OLLAMA_URL = "http://localhost:11434"
MODEL = "qwen2.5:3b"

# Model parameters
options = {
    "temperature": 0.0,        # Deterministic responses
    "num_predict": 150,        # Max tokens per response (300 for lists)
    "num_ctx": 2048,           # Context window (adaptive: 1024-3072)
    "num_thread": 6,           # Physical cores only (Ryzen 5 2600)
    "num_batch": 512,          # Large batch → lower TTFT
    "repeat_penalty": 1.1,
}

CARPETA_FUENTES = "/opt/siaa/fuentes"

# Document structure
doc_entry = {
    "ruta": "/opt/siaa/fuentes/acuerdo_psaa16.md",
    "nombre_original": "acuerdo_psaa16.md",
    "contenido": "...",
    "palabras": set(),           # Tokenized vocabulary
    "tamano": 45231,             # Character count
    "coleccion": "general",
    "token_count": Counter(),    # Term frequency index
    "total_tokens": 1542,
    "tokens_nombre": {"acuerdo", "psaa16"},
    "num_chunks": 38,            # Pre-calculated chunks
}

CACHE_MAX_ENTRADAS = 200    # Maximum entries
CACHE_TTL_SEGUNDOS = 3600   # 1 hour TTL
CACHE_SOLO_DOC = True       # Only cache documentary queries

# Cache entry structure
{
    "respuesta": "El SIERJU es un sistema de información...",
    "cita": "📄 Fuente: ACUERDO PSAA16-10476",
    "ts": 1709856234.5,
    "hits": 12,
}

ollama_estado = {
    "disponible": False,
    "ultimo_check": 0,
    "fallos": 0,
    "warmup_done": None
}

def verificar_ollama() -> bool:
    try:
        r = requests.get(f"{OLLAMA_URL}/api/tags", timeout=TIMEOUT_HEALTH)
        ok = (r.status_code == 200)
    except Exception:
        ok = False
    
    with ollama_lock:
        ollama_estado["disponible"] = ok
        ollama_estado["ultimo_check"] = time.time()
        ollama_estado["fallos"] = 0 if ok else ollama_estado["fallos"] + 1
        warmup_pendiente = ok and ollama_estado["warmup_done"] is None
    
    # Warm-up: load model into RAM on first success
    if warmup_pendiente:
        print(f"  [Ollama] Precargando {MODEL} en RAM...", flush=True)
        requests.post(
            f"{OLLAMA_URL}/api/chat",
            json={"model": MODEL, "messages": [{"role": "user", "content": "ok"}],
                  "stream": False, "options": {"num_predict": 1, "num_ctx": 64}}
        )
        ollama_estado["warmup_done"] = True
    
    return ok

# Background monitoring thread
def _monitor_loop():
    while True:
        verificar_ollama()
        time.sleep(15)

threading.Thread(target=_monitor_loop, daemon=True).start()

def registrar_consulta(
    tipo: str,          # "CONV", "DOC", "CACHE_HIT", "ERROR"
    pregunta: str,
    respuesta: str,
    docs: list,
    ctx_chars: int,
    tiempo_seg: float,
    cache_hit: bool = False,
):
    # Detect issues automatically
    no_encontro = "no encontré esa información" in respuesta.lower()
    habia_docs = len(docs) > 0 and ctx_chars > 100
    
    if no_encontro and habia_docs:
        alerta = "POSIBLE_ALUCINACION"   # Had docs but said "not found"
    elif no_encontro and not habia_docs:
        alerta = "SIN_CONTEXTO"           # No docs available (correct)
    elif tipo == "ERROR":
        alerta = "ERROR"
    else:
        alerta = "OK"
    
    entrada = {
        "ts": time.strftime("%Y-%m-%dT%H:%M:%S"),
        "tipo": "CACHE_HIT" if cache_hit else tipo,
        "alerta": alerta,
        "pregunta": pregunta[:200],
        "respuesta": respuesta[:300],
        "docs": docs,
        "ctx_chars": ctx_chars,
        "tiempo_s": round(tiempo_seg, 2),
    }
    
    # Write to JSONL with rotation at 5000 lines
    with _log_lock:
        with open(LOG_ARCHIVO, "a", encoding="utf-8") as f:
            f.write(json.dumps(entrada, ensure_ascii=False) + "\n")