import json, os, logging, base64, time, hashlib, codecs, re, http.client, ssl
from urllib.parse import urlparse
from searx import network
try:
    from searx.network import get_network
except ImportError:
    get_network = None  # Graceful fallback for test/demo environments
from flask import Response, request, abort, jsonify
from searx.plugins import Plugin, PluginInfo
from searx.result_types import EngineResults
from flask_babel import gettext
from markupsafe import Markup

logger = logging.getLogger(__name__)

TOKEN_EXPIRY_SEC = 3600
STREAM_CHUNK_SIZE = 128
STREAM_TIMEOUT_SEC = 60

def _env_flag(name: str, default: bool = False) -> bool:
    v = os.getenv(name)
    if v is None:
        return default
    return v.strip().lower() in ('1', 'true', 'yes', 'on')
def _get_streaming_connection(url: str):
    parsed = urlparse(url)
    host = parsed.hostname
    port = parsed.port or (443 if parsed.scheme == 'https' else 80)
    path = parsed.path + ('?' + parsed.query if parsed.query else '')
    
    verify_ssl = True
    if get_network is not None:
        try:
            net = get_network()
            verify_ssl = getattr(net, 'verify', True)
        except Exception:
            pass
    
    if parsed.scheme == 'https':
        ctx = ssl.create_default_context() if verify_ssl else ssl._create_unverified_context()
        conn = http.client.HTTPSConnection(host, port, timeout=STREAM_TIMEOUT_SEC, context=ctx)
    else:
        conn = http.client.HTTPConnection(host, port, timeout=STREAM_TIMEOUT_SEC)
    
    return conn, path



PLUGIN_NAME = "AI Answers"
DEFAULT_TABS = "general,science,it,news"

PROVIDER_PRESETS = {
    'openai':     {'url': 'https://api.openai.com/v1/chat/completions',       'model': 'gpt-4o-mini'},
    'openrouter': {'url': 'https://openrouter.ai/api/v1/chat/completions',    'model': 'google/gemma-3-27b-it:free'},
    'ollama':     {'url': 'http://localhost:11434/v1/chat/completions',       'model': 'llama3.2'},
    'localai':    {'url': 'http://localhost:8080/v1/chat/completions',        'model': 'gpt-4'},
    'lmstudio':   {'url': 'http://localhost:1234/v1/chat/completions',        'model': 'local-model'},
    'gemini':     {'url': 'https://generativelanguage.googleapis.com/v1beta/models/{model}:streamGenerateContent', 'model': 'gemma-3-27b-it'},
    'azure':      {'url': None,                                               'model': 'azure-deployment'},
    'huggingface': {'url': 'https://api-inference.huggingface.co/models/{model}/v1/chat/completions', 'model': 'meta-llama/Meta-Llama-3-8B-Instruct'}
}

# UI assets (inlined for single-file install)

INTERACTIVE_CSS = '''
                        @keyframes sxng-fade-in-up {
                            0% { opacity: 0; transform: translateY(10px); }
                            100% { opacity: 1; transform: translateY(0); }
                        }
                        .sxng-footer {
                            display: flex;
                            align-items: center;
                            gap: 0.5rem;
                            margin-top: 0.5rem;
                            opacity: 0;
                            animation: sxng-fade-in-up 0.5s ease-out forwards;
                        }
                        .sxng-btn {
                            display: inline-flex;
                            align-items: center;
                            justify-content: center;
                            width: 32px;
                            height: 32px;
                            padding: 0;
                            border: 1px solid transparent;
                            border-radius: 6px;
                            background: transparent;
                            color: var(--color-base-font, #333);
                            cursor: pointer;
                            transition: all 0.2s ease;
                            opacity: 0.6;
                        }
                        .sxng-btn:hover {
                            background: var(--color-base-background-hover, rgba(0,0,0,0.05));
                            color: var(--color-result-link, #5e81ac);
                            opacity: 1;
                            transform: translateY(-1px);
                        }
                        .sxng-btn svg { width: 18px; height: 18px; fill: currentColor; }
                        .sxng-input-wrapper {
                            flex-grow: 1;
                            display: flex;
                            align-items: center;
                            margin: 0 0.5rem;
                            position: relative;
                        }
                        .sxng-input {
                            width: 100%;
                            background: transparent;
                            border: none;
                            color: var(--color-base-font, #333);
                            font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial, sans-serif;
                            font-size: 16px;
                            padding: 0.5rem 2.5rem 0.5rem 0;
                            opacity: 0.8;
                            transition: opacity 0.2s;
                        }
                        .sxng-input:focus { outline: none; opacity: 1; }
                        .sxng-input::placeholder { color: var(--color-base-font, #333); opacity: 0.35; }
                        .sxng-input-line {
                            position: absolute;
                            bottom: 0;
                            left: 0;
                            width: 0;
                            height: 1px;
                            background: var(--color-result-link, #5e81ac);
                            transition: width 0.3s ease;
                        }
                        .sxng-input:focus + .sxng-input-line { width: 100%; }
                        .sxng-user-msg {
                            display: block;
                            width: fit-content;
                            max-width: 80%;
                            margin: 0.75rem 0 0.75rem auto;
                            padding: 0.25rem 0.6rem 0.25rem 0;
                            border-right: 2px solid var(--color-result-link, #5e81ac);
                            text-align: right;
                            font-size: 0.85rem;
                            line-height: 1.4;
                            opacity: 0.55;
                            animation: sxng-fade-in-up 0.3s ease-out forwards;
                        }
                        .sxng-input-submit {
                            all: unset;
                            position: absolute;
                            right: 0;
                            top: 50%;
                            transform: translateY(-50%);
                            display: inline-flex;
                            align-items: center;
                            justify-content: center;
                            width: 32px;
                            height: 32px;
                            padding: 0;
                            background: transparent !important;
                            border: none !important;
                            border-radius: 6px;
                            color: var(--color-base-font, #333);
                            cursor: pointer;
                            opacity: 0.3;
                            transition: all 0.2s ease;
                        }
                        .sxng-input-wrapper:focus-within .sxng-input-submit,
                        .sxng-input-submit:hover { 
                            opacity: 1; 
                            color: var(--color-result-link, #5e81ac); 
                            background: var(--color-base-background-hover, rgba(0,0,0,0.05)) !important;
                        }
                        .sxng-input-submit svg { width: 18px; height: 18px; fill: currentColor; }
                        .sxng-input-submit svg { width: 18px; height: 18px; fill: currentColor; }
'''

INTERACTIVE_HTML = '''
                    <div id="sxng-footer" class="sxng-footer" style="display:none;">
                        <button class="sxng-btn" id="btn-copy" title="Copy to clipboard">
                            <svg viewBox="0 0 24 24"><path d="M16 1H4C2.9 1 2 1.9 2 3V17H4V3H16V1M19 5H8C6.9 5 6 5.9 6 7V21C6 22.1 6.9 23 8 23H19C20.1 23 21 22.1 21 21V7C21 5.9 20.1 5 19 5M19 21H8V7H19V21Z"/></svg>
                        </button>
                        <button class="sxng-btn" id="btn-regen" title="Regenerate answer">
                            <svg viewBox="0 0 24 24"><path d="M17.65 6.35C16.2 4.9 14.21 4 12 4C7.58 4 4.01 7.58 4.01 12C4.01 16.42 7.58 20 12 20C15.73 20 18.84 17.45 19.73 14H17.65C16.83 16.33 14.61 18 12 18C8.69 18 6 15.31 6 12C6 8.69 8.69 6 12 6C13.66 6 15.14 6.69 16.22 7.78L13 11H20V4L17.65 6.35Z"/></svg>
                        </button>
                        <form id="sxng-action-form" class="sxng-input-wrapper" onsubmit="event.preventDefault();">
                            <input type="text" id="sxng-action-input" class="sxng-input" placeholder="Ask..." aria-label="Ask follow-up" autocomplete="off">
                            <div class="sxng-input-line"></div>
                            <button type="submit" id="btn-action" class="sxng-input-submit" title="Send / Continue">
                                <svg viewBox="0 0 24 24"><path d="M19,7V11H5.83L9.41,7.41L8,6L2,12L8,18L9.41,16.59L5.83,13H21V7H19Z"/></svg>
                            </button>
                        </form>
                    </div>
'''

CITATION_HELPER_JS = '''
                        function renderCitations(text, urls) {
                            const fragment = document.createDocumentFragment();
                            const re = /\[(\d{1,2}(?:\s*,\s*\d{1,2})*)\]/g;
                            let lastIdx = 0;
                            const matches = [...text.matchAll(re)];
                            
                            matches.forEach(match => {
                                if (match.index > lastIdx) {
                                    const s = document.createElement('span');
                                    s.className = 'sxng-chunk';
                                    // Preserve whitespace by not trimming
                                    s.textContent = text.substring(lastIdx, match.index);
                                    fragment.appendChild(s);
                                }
                                match[1].split(/\s*,\s*/).forEach(n => {
                                    const idx = parseInt(n.trim());
                                    if (idx >= 1 && idx <= urls.length) {
                                        const url = urls[idx-1];
                                        if (url) {
                                            const a = document.createElement('a');
                                            a.href = url;
                                            a.target = '_blank';
                                            a.style.cssText = 'text-decoration:none;color:var(--color-result-link);font-weight:bold;';
                                            a.textContent = `[${n.trim()}]`;
                                            a.className = 'sxng-chunk';
                                            fragment.appendChild(a);
                                        } else {
                                            const s = document.createElement('span');
                                            s.className = 'sxng-chunk';
                                            s.textContent = `[${n.trim()}]`;
                                            fragment.appendChild(s);
                                        }
                                    } else {
                                        const s = document.createElement('span');
                                        s.className = 'sxng-chunk';
                                        s.textContent = `[${n.trim()}]`;
                                        fragment.appendChild(s);
                                    }
                                });
                                lastIdx = match.index + match[0].length;
                            });
                            
                            if (lastIdx < text.length) {
                                const s = document.createElement('span');
                                s.className = 'sxng-chunk';
                                // Preserve whitespace by not trimming
                                s.textContent = text.substring(lastIdx);
                                fragment.appendChild(s);
                            }
                            return fragment;
                        }
'''

INTERACTIVE_JS = '''
                        const footer = document.getElementById('sxng-footer');
                        const input = document.getElementById('sxng-action-input');
                        // Inherited from outer scope: box, data, conversation

                        // Theme detection: inherit host accent color
                        if (window.getComputedStyle && box) {
                            try {
                                const docStyles = getComputedStyle(document.documentElement);
                                let accent = docStyles.getPropertyValue('--color-result-link').trim();
                                if (!accent) {
                                    const a = document.createElement('a');
                                    document.body.appendChild(a);
                                    accent = getComputedStyle(a).color;
                                    document.body.removeChild(a);
                                }
                                if (accent) {
                                    box.style.setProperty('--color-result-link', accent);
                                    box.style.setProperty('--sxng-ai-accent', accent);
                                }
                            } catch(e) {}
                        }

                        // Persist conversation state to URL hash (stateless)
                        const updateState = () => {
                            try {
                                const state = {
                                    t: conversation.turns.map(t => ({
                                        r: t.role === 'user' ? 'u' : 'a',
                                        c: t.content.replace(/\s+/g, ' ').trim()
                                    })),
                                    u: urls
                                };
                                const b64 = btoa(encodeURIComponent(JSON.stringify(state)).replace(/%([0-9A-F]{2})/g, (m,p)=>String.fromCharCode('0x'+p)));
                                history.replaceState(null, null, '#ai=' + b64);
                            } catch(e) {}
                        };

                        if (location.hash.includes('ai=')) {
                            try {
                                const b64 = location.hash.split('ai=')[1];
                                const json = decodeURIComponent(atob(b64).split('').map(c => '%' + ('00' + c.charCodeAt(0).toString(16)).slice(-2)).join(''));
                                const state = JSON.parse(json);
                                if (state.t && state.t.length > 0) {
                                    // Restore URLs for citation indexing
                                    if (state.u && Array.isArray(state.u)) {
                                        urls = state.u;
                                    }
                                    
                                    conversation.turns = state.t.map(t => ({
                                        role: t.r === 'u' ? 'user' : 'assistant',
                                        content: t.c.trim(),
                                        ts: 0
                                    }));
                                    
                                    // Helper function to inject citations into text
                                    const injectCitations = (text) => {
                                        return renderCitations(text, urls);
                                    };
                                    
                                    data.innerHTML = '';
                                    conversation.turns.forEach((turn, i) => {
                                        if (turn.role === 'user') {
                                            if (turn.content !== conversation.originalQuery) {
                                                const u = document.createElement('span');
                                                u.className = 'sxng-user-msg';
                                                u.textContent = turn.content;
                                                data.appendChild(u);
                                                const clr = document.createElement('div');
                                                clr.style.clear = 'both';
                                                data.appendChild(clr);
                                            }
                                        } else {
                                            // Inject citations for assistant responses
                                            data.appendChild(injectCitations(turn.content));
                                        }
                                    });
                                    box.style.display = 'block';
                                    if(wrapper) wrapper.style.display = '';
                                    if(footer && is_interactive) footer.style.display = 'flex';
                                    restored = true;
                                }
                            } catch(e) { console.warn('Restore failed', e); }
                        }
                        document.getElementById('btn-copy').onclick = async (e) => {
                            const btn = e.currentTarget;
                            const originalContent = btn.innerHTML;
                            const text = Array.from(data.childNodes)
                                .filter(n => n.nodeType === 3 || n.tagName === 'SPAN')
                                .map(n => n.textContent)
                                .join('');
                            await navigator.clipboard.writeText(text);
                            btn.innerHTML = '<svg viewBox="0 0 24 24" style="color:#a3be8c;"><path d="M9 16.17L4.83 12L3.41 13.41L9 19L21 7L19.59 5.59L9 16.17Z"/></svg>';
                            setTimeout(() => btn.innerHTML = originalContent, 2000);
                        };

                        document.getElementById('btn-regen').onclick = () => {
                            data.innerHTML = '<span class="sxng-cursor"></span>';
                            footer.style.display = 'none';
                            startStream();
                        };

                        const handleAction = async (e) => {
                            if (e) e.preventDefault();
                            const val = input.value.trim();
                            
                            conversation.turns.push({role: 'user', content: val, ts: Date.now()});
                            updateState();
                            
                            const currentText = conversation.turns.slice(0, -1).slice(-6)
                                .map(t => (t.role === 'user' ? 'Q' : 'A') + ': ' + t.content)
                                .join('\\n\\n');

                            input.value = '';
                            input.blur();
                            footer.style.display = 'none';

                            if (val) {
                                const cursor = data.querySelector('.sxng-cursor');
                                if (cursor) cursor.remove();
                                const userMsg = document.createElement('span');
                                userMsg.className = 'sxng-user-msg';
                                userMsg.textContent = val;
                                data.appendChild(userMsg);
                                const clr = document.createElement('div');
                                clr.style.clear = 'both';
                                data.appendChild(clr);

                                const newCursor = document.createElement('span');
                                newCursor.className = 'sxng-cursor';
                                data.appendChild(newCursor);
                                
                                const synthesized = synthesizeQuery(q_init, val);
                                let auxContext = null;
                                try {
                                    const auxData = await fetch('/ai-auxiliary-search', {
                                        method: 'POST',
                                        headers: {'Content-Type': 'application/json'},
                                        body: JSON.stringify({query: synthesized, lang: lang_init, offset: urls.length})
                                    }).then(r => r.json());
                                    if (auxData.context) {
                                        const originalBackground = conversation.originalContext.substring(0, 1500);
                                        auxContext = `FRESH SOURCES (most relevant):\\n${auxData.context}\\n\\nBACKGROUND (for reference):\\n${originalBackground}`;
                                        if (auxData.new_urls && Array.isArray(auxData.new_urls)) {
                                            urls = urls.concat(auxData.new_urls);
                                        }
                                    }
                                } catch (err) {}
                                
                                await startStream(val, currentText, auxContext);
                                updateState();
                            } else {
                                const cursor = data.querySelector('.sxng-cursor');
                                if (cursor) cursor.remove();
                                data.appendChild(document.createElement('br'));
                                data.appendChild(document.createElement('br'));
                                const newCursor = document.createElement('span');
                                newCursor.className = 'sxng-cursor';
                                data.appendChild(newCursor);
                                await startStream("Continue", currentText);
                                updateState();
                            }
                        };

                        document.getElementById('sxng-action-form').onsubmit = handleAction;
                        input.onfocus = () => {
                            setTimeout(() => {
                                input.scrollIntoView({behavior: 'smooth', block: 'center'});
                            }, 300);
                        };

                        const _origStream = startStream;
                        startStream = async function(...args) {
                            if (args.length === 0 && restored) return;
                            await _origStream.apply(this, args);
                            if (args.length === 0) updateState();
                        };
'''


import typing
if typing.TYPE_CHECKING:
    from searx.search import SearchWithPlugins
    from searx.extended_types import SXNG_Request
    from . import PluginCfg

class SXNGPlugin(Plugin):
    id = "ai_answers"

    def __init__(self, plg_cfg: "PluginCfg"):
        super().__init__(plg_cfg)
        self.info = PluginInfo(
            id=self.id,
            name=gettext(f"{PLUGIN_NAME} Plugin"),
            description=gettext("Live AI search answers using LLM providers."),
            preference_section="general",
        )
        self._load_config()



    def _ollama_unload_model(self) -> None:
        """
        Force-unload an Ollama model after a response by calling the native /api/chat endpoint:
          {"model": "...", "messages": [], "keep_alive": 0}
        """
        try:
            if self.provider != 'ollama':
                return
            if not getattr(self, 'ollama_unload_after', False):
                return
            unload_url = (getattr(self, 'ollama_unload_url', '') or '').strip()
            if not unload_url:
                return

            conn = None
            try:
                conn, path = _get_streaming_connection(unload_url)
                payload = json.dumps({
                    "model": self.model,
                    "messages": [],
                    "keep_alive": 0
                })
                headers = {"Content-Type": "application/json"}
                # Optional: if Ollama is behind auth, reuse LLM_KEY
                if self.api_key and self.api_key not in ('none', 'ollama'):
                    headers["Authorization"] = f"Bearer {self.api_key}"
                conn.request("POST", path, body=payload, headers=headers)
                res = conn.getresponse()
                res.read()  # drain
                if res.status >= 400:
                    logger.warning(f"{PLUGIN_NAME}: Ollama unload failed: {res.status} {res.reason}")
            finally:
                if conn:
                    conn.close()
        except Exception as e:
            logger.warning(f"{PLUGIN_NAME}: Ollama unload error: {e}")

    def _load_config(self):
        self.interactive = os.getenv('LLM_INTERACTIVE', 'true').lower().strip() in ('true', '1', 'yes', 'on')
        raw_provider = os.getenv('LLM_PROVIDER', '').lower().strip()
        
        raw_url = os.getenv('LLM_URL', '').strip()
        if not raw_provider and raw_url:
            url_lower = raw_url.lower()
            if 'openai.com' in url_lower:
                raw_provider = 'openai'
            elif 'openrouter.ai' in url_lower:
                raw_provider = 'openrouter'
            elif ':11434' in url_lower:
                raw_provider = 'ollama'
            elif 'generativelanguage.googleapis.com' in url_lower:
                raw_provider = 'gemini'
            elif 'openai.azure.com' in url_lower or '.azure.com' in url_lower:
                raw_provider = 'azure'
            elif 'huggingface.co' in url_lower:
                raw_provider = 'huggingface'
            else:
                # Unknown URL fallback to OpenAI-compatible
                raw_provider = 'openai'
                logger.info(f"{PLUGIN_NAME}: Using OpenAI-compatible mode for custom URL")
        
        if not raw_provider:
            self.provider = ''
            self.model = ''
            self.is_gemini = False
            self.api_key = ''
            return
        
        if raw_provider not in PROVIDER_PRESETS:
            logger.warning(f"{PLUGIN_NAME}: Unknown provider '{raw_provider}', falling back to 'openai'")
        self.provider = raw_provider if raw_provider in PROVIDER_PRESETS else 'openai'
        self.is_gemini = (self.provider == 'gemini')
        preset = PROVIDER_PRESETS[self.provider]

        self.api_key = os.getenv('LLM_KEY', '')
        if not self.api_key and self.provider in ('ollama', 'localai', 'lmstudio'):
            self.api_key = 'none'
        self.api_key = self.api_key.strip()

        self.model = os.getenv('LLM_MODEL', preset['model']).strip()

        try:
            self.max_tokens = int(os.getenv('LLM_MAX_TOKENS', 500))
        except ValueError:
            self.max_tokens = 500
        try:
            self.temperature = float(os.getenv('LLM_TEMPERATURE', 0.2))
        except ValueError:
            self.temperature = 0.2
        try:
            self.context_deep_count = max(0, int(os.getenv('LLM_CONTEXT_DEEP_COUNT', 5)))
        except ValueError:
            self.context_deep_count = 5
        try:
            self.context_shallow_count = max(0, int(os.getenv('LLM_CONTEXT_SHALLOW_COUNT', 15)))
        except ValueError:
            self.context_shallow_count = 15

        self.allowed_tabs = set(t.strip() for t in os.getenv('LLM_TABS', DEFAULT_TABS).split(','))
        
        preset_url = preset['url']
        if preset_url and '{model}' in preset_url:
            preset_url = preset_url.format(model=self.model)
        
        raw_url = os.getenv('LLM_URL', '').strip() or preset_url
        if not raw_url.startswith(('http://', 'https://')):
            raw_url = f"https://{raw_url}"
        self.endpoint_url = raw_url
        

        # Ollama: optional "unload after response" behavior (plugin-specific).
        # Enable with:
        #   LLM_OLLAMA_UNLOAD_AFTER=1
        # Optional override:
        #   LLM_OLLAMA_UNLOAD_URL=http://<host>:11434/api/chat
        self.ollama_unload_after = (
            _env_flag('LLM_OLLAMA_UNLOAD_AFTER', False) or _env_flag('OLLAMA_UNLOAD_AFTER', False)
        )
        self.ollama_unload_url = (os.getenv('LLM_OLLAMA_UNLOAD_URL') or os.getenv('OLLAMA_UNLOAD_URL') or '').strip()
        if self.provider == 'ollama' and self.ollama_unload_after and not self.ollama_unload_url:
            try:
                p = urlparse(self.endpoint_url)
                scheme = p.scheme or 'http'
                host = p.hostname or 'localhost'
                port = p.port
                netloc = f"{host}:{port}" if port else host
                self.ollama_unload_url = f"{scheme}://{netloc}/api/chat"
            except Exception:
                self.ollama_unload_url = "http://localhost:11434/api/chat"
        if self.api_key:
            self.secret = os.getenv('SXNG_LLM_SECRET') or hashlib.sha256(self.api_key.encode()).hexdigest()
        else:
             self.secret = os.getenv('SXNG_LLM_SECRET', '')

    def _parse_aux_results(self, raw_results, raw_infoboxes, raw_answers):
        results = []
        limit = self.context_deep_count + self.context_shallow_count
        for r in raw_results[:limit]:
            results.append({
                'title': r.get('title', ''),
                'content': r.get('content', ''),
                'url': r.get('url', ''),
                'publishedDate': r.get('publishedDate', '')
            })
        
        # SearXNG already merges infoboxes by ID - take first with full content
        infoboxes = []
        for ib in raw_infoboxes[:1]:
            infoboxes.append({
                'name': ib.get('infobox', '') or ib.get('title', ''),
                'content': ib.get('content', '')[:2000],
                'attributes': ib.get('attributes', [])
            })
            
        # Only extract simple Answer types (skip Translations, WeatherAnswer etc.)
        answers = []
        for a in list(raw_answers)[:2]:
            if hasattr(a, 'answer') and isinstance(getattr(a, 'answer', None), str):
                answers.append(a.answer)
            elif isinstance(a, dict) and a.get('answer'):
                answers.append(str(a['answer']))
                   
        return results, infoboxes, answers



    def init(self, app):
        if not self.provider:
            return

        @app.route('/ai-auxiliary-search', methods=['POST'])
        def ai_auxiliary_search():
            if not self.api_key:
                abort(403)
            
            data = request.json or {}
            query = data.get('query', '').strip()
            lang = data.get('lang', 'all')
            categories = data.get('categories', 'general')
            offset = data.get('offset', 0)
            if not query:
                return jsonify({'results': []})
            
            # Direct kernel access (bypasses HTTP loopback)
            try:
                from searx.search import SearchWithPlugins
                from searx.search.models import SearchQuery
                from searx.query import RawTextQuery
                from searx.webadapter import get_engineref_from_category_list
                
                preferences = getattr(request, 'preferences', None)
                disabled_engines = preferences.engines.get_disabled() if preferences else []
                rtq = RawTextQuery(query, disabled_engines)
                if isinstance(categories, str):
                    category_list = [c.strip() for c in categories.split(',') if c.strip()]
                else:
                    category_list = categories or ['general']
                
                enginerefs = get_engineref_from_category_list(category_list, disabled_engines)
                sq = SearchQuery(
                    query=rtq.getQuery(),
                    engineref_list=enginerefs,
                    lang=lang,
                    pageno=1,
                )
                # Empty plugins list prevents recursion
                search_obj = SearchWithPlugins(sq, request, user_plugins=[])
                result_container = search_obj.search()
                
                raw_results = result_container.get_ordered_results()
                raw_infoboxes = getattr(result_container, 'infoboxes', [])
                raw_answers = getattr(result_container, 'answers', [])
                
                results, infoboxes, answers = self._parse_aux_results(raw_results, raw_infoboxes, raw_answers)
                
                context_str, new_urls = self._assemble_context(results, infoboxes, answers, offset)

                return jsonify({
                    'context': context_str,
                    'new_urls': new_urls,
                    'results': results, 
                    'infoboxes': infoboxes,
                    'answers': answers,
                    'query': query
                })

            except ImportError:
                try:
                    search_url = f'{request.url_root}search'
                    params = {
                        'q': query,
                        'format': 'json',
                        'categories': categories,
                        'language': lang
                    }
                    
                    headers = {
                        'X-AI-Auxiliary': '1',
                        'Accept-Language': request.headers.get('Accept-Language', '')
                    }
                    
                    
                    res = network.get(search_url, params=params, headers=headers, timeout=2)
                    search_data = res.json()
                        
                    

                    results, infoboxes, answers = self._parse_aux_results(
                        search_data.get('results', []),
                        search_data.get('infoboxes', []),
                        search_data.get('answers', [])
                    )
                    
                    context_str, new_urls = self._assemble_context(results, infoboxes, answers, offset)

                    return jsonify({
                        'context': context_str,
                        'new_urls': new_urls,
                        'results': results, 
                        'infoboxes': infoboxes,
                        'answers': answers,
                        'query': query
                    })
                except Exception as e:
                    return jsonify({'results': [], 'error': str(e)})
            except Exception as e:
                return jsonify({'results': [], 'error': str(e)})

        @app.route('/ai-stream', methods=['POST'])
        def handle_ai_stream():
            data = request.json or {}
            if data.get('warmup'):
                return Response('', status=204)
            
            token = data.get('tk', '')
            q = data.get('q', '')
            lang = data.get('lang', 'all')
            
            try:
                ts, sig = token.rsplit('.', 1)
                expected = hashlib.sha256(f"{ts}{self.secret}".encode()).hexdigest()
                if sig != expected or (time.time() - float(ts)) > TOKEN_EXPIRY_SEC:
                    abort(403)
            except (ValueError, KeyError, AttributeError):
                abort(403)

            context_text = data.get('context', '')
            prev_answer = (data.get('prev_answer') or '')[-4000:]
            
            if not self.api_key:
                return Response("Missing API key or query", status=400)
            
            today = time.strftime("%Y-%m-%d")
            target_words = int(self.max_tokens * 0.2)
            lang_instruction = f" Respond in {lang}." if lang not in ('all', 'auto') else ""

            SYSTEM = f"You are a search synthesis engine. Direct, grounded, citation-accurate. Today is {today}.{lang_instruction}"
            max_source_idx = 0
            if context_text:
                indices = re.findall(r'\[(\d+)\]', context_text)
                if indices:
                    max_source_idx = max(map(int, indices))

            CORE_RULES = [
                "DENSITY 4/5: Expert-briefing level. No filler, no transitions. Every sentence = new information.",
                f"BREVITY: {target_words} words max. Complete, not verbose.",
                f"CITATIONS: Cite format is [n] for facts from grounding sources",
                "NO HEDGE: State answers confidently. Note uncertainty only if critical.",
            ]

            if q == "Continue":
                task = "CONTINUE: Pick up exactly where previous answer stopped. No repetition. Seamless flow."
            elif prev_answer:
                task = "FOLLOW-UP: Address the new question using prior context. Prioritize the new query."
            else:
                task = "ANSWER FIRST: Lead with the direct answer. No preamble, no context-setting."

            grounding = "GROUNDING: KNOWLEDGE GRAPH > DEEP > SHALLOW." if context_text else "GROUNDING: No sources available. Use knowledge and note 'based on general knowledge'."
            history_rule = "HISTORY: Refer to prior exchange for context. Ideally, do not repeat any claims." if prev_answer else None

            instructions = [task] + CORE_RULES + [grounding]
            if history_rule:
                instructions.append(history_rule)

            numbered_instructions = "\n".join(f"{i+1}. {r}" for i, r in enumerate(instructions))
            prompt = f"""<system>{SYSTEM}</system>

<sources>
{context_text or 'None.'}
</sources>

<history>
{prev_answer or 'None.'}
</history>

<query>{q}</query>

<instructions>
{numbered_instructions}
</instructions>

<answer>"""

            def stream_gemini():
                if '?' in self.endpoint_url:
                    url = f"{self.endpoint_url}&key={self.api_key}"
                else:
                    url = f"{self.endpoint_url}?key={self.api_key}"

                conn = None
                try:
                    conn, path = _get_streaming_connection(url)
                    payload = json.dumps({"contents": [{"parts": [{"text": prompt}]}], "generationConfig": {"maxOutputTokens": self.max_tokens, "temperature": self.temperature, "stopSequences": ["</answer>"]}})
                    conn.request("POST", path, body=payload, headers={"Content-Type": "application/json"})
                    res = conn.getresponse()
                    
                    if res.status != 200:
                        logger.error(f"{PLUGIN_NAME}: Gemini API {res.status}")
                        return

                    decoder = json.JSONDecoder()
                    buffer = ""
                    while True:
                        chunk = res.read(STREAM_CHUNK_SIZE)
                        if not chunk: break
                        buffer += chunk.decode('utf-8', errors='replace')
                        while buffer:
                            buffer = buffer.lstrip()
                            if not buffer: break
                            try:
                                obj, idx = decoder.raw_decode(buffer)
                                items = obj if isinstance(obj, list) else [obj]
                                for item in items:
                                    candidates = item.get('candidates', [])
                                    if candidates:
                                        content = candidates[0].get('content', {})
                                        parts = content.get('parts', [])
                                        if parts:
                                            text = parts[0].get('text', '')
                                            if text: yield text
                                buffer = buffer[idx:]
                            except json.JSONDecodeError: break
                except Exception as e:
                    logger.error(f"{PLUGIN_NAME}: Gemini stream error: {e}")
                finally:
                    if conn: conn.close()

            def stream_openai_compatible():
                conn = None
                try:
                    conn, path = _get_streaming_connection(self.endpoint_url)
                    payload = json.dumps({
                        "model": self.model,
                        "messages": [{"role": "user", "content": prompt}],
                        "stream": True,
                        "max_tokens": self.max_tokens,
                        "temperature": self.temperature,
                        "stop": ["</answer>"]
                    })
                    headers = {
                        "Content-Type": "application/json",
                        "HTTP-Referer": "https://github.com/searxng/searxng",
                        "X-Title": "SearXNG"
                    }
                    if self.provider == 'azure':
                        headers['api-key'] = self.api_key
                    else:
                        headers['Authorization'] = f"Bearer {self.api_key}"
                    conn.request("POST", path, body=payload, headers=headers)
                    res = conn.getresponse()

                    if res.status != 200:
                        logger.error(f"{PLUGIN_NAME}: {self.provider} API {res.status}")
                        return

                    decoder = json.JSONDecoder()
                    buffer = b""
                    while True:
                        chunk = res.read(STREAM_CHUNK_SIZE)
                        if not chunk: break
                        buffer += chunk
                        while b"\n" in buffer:
                            line_bytes, buffer = buffer.split(b"\n", 1)
                            line = line_bytes.decode('utf-8', errors='replace')
                            if line.startswith("data: "):
                                data_str = line[6:].strip()
                                if data_str == "[DONE]": return
                                try:
                                    obj, _ = decoder.raw_decode(data_str)
                                    content = obj.get("choices", [{}])[0].get("delta", {}).get("content", "")
                                    if content: yield content
                                except json.JSONDecodeError:
                                    pass
                except Exception as e:
                    logger.error(f"{PLUGIN_NAME}: {self.provider} stream error: {e}")
                finally:
                    if conn: conn.close()

            generator = stream_gemini if self.is_gemini else stream_openai_compatible


            # If configured, force-unload Ollama model right after finishing the stream.

            # This uses the native /api/chat endpoint with keep_alive=0.

            if self.provider == 'ollama' and getattr(self, 'ollama_unload_after', False):

                gen_fn = generator

                def generator():

                    try:

                        yield from gen_fn()

                    finally:

                        self._ollama_unload_model()
            return Response(generator(), mimetype='text/event-stream', headers={
                'X-Accel-Buffering': 'no',
                'Cache-Control': 'no-cache, no-store',
                'Connection': 'keep-alive',
                'Transfer-Encoding': 'chunked',
                'Content-Encoding': 'identity',
            })
        return True

    def _assemble_context(self, raw_results, infoboxes, answers, offset=0) -> tuple[str, list]:
        """Builds context string from normalized search data. Returns (context_str, urls)."""
        context_parts = []
        result_urls = []
        
        # Knowledge graph
        knowledge_graph_lines = []
        for ib in infoboxes:
            ib_name = ib.get('name', '') or ib.get('infobox', '') or ib.get('title', '')
            ib_content = str(ib.get('content', '')).replace('\n', ' ').strip()
            
            if ib_name:
                parts = [f"INFOBOX [{ib_name}]:"]
                if ib_content:
                    parts.append(ib_content)
                for attr in ib.get('attributes', []):
                    attr_label = attr.get('label', '')
                    attr_value = attr.get('value', '')
                    if attr_label and attr_value:
                        parts.append(f"  {attr_label}: {attr_value}")
                
                knowledge_graph_lines.append(" ".join(parts) if len(parts) == 2 else "\n".join(parts))

        for ans_text in answers:
            if ans_text and not str(ans_text).startswith('<'):
                knowledge_graph_lines.append(f"ANSWER: {str(ans_text)[:300]}")
        
        if knowledge_graph_lines:
            context_parts.append("KNOWLEDGE GRAPH:\n" + "\n".join(knowledge_graph_lines))
        
        # Deep sources: full content
        deep_lines = []
        for i, r in enumerate(raw_results[:self.context_deep_count]):
            url = r.get('url', '')
            result_urls.append(url)
            domain = urlparse(url).netloc.replace('www.', '')
            date = r.get('publishedDate')
            date_str = f" ({date})" if date else ""
            title = (r.get('title') or "").replace('\n', ' ').strip()
            content = str(r.get('content', '')).replace('\n', ' ').strip()[:800]
            idx = i + 1 + offset
            deep_lines.append(f"[{idx}] {domain}{date_str}: {title}: {content}")
        
        if deep_lines:
            context_parts.append("DEEP SOURCES:\n" + "\n".join(deep_lines))
            
        # Shallow sources: headlines only
        if self.context_shallow_count > 0:
            shallow_lines = []
            start_idx = self.context_deep_count
            end_idx = self.context_deep_count + self.context_shallow_count
            for i, r in enumerate(raw_results[start_idx:end_idx]):
                url = r.get('url', '')
                result_urls.append(url)
                domain = urlparse(url).netloc.replace('www.', '')
                title = (r.get('title') or '').replace('\n', ' ').strip()[:60]
                idx = i + 1 + start_idx + offset
                shallow_lines.append(f"[{idx}] {domain}: {title}")
            
            if shallow_lines:
                context_parts.append("SHALLOW SOURCES (headlines):\n" + "\n".join(shallow_lines))
        
        return "\n\n".join(context_parts), result_urls

    def post_search(self, request: "SXNG_Request", search: "SearchWithPlugins") -> EngineResults:
        results = EngineResults()
        try:
            if request and hasattr(request, 'headers') and request.headers.get('X-AI-Auxiliary'):
                return results
            
            current_tabs = set(search.search_query.categories)
            if not current_tabs: current_tabs = {'general'}

            if not self.active or not self.api_key or search.search_query.pageno > 1 or not self.allowed_tabs.intersection(current_tabs):
                return results

            raw_results = search.result_container.get_ordered_results()
            raw_infoboxes = getattr(search.result_container, 'infoboxes', [])
            raw_answers = getattr(search.result_container, 'answers', [])
            
            # Normalize for unified context assembly
            _, infoboxes, answers = self._parse_aux_results(raw_results, raw_infoboxes, raw_answers)
            context_str, _ = self._assemble_context(raw_results, infoboxes, answers)


            ts = str(int(time.time()))
            q_clean = search.search_query.query.strip()
            lang = search.search_query.lang
            sig = hashlib.sha256(f"{ts}{self.secret}".encode()).hexdigest()
            tk = f"{ts}.{sig}"
            
            b64_context = base64.b64encode(context_str.encode('utf-8')).decode('utf-8')
            js_q = json.dumps(q_clean)
            js_lang = json.dumps(lang)
            total_context_count = self.context_deep_count + self.context_shallow_count
            js_urls = json.dumps([r.get('url') for r in raw_results[:total_context_count]])

            is_interactive = self.interactive
            
            interactive_css = INTERACTIVE_CSS if is_interactive else ''
            interactive_html = INTERACTIVE_HTML if is_interactive else ''
            interactive_js_init = INTERACTIVE_JS if is_interactive else ''

            interactive_js_complete = "footer.style.display = 'flex';" if is_interactive else ''
            stream_fn_sig = 'async function startStream(overrideQ = null, prevAnswer = null, auxContext = null)'
            stream_q = 'overrideQ || q_init' if is_interactive else 'q_init'
            stream_body = f'''prev_answer: prevAnswer''' if is_interactive else ''

            html_payload = f'''
                <article id="sxng-stream-box" class="answer" style="display:none; margin: 1rem 0;">
                    <style>
                        @keyframes sxng-fade-pulse {{
                            0%, 100% {{ opacity: 0.1; }}
                            50% {{ opacity: 1; }}
                        }}
                        @keyframes sxng-fade-in {{
                            from {{ opacity: 0; }}
                            to {{ opacity: 1; }}
                        }}
                        #sxng-stream-data {{
                            position: relative;
                            margin: 0;
                            min-height: 1.5em;
                        }}
                        .sxng-cursor {{
                            display: inline-block;
                            width: 0.6em;
                            height: 1.2em;
                            background: var(--color-result-link-visited, var(--color-result-link, #b48ead));
                            vertical-align: text-bottom;
                            animation: sxng-fade-pulse 1s ease-in-out infinite;
                            margin-right: 0.2rem;
                            border-radius: 2px;
                        }}
                        .sxng-chunk {{
                            opacity: 1;
                        }}
                        @media (min-width: 769px) {{
                            .sxng-chunk {{
                                animation: sxng-fade-in 0.3s ease-out;
                            }}
                        }}
                        {interactive_css}
                    </style>
                    <p id="sxng-stream-data" style="white-space: pre-wrap; color: var(--color-result-description); font-size: 0.95rem; margin:0;"><span class="sxng-cursor"></span></p>
                    {interactive_html}
                    <script>
                    (async () => {{
                        const is_interactive = {'true' if is_interactive else 'false'};
                        const q_init = {js_q};
                        const lang_init = {js_lang};
                        let urls = {js_urls};
                        const b64_init = "{b64_context}";
                        const tk_init = "{tk}";
                        const conversation = {{
                            originalQuery: q_init,
                            originalContext: new TextDecoder().decode(Uint8Array.from(atob(b64_init), c => c.charCodeAt(0))),
                            originalSources: [...urls],
                            turns: [{{role: 'user', content: q_init, ts: Date.now()}}]
                        }};
                        const box = document.getElementById('sxng-stream-box');
                        const data = document.getElementById('sxng-stream-data');
                        const wrapper = box.closest('.answer');
                        if (wrapper) wrapper.style.display = 'none';
                        let restored = false;
                        let isStreaming = false;
                        
                        {CITATION_HELPER_JS}

                        {interactive_js_init}
                        function synthesizeQuery(original, followup) {{
                            // combine with original for context, stripping generic question starters
                            const cleanOrig = original.replace(/^(what|how|why|when|where|who|which|is|are|can|does|do)(\s+(is|are|do|does|can|to|a|an|the))?\s+/i, '');
                            const origWords = cleanOrig.split(' ').slice(0, 12);
                            return `${{origWords.join(' ')}} ${{followup}}`.trim();
                        }}



                        {stream_fn_sig} {{
                            if (isStreaming) {{
                                console.warn('[AI Answers] Stream already in progress, ignoring duplicate call');
                                return;
                            }}
                            
                            isStreaming = true;
                            try {{
                                const ctx = auxContext || conversation.originalContext;
                                if (wrapper) wrapper.style.display = '';
                                box.style.display = 'block';

                                const controller = new AbortController();
                                const timeoutId = setTimeout(() => controller.abort(), 60000);
                                const finalQ = {stream_q};
                                
                                const bodyObj = {{ q: finalQ, lang: lang_init, context: ctx, tk: tk_init{', ' + stream_body if stream_body else ''} }};
                                const res = await fetch('/ai-stream', {{
                                    method: 'POST',
                                    headers: {{ 'Content-Type': 'application/json' }},
                                    body: JSON.stringify(bodyObj),
                                    signal: controller.signal
                                }});

                                clearTimeout(timeoutId);
                                if (!res.ok) {{
                                    const errSpan = document.createElement('span');
                                    errSpan.style.color = '#bf616a';
                                    errSpan.textContent = "Error: " + res.statusText;
                                    data.appendChild(errSpan);
                                    return;
                                }}

                                const reader = res.body.getReader();
                                const decoder = new TextDecoder();
                                let cursor = data.querySelector('.sxng-cursor');
                                if (!cursor) {{
                                    cursor = document.createElement('span');
                                    cursor.className = 'sxng-cursor';
                                    data.appendChild(cursor);
                                }}

                                let started = false;
                                let pendingSpace = '';
                                let lastScrollKick = 0;
                                let collectedResponse = '';

                                let buffer = '';
                                const flushBuffer = (force = false) => {{
                                    if (!buffer) return;
                                    
                                    if (force) {{
                                        const fragment = renderCitations(buffer, urls);
                                        if (cursor) cursor.before(fragment);
                                        else data.appendChild(fragment);
                                        buffer = '';
                                        return;
                                    }}

                                    while (true) {{
                                        const match = buffer.match(/(\\[\\d+(?:,\\s*\\d+)*\\])/);
                                        
                                        if (!match) break;
                                        
                                        const preText = buffer.substring(0, match.index);
                                        if (preText) {{
                                            const s = document.createElement('span');
                                            s.className = 'sxng-chunk';
                                            s.textContent = preText;
                                            cursor.before(s);
                                        }}

                                        const citationText = match[0];
                                        const fragment = renderCitations(citationText, urls);
                                        cursor.before(fragment);

                                        buffer = buffer.substring(match.index + match[0].length);
                                    }}

                                    const openIdx = buffer.lastIndexOf('[');
                                    if (openIdx === -1) {{
                                        if (buffer) {{
                                            const s = document.createElement('span');
                                            s.className = 'sxng-chunk';
                                            s.textContent = buffer;
                                            cursor.before(s);
                                            buffer = '';
                                        }}
                                    }} else {{
                                        const safeChunk = buffer.substring(0, openIdx);
                                        if (safeChunk) {{
                                            const s = document.createElement('span');
                                            s.className = 'sxng-chunk';
                                            s.textContent = safeChunk;
                                            cursor.before(s);
                                        }}
                                        buffer = buffer.substring(openIdx);
                                        
                                        if (buffer.length > 50) {{
                                            const s = document.createElement('span');
                                            s.className = 'sxng-chunk';
                                            s.textContent = buffer[0];
                                            cursor.before(s);
                                            buffer = buffer.substring(1);
                                        }}
                                    }}
                                }};

                                while (true) {{
                                    const {{done, value}} = await reader.read();
                                    if (done) break;

                                    const chunk = decoder.decode(value, {{stream: true}});
                                    if (chunk) {{
                                        collectedResponse += chunk;
                                        
                                        // Initial scroll/focus logic (only once)
                                        if (!started) {{
                                            const trimmed = chunk.replace(/^[\\s.,;:!?]+/, '');
                                            if (!trimmed && !collectedResponse.trim()) continue; // Skip leading garbage
                                            if (cursor && !cursor.isConnected) data.appendChild(cursor);
                                            started = true;
                                        }}

                                        buffer += chunk;
                                        flushBuffer(false);

                                        const now = Date.now();  // Periodic repaint for mobile
                                        if (now - lastScrollKick > 500) {{
                                            lastScrollKick = now;
                                            void window.getComputedStyle(data).opacity;
                                        }}
                                    }}
                                }}
                                
                                // Final flush of any partial text (e.g. unclosed brackets)
                                flushBuffer(true);
                                
                                if (cursor) cursor.remove();

                                // Cleanup trailing newlines
                                let last = data.lastChild;
                                while (last) {{
                                    if (last.textContent && last.textContent.trim().length === 0) {{
                                        const prev = last.previousSibling;
                                        last.remove();
                                        last = prev;
                                    }} else {{
                                        if (last.textContent) last.textContent = last.textContent.trimEnd();
                                        break;
                                    }}
                                }}

                                if (!started) {{
                                    const cursor = data.querySelector('.sxng-cursor');
                                    if (cursor) cursor.remove();
                                    const errSpan = document.createElement('span');
                                    errSpan.style.color = '#bf616a';
                                    errSpan.textContent = 'No response received. Check API configuration.';
                                    data.appendChild(errSpan);
                                    return;
                                }}

                                {interactive_js_complete}

                                if (collectedResponse) {{
                                    conversation.turns.push({{role: 'assistant', content: collectedResponse.trim(), ts: Date.now()}});
                                }}

                            }} catch (e) {{
                                console.error(e);
                                if (box.parentElement) box.parentElement.remove();
                                else box.remove();
                            }} finally {{
                                // Always release lock, even on error
                                isStreaming = false;
                            }}
                        }}

                        // Warmup handshake
                        fetch('/ai-stream', {{
                            method: 'POST',
                            headers: {{'Content-Type': 'application/json'}},
                            body: JSON.stringify({{warmup: true}}),
                            keepalive: true
                        }}).catch(() => {{}});

                        if (!restored) startStream();
                    }})();
                    </script>
                </article>
            '''
            search.result_container.answers.add(results.types.Answer(answer=Markup(html_payload)))
        except Exception as e:
            logger.error(f"{PLUGIN_NAME}: {e}")
        return results