, commit_message="Update")
         return True
     except: return False
 
@@ -125,16 +185,16 @@ def expand_query_ai(query, model_id="llama-3.3-70b-versatile"):
     try:
         client = Groq(api_key=GROQ_KEY)
         resp = client.chat.completions.create(model=model_id,
-            messages=[{"role":"system","content":"Biomedical PubMed search expert for SJSU CardioLab. Convert query to optimized MeSH terms and technical keywords. Focus on: mechanical heart valves, hemodynamics, blood flow, PIV, thrombogenicity, FSI, CFD, microfluidics, CKD, creatinine. Return ONLY search terms no explanation."},
-            {"role":"user","content":"Optimize for PubMed: "+query}],max_tokens=80)
+            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=8):
+def fetch_pubmed(query, n=6):
     try:
-        biofilter = " AND (heart valve OR hemodynamics OR microfluidic OR thrombogen OR creatinine OR PIV OR CFD OR CKD OR fluid structure)"
+        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":query+biofilter,"retmax":n,"retmode":"json","sort":"date","field":"tiab"},timeout=12)
+            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",
@@ -154,7 +214,7 @@ def fetch_pubmed(query, n=8):
         return results
     except: return []
 
-def fetch_scholar(query, n=8):
+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)
@@ -163,23 +223,22 @@ def fetch_scholar(query, n=8):
         for p in papers:
             year = p.get("year",0) or 0
             if int(year) < 2015: continue
-            results.append({"source":"Semantic Scholar","title":p.get("title",""),"year":str(year),
+            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=6):
+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","resulttype":"core"},timeout=12)
+            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","")
+            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,
@@ -187,106 +246,38 @@ def fetch_europe_pmc(query, n=6):
         return results
     except: return []
 
-def fetch_crossref(query, n=5):
-    try:
-        r = requests.get("https://api.crossref.org/works",
-            params={"query":query,"rows":n,"sort":"relevance","select":"title,DOI,published"},timeout=12)
-        items = r.json().get("message",{}).get("items",[])
-        results = []
-        for item in items:
-            title = item.get("title",[""])[0] if item.get("title") else ""
-            doi = item.get("DOI","")
-            pub = item.get("published",{}).get("date-parts",[[""]])[0]
-            year = str(pub[0]) if pub else ""
-            if year and int(year) < 2015: continue
-            if not doi: continue
-            results.append({"source":"CrossRef","title":title,"year":year,
-                "url":"https://doi.org/"+doi,"citations":"N/A"})
-        return results
-    except: return []
-
-def fetch_sjsu_scholarworks(query, n=6):
-    try:
-        # SJSU ScholarWorks Digital Commons search
-        r = requests.get("https://scholarworks.sjsu.edu/do/search/",
-            params={"q":query,"start":"0","context":"6781027","format":"json"},
-            timeout=12, headers={"User-Agent":"CardioLab-AI/1.0"})
-        results = []
-        if r.status_code == 200:
-            try:
-                data = r.json()
-                docs = data.get("response",{}).get("docs",[])
-                for doc in docs[:n]:
-                    title = doc.get("title","")
-                    year = str(doc.get("publication_date",""))[:4]
-                    url = doc.get("url","") or "https://scholarworks.sjsu.edu/"
-                    if title:
-                        results.append({"source":"SJSU ScholarWorks","title":str(title),"year":year,
-                            "url":url,"citations":"SJSU"})
-            except: pass
-        if not results:
-            # Fallback: provide direct SJSU search link
-            search_url = "https://scholarworks.sjsu.edu/do/search/?q="+requests.utils.quote(query)+"&context=6781027"
-            results.append({"source":"SJSU ScholarWorks",
-                "title":"Click to search SJSU ScholarWorks for: "+query,
-                "year":"","url":search_url,"citations":"SJSU"})
-        return results
-    except:
-        search_url = "https://scholarworks.sjsu.edu/do/search/?q="+requests.utils.quote(query)
-        return [{"source":"SJSU ScholarWorks","title":"Search SJSU ScholarWorks: "+query,
-            "year":"","url":search_url,"citations":"SJSU"}]
-
-def rank_with_ai(query, results, model_id="llama-3.3-70b-versatile"):
-    if not GROQ_KEY or not results: return results
-    try:
-        client = Groq(api_key=GROQ_KEY)
-        papers_text = chr(10).join([str(i+1)+". "+r["title"]+" ("+r["year"]+")" for i,r in enumerate(results[:15])])
-        resp = client.chat.completions.create(model=model_id,
-            messages=[{"role":"system","content":"Biomedical research expert. Rank papers by relevance to query. Return ONLY numbers comma separated. Example: 3,1,5,2,4"},
-            {"role":"user","content":"Query: "+query+chr(10)+"Papers:"+chr(10)+papers_text}],max_tokens=60)
-        order_text = resp.choices[0].message.content.strip()
-        order = [int(x.strip())-1 for x in order_text.split(",") if x.strip().isdigit()]
-        ranked = [results[i] for i in order if i < len(results)]
-        rest = [r for i,r in enumerate(results) if i not in order]
-        return ranked + rest
-    except: return results
-
 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=8)
-    r2 = fetch_scholar(expanded, n=8)
-    r3 = fetch_europe_pmc(expanded, n=6)
-    r4 = fetch_crossref(expanded, n=5)
-    r5 = fetch_sjsu_scholarworks(query, n=6)
-    all_results = r1 + r2 + r3 + r4 + r5
+    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)
-    ranked = rank_with_ai(query, unique, model_id)
-    out = "QUERY: "+query+chr(10)
-    out += "AI MODEL: "+search_model+chr(10)
-    out += "AI EXPANDED: "+expanded+chr(10)
-    out += "SOURCES: PubMed + Semantic Scholar + Europe PMC + CrossRef + SJSU ScholarWorks"+chr(10)
-    out += "TOTAL UNIQUE PAPERS: "+str(len(ranked))+chr(10)
+    out = "QUERY: "+query+chr(10)+"AI EXPANDED: "+expanded+chr(10)
     out += "="*45+chr(10)+chr(10)
-    groups = {"PubMed":[],"Semantic Scholar":[],"Europe PMC":[],"CrossRef":[],"SJSU ScholarWorks":[]}
-    for r in ranked[:25]:
+    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+" ("+str(len(papers))+" papers) ---"+chr(10)
+        out += "--- "+source+" ---"+chr(10)
         for p in papers:
             out += p["title"][:85]+" ("+p["year"]+")"
-            if p["citations"] not in ("N/A","SJSU",""): out += " | "+p["citations"]+" citations"
+            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(query, n=3):
+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)
@@ -295,7 +286,7 @@ def get_pubmed(query, n=3):
         return chr(10).join(["https://pubmed.ncbi.nlm.nih.gov/"+i for i in ids])
     except: return ""
 
-# ── CHAT ───────────────────────────────────────────────────────────
+# ── CHAT WITH RAG ──────────────────────────────────────────────────
 def research_chat(message, history, chat_model="Llama 3.3 70B (Best)"):
     if not GROQ_KEY:
         history.append({"role":"user","content":message})
@@ -304,14 +295,28 @@ 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)
-        msgs = [{"role":"system","content":"You are CardioLab AI for SJSU Biomedical Engineering. Expert in MHV MCL PIV TGT uPAD CKD FSI. Remember conversation. Never invent URLs. "+KNOWHOW}]
+        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)+
+                "SJSU CARDIOLAB PAPERS:"+chr(10)+paper_context+chr(10)+chr(10)+
+                "ADDITIONAL KNOWLEDGE: "+KNOWHOW)
+        else:
+            system_prompt = "You are CardioLab AI for SJSU Biomedical Engineering. Expert in MHV MCL PIV TGT uPAD CKD FSI. "+KNOWHOW
+        msgs = [{"role":"system","content":system_prompt}]
         for item in history:
             if isinstance(item, dict): msgs.append({"role":item["role"],"content":item["content"]})
         msgs.append({"role":"user","content":message})
-        resp = client.chat.completions.create(model=model_id,messages=msgs,max_tokens=700)
+        resp = client.chat.completions.create(model=model_id, messages=msgs, max_tokens=800)
         answer = resp.choices[0].message.content
-        pubmed = get_pubmed(message, n=3)
-        if pubmed: answer += chr(10)+chr(10)+"PubMed:"+chr(10)+pubmed
+        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":answer})
         return "", history
@@ -328,7 +333,10 @@ def voice_chat(audio, history):
         client = Groq(api_key=GROQ_KEY)
         with open(audio, "rb") as f:
             tx = client.audio.transcriptions.create(file=("audio.wav", f, "audio/wav"), model="whisper-large-v3")
-        msgs = [{"role":"system","content":"You are CardioLab AI. "+KNOWHOW}]
+        paper_context, _ = search_papers(tx.text, n=3)
+        system = "You are CardioLab AI. "+KNOWHOW
+        if paper_context: system = "You are CardioLab AI. Use these SJSU papers:"+chr(10)+paper_context+chr(10)+KNOWHOW
+        msgs = [{"role":"system","content":system}]
         for item in history:
             if isinstance(item, dict): msgs.append({"role":item["role"],"content":item["content"]})
         msgs.append({"role":"user","content":tx.text})
@@ -340,7 +348,6 @@ def voice_chat(audio, history):
         history.append({"role":"assistant","content":"Voice error: "+str(e)})
         return history
 
-# ── ANALYSIS TOOLS ─────────────────────────────────────────────────
 def analyze_upad_photo(image):
     if image is None: return None, "Upload a uPAD photo first."
     try:
@@ -393,8 +400,8 @@ def analyze_piv_csv(file,theme="White"):
                 xp=xv.values if tc else x
                 ax.plot(xp,df[sc2],color="#0057a8",linewidth=2.5,marker="s",markersize=5)
                 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.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)
         def psc(ax):
             if vc and sc2:
@@ -414,20 +421,18 @@ def analyze_piv_csv(file,theme="White"):
             st+="="*20+chr(10)+("OVERALL: HIGH RISK" if risk else "OVERALL: LOW RISK")
             ax.text(0.05,0.97,st,transform=ax.transAxes,color=fg,fontsize=10,va="top",fontfamily="monospace",
                 bbox=dict(boxstyle="round,pad=0.8",facecolor=pb,edgecolor=bc,linewidth=2.5))
-        i1=mk_chart(pv,"Velocity Profile",bg,fg,gc,ac,pb)
-        i2=mk_chart(ps,"Wall Shear Stress",bg,fg,gc,ac,pb)
-        i3=mk_chart(psc,"Velocity vs Shear",bg,fg,gc,ac,pb)
-        i4=mk_chart(psum,"Clinical Summary",bg,fg,gc,ac,pb)
+        i1=mk_chart(pv,"Velocity Profile",bg,fg,gc,ac,pb); i2=mk_chart(ps,"Wall Shear Stress",bg,fg,gc,ac,pb)
+        i3=mk_chart(psc,"Velocity vs Shear",bg,fg,gc,ac,pb); i4=mk_chart(psum,"Clinical Summary",bg,fg,gc,ac,pb)
         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":"PIV expert SJSU CardioLab. Analyze stats give clinical interpretation."},
-                    {"role":"user","content":"PIV from 27mm SJM Regent MHV 70bpm:"+chr(10)+df.describe().to_string()[:500]}],max_tokens=250)
+                    messages=[{"role":"system","content":"PIV expert SJSU CardioLab."},
+                    {"role":"user","content":"PIV from 27mm SJM Regent:"+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,"PIV: "+str(len(df))+" rows | "+", ".join(df.columns.tolist())+ai
+        return i1,i2,i3,i4,"PIV: "+str(len(df))+" rows"+ai
     except Exception as e: return None,None,None,None,"Error: "+str(e)
 
 def analyze_tgt_csv(file,theme="White"):
@@ -457,8 +462,8 @@ def analyze_tgt_csv(file,theme="White"):
                     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)
-                    mv=round(float(np.max(yp)),2); st="HIGH" if mv>lim else "NORMAL"
-                    ax.set_title(title+chr(10)+"Max: "+str(mv)+"  Status: "+st,color=fg,fontweight="bold",fontsize=12)
+                    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")
@@ -469,11 +474,11 @@ def analyze_tgt_csv(file,theme="White"):
             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 LOW MODERATE or HIGH."},
+                    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)
                 ai=chr(10)+"AI: "+resp.choices[0].message.content
             except: pass
-        return i1,i2,i3,i4,"TGT: "+str(len(df))+" rows | "+", ".join(df.columns.tolist())+ai
+        return i1,i2,i3,i4,"TGT: "+str(len(df))+" rows"+ai
     except Exception as e: return None,None,None,None,"Error: "+str(e)
 
 def generate_image(prompt):
@@ -514,12 +519,20 @@ def tgt_manual(t,p,h,pl,tm):
 # ── UI ─────────────────────────────────────────────────────────────
 with gr.Blocks(title="CardioLab AI - SJSU", css=CSS) as demo:
     gr.HTML(HEADER)
-    with gr.Tabs():
 
+    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>''')
+
+    with gr.Tabs():
         with gr.Tab("Chat"):
             with gr.Row():
                 with gr.Column(scale=1, min_width=200):
-                    gr.HTML('''<div style="background:#202123;padding:10px;border-radius:8px;margin-bottom:6px;"><div style="color:#e8a020;font-weight:700;font-size:0.85em;">SJSU CARDIOLAB</div><div style="color:#9ca3af;font-size:0.7em;">Conversations</div></div>''')
+                    gr.HTML('''<div style="background:#202123;padding:10px;border-radius:8px;margin-bottom:6px;">
+                    <div style="color:#e8a020;font-weight:700;font-size:0.85em;">SJSU CARDIOLAB</div>
+                    <div style="color:#9ca3af;font-size:0.7em;">Conversations</div></div>''')
                     new_chat_btn = gr.Button("New Chat", variant="secondary")
                     session_dropdown = gr.Dropdown(choices=get_session_list(), label="Saved Sessions", interactive=True)
                     load_btn = gr.Button("Load Session", variant="primary")
@@ -531,7 +544,7 @@ with gr.Blocks(title="CardioLab AI - SJSU", css=CSS) as demo:
                 with gr.Column(scale=4):
                     chatbot = gr.Chatbot(label="", height=480, show_label=False, container=False)
                     with gr.Row():
-                        msg_box = gr.Textbox(placeholder="Message CardioLab AI...", label="", lines=2, scale=4, container=False)
+                        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")
                             send_btn = gr.Button("Send", variant="primary")
@@ -554,33 +567,28 @@ 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 5 sources: PubMed + Semantic Scholar + Europe PMC + CrossRef + SJSU ScholarWorks")
+            gr.Markdown("### Search PubMed + Semantic Scholar + Europe PMC + SJSU ScholarWorks")
             with gr.Row():
-                search_input = gr.Textbox(placeholder="e.g. unsteady flow bileaflet mechanical heart valve hemodynamics", label="Research Topic", scale=3)
+                search_input = gr.Textbox(placeholder="e.g. bileaflet mechanical heart valve hemodynamics thrombogenicity", 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 5 Sources", variant="primary", 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.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)
-            gr.Markdown("**Try:** `bileaflet heart valve thrombogenicity` | `uPAD microfluidic creatinine CKD` | `PIV hemodynamics prosthetic valve` | `SJSU CardioLab biomedical`")
 
         with gr.Tab("PIV CSV"):
-            gr.Markdown("Upload PIV CSV - 4 separate charts + AI clinical analysis")
             with gr.Row():
                 piv_file = gr.File(label="Upload PIV CSV", file_types=[".csv"], scale=3)
                 piv_theme = gr.Radio(["White","Dark"], value="White", label="Theme", scale=1)
             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 Profile",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="Velocity 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"):
-            gr.Markdown("Upload TGT CSV - blood biomarker charts + thrombogenicity assessment")
             with gr.Row():
                 tgt_file = gr.File(label="Upload TGT CSV", file_types=[".csv"], scale=3)
                 tgt_theme = gr.Radio(["White","Dark"], value="White", label="Theme", scale=1)
@@ -610,7 +618,7 @@ with gr.Blocks(title="CardioLab AI - SJSU", css=CSS) as demo:
 
         with gr.Tab("AI Image"):
             with gr.Row():
-                img_prompt = gr.Textbox(placeholder="e.g. 27mm bileaflet mechanical heart valve cross section", label="Describe the image", lines=2, scale=4)
+                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_status = gr.Textbox(label="Status", lines=1)
@@ -621,24 +629,19 @@ with gr.Blocks(title="CardioLab AI - SJSU", css=CSS) as demo:
         with gr.Tab("PIV Manual"):
             with gr.Row():
                 with gr.Column():
-                    v=gr.Number(label="Max Velocity m/s",value=1.8,info="Normal: 0.5-2.0")
-                    s=gr.Number(label="Wall Shear Stress Pa",value=6.5,info="Normal: <5")
-                    h=gr.Number(label="Heart Rate bpm",value=72,info="Normal: 60-100")
-                    piv_out=gr.Textbox(label="Result",lines=4)
+                    v=gr.Number(label="Max Velocity m/s",value=1.8); s=gr.Number(label="Wall Shear Pa",value=6.5)
+                    h=gr.Number(label="Heart Rate bpm",value=72); piv_out=gr.Textbox(label="Result",lines=4)
                     gr.Button("Analyze PIV",variant="primary").click(piv_manual,inputs=[v,s,h],outputs=piv_out)
 
         with gr.Tab("TGT Manual"):
             with gr.Row():
                 with gr.Column():
-                    t1=gr.Number(label="TAT ng/mL",value=18,info="Normal: <8")
-                    t2=gr.Number(label="PF1.2 nmol/L",value=2.5,info="Normal: <2.0")
-                    t3=gr.Number(label="Free Hemoglobin mg/L",value=60,info="Normal: <20")
-                    t4=gr.Number(label="Platelet Count",value=140,info="Normal: >150")
-                    t5=gr.Number(label="Time minutes",value=40)
-                    out2=gr.Textbox(label="Result",lines=6)
+                    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)
                     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 | SJSU Biomedical Engineering | Built on <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 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>""")
 
 demo.launch()