vornith stability: auto-retry+reload, k=2 prompt budget, budget-sized warm ping, web auto-retry; revert keep_alive

app.py

@@ -68,10 +68,12 @@ def api_page():
 
 @app.route("/health")
 def health():
+    # No generation probe here: a tiny probe prompt on the shared resident instance
+    # is the vornith corruption trigger. Model presence via /api/tags (no GPU work).
     try:
-        r = http.post(f"{CFG['ollama_url']}/api/generate", stream=True, timeout=4,
+        r = http.get(f"{CFG['ollama_url']}/api/tags", timeout=4)
-                      json={"model": CFG["model"], "prompt": "ping", "stream": True, "options": {"num_predict": 1}})
+        models = [m.get("name") for m in r.json().get("models", [])]
-        llm = r.status_code == 200
+        llm = CFG["model"] in models
     except Exception:
         llm = False
     return jsonify(status="ok", llm=llm, model=CFG["model"], **core.get_stats())
@@ -122,6 +124,14 @@ def chat():
     return Response(generate(), mimetype="text/event-stream",
                     headers={"Cache-Control": "no-cache", "X-Accel-Buffering": "no"})
 
+@app.route("/api/warm")
+def api_warm():
+    """Heartbeat target for the systemd timer; also a manual health-probe for the model lane."""
+    healthy, sample = core.warm_model()
+    if not healthy and "CUDA" not in sample:
+        core.unload_model()  # corrupt instance: drop it; next warm reloads clean
+    return jsonify(healthy=healthy, sample=sample, model=CFG["model"])
+
 @app.route("/api/remaining")
 def remaining():
     ip = core.client_ip(request)

config.example.json

@@ -1,8 +1,9 @@
 {
-  "model": "ornith-1.5:9b-64k",
+  "model": "vornith:latest",
   "ollama_url": "http://10.30.20.29:11434",
   "num_ctx": 8192,
-  "rag_k": 6,
+  "rag_k": 2,
+  "num_predict": 1100,
   "port": 8012,
   "mcp_port": 8012,
   "base_url": "https://draco.thetempleofdoom.com",

deploy/draco-warm.service

@@ -0,0 +1,6 @@
+[Unit]
+Description=DRACO model keep-warm ping
+
+[Service]
+Type=oneshot
+ExecStart=/usr/bin/curl -s -m 130 http://127.0.0.1:8012/api/warm

deploy/draco-warm.timer

@@ -0,0 +1,9 @@
+[Unit]
+Description=DRACO keep model warm every 4 minutes
+
+[Timer]
+OnBootSec=60
+OnUnitActiveSec=240
+
+[Install]
+WantedBy=timers.target

deploy/draco.service

@@ -5,7 +5,7 @@ After=network.target
 [Service]
 Type=simple
 WorkingDirectory=/opt/draco
-ExecStart=/opt/draco/venv/bin/gunicorn -w 2 --threads 8 -b 127.0.0.1:8012 --timeout 300 app:app
+ExecStart=/opt/draco/venv/bin/gunicorn -w 2 --threads 8 -b 127.0.0.1:8012 --timeout 300 --graceful-timeout 10 app:app
 Restart=always
 RestartSec=5
 Environment=PYTHONUNBUFFERED=1

draco_core.py

@@ -20,6 +20,7 @@ CREATE TABLE IF NOT EXISTS users (
     api_key VARCHAR(64) UNIQUE NOT NULL,
     credits INTEGER DEFAULT 0,
     free_used INTEGER DEFAULT 0,
+    total_calls INTEGER DEFAULT 0,
     is_admin INTEGER DEFAULT 0,
     created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
     last_used_at TIMESTAMP
@@ -58,6 +59,11 @@ def db():
 def init_db():
     c = db()
     c.executescript(SCHEMA)
+    try:  # in-place migration for DBs created before total_calls existed
+        c.execute("SELECT total_calls FROM users LIMIT 1")
+    except sqlite3.OperationalError:
+        c.execute("ALTER TABLE users ADD COLUMN total_calls INTEGER DEFAULT 0")
+        c.commit()
     c.execute("INSERT OR IGNORE INTO users (email, api_key, credits, is_admin) VALUES (?,?,?,1)",
               ("admin@draco.local", CFG["admin_key"], 999999))
     c.commit()
@@ -127,12 +133,17 @@ SYSTEM = (
     "Answer directly with no reasoning preamble and no meta commentary."
 )
 
-def build_prompt(question, k=6):
+def build_prompt(question, k=None):
+    """RAG prompt. PROMPT BUDGET RULE: all prompts must land in the same size class
+    (~600 tok). vornith (linear-attn+MTP hybrid) corrupts when one resident instance
+    receives mixed tiny/long prompts; uniform bounded prompts are proven stable."""
+    k = k or CFG.get("rag_k", 2)
+    k = min(k, 3)  # hard cap: 3 x 1200-char excerpts ~= clean-zone prompt
     hits = search_library(question, k=k)
     if hits:
         blocks = []
         for n, h in enumerate(hits, 1):
-            blocks.append(f"[{n}] {h['title']} ({h['category']})\n{h['text'][:1400]}")
+            blocks.append(f"[{n}] {h['title']} ({h['category']})\n{h['text'][:1200]}")
         ctx = "\n\n".join(blocks)
         prompt = f"{SYSTEM}\n\nBOOK EXCERPTS:\n{ctx}\n\nQUESTION: {question}\n\nANSWER (cite [n]):"
     else:
@@ -314,6 +325,9 @@ def stream_ollama(prompt):
     """Yield (channel, piece) tuples: channel 'think' or 'answer'."""
     r = http.post(f"{CFG['ollama_url']}/api/generate",
                   json={"model": CFG["model"], "prompt": prompt, "stream": True,
+                        # RESIDENT model: cold-load prefill >1k tokens CUDA-crashes on this
+                        # hybrid arch. Corruption (????? output) is handled by detection +
+                        # auto-reload in the app layer instead.
                         "options": {"temperature": 0.4, "num_predict": CFG.get("num_predict", 1100),
                                     "num_ctx": CFG.get("num_ctx", 8192)}},
                   timeout=(5, None), stream=True)
@@ -328,24 +342,55 @@ def stream_ollama(prompt):
         yield channel, piece
 
 def ask_ollama(prompt):
-    """Non-streaming RAG answer (reasoning stripped, tagged or not)."""
+    """Non-streaming RAG answer with auto-recovery: on degenerate/CUDA failure,
+    unload the model, reload fresh, retry once before surfacing an error."""
+    last_err = None
+    for attempt in (1, 2):
+        try:
             r = http.post(f"{CFG['ollama_url']}/api/generate",
                           json={"model": CFG["model"], "prompt": prompt, "stream": False,
                                 "options": {"temperature": 0.4, "num_predict": CFG.get("num_predict", 1100),
                                             "num_ctx": CFG.get("num_ctx", 8192)}},
                           timeout=180)
             r.raise_for_status()
-    _, answer = strip_cot(r.json().get("response", ""))
+            raw = r.json().get("response", "")
-    if degenerate(answer):
+            if raw and degenerate(raw):
-        unload_model()
+                last_err = RuntimeError("degenerate output")
-        raise RuntimeError("model returned degenerate output — model unloaded, retry")
+            else:
+                _, answer = strip_cot(raw)
                 return answer
+        except http.HTTPError as e:
+            last_err = e
+        except Exception as e:
+            last_err = e
+        unload_model()  # force clean reload for the next attempt
+        time.sleep(2)
+    raise RuntimeError(f"model unstable after retry: {last_err}")
 
 def degenerate(s):
     """vornith VRAM/session corruption signature: run of '?' chars."""
     s = s.strip()
     return len(s) >= 40 and s.count("?") / len(s) > 0.4
 
+def warm_model():
+    """Budget-sized heartbeat: keeps the model resident AND exercised at the canonical
+    prompt size. Tiny prompts (<50 tok) on the shared instance are the corruption
+    trigger, so the ping itself must be RAG-sized."""
+    try:
+        excerpt = ("The utility of a uniform prompt budget is that the model never "
+                   "encounters a context-length distribution shift between requests. " * 9)
+        prompt = (f"{SYSTEM}\n\nBOOK EXCERPTS:\n[1] Warmup Excerpt (maintenance)\n{excerpt}"
+                  "\n\nQUESTION: Reply with exactly: ok\n\nANSWER (cite [n]):")
+        r = http.post(f"{CFG['ollama_url']}/api/generate",
+                      json={"model": CFG["model"], "prompt": prompt,
+                            "stream": False, "options": {"num_predict": 4}},
+                      timeout=120)
+        sample = r.json().get("response", "") if r.status_code == 200 else ""
+        healthy = r.status_code == 200 and not degenerate(sample)
+        return healthy, sample[:120]
+    except Exception as e:
+        return False, str(e)[:120]
+
 def unload_model():
     """Drop the model from VRAM so the next request reloads clean."""
     try:

pages.py

@@ -140,12 +140,14 @@ async function ask(text){
  add('you','user').textContent=text;
  const body=add('draco','assistant');body.innerHTML='<span class=cursor></span>';
  const srcs=document.createElement('div');srcs.className='srcs';srcs.style.display='none';
+ for(let attempt=1;attempt<=3;attempt++){
+  let hadTokens=false,hadError=null,acc='';
   try{
    const r=await fetch('/api/chat',{method:'POST',headers:{'Content-Type':'application/json'},
     body:JSON.stringify({q:text})});
    if(!r.ok){const t=await r.text();let m='HTTP '+r.status;try{m=JSON.parse(t).error||m}catch(e){}
     body.textContent='⚠ '+m;busy=false;go.disabled=false;return;}
-  const reader=r.body.getReader(),dec=new TextDecoder();let buf='',acc='',thinkEl=null;
+   const reader=r.body.getReader(),dec=new TextDecoder();let buf='',thinkEl=null;
    while(true){const{value,done}=await reader.read();if(done)break;
     buf+=dec.decode(value,{stream:true});const lines=buf.split('\\n');buf=lines.pop()||'';
     for(const line of lines){if(!line.startsWith('data: '))continue;
@@ -159,14 +161,19 @@ async function ask(text){
        body.parentNode.insertBefore(thinkEl,body);}
       thinkEl.querySelector('.tbody').textContent+=ev.text;
       log.scrollTop=log.scrollHeight;
-    }else if(ev.type==='token'){acc+=ev.text;body.innerHTML='';
+     }else if(ev.type==='token'){acc+=ev.text;hadTokens=true;body.innerHTML='';
       body.appendChild(document.createTextNode(acc));body.appendChild(document.createElement('span')).className='cursor';
       log.scrollTop=log.scrollHeight;
-    }else if(ev.type==='error'){acc+='\n⚠ '+ev.text;}
+     }else if(ev.type==='error'){hadError=ev.text;}
     }}
- }catch(e){acc+='\n⚠ connection lost';}
+  }catch(e){hadError='connection lost';}
- body.innerHTML='';body.appendChild(document.createTextNode(acc||'(no output)'));
+  if(hadTokens&&!hadError){break;}
+  if(attempt<3){body.innerHTML='<span class=cursor></span>';await new Promise(r=>setTimeout(r,1500));}
+  else if(hadError){body.textContent='⚠ '+hadError;}
+  else break;
+ }
  const cur=body.querySelector('.cursor');if(cur)cur.remove();
+ if(!body.textContent.trim())body.textContent='(no output)';
  body.appendChild(srcs);
  const left=document.getElementById('left');
  fetch('/api/remaining').then(r=>r.json()).then(d=>left.textContent=d.remaining+' free questions left today').catch(()=>{});