app/core/ai.py

"""Multi-provider AI engine with smart task routing.

Runtime chain: Groq -> Cerebras -> OpenRouter -> Mistral -> Ollama.
Task hints route to the best model for the job:
  - "arabic"   → large models (70B+) for Arabic NLP quality
  - "code"     → code-optimized models
  - "fast"     → smallest/fastest model available
  - "default"  → standard free-tier chain
"""
import json, logging, os, re, requests

logger = logging.getLogger(__name__)
_OLLAMA_BASE = "http://localhost:11434"

_PROVIDER_URLS = {
    "groq":       "https://api.groq.com/openai/v1/chat/completions",
    "cerebras":   "https://api.cerebras.ai/v1/chat/completions",
    "openrouter": "https://openrouter.ai/api/v1/chat/completions",
    "mistral":    "https://api.mistral.ai/v1/chat/completions",
    "openai":     "https://api.openai.com/v1/chat/completions",
    "cohere":     "https://api.cohere.com/v2/chat",
}

# ── Model tiers per provider ──
_FREE_MODELS = {
    "groq":       "llama-3.1-8b-instant",
    "cerebras":   "llama3.1-8b",
    "openrouter": "google/gemma-3-12b-it:free",
    "mistral":    "mistral-small-latest",
    "cohere":     "command-r-08-2024",
}
_PREMIUM_MODELS = {
    "groq":       "llama-3.3-70b-versatile",
    "cerebras":   "qwen-3-235b-a22b-instruct-2507",
    "openrouter": "google/gemma-3-27b-it:free",
    "mistral":    "mistral-medium-latest",
    "openai":     "gpt-4o-mini",
    "cohere":     "command-r-08-2024",
}

# ── Task-specific model routing ──
# Maps task hints to the best model per provider.
# "arabic" needs large models for Arabic morphology, grammar, dialect awareness.
# "code" needs code-tuned models for test generation, SQL, schema analysis.
# "fast" uses smallest models for quick responses.
_TASK_MODELS = {
    "arabic": {
        "groq":       "llama-3.3-70b-versatile",
        "cerebras":   "qwen-3-235b-a22b-instruct-2507",
        "openrouter": "google/gemma-3-27b-it:free",
        "mistral":    "mistral-medium-latest",
        "cohere":     "command-r7b-arabic-02-2025",
    },
    "code": {
        "groq":       "llama-3.3-70b-versatile",
        "cerebras":   "qwen-3-235b-a22b-instruct-2507",
        "openrouter": "google/gemma-3-27b-it:free",
        "mistral":    "mistral-medium-latest",
        "cohere":     "command-r-08-2024",
    },
    "fast": {
        "groq":       "llama-3.1-8b-instant",
        "cerebras":   "llama3.1-8b",
        "openrouter": "google/gemma-3-12b-it:free",
        "mistral":    "mistral-small-latest",
        "cohere":     "command-r-08-2024",
    },
}

# ── Task-specific provider priority ──
_TASK_PRIORITY = {
    "arabic":  ["cerebras", "groq", "openrouter", "cohere", "mistral"],
    "code":    ["groq", "cerebras", "openrouter", "cohere", "mistral"],
    "fast":    ["cerebras", "groq", "openrouter", "cohere", "mistral"],
    "default": ["groq", "cerebras", "openrouter", "cohere", "mistral"],
}

_CHAIN_CFG = {
    "groq":       {"key_env": "GROQ_API_KEY",       "timeout": 30, "extra": {}},
    "cerebras":   {"key_env": "CEREBRAS_API_KEY",   "timeout": 30, "extra": {}},
    "openrouter": {"key_env": "OPENROUTER_API_KEY", "timeout": 45,
                   "extra": {"HTTP-Referer": "https://github.com/Moealsarraj", "X-Title": "AI Tools"}},
    "mistral":    {"key_env": "MISTRAL_API_KEY",    "timeout": 40, "extra": {}},
    "cohere":     {"key_env": "COHERE_API_KEY",     "timeout": 45, "extra": {}},
}

# Build available providers (those with valid keys)
_AVAILABLE = {}
for _name, _cfg in _CHAIN_CFG.items():
    _k = os.environ.get(_cfg["key_env"], "")
    if _k:
        _AVAILABLE[_name] = {
            "name":    _name,
            "url":     _PROVIDER_URLS[_name],
            "key":     _k,
            "timeout": _cfg["timeout"],
            "extra":   _cfg["extra"],
        }

# Ollama fallback
_OLLAMA_PROVIDER = None
try:
    _r = requests.get(f"{_OLLAMA_BASE}/api/tags", timeout=3)
    if _r.status_code == 200:
        _installed = [m["name"] for m in _r.json().get("models", [])]
        if _installed:
            _OLLAMA_PROVIDER = {"name": "ollama", "model": _installed[0]}
except Exception:
    pass

# ── Google Gemini (special API format) ──
_GEMINI_KEY = os.environ.get("GEMINI_API_KEY", "")
if _GEMINI_KEY:
    _AVAILABLE["gemini"] = {
        "name": "gemini",
        "url": "https://generativelanguage.googleapis.com/v1beta/models",
        "key": _GEMINI_KEY,
        "timeout": 60,
        "extra": {},
    }
    _FREE_MODELS["gemini"] = "gemini-2.0-flash"
    _PREMIUM_MODELS["gemini"] = "gemini-2.0-flash"
    for task in _TASK_MODELS:
        _TASK_MODELS[task]["gemini"] = "gemini-2.0-flash"
    for task in _TASK_PRIORITY:
        if "gemini" not in _TASK_PRIORITY[task]:
            _TASK_PRIORITY[task].insert(2, "gemini")

_AI_AVAILABLE = bool(_AVAILABLE or _OLLAMA_PROVIDER)


def _post_gemini(key: str, model: str, messages: list, max_tokens: int, timeout: int = 60) -> str:
    """Call Google Gemini API (non-OpenAI format)."""
    # Convert OpenAI message format to Gemini format
    contents = []
    system_text = ""
    for msg in messages:
        role = msg["role"]
        if role == "system":
            system_text = msg["content"]
            continue
        contents.append({
            "role": "user" if role == "user" else "model",
            "parts": [{"text": msg["content"]}],
        })

    body = {
        "contents": contents,
        "generationConfig": {"maxOutputTokens": max_tokens},
    }
    if system_text:
        body["systemInstruction"] = {"parts": [{"text": system_text}]}

    url = f"https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent?key={key}"
    r = requests.post(url, json=body, timeout=timeout)
    r.raise_for_status()
    data = r.json()
    return _clean(data["candidates"][0]["content"]["parts"][0]["text"])


def get_available_providers() -> list[dict]:
    """Return list of available providers with their model info."""
    providers = []
    for name, prov in _AVAILABLE.items():
        providers.append({
            "name": name,
            "model_free": _FREE_MODELS.get(name, ""),
            "model_premium": _PREMIUM_MODELS.get(name, ""),
        })
    return providers


def call_ai_single(provider_name: str, messages: list, system: str = "",
                   max_tokens: int = 2048, use_premium: bool = True) -> str:
    """Call a specific provider directly (no fallback chain)."""
    if provider_name not in _AVAILABLE:
        raise ValueError(f"Provider {provider_name!r} not available")
    prov = _AVAILABLE[provider_name]
    models = _PREMIUM_MODELS if use_premium else _FREE_MODELS
    model = models.get(provider_name, prov.get("model", ""))
    if system:
        messages = [{"role": "system", "content": system}] + messages
    if provider_name == "gemini":
        return _post_gemini(prov["key"], model, messages, max_tokens, prov["timeout"])
    if provider_name == "cohere":
        return _post_cohere(prov["key"], model, messages, max_tokens, prov["timeout"])
    return _post_openai(
        prov["url"], prov["key"], model,
        messages, max_tokens, prov["extra"], prov["timeout"]
    )


_RE_THINK = re.compile(r"<think>.*?</think>", re.DOTALL)
_RE_OPEN  = re.compile(r"^```[a-z]*\n?", re.MULTILINE)
_RE_CLOSE = re.compile(r"\n?```$", re.MULTILINE)

def _clean(raw: str) -> str:
    raw = _RE_THINK.sub("", raw).strip()
    raw = _RE_OPEN.sub("", raw)
    return _RE_CLOSE.sub("", raw).strip()

def _post_openai(url, key, model, messages, max_tokens, extra_headers, timeout=60):
    headers = {"Authorization": f"Bearer {key}", "Content-Type": "application/json"}
    headers.update(extra_headers)
    r = requests.post(url, headers=headers,
        json={"model": model, "messages": messages, "max_tokens": max_tokens},
        timeout=timeout)
    r.raise_for_status()
    return _clean(r.json()["choices"][0]["message"]["content"])


def _post_cohere(key: str, model: str, messages: list, max_tokens: int, timeout: int = 45) -> str:
    """Call Cohere V2 Chat API."""
    headers = {"Authorization": f"Bearer {key}", "Content-Type": "application/json"}
    r = requests.post("https://api.cohere.com/v2/chat",
        headers=headers,
        json={"model": model, "messages": messages, "max_tokens": max_tokens},
        timeout=timeout)
    r.raise_for_status()
    data = r.json()
    # V2 returns content as list of blocks
    content = data.get("message", {}).get("content", [])
    if content and isinstance(content, list):
        return _clean(content[0].get("text", ""))
    return _clean(str(data))


def _build_chain(task_hint: str) -> list[dict]:
    """Build an ordered provider chain for the given task hint."""
    hint = task_hint if task_hint in _TASK_PRIORITY else "default"
    priority = _TASK_PRIORITY[hint]
    models = _TASK_MODELS.get(hint, _FREE_MODELS)

    chain = []
    for name in priority:
        if name in _AVAILABLE:
            prov = _AVAILABLE[name].copy()
            prov["model"] = models.get(name, _FREE_MODELS.get(name, ""))
            chain.append(prov)
    return chain


def call_ai(messages: list, system: str = "", max_tokens: int = 2048,
            api_key_row: dict | None = None, task_hint: str = "default") -> str:
    """Call AI with smart task-based routing.

    task_hint: "arabic" | "code" | "fast" | "default"
    """
    if system:
        messages = [{"role": "system", "content": system}] + messages
    # Custom API key path (used by e.g. Wasit/Amin integrations)
    if api_key_row:
        provider = api_key_row.get("provider", "openai")
        key      = api_key_row["key"]
        url      = api_key_row.get("url") or _PROVIDER_URLS.get(provider, "")
        model    = api_key_row.get("model") or _PREMIUM_MODELS.get(provider, "gpt-4o-mini")
        if not url:
            raise ValueError(f"No endpoint known for provider {provider!r}")
        if provider == "claude":
            r = requests.post("https://api.anthropic.com/v1/messages",
                headers={"x-api-key": key, "anthropic-version": "2023-06-01",
                         "content-type": "application/json"},
                json={"model": "claude-sonnet-4-6", "max_tokens": max_tokens, "messages": messages},
                timeout=60)
            r.raise_for_status()
            return _clean(r.json()["content"][0]["text"])
        return _post_openai(url, key, model, messages, max_tokens, {})
    if not _AI_AVAILABLE:
        raise RuntimeError("No AI provider. Set GROQ_API_KEY or similar in .env")
    # Ollama-only path
    if not _AVAILABLE and _OLLAMA_PROVIDER:
        r = requests.post(f"{_OLLAMA_BASE}/api/chat",
            json={"model": _OLLAMA_PROVIDER["model"], "messages": messages, "stream": False},
            timeout=120)
        r.raise_for_status()
        return _clean(r.json()["message"]["content"])
    # Smart task-routed chain
    chain = _build_chain(task_hint)
    if not chain:
        chain = _build_chain("default")

    last_exc = None
    for prov in chain:
        try:
            logger.debug("Trying %s/%s for task=%s", prov["name"], prov["model"], task_hint)
            if prov["name"] == "gemini":
                return _post_gemini(prov["key"], prov["model"], messages, max_tokens, prov["timeout"])
            if prov["name"] == "cohere":
                return _post_cohere(prov["key"], prov["model"], messages, max_tokens, prov["timeout"])
            return _post_openai(
                prov["url"], prov["key"], prov["model"],
                messages, max_tokens, prov["extra"], prov["timeout"]
            )
        except requests.exceptions.HTTPError as e:
            status = e.response.status_code if e.response is not None else 0
            if status in (402, 429, 503, 502):
                logger.debug("Provider %s returned %s, trying next", prov["name"], status)
                last_exc = e
                continue
            raise
        except (requests.exceptions.ConnectionError,
                requests.exceptions.Timeout) as e:
            last_exc = e
            continue
    # Try Ollama as last resort
    if _OLLAMA_PROVIDER:
        r = requests.post(f"{_OLLAMA_BASE}/api/chat",
            json={"model": _OLLAMA_PROVIDER["model"], "messages": messages, "stream": False},
            timeout=120)
        r.raise_for_status()
        return _clean(r.json()["message"]["content"])
    raise last_exc or RuntimeError("All AI providers failed or rate-limited")

def _repair_json(text: str) -> str:
    """Escape literal control characters inside JSON string values."""
    result = []
    in_str = False
    esc = False
    for c in text:
        if esc:
            result.append(c)
            esc = False
            continue
        if c == '\\' and in_str:
            result.append(c)
            esc = True
            continue
        if c == '"':
            in_str = not in_str
            result.append(c)
            continue
        if in_str and c == '\n':
            result.append('\\n')
            continue
        if in_str and c == '\r':
            result.append('\\r')
            continue
        if in_str and c == '\t':
            result.append('\\t')
            continue
        result.append(c)
    return ''.join(result)

def _extract_json(raw: str):
    """Try progressively harder to extract valid JSON from raw text."""
    raw = raw.strip()
    # Direct parse
    try:
        return json.loads(raw)
    except json.JSONDecodeError:
        pass
    # Repair literal newlines inside strings then retry
    repaired = _repair_json(raw)
    try:
        return json.loads(repaired)
    except json.JSONDecodeError:
        pass
    # Find first { or [ then walk to find matching closer
    for source in (repaired, raw):
        for start_ch, end_ch in [('{', '}'), ('[', ']')]:
            idx = source.find(start_ch)
            if idx == -1:
                continue
            depth = 0
            in_str = False
            esc = False
            for i in range(idx, len(source)):
                c = source[i]
                if esc:
                    esc = False
                    continue
                if c == '\\' and in_str:
                    esc = True
                    continue
                if c == '"':
                    in_str = not in_str
                    continue
                if in_str:
                    continue
                if c == start_ch:
                    depth += 1
                elif c == end_ch:
                    depth -= 1
                    if depth == 0:
                        candidate = source[idx:i+1]
                        try:
                            return json.loads(candidate)
                        except json.JSONDecodeError:
                            break
    raise ValueError(f"AI returned non-JSON: {raw[:200]}")

def call_ai_json(messages: list, system: str = "", max_tokens: int = 2048,
                 api_key_row: dict | None = None, task_hint: str = "default") -> dict | list:
    raw = call_ai(messages, system=system, max_tokens=max_tokens,
                  api_key_row=api_key_row, task_hint=task_hint)
    return _extract_json(raw)