api/vision.py

"""
vision.py — Generazione e analisi immagini + ricerca immagini.

Endpoints:
  POST /api/vision/generate  — FLUX.1-schnell (HF Inference API)
  POST /api/vision/analyze   — Groq llama-3.2-vision / GPT-4o-mini / BLIP fallback
  GET  /api/vision/search    — Pexels > Pixabay > Unsplash Source (zero API key)

Problematiche HF Inference API:
  - 503 "loading": cold-start fino a 60s → retry con backoff
  - Output generate: raw bytes PNG (non JSON)
  - BLIP: captioning solo, non risponde a domande aperte
  - Rate limit senza HF_TOKEN: ~10 req/hr per IP

Fallback chain analyze_image:
  1. Groq llama-3.2-11b-vision-preview (free tier, veloce, richiede GROQ_API_KEY)
  2. GPT-4o-mini vision (richiede OPENAI_API_KEY)
  3. BLIP-large captioning (HF Inference, libero ma solo didascalia)
"""
import asyncio, base64, os, httpx, logging
from fastapi import APIRouter
from pydantic import BaseModel

router  = APIRouter(prefix="/api/vision", tags=["vision"])
_logger = logging.getLogger("vision")

_HF_API     = "https://api-inference.huggingface.co"
_USER_AGENT = "Mozilla/5.0 (compatible; AgenteAI/3.0)"

_MODEL_MAP: dict[str, str] = {
    "FLUX.1-schnell": "black-forest-labs/FLUX.1-schnell",
    "FLUX.1-dev":     "black-forest-labs/FLUX.1-dev",
    "sdxl":           "stabilityai/stable-diffusion-xl-base-1.0",
    "flux":           "black-forest-labs/FLUX.1-schnell",
    "flux-schnell":   "black-forest-labs/FLUX.1-schnell",
}


def _hf_headers(content_type: str = "application/json") -> dict:
    token = os.getenv("HF_TOKEN", "")
    h = {"User-Agent": _USER_AGENT, "Content-Type": content_type}
    if token:
        h["Authorization"] = f"Bearer {token}"
    return h


# ─── Models ───────────────────────────────────────────────────────────────────

class GenerateImageRequest(BaseModel):
    prompt:          str
    negative_prompt: str = ""
    width:           int = 512
    height:          int = 512
    steps:           int = 4
    model:           str = "FLUX.1-schnell"
    save_path:       str = ""


class AnalyzeImageRequest(BaseModel):
    url:          str = ""
    base64_image: str = ""
    question:     str = "Descrivi questa immagine in dettaglio in italiano."


# ─── /generate ────────────────────────────────────────────────────────────────

@router.post("/generate")
async def generate_image(req: GenerateImageRequest):
    """
    Genera immagine via HF Inference API (FLUX.1-schnell default).

    Strategia retry:
      - Se il modello è in cold-start (503), aspetta estimated_time (max 45s)
        e riprova una volta sola. Due tentativi totali.
      - HF restituisce raw bytes PNG — non JSON.
      - steps ottimali FLUX.1-schnell: 4 (veloce) – 8 (qualità).
    """
    model_id = _MODEL_MAP.get(req.model, "black-forest-labs/FLUX.1-schnell")
    url      = f"{_HF_API}/models/{model_id}"

    payload: dict = {"inputs": req.prompt.strip()[:400]}
    params: dict = {"num_inference_steps": min(max(req.steps, 1), 8)}
    if req.width  != 512: params["width"]  = min(max(req.width,  256), 1024)
    if req.height != 512: params["height"] = min(max(req.height, 256), 1024)
    if req.negative_prompt:
        params["negative_prompt"] = req.negative_prompt[:200]
    payload["parameters"] = params

    async with httpx.AsyncClient(timeout=90) as client:
        for attempt in range(2):
            try:
                r = await client.post(url, headers=_hf_headers(), json=payload)

                if r.status_code == 200:
                    b64 = base64.b64encode(r.content).decode()
                    return {
                        "ok": True, "image_b64": b64, "mime": "image/png",
                        "model": req.model, "prompt": req.prompt[:100],
                    }

                if r.status_code == 503 and attempt == 0:
                    try:
                        wait = min(float(r.json().get("estimated_time", 20)), 45)
                    except Exception:
                        wait = 20
                    _logger.info("HF model loading, waiting %.0fs…", wait)
                    await asyncio.sleep(wait)
                    continue

                try:
                    err = r.json().get("error", r.text[:200])
                except Exception:
                    err = r.text[:200]
                return {
                    "ok": False, "error": f"HF API {r.status_code}: {err}",
                    "hint": "Aggiungi HF_TOKEN nelle variabili d'ambiente per più richieste/ora.",
                }

            except httpx.TimeoutException:
                return {"ok": False, "error": "Timeout 90s — modello in cold-start. Riprova tra 30s."}
            except Exception as e:
                return {"ok": False, "error": str(e)[:300]}

    return {"ok": False, "error": "Impossibile generare dopo 2 tentativi."}


# ─── /analyze ─────────────────────────────────────────────────────────────────

@router.post("/analyze")
async def analyze_image(req: AnalyzeImageRequest):
    """
    Analizza immagine con vision LLM.

    Chain:
      1. Groq llama-3.2-11b-vision (free tier, 30 img/min)
      2. GPT-4o-mini vision
      3. BLIP-large captioning (HF, puro captioning senza Q&A)
    """
    # Scarica immagine se URL
    image_b64  = req.base64_image
    image_mime = "image/jpeg"

    if not image_b64 and req.url:
        try:
            async with httpx.AsyncClient(timeout=15, follow_redirects=True) as c:
                r = await c.get(req.url, headers={"User-Agent": _USER_AGENT})
                if r.status_code == 200:
                    ct = r.headers.get("content-type", "")
                    image_mime = "image/png" if "png" in ct else "image/webp" if "webp" in ct else "image/jpeg"
                    image_b64  = base64.b64encode(r.content).decode()
                else:
                    return {"ok": False, "error": f"Download immagine fallito: HTTP {r.status_code}"}
        except Exception as e:
            return {"ok": False, "error": f"Errore download: {str(e)[:200]}"}

    if not image_b64:
        return {"ok": False, "error": "Nessuna immagine (url o base64_image richiesti)."}

    img_data_url = f"data:{image_mime};base64,{image_b64}"
    question     = req.question.strip() or "Descrivi questa immagine in dettaglio in italiano."
    vision_body_msgs = [{"role": "user", "content": [
        {"type": "image_url", "image_url": {"url": img_data_url}},
        {"type": "text", "text": question},
    ]}]

    # 1. Groq vision (free tier)
    _groq_key = os.getenv("GROQ_API_KEY", "")
    if _groq_key:
        try:
            async with httpx.AsyncClient(timeout=30) as c:
                r = await c.post(
                    "https://api.groq.com/openai/v1/chat/completions",
                    headers={"Authorization": f"Bearer {_groq_key}", "Content-Type": "application/json"},
                    json={"model": "llama-3.2-11b-vision-preview", "max_tokens": 600,
                          "messages": vision_body_msgs},
                )
                if r.status_code == 200:
                    # S750-GAP-H: guard choices[] — Groq può ritornare {"error":"rate_limit"}
                    _chs = r.json().get("choices") or []
                    _desc = (_chs[0].get("message",{}).get("content") or "") if _chs else ""
                    if _desc:
                        return {"ok": True, "description": _desc, "provider": "llama-3.2-vision"}
        except Exception as _e:
            _logger.debug("analyze_image: groq vision failed (%s)", type(_e).__name__)

    # 2. Gemini Vision (free tier — GEMINI_API_KEY da aistudio.google.com)
    # GAP-TOOL-2-fix: Gemini 1.5 Flash supporta vision, è gratuito su AI Studio, non richiede dominio.
    # Inserito prima di GPT-4o-mini (paid) come primo fallback gratuito di Groq.
    _gemini_key = os.getenv("GEMINI_API_KEY", "")
    if _gemini_key:
        try:
            _g_payload = {
                "contents": [{
                    "parts": [
                        {"inline_data": {"mime_type": image_mime, "data": image_b64}},
                        {"text": question},
                    ]
                }],
                "generationConfig": {"maxOutputTokens": 600},
            }
            async with httpx.AsyncClient(timeout=30) as c:
                r = await c.post(
                    f"https://generativelanguage.googleapis.com/v1beta/models/gemini-1.5-flash:generateContent?key={_gemini_key}",
                    headers={"Content-Type": "application/json"},
                    json=_g_payload,
                )
                if r.status_code == 200:
                    _cands = r.json().get("candidates") or []
                    _parts = (_cands[0].get("content", {}).get("parts") or []) if _cands else []
                    _desc_g = next((p.get("text", "") for p in _parts if "text" in p), "")
                    if _desc_g:
                        return {"ok": True, "description": _desc_g, "provider": "gemini-1.5-flash"}
        except Exception as _e:
            _logger.debug("analyze_image: gemini vision failed (%s)", type(_e).__name__)

    # 3. OpenAI GPT-4o-mini vision
    _openai_key  = os.getenv("OPENAI_API_KEY", "")
    _openai_base = os.getenv("OPENAI_BASE_URL", "https://api.openai.com/v1").rstrip("/")
    if _openai_key:
        try:
            async with httpx.AsyncClient(timeout=30) as c:
                r = await c.post(
                    f"{_openai_base}/chat/completions",
                    headers={"Authorization": f"Bearer {_openai_key}", "Content-Type": "application/json"},
                    json={"model": "gpt-4o-mini", "max_tokens": 600, "messages": vision_body_msgs},
                )
                if r.status_code == 200:
                    # S750-GAP-H: guard choices[] — provider può ritornare {"error":...}
                    _chs2 = r.json().get("choices") or []
                    _desc2 = (_chs2[0].get("message",{}).get("content") or "") if _chs2 else ""
                    if _desc2:
                        return {"ok": True, "description": _desc2, "provider": "gpt-4o-mini"}
        except Exception as _e:
            _logger.debug("analyze_image: openai vision failed (%s)", type(_e).__name__)

    # 4. HF BLIP-large (captioning only — ultimo fallback)
    try:
        img_bytes = base64.b64decode(image_b64)
        async with httpx.AsyncClient(timeout=30) as c:
            r = await c.post(
                f"{_HF_API}/models/Salesforce/blip-image-captioning-large",
                headers={k: v for k, v in _hf_headers("application/octet-stream").items()},
                content=img_bytes,
            )
            if r.status_code == 200:
                results = r.json()
                caption = (results[0].get("generated_text", "") if isinstance(results, list) and results else "")
                if caption:
                    note = ("\n\n_BLIP fornisce solo didascalia base. Per Q&A su immagini, "
                            "aggiungi GROQ_API_KEY (gratuito su console.groq.com)._")
                    return {"ok": True, "description": caption + note, "provider": "blip-large"}
            elif r.status_code == 503:
                return {"ok": False, "error": "BLIP in avvio (cold-start ~30s). Riprova tra qualche secondo.",
                        "hint": "Aggiungi GROQ_API_KEY per analisi rapida e senza limiti di cold-start."}
    except Exception as _e:
        _logger.debug("analyze_image: blip captioning failed (%s)", type(_e).__name__)

    return {
        "ok": False, "error": "Analisi immagini non disponibile.",
        "hint": ("Aggiungi GROQ_API_KEY (free su console.groq.com) o OPENAI_API_KEY "
                 "nelle variabili del tuo HF Space."),
    }


# ─── /search ──────────────────────────────────────────────────────────────────

@router.get("/search")
async def search_images(query: str, limit: int = 5, safe: bool = True):
    """
    Ricerca immagini: Pexels > Pixabay > Unsplash Source fallback.

    Pexels API (PEXELS_API_KEY, gratuita 200 req/hr):
      https://www.pexels.com/api/

    Pixabay API (PIXABAY_API_KEY, gratuita 500 req/hr):
      https://pixabay.com/api/docs/

    Unsplash Source (zero API key — fallback):
      URL deterministica ma rilevante per la query.
      Non è una vera ricerca — genera varianti random per lo stesso query.
    """
    n = min(max(int(limit), 1), 10)

    # Pexels
    _pexels = os.getenv("PEXELS_API_KEY", "")
    if _pexels:
        try:
            async with httpx.AsyncClient(timeout=10) as c:
                r = await c.get(
                    "https://api.pexels.com/v1/search",
                    headers={"Authorization": _pexels},
                    params={"query": query, "per_page": n, "size": "medium"},
                )
                if r.status_code == 200:
                    photos = r.json().get("photos", [])
                    return {"ok": True, "source": "pexels", "results": [
                        {"url": p["src"]["medium"], "thumb": p["src"]["tiny"],
                         "alt": p.get("alt", query), "source": "Pexels",
                         "page_url": p["url"], "author": p["photographer"]}
                        for p in photos
                    ]}
        except Exception as _exc:
            _logger.debug("[vision] silenced %s", type(_exc).__name__)  # noqa: BLE001

    # Pixabay
    _pixabay = os.getenv("PIXABAY_API_KEY", "")
    if _pixabay:
        try:
            async with httpx.AsyncClient(timeout=10) as c:
                r = await c.get(
                    "https://pixabay.com/api/",
                    params={"key": _pixabay, "q": query, "per_page": n,
                            "image_type": "photo", "safesearch": "true" if safe else "false"},
                )
                if r.status_code == 200:
                    return {"ok": True, "source": "pixabay", "results": [
                        {"url": h["webformatURL"], "thumb": h["previewURL"],
                         "alt": h.get("tags", query), "source": "Pixabay",
                         "page_url": h["pageURL"], "author": h["user"]}
                        for h in r.json().get("hits", [])
                    ]}
        except Exception as _exc:
            _logger.debug("[vision] silenced %s", type(_exc).__name__)  # noqa: BLE001

    # Unsplash Source fallback (zero key — random ma rilevante)
    sq = query.replace(" ", ",")[:80]
    return {
        "ok": True, "source": "unsplash_source",
        "note": "Usa PEXELS_API_KEY o PIXABAY_API_KEY per ricerca precisa.",
        "results": [
            {"url":     f"https://source.unsplash.com/600x400/?{sq}&sig={i}",
             "thumb":   f"https://source.unsplash.com/200x150/?{sq}&sig={i}",
             "alt":     f"{query} — immagine {i + 1}",
             "source":  "Unsplash", "page_url": f"https://unsplash.com/s/photos/{sq}",
             "author":  "Unsplash"}
            for i in range(n)
        ],
    }


class ScreenshotRequest(BaseModel):
    url:        str
    width:      int   = 1280
    height:     int   = 900
    mobile:     bool  = False
    wait:       float = 2.0    # secondi di attesa dopo il caricamento
    full_page:  bool  = False


@router.post("/screenshot")
async def screenshot_url(req: ScreenshotRequest):
    """
    Screenshot Playwright (Chromium headless) di qualsiasi URL.
    Restituisce PNG come data URL base64.
    Fallback automatico a Microlink se Playwright non disponibile.
    """
    try:
        from playwright.async_api import async_playwright  # noqa: PLC0415
    except ImportError:
        return {"ok": False, "error": "playwright non installato sul server.", "fallback": "microlink"}

    w = min(max(int(req.width) or 1280, 320), 2560)
    h = min(max(int(req.height) or 900, 240), 1440)

    try:
        async with async_playwright() as p:
            browser = await p.chromium.launch(
                headless=True,
                args=["--no-sandbox", "--disable-dev-shm-usage", "--disable-gpu", "--single-process"],
            )
            ctx = await browser.new_context(
                viewport={"width": 375 if req.mobile else w, "height": 812 if req.mobile else h},
                device_scale_factor=2 if req.mobile else 1,
                user_agent=(
                    "Mozilla/5.0 (iPhone; CPU iPhone OS 17_0 like Mac OS X) AppleWebKit/605.1.15"
                    if req.mobile else
                    "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 Chrome/120 Safari/537.36"
                ),
            )
            page = await ctx.new_page()
            await page.goto(str(req.url), wait_until="domcontentloaded", timeout=20_000)
            if req.wait > 0:
                await asyncio.sleep(min(float(req.wait), 5.0))
            png_bytes = await page.screenshot(full_page=req.full_page, type="png")
            await browser.close()

        data_url = "data:image/png;base64," + base64.b64encode(png_bytes).decode()
        return {
            "ok": True,
            "image_url": data_url,
            "width":  375 if req.mobile else w,
            "height": 812 if req.mobile else h,
            "device": "mobile" if req.mobile else "desktop",
        }
    except Exception as e:
        _logger.warning("screenshot %s: playwright failed: %s — trying Microlink", req.url, str(e)[:80])
        # S601: Microlink fallback — chiamata reale invece di semplice errore
        try:
            import httpx as _httpx_ml
            async with _httpx_ml.AsyncClient(timeout=15) as _mc:
                _ml_r = await _mc.get(
                    "https://api.microlink.io/",
                    params={
                        "url": str(req.url),
                        "screenshot": "true",
                        "meta":       "false",
                        "embed":      "screenshot.url",
                        "viewport.width":  375 if req.mobile else w,
                        "viewport.height": 812 if req.mobile else h,
                    },
                    headers={"User-Agent": "agente-ai/3.2"},
                )
                if _ml_r.status_code == 200:
                    _ml_data = _ml_r.json()
                    _img_url = (_ml_data.get("data") or {}).get("screenshot", {}).get("url") or _ml_data.get("data")
                    if isinstance(_img_url, str) and _img_url.startswith("http"):
                        # Download image and convert to base64 data URL
                        _img_resp = await _mc.get(_img_url, timeout=10)
                        if _img_resp.status_code == 200:
                            _ct = _img_resp.headers.get("content-type", "image/jpeg")
                            _data_url = f"data:{_ct};base64," + base64.b64encode(_img_resp.content).decode()
                            return {
                                "ok":       True,
                                "image_url": _data_url,
                                "width":    375 if req.mobile else w,
                                "height":   812 if req.mobile else h,
                                "device":   "mobile" if req.mobile else "desktop",
                                "source":   "microlink",
                            }
        except Exception as _ml_exc:
            _logger.warning("screenshot %s: microlink also failed: %s", req.url, str(_ml_exc)[:80])
        return {"ok": False, "error": str(e)[:200], "fallback": "microlink_unavailable"}


# ── PDF Generation ─────────────────────────────────────────────────────────────

class PdfRequest(BaseModel):
    content:  str
    filename: str                  = "documento.pdf"
    format:   str                  = "html"   # html | markdown | text


@router.post("/pdf")
async def create_pdf(req: PdfRequest):
    """
    S601: Genera PDF da HTML, Markdown o testo.
    Tenta WeasyPrint → reportlab → risposta JSON+base64 del contenuto raw.
    Restituisce {ok, pdf_b64, filename, size_bytes} oppure {ok: false, error}.
    """
    import base64 as _b64
    _original_len = len(req.content)
    content       = req.content[:80_000]
    _truncated    = _original_len > 80_000  # S724-PDF-3: segnala troncamento nel response

    # 1. WeasyPrint (miglior qualità HTML→PDF)
    try:
        from weasyprint import HTML as _WP_HTML, CSS as _WP_CSS  # noqa: PLC0415
        if req.format == "html":
            _html = content
        elif req.format == "markdown":
            try:
                import markdown as _md  # noqa: PLC0415
                _html = _md.markdown(content, extensions=["tables", "fenced_code"])
            except ImportError:
                _html = f"<pre>{content}</pre>"
        else:
            _html = f"<pre style='font-family:monospace;white-space:pre-wrap'>{content}</pre>"
        _pdf_bytes = _WP_HTML(string=_html).write_pdf()
        _r = {"ok": True, "pdf_b64": _b64.b64encode(_pdf_bytes).decode(),
               "filename": req.filename, "size_bytes": len(_pdf_bytes), "engine": "weasyprint"}
        if _truncated: _r.update({"truncated": True, "original_length": _original_len})  # S724-PDF-3
        return _r
    except ImportError:
        pass  # WeasyPrint non installato — prova reportlab
    except Exception as _wp_exc:
        _logger.warning("pdf weasyprint: %s", str(_wp_exc)[:120])

    # 2. reportlab (fallback leggero)
    try:
        from reportlab.lib.pagesizes import A4  # noqa: PLC0415
        from reportlab.platypus import SimpleDocTemplate, Paragraph, Spacer  # noqa: PLC0415
        from reportlab.lib.styles import getSampleStyleSheet  # noqa: PLC0415
        import io as _io
        _buf = _io.BytesIO()
        _doc = SimpleDocTemplate(_buf, pagesize=A4)
        _styles = getSampleStyleSheet()
        _story = []
        for line in content.split("\n"):
            if line.strip():
                _story.append(Paragraph(line.replace("<", "&lt;").replace(">", "&gt;"), _styles["Normal"]))
            else:
                _story.append(Spacer(1, 12))
        _doc.build(_story)
        _pdf_bytes = _buf.getvalue()
        _r = {"ok": True, "pdf_b64": _b64.b64encode(_pdf_bytes).decode(),
               "filename": req.filename, "size_bytes": len(_pdf_bytes), "engine": "reportlab"}
        if _truncated: _r.update({"truncated": True, "original_length": _original_len})  # S724-PDF-3
        return _r
    except ImportError:
        pass
    except Exception as _rl_exc:
        _logger.warning("pdf reportlab: %s", str(_rl_exc)[:120])

    # 3. Graceful degradation — restituisce il contenuto raw encodato
    _raw = content.encode("utf-8")
    return {
        "ok":        False,
        "error":     "Nessun engine PDF disponibile (WeasyPrint/reportlab non installati). Contenuto allegato come testo.",
        "content_b64": _b64.b64encode(_raw).decode(),
        "filename":  req.filename.replace(".pdf", ".txt"),
        "size_bytes": len(_raw),
    }