#include <algorithm>
#include <atomic>
#include <cctype>
#include <cinttypes>
#include <cstring>
#include <fstream>
#include <iostream>
#include <iterator>
#include <map>
#include <mutex>
#include <optional>
#include <sstream>
#include <stdexcept>
#include <string>
#include <thread>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include <string>

#ifdef _OPENMP
#include <omp.h>
#else
inline int omp_get_max_threads(){ return 1; }
inline int omp_get_thread_num(){ return 0; }
#endif

extern unsigned char dictionary_json[];
extern unsigned int dictionary_json_len;


static inline bool is_space(char c){ return std::isspace(static_cast<unsigned char>(c)) != 0; }
static inline char to_low(char c){ return static_cast<char>(std::tolower(static_cast<unsigned char>(c))); }
static inline void safe_flush(std::ostream &os){ os.flush(); }

struct DictionaryEntry {
    std::string pos;
    std::string word;
    std::vector<std::string> definitions;
};

static std::vector<DictionaryEntry> global_dictionary_entries;
static std::unordered_map<std::string, std::vector<std::string>> global_def_tokens_cache;
static std::unordered_map<std::string, std::vector<std::string>> global_pos_cache;

static inline bool is_word_char_for_key(char c){
    unsigned char uc = static_cast<unsigned char>(c);
    return std::isalnum(uc) != 0 || c == '\'' || c == '-';
}

static std::string normalize_dictionary_key(const std::string &s){
    size_t b = 0, e = s.size();
    while (b < e && !is_word_char_for_key(s[b])) ++b;
    while (e > b && !is_word_char_for_key(s[e - 1])) --e;

    std::string out;
    out.reserve(e - b);
    for (size_t i = b; i < e; ++i) out.push_back(to_low(s[i]));
    return out;
}

static std::string normalize_pos_tag(const std::string &s){
    std::string out;
    out.reserve(s.size());
    for (char c : s){
        unsigned char uc = static_cast<unsigned char>(c);
        if (std::isalpha(uc) != 0) out.push_back(to_low(c));
    }
    return out;
}

enum class PosClass {
    Unknown,
    Noun,
    Verb,
    Adj,
    Adv,
    Pron,
    Prep,
    Conj,
    Det,
    Num,
    Interj
};

static PosClass pos_class_from_tag(const std::string &tag){
    if (tag == "n" || tag == "noun") return PosClass::Noun;
    if (tag == "v" || tag == "verb" || tag == "part" || tag == "participle" || tag == "p") return PosClass::Verb;
    if (tag == "a" || tag == "adj" || tag == "adjective") return PosClass::Adj;
    if (tag == "adv" || tag == "adverb") return PosClass::Adv;
    if (tag == "pron" || tag == "pronoun") return PosClass::Pron;
    if (tag == "prep" || tag == "preposition") return PosClass::Prep;
    if (tag == "conj" || tag == "conjunction") return PosClass::Conj;
    if (tag == "art" || tag == "article" || tag == "det" || tag == "determiner") return PosClass::Det;
    if (tag == "num" || tag == "number") return PosClass::Num;
    if (tag == "interj" || tag == "interjection") return PosClass::Interj;
    return PosClass::Unknown;
}

static bool has_pos_class(const std::vector<std::string> &tags, PosClass cls){
    for (const auto &t : tags){
        if (pos_class_from_tag(t) == cls) return true;
    }
    return false;
}

static const std::vector<std::string> &dictionary_pos_for_token(const std::string &surface){
    static const std::vector<std::string> empty;
    auto key = normalize_dictionary_key(surface);
    if (key.empty()) return empty;

    auto it = global_pos_cache.find(key);
    return (it == global_pos_cache.end()) ? empty : it->second;
}

static bool first_alpha_is_upper(const std::string &s){
    for (char c : s){
        unsigned char uc = static_cast<unsigned char>(c);
        if (std::isalpha(uc) != 0) return std::isupper(uc) != 0;
    }
    return false;
}

static bool first_alpha_is_lower(const std::string &s){
    for (char c : s){
        unsigned char uc = static_cast<unsigned char>(c);
        if (std::isalpha(uc) != 0) return std::islower(uc) != 0;
    }
    return false;
}

static bool is_sentence_boundary_token(const std::string &s){
    if (s.empty()) return false;
    char c = s.back();
    return c == '.' || c == '!' || c == '?';
}

static bool is_open_punct_token(const std::string &s){
    return s == "(" || s == "[" || s == "{" || s == "\"" || s == "'";
}

static bool is_punctuation_only_token(const std::string &s){
    if (s.empty()) return false;
    for (char c : s){
        unsigned char uc = static_cast<unsigned char>(c);
        if (std::isalnum(uc) != 0) return false;
    }
    return true;
}

static std::string to_lower_ascii(std::string s) {
    std::transform(s.begin(), s.end(), s.begin(),
                   [](unsigned char c) { return static_cast<char>(std::tolower(c)); });
    return s;
}

static bool is_common_determiner(const std::string &s) {
    static const std::unordered_set<std::string> words = {
        "a", "an", "the", "this", "that", "these", "those",
        "my", "your", "his", "her", "its", "our", "their",
        "all", "some", "any", "each", "every", "either", "neither",
        "no", "many", "much", "more", "most", "few", "fewer", "less",
        "little", "several", "both", "another", "other", "such",
        "own", "same", "certain", "what", "which", "whose",
        "whatever", "whichever", "enough", "various", "particular"
    };

    return words.find(to_lower_ascii(s)) != words.end();
}

static bool is_common_preposition(const std::string &s) {
    static const std::unordered_set<std::string> words = {
        "of", "in", "on", "at", "by", "for", "from", "with",
        "into", "onto", "about", "over", "under", "after", "before",
        "between", "through", "during", "without", "within", "across",
        "against", "among", "around", "as", "like", "per", "via",
        "toward", "towards", "upon", "beneath", "beside", "besides",
        "beyond", "inside", "outside", "near", "past", "since",
        "until", "till", "up", "down", "off", "out", "underneath",
        "amid", "amidst", "despite", "except", "regarding", "concerning",
        "throughout", "amongst", "opposite", "along", "behind", "ahead",
        "above", "below", "plus", "minus"
    };

    return words.find(to_lower_ascii(s)) != words.end();
}

static bool is_common_aux_or_modal(const std::string &s) {
    static const std::unordered_set<std::string> words = {
        "to", "be", "am", "is", "are", "was", "were", "been", "being",
        "have", "has", "had", "do", "does", "did",
        "can", "could", "may", "might", "must", "shall", "should", "will",
        "would", "ought", "need", "dare", "used",
        "cannot", "can't", "won't", "wouldn't", "shouldn't", "couldn't",
        "mustn't", "mayn't", "mightn't", "shan't"
    };

    return words.find(to_lower_ascii(s)) != words.end();
}

static bool begins_with_vowel_sound(const std::string &s){
    if (s.empty()) return false;

    const std::string t = to_lower_ascii(s);

    static const std::unordered_set<std::string> vowel_sound_exceptions = {
        "hour", "hours", "honest", "honesty", "honor", "honour", "honorable", "honourable",
        "heir", "heiress", "herb", "herbal", "herbalist", "homage", "hono(u)rary"
    };

    static const std::unordered_set<std::string> consonant_sound_prefixes = {
        "uni", "univ", "unit", "use", "user", "euro", "eu", "one", "once", "ouija",
        "ufo", "ut", "uk", "ubiq", "ewe", "eul", "eup", "x", "y"
    };

    for (const auto &w : vowel_sound_exceptions) {
        if (t.rfind(w, 0) == 0) {
            return true;
        }
    }

    for (const auto &p : consonant_sound_prefixes) {
        if (t.rfind(p, 0) == 0) {
            return false;
        }
    }

    if (t.empty()) return false;
    char c = t[0];
    return c == 'a' || c == 'e' || c == 'i' || c == 'o' || c == 'u';
}

static double english_rule_bonus(const std::string &context_tok, const std::string &cand){
    const std::string ctx_key = normalize_dictionary_key(context_tok);
    const std::string cand_key = normalize_dictionary_key(cand);

    const auto &ctx_tags = dictionary_pos_for_token(context_tok);
    const auto &cand_tags = dictionary_pos_for_token(cand);

    const bool sentence_start = context_tok.empty() || is_sentence_boundary_token(context_tok) || is_open_punct_token(context_tok);

    const bool cand_nounish = has_pos_class(cand_tags, PosClass::Noun) ||
                               has_pos_class(cand_tags, PosClass::Adj) ||
                               has_pos_class(cand_tags, PosClass::Pron) ||
                               has_pos_class(cand_tags, PosClass::Num);

    const bool cand_verbish = has_pos_class(cand_tags, PosClass::Verb);
    const bool cand_advish  = has_pos_class(cand_tags, PosClass::Adv);
    const bool cand_prepish = has_pos_class(cand_tags, PosClass::Prep);
    const bool cand_detish  = has_pos_class(cand_tags, PosClass::Det);

    double bonus = 0.0;

    if (!cand_key.empty()){
        if (sentence_start){
            bonus += first_alpha_is_upper(cand) ? 0.22 : -0.08;
        } else if (first_alpha_is_upper(cand)){
            bonus -= 0.03;
        }
    }

    if (ctx_key == "a" || ctx_key == "an"){
        const bool vowel = begins_with_vowel_sound(cand_key.empty() ? cand : cand_key);
        bonus += ((ctx_key == "an") == vowel) ? 0.28 : -0.18;
    }

    const bool ctx_det = has_pos_class(ctx_tags, PosClass::Det) || is_common_determiner(ctx_key);
    const bool ctx_prep = has_pos_class(ctx_tags, PosClass::Prep) || is_common_preposition(ctx_key);
    const bool ctx_aux = is_common_aux_or_modal(ctx_key);

    if (ctx_det){
        if (cand_nounish) bonus += 0.20;
        if (cand_verbish || cand_advish || cand_prepish) bonus -= 0.08;
    }

    if (ctx_prep){
        if (cand_nounish) bonus += 0.16;
        if (cand_verbish) bonus -= 0.06;
    }

    if (ctx_aux){
        if (cand_verbish) bonus += 0.18;
        if (cand_detish) bonus -= 0.04;
    }

    if (has_pos_class(ctx_tags, PosClass::Pron) || has_pos_class(ctx_tags, PosClass::Noun)){
        if (cand_verbish) bonus += 0.05;
    }

    if (!context_tok.empty() && (context_tok.back() == ',' || context_tok.back() == ';' || context_tok.back() == ':')){
        if (!cand.empty() && first_alpha_is_lower(cand)) bonus += 0.04;
    }

    if (is_punctuation_only_token(cand)){
        if (sentence_start) bonus -= 0.05;
        else if (!context_tok.empty() && std::isalnum(static_cast<unsigned char>(context_tok.back())) != 0) bonus += 0.03;
    }

    if (is_sentence_boundary_token(cand)) bonus += 0.06;

    return bonus;
}

static std::vector<std::string> tokenize_whitespace(const std::string &s){
    std::istringstream iss(s);
    std::vector<std::string> out;
    std::string t;
    while (iss >> t) out.push_back(t);
    return out;
}

// Static helper to check dictionary presence using your existing pos cache
static inline bool check_dictionary(const std::string &word) {
    return !dictionary_pos_for_token(word).empty();
}

// Morphological lemmatizer to convert sub-tokens to their dictionary-defined base form.
// Returns std::optional<std::string>: contains the valid base form if found in the dictionary,
// or std::nullopt if the token has no definition and should be ignored.
static std::optional<std::string> get_valid_base_form(const std::string &word) {
    if (word.empty()) return std::nullopt;

    // 1. BEFORE extracting/stemming, check if the original form is already in the dictionary
    if (check_dictionary(word)) {
        return word;
    }

    // 2. Continuous verb form (-ing)
    if (word.length() > 3 && word.substr(word.length() - 3) == "ing") {
        // e.g., playing -> play
        std::string base1 = word.substr(0, word.length() - 3);
        if (check_dictionary(base1)) return base1;

        // e.g., making -> make
        std::string base2 = base1 + "e";
        if (check_dictionary(base2)) return base2;

        // e.g., running -> run (doubled consonant check)
        if (base1.length() >= 2 && base1.back() == base1[base1.length() - 2]) {
            std::string base3 = base1.substr(0, base1.length() - 1);
            if (check_dictionary(base3)) return base3;
        }
    }

    // 3. Past tense / Participle form (-ed)
    if (word.length() > 2 && word.substr(word.length() - 2) == "ed") {
        // e.g., baked -> bake
        std::string base1 = word.substr(0, word.length() - 1);
        if (check_dictionary(base1)) return base1;

        // e.g., played -> play
        std::string base2 = word.substr(0, word.length() - 2);
        if (check_dictionary(base2)) return base2;

        // e.g., hopped -> hop (doubled consonant check)
        if (base2.length() >= 2 && base2.back() == base2[base2.length() - 2]) {
            std::string base3 = base2.substr(0, base2.length() - 1);
            if (check_dictionary(base3)) return base3;
        }
    }

    // 4. Superlative form (-est)
    if (word.length() > 3 && word.substr(word.length() - 3) == "est") {
        // e.g., simplest -> simple (strip 'st')
        std::string base1 = word.substr(0, word.length() - 2);
        if (check_dictionary(base1)) return base1;

        // e.g., fastest -> fast
        std::string base2 = word.substr(0, word.length() - 3);
        if (check_dictionary(base2)) return base2;

        // e.g., happiest -> happy (-iest -> -y)
        if (word.length() > 4 && word.substr(word.length() - 4) == "iest") {
            std::string base_y = word.substr(0, word.length() - 4) + "y";
            if (check_dictionary(base_y)) return base_y;
        }

        // e.g., biggest -> big (doubled consonant check)
        if (base2.length() >= 2 && base2.back() == base2[base2.length() - 2]) {
            std::string base3 = base2.substr(0, base2.length() - 1);
            if (check_dictionary(base3)) return base3;
        }
    }

    // 5. Comparative form (-er)
    if (word.length() > 2 && word.substr(word.length() - 2) == "er") {
        // e.g., simpler -> simple (strip 'r')
        std::string base1 = word.substr(0, word.length() - 1);
        if (check_dictionary(base1)) return base1;

        // e.g., faster -> fast
        std::string base2 = word.substr(0, word.length() - 2);
        if (check_dictionary(base2)) return base2;

        // e.g., happier -> happy (-ier -> -y)
        if (word.length() > 3 && word.substr(word.length() - 3) == "ier") {
            std::string base_y = word.substr(0, word.length() - 3) + "y";
            if (check_dictionary(base_y)) return base_y;
        }

        // e.g., bigger -> big (doubled consonant check)
        if (base2.length() >= 2 && base2.back() == base2[base2.length() - 2]) {
            std::string base3 = base2.substr(0, base2.length() - 1);
            if (check_dictionary(base3)) return base3;
        }
    }

    // 6. Adverbial form (-ly)
    if (word.length() > 2 && word.substr(word.length() - 2) == "ly") {
        // e.g., quickly -> quick
        std::string base1 = word.substr(0, word.length() - 2);
        if (check_dictionary(base1)) return base1;

        // e.g., happily -> happy (-ily -> -y)
        if (word.length() > 3 && word.substr(word.length() - 3) == "ily") {
            std::string base2 = word.substr(0, word.length() - 3) + "y";
            if (check_dictionary(base2)) return base2;
        }

        // e.g., basically -> basic (-ally -> -ic)
        if (word.length() > 4 && word.substr(word.length() - 4) == "ally") {
            std::string base3 = word.substr(0, word.length() - 4);
            if (check_dictionary(base3)) return base3;

            std::string base4 = base3 + "al";
            if (check_dictionary(base4)) return base4;
        }

        // e.g., gently -> gentle (-ly -> -le)
        if (word.back() == 'y' && word.length() > 2 && word[word.length() - 2] == 'l') {
            std::string base5 = word.substr(0, word.length() - 1) + "e";
            if (check_dictionary(base5)) return base5;
        }
    }

    // 7. Noun suffix (-ness)
    if (word.length() > 4 && word.substr(word.length() - 4) == "ness") {
        // e.g., sadness -> sad
        std::string base1 = word.substr(0, word.length() - 4);
        if (check_dictionary(base1)) return base1;

        // e.g., happiness -> happy (-iness -> -y)
        if (base1.length() > 1 && base1.back() == 'i') {
            std::string base2 = base1.substr(0, base1.length() - 1) + "y";
            if (check_dictionary(base2)) return base2;
        }
    }

    // 8. Noun suffix (-ment)
    if (word.length() > 4 && word.substr(word.length() - 4) == "ment") {
        // e.g., development -> develop
        std::string base1 = word.substr(0, word.length() - 4);
        if (check_dictionary(base1)) return base1;
    }

    // 9. Plural / 3rd-person singular form (-ies -> -y)
    if (word.length() > 3 && word.substr(word.length() - 3) == "ies") {
        std::string base = word.substr(0, word.length() - 3) + "y";
        if (check_dictionary(base)) return base;
    }

    // 10. Plural / 3rd-person singular form (-ves -> -f / -fe)
    if (word.length() > 3 && word.substr(word.length() - 3) == "ves") {
        std::string base1 = word.substr(0, word.length() - 3) + "f";
        if (check_dictionary(base1)) return base1;

        std::string base2 = word.substr(0, word.length() - 3) + "fe";
        if (check_dictionary(base2)) return base2;
    }

    // 11. Plural form (-es)
    if (word.length() > 2 && word.substr(word.length() - 2) == "es") {
        std::string base = word.substr(0, word.length() - 2);
        if (check_dictionary(base)) return base;
    }

    // 12. Plural form (-s)
    if (word.length() > 1 && word.back() == 's' && word[word.length() - 2] != 's') {
        std::string base = word.substr(0, word.length() - 1);
        if (check_dictionary(base)) return base;
    }

    // No dictionary-defined base form found
    return std::nullopt;
}

static std::vector<std::string> tokenize_others(const std::string &s) {
    // Step A: Check if the entire input string already has a dictionary definition.
    if (check_dictionary(s)) {
        std::string lower_s;
        lower_s.reserve(s.size());
        for (char c : s) {
            lower_s.push_back(to_low(c));
        }
        return { lower_s };
    }

    std::vector<std::string> out;
    std::string cur;

    // Lambda to flush the accumulated token buffer, validate it, and add to output
    auto flush = [&]() {
        if (!cur.empty()) {
            auto valid_base = get_valid_base_form(cur);
            if (valid_base.has_value()) {
                out.push_back(valid_base.value());
            }
            cur.clear();
        }
    };

    for (size_t i = 0; i < s.size(); ++i) {
        unsigned char uc = static_cast<unsigned char>(s[i]);
        char ch = static_cast<char>(uc);

        // Split on spaces, underscores, hyphens, and slashes
        if (ch == '_' || ch == '-' || ch == '/' || std::isspace(uc)) {
            flush();
            continue;
        }

        const bool is_upper = std::isupper(uc) != 0;
        bool camel_boundary = false;

        if (is_upper && !cur.empty()) {
            unsigned char prev = static_cast<unsigned char>(s[i - 1]);
            const bool prev_lower_or_digit = (std::islower(prev) != 0) || (std::isdigit(prev) != 0);
            const bool prev_upper = std::isupper(prev) != 0;
            const bool next_lower =
                (i + 1 < s.size()) && (std::islower(static_cast<unsigned char>(s[i + 1])) != 0);

            camel_boundary = prev_lower_or_digit || (prev_upper && next_lower);
        }

        // Split on transitions between alphabetical characters and digits (e.g., "book5read")
        bool digit_boundary = false;
        if (!cur.empty()) {
            unsigned char prev = static_cast<unsigned char>(s[i - 1]);
            bool prev_digit = std::isdigit(prev) != 0;
            bool curr_digit = std::isdigit(uc) != 0;
            if (prev_digit != curr_digit) {
                digit_boundary = true;
            }
        }

        if (camel_boundary || digit_boundary) {
            flush();
        }

        if (std::isalnum(uc) != 0) {
            cur.push_back(static_cast<char>(std::tolower(uc)));
        } else {
            flush();
        }
    }

    flush();
    return out;
}

using StrPtr = const std::string*;
using TokenId = std::uint32_t;
static constexpr TokenId TOKEN_ID_INVALID = 0xFFFFFFFFu;

static inline std::size_t popcount_u64(std::uint64_t x){
#if defined(_MSC_VER)
    return static_cast<std::size_t>(__popcnt64(x));
#else
    return static_cast<std::size_t>(__builtin_popcountll(static_cast<unsigned long long>(x)));
#endif
}

struct StringInterner {
    std::unordered_set<std::string> pool;
    std::unordered_map<std::string, TokenId> id_by_string;
    std::vector<const std::string*> string_by_id;
    mutable std::mutex m;

    const std::string* intern(const std::string &s){
        std::lock_guard<std::mutex> lk(m);

        auto [it, inserted] = pool.emplace(s);
        if (inserted){
            const TokenId id = static_cast<TokenId>(string_by_id.size());
            id_by_string.emplace(*it, id);
            string_by_id.push_back(&*it);
        }

        return &*it;
    }

    TokenId id_of(const std::string &s) const {
        std::lock_guard<std::mutex> lk(m);

        auto it = id_by_string.find(s);
        return (it == id_by_string.end()) ? TOKEN_ID_INVALID : it->second;
    }

    TokenId id_of(const std::string *p) const {
        return p ? id_of(*p) : TOKEN_ID_INVALID;
    }

    const std::string* ptr_from_id(TokenId id) const {
        std::lock_guard<std::mutex> lk(m);

        return (id < string_by_id.size()) ? string_by_id[(size_t)id] : nullptr;
    }

    std::size_t size() const {
        std::lock_guard<std::mutex> lk(m);
        return string_by_id.size();
    }
};

static inline void bitset_set(std::uint64_t *bits, std::size_t words, TokenId id){
    if (id == TOKEN_ID_INVALID) return;
    const std::size_t idx = static_cast<std::size_t>(id >> 6);
    if (idx >= words) return;
    bits[idx] |= (1ULL << (id & 63u));
}

static inline std::size_t bitset_count(const std::uint64_t *bits, std::size_t words){
    std::size_t total = 0;
    for (std::size_t i = 0; i < words; ++i) total += popcount_u64(bits[i]);
    return total;
}

static inline std::size_t bitset_intersection_count(const std::uint64_t *a, const std::uint64_t *b, std::size_t words){
    std::size_t total = 0;
    for (std::size_t i = 0; i < words; ++i) total += popcount_u64(a[i] & b[i]);
    return total;
}

static void build_def_tokens_cache(){
    global_def_tokens_cache.clear();
    global_pos_cache.clear();

    global_def_tokens_cache.reserve(global_dictionary_entries.size());
    global_pos_cache.reserve(global_dictionary_entries.size());

    for (const auto &entry : global_dictionary_entries){
        // 1. Tokenize definitions once per entry to avoid redundant processing
        std::vector<std::string> all_def_toks;
        for (const auto &def : entry.definitions){
            auto toks = tokenize_others(def);
            all_def_toks.insert(all_def_toks.end(), toks.begin(), toks.end());
        }

        std::string pos = normalize_pos_tag(entry.pos);

        // 2. Split entry.word by semicolons or commas
        std::vector<std::string> sub_words;
        std::string current_sub;
        for (char c : entry.word) {
            if (c == ';' || c == ',') {
                if (!current_sub.empty()) {
                    sub_words.push_back(current_sub);
                    current_sub.clear();
                }
            } else {
                current_sub.push_back(c);
            }
        }
        if (!current_sub.empty()) {
            sub_words.push_back(current_sub);
        }

        // 3. Register each sub-word to the cache
        for (size_t i = 0; i < sub_words.size(); ++i) {
            const std::string key = normalize_dictionary_key(sub_words[i]);
            if (key.empty()) continue;

            if (!pos.empty()) {
                // Move on the final sub-word, copy otherwise
                if (i == sub_words.size() - 1) {
                    global_pos_cache[key].push_back(std::move(pos));
                } else {
                    global_pos_cache[key].push_back(pos);
                }
            }

            auto &defs = global_def_tokens_cache[key];
            defs.insert(defs.end(), all_def_toks.begin(), all_def_toks.end());
        }
    }

    // 4. Keep existing sorting and deduplication logic intact
    for (auto &pr : global_def_tokens_cache){
        auto &v = pr.second;
        std::sort(v.begin(), v.end());
        v.erase(std::unique(v.begin(), v.end()), v.end());
    }

    for (auto &pr : global_pos_cache){
        auto &v = pr.second;
        std::sort(v.begin(), v.end());
        v.erase(std::unique(v.begin(), v.end()), v.end());
    }
}

struct PtrHash { size_t operator()(StrPtr p) const noexcept { return std::hash<StrPtr>()(p); } };
struct PtrEq   { bool operator()(StrPtr a, StrPtr b) const noexcept { return a == b; } };

using NextSet = std::vector<StrPtr>;

struct NgramHash {
    std::size_t operator()(const std::vector<StrPtr>& v) const noexcept {
        std::size_t seed = v.size();
        for(auto& p : v) {
            seed ^= std::hash<StrPtr>()(p) + 0x9e3779b9 + (seed << 6) + (seed >> 2);
        }
        return seed;
    }
};

struct KnowledgeBase {
    StringInterner interner;
    // Modified to use a vector of StrPtr (N-gram) as the key
    std::unordered_map<std::vector<StrPtr>, NextSet, NgramHash> next;
    mutable std::mutex m;

    std::unordered_map<StrPtr, std::vector<StrPtr>, PtrHash, PtrEq> def_index;
    mutable std::mutex def_m;
    int def_depth = 0;

    void add_ngram(const std::vector<StrPtr>& ctx, StrPtr v){
        for (auto p : ctx) ensure_def_for_interned(p);
        ensure_def_for_interned(v);

        std::lock_guard<std::mutex> lk(m);
        auto &vec = next[ctx];
        for (auto p : vec) if (p == v) return;
        vec.push_back(v);
    }

    void set_def_depth(int D){
        std::lock_guard<std::mutex> lk(def_m);
        if (D != def_depth){
            def_index.clear();
            def_depth = D;
        }
    }

    void ensure_def_for_interned(StrPtr wp){
        if (wp == nullptr || def_depth <= 0) return;

        {
            std::lock_guard<std::mutex> lk(def_m);
            if (def_index.find(wp) != def_index.end()) return;
        }

        std::unordered_set<StrPtr, PtrHash, PtrEq> acc;
        std::vector<StrPtr> frontier;

        const std::string start_key = normalize_dictionary_key(*wp);
        if (!start_key.empty()){
            auto it_def = global_def_tokens_cache.find(start_key);
            if (it_def != global_def_tokens_cache.end()){
                for (const auto &tok : it_def->second){
                    StrPtr tp = interner.intern(tok);
                    if (acc.insert(tp).second) frontier.push_back(tp);
                }
            }
        }

        for (int depth = 1; depth < def_depth && !frontier.empty(); ++depth){
            std::vector<StrPtr> next_frontier;

            for (StrPtr w : frontier){
                const std::string key = normalize_dictionary_key(*w);
                if (key.empty()) continue;

                auto it2 = global_def_tokens_cache.find(key);
                if (it2 == global_def_tokens_cache.end()) continue;

                for (const auto &tok : it2->second){
                    StrPtr tp = interner.intern(tok);
                    if (acc.insert(tp).second) next_frontier.push_back(tp);
                }
            }

            frontier.swap(next_frontier);
        }

        std::vector<StrPtr> out;
        out.reserve(acc.size());
        for (StrPtr p : acc) out.push_back(p);

        {
            std::lock_guard<std::mutex> lk(def_m);
            def_index.emplace(wp, std::move(out));
        }
    }

    std::optional<NextSet> lookup_ngram(const std::vector<StrPtr>& ctx) const {
        std::lock_guard<std::mutex> lk(m);
        auto it = next.find(ctx);
        if (it == next.end()) return std::nullopt;
        return it->second;
    }
};

static void learn_tokens_ngram(KnowledgeBase &kb, const std::vector<std::string>& tokens, int n_gram_size) {
    std::vector<StrPtr> window;
    for (const auto& tok : tokens) {
        StrPtr tp = kb.interner.intern(tok);
        for (size_t j = 0; j < window.size(); ++j) {
            std::vector<StrPtr> ctx(window.begin() + j, window.end());
            kb.add_ngram(ctx, tp);
        }
        window.push_back(tp);
        if (window.size() > static_cast<size_t>(n_gram_size)) window.erase(window.begin());
    }
}

static std::vector<StrPtr>
intern_tokens(KnowledgeBase &kb, const std::vector<std::string> &tokens)
{
    std::vector<StrPtr> out;
    out.reserve(tokens.size());
    for (const auto &t : tokens) out.push_back(kb.interner.intern(t));
    return out;
}

static inline bool json_valid_index(size_t i, size_t n){ return i < n; }

static std::string parse_quoted_string(const std::string &text, size_t &i){
    std::string out;
    if (!json_valid_index(i, text.size()) || text[i] != '"') throw std::runtime_error("expected '\"'");
    ++i;
    while (json_valid_index(i, text.size())){
        char c = text[i++];
        if (c == '"') break;
        if (c == '\\'){
            if (!json_valid_index(i, text.size())) break;
            char e = text[i++];
            if (e=='n') out.push_back('\n');
            else if (e=='t') out.push_back('\t');
            else out.push_back(e);
        } else out.push_back(c);
    }
    return out;
}

static void skip_spaces(const std::string &s, size_t &i){
    while (json_valid_index(i, s.size()) && is_space(s[i])) ++i;
}

static void skip_json_value(const std::string &s, size_t &i);

static std::vector<std::string> parse_json_string_array(const std::string &text, size_t &i){
    std::vector<std::string> out;
    if (!json_valid_index(i, text.size()) || text[i] != '[') return out;

    ++i;
    while (true){
        skip_spaces(text, i);
        if (!json_valid_index(i, text.size())) break;
        if (text[i] == ']'){ ++i; break; }

        if (text[i] == '"') out.push_back(parse_quoted_string(text, i));
        else skip_json_value(text, i);

        skip_spaces(text, i);
        if (json_valid_index(i, text.size()) && text[i] == ','){ ++i; continue; }
        if (json_valid_index(i, text.size()) && text[i] == ']'){ ++i; break; }
    }

    return out;
}

static void skip_json_value(const std::string &s, size_t &i){
    skip_spaces(s, i);
    if (!json_valid_index(i, s.size())) return;

    if (s[i] == '"'){
        (void)parse_quoted_string(s, i);
        return;
    }

    if (s[i] == '['){
        ++i;
        while (json_valid_index(i, s.size())){
            skip_spaces(s, i);
            if (!json_valid_index(i, s.size())) break;
            if (s[i] == ']'){ ++i; break; }
            skip_json_value(s, i);
            skip_spaces(s, i);
            if (json_valid_index(i, s.size()) && s[i] == ','){ ++i; continue; }
            if (json_valid_index(i, s.size()) && s[i] == ']'){ ++i; break; }
        }
        return;
    }

    if (s[i] == '{'){
        ++i;
        while (json_valid_index(i, s.size())){
            skip_spaces(s, i);
            if (!json_valid_index(i, s.size())) break;
            if (s[i] == '}'){ ++i; break; }
            if (s[i] == '"'){
                (void)parse_quoted_string(s, i);
                skip_spaces(s, i);
                if (json_valid_index(i, s.size()) && s[i] == ':') ++i;
                skip_json_value(s, i);
                skip_spaces(s, i);
                if (json_valid_index(i, s.size()) && s[i] == ','){ ++i; continue; }
                if (json_valid_index(i, s.size()) && s[i] == '}'){ ++i; break; }
            } else {
                ++i;
            }
        }
        return;
    }

    while (json_valid_index(i, s.size())){
        char c = s[i];
        if (c == ',' || c == ']' || c == '}' || is_space(c)) break;
        ++i;
    }
}

static std::vector<DictionaryEntry> parse_dictionary_json(){
    std::vector<DictionaryEntry> dict;
    if (dictionary_json_len == 0) return dict;

    std::string text;
    text.reserve(dictionary_json_len);
    for (unsigned int b = 0; b < dictionary_json_len; ++b){
        text.push_back(static_cast<char>(dictionary_json[b]));
    }

    size_t i = 0;
    skip_spaces(text, i);
    if (!json_valid_index(i, text.size()) || text[i] != '[') return dict;
    ++i;

    while (true){
        skip_spaces(text, i);
        if (!json_valid_index(i, text.size())) break;
        if (text[i] == ']'){ ++i; break; }
        if (text[i] != '{'){
            skip_json_value(text, i);
            skip_spaces(text, i);
            if (json_valid_index(i, text.size()) && text[i] == ','){ ++i; continue; }
            if (json_valid_index(i, text.size()) && text[i] == ']'){ ++i; break; }
            continue;
        }

        ++i;
        DictionaryEntry entry;

        while (true){
            skip_spaces(text, i);
            if (!json_valid_index(i, text.size())) break;
            if (text[i] == '}'){ ++i; break; }

            std::string field = parse_quoted_string(text, i);
            skip_spaces(text, i);
            if (!json_valid_index(i, text.size()) || text[i] != ':') break;
            ++i;
            skip_spaces(text, i);

            if (field == "word"){
                entry.word = parse_quoted_string(text, i);
            } else if (field == "pos"){
                entry.pos = parse_quoted_string(text, i);
            } else if (field == "definitions"){
                entry.definitions = parse_json_string_array(text, i);
            } else {
                skip_json_value(text, i);
            }

            skip_spaces(text, i);
            if (json_valid_index(i, text.size()) && text[i] == ','){ ++i; continue; }
            if (json_valid_index(i, text.size()) && text[i] == '}'){ ++i; break; }
        }

        if (!entry.word.empty()) dict.push_back(std::move(entry));

        skip_spaces(text, i);
        if (json_valid_index(i, text.size()) && text[i] == ','){ ++i; continue; }
        if (json_valid_index(i, text.size()) && text[i] == ']'){ ++i; break; }
    }

    return dict;
}

static std::string best_candidate_by_similarity(
    const NextSet &cands,
    const std::vector<StrPtr> &prompt_ptrs,
    const std::vector<StrPtr> &resp_ptrs,
    const std::unordered_map<StrPtr, std::vector<StrPtr>, PtrHash, PtrEq> &def_index,
    const StringInterner &interner,
    const std::unordered_map<std::string,int> &recent_counts,
    double repeat_penalty,
    const std::string &context_tok)
{
    if (cands.empty()) return std::string();
    if (cands.size() == 1) return *cands[0];

    const std::size_t words = std::max<std::size_t>(1, (interner.size() + 63u) / 64u);

    std::vector<std::uint64_t> agg_words(words, 0ULL);

    auto add_token_and_defs = [&](StrPtr t){
        if (!t) return;

        // Process the main token and its direct definitions
        TokenId tid = interner.id_of(t);
        if (tid != TOKEN_ID_INVALID) bitset_set(agg_words.data(), words, tid);

        auto it = def_index.find(t);
        if (it != def_index.end()){
            for (StrPtr d : it->second){
                TokenId did = interner.id_of(d);
                if (did != TOKEN_ID_INVALID) bitset_set(agg_words.data(), words, did);
            }
        }

        // Tokenize the expanded dictionary string to process sub-tokens
        std::vector<std::string> sub_tokens = tokenize_others(*t);
        for (const std::string& sub_tok : sub_tokens) {

            // Use id_of to check if the sub-token exists without adding a new string to the pool
            TokenId sub_tid = interner.id_of(sub_tok);

            if (sub_tid != TOKEN_ID_INVALID) {
                // Add the sub-token itself to the bitset
                bitset_set(agg_words.data(), words, sub_tid);

                // Retrieve the interned pointer so we can look it up in def_index
                StrPtr sub_ptr = interner.ptr_from_id(sub_tid);
                if (sub_ptr) {
                    auto sub_it = def_index.find(sub_ptr);
                    if (sub_it != def_index.end()) {
                        // Add the sub-token's definitions to the bitset
                        for (StrPtr d : sub_it->second) {
                            TokenId did = interner.id_of(d);
                            if (did != TOKEN_ID_INVALID) bitset_set(agg_words.data(), words, did);
                        }
                    }
                }
            }
        }
    };

    for (StrPtr t : prompt_ptrs) add_token_and_defs(t);
    for (StrPtr t : resp_ptrs) add_token_and_defs(t);

    const std::size_t agg_count = bitset_count(agg_words.data(), words);

    const std::size_t M = cands.size();
    std::vector<std::uint64_t> cand_words(M * words, 0ULL);
    std::vector<std::size_t> cand_counts(M, 0);

    for (std::size_t i = 0; i < M; ++i){
        std::uint64_t *row = cand_words.data() + i * words;
        const StrPtr cand = cands[i];

        // Tag the candidate word itself
        TokenId cid = interner.id_of(cand);
        if (cid != TOKEN_ID_INVALID) bitset_set(row, words, cid);

        // Expand the candidate's direct definition
        auto it = def_index.find(cand);
        if (it != def_index.end()){
            for (StrPtr d : it->second){
                TokenId did = interner.id_of(d);
                if (did != TOKEN_ID_INVALID) bitset_set(row, words, did);
            }
        }

        // Tokenize the candidate to process sub-tokens (e.g., camelCase, hyphens)
        std::vector<std::string> sub_tokens = tokenize_others(*cand);
        for (const std::string& sub_tok : sub_tokens) {

            // Check if the sub-token exists in our knowledge base
            TokenId sub_tid = interner.id_of(sub_tok);

            if (sub_tid != TOKEN_ID_INVALID) {
                // Add the sub-token itself to the candidate's semantic fingerprint (row)
                bitset_set(row, words, sub_tid);

                // Convert the ID back to a pointer so we can check the dictionary
                StrPtr sub_ptr = interner.ptr_from_id(sub_tid);
                if (sub_ptr) {
                    auto sub_it = def_index.find(sub_ptr);
                    if (sub_it != def_index.end()) {
                        // Add the sub-token's deeper definitions to the candidate's row
                        for (StrPtr d : sub_it->second) {
                            TokenId did = interner.id_of(d);
                            if (did != TOKEN_ID_INVALID) bitset_set(row, words, did);
                        }
                    }
                }
            }
        }

        // Finally, count the bits for the Jaccard similarity math
        cand_counts[i] = bitset_count(row, words);
    }

    std::vector<double> scores(M, 0.0);

#if defined(_OPENMP) && defined(CHATIPC_ENABLE_OMP_TARGET)
    const bool use_target = (omp_get_num_devices() > 0) && (M >= 256);
#else
    const bool use_target = false;
#endif

    if (use_target){
        std::uint64_t *agg_ptr = agg_words.data();
        std::uint64_t *cand_ptr = cand_words.data();
        std::size_t *count_ptr = cand_counts.data();
        double *score_ptr = scores.data();
        const std::size_t cand_words_total = cand_words.size();

        #pragma omp target data map(to: agg_ptr[0:words], cand_ptr[0:cand_words_total], count_ptr[0:M]) map(from: score_ptr[0:M])
        {
            #pragma omp target teams distribute parallel for
            for (ptrdiff_t i = 0; i < static_cast<ptrdiff_t>(M); ++i){
                const std::uint64_t *row = cand_ptr + static_cast<std::size_t>(i) * words;

                std::size_t inter = 0;
                for (std::size_t w = 0; w < words; ++w){
                    inter += popcount_u64(agg_ptr[w] & row[w]);
                }

                const std::size_t uni = agg_count + count_ptr[(size_t)i] - inter;
                score_ptr[(size_t)i] = uni ? static_cast<double>(inter) / static_cast<double>(uni) : 0.0;
            }
        }
    } else {
        #ifdef _OPENMP
        #pragma omp parallel for schedule(static)
        #endif
        for (ptrdiff_t i = 0; i < static_cast<ptrdiff_t>(M); ++i){
            const std::uint64_t *row = cand_words.data() + static_cast<std::size_t>(i) * words;

            const std::size_t inter = bitset_intersection_count(agg_words.data(), row, words);
            const std::size_t uni = agg_count + cand_counts[(size_t)i] - inter;
            scores[(size_t)i] = uni ? static_cast<double>(inter) / static_cast<double>(uni) : 0.0;
        }
    }

    double best = -1e9;
    std::string best_tok;

    for (std::size_t i = 0; i < M; ++i){
        const std::string &tok = *cands[i];
        const std::string tok_key = normalize_dictionary_key(tok);
        const std::string count_key = tok_key.empty() ? tok : tok_key;

        auto rc_it = recent_counts.find(count_key);
        const int cnt = (rc_it == recent_counts.end()) ? 0 : rc_it->second;

        const double adjusted =
            scores[i] +
            english_rule_bonus(context_tok, tok) -
            repeat_penalty * static_cast<double>(cnt);

        if (adjusted > best || (adjusted == best && tok < best_tok)){
            best = adjusted;
            best_tok = tok;
        }
    }

    return best_tok;
}

static std::vector<std::string> construct_response(KnowledgeBase &kb,
                                                   const std::vector<std::string> &prompt_toks,
                                                   size_t response_maxlen,
                                                   double repeat_penalty,
                                                   int n_gram_size)
{
    std::vector<std::string> resp;
    if (prompt_toks.empty() || response_maxlen == 0) return resp;

    auto prompt_ptrs = intern_tokens(kb, prompt_toks);
    std::vector<StrPtr> resp_ptrs;
    std::unordered_map<std::string, int> recent_counts;

    // The context combines prompt and generated text
    std::vector<StrPtr> context = prompt_ptrs;

    auto would_create_2_cycle = [&](const std::string &cand) -> bool {
        if (resp.size() < 3) return false;
        return normalize_dictionary_key(cand) == normalize_dictionary_key(resp[resp.size() - 2]) &&
               normalize_dictionary_key(resp.back()) == normalize_dictionary_key(resp[resp.size() - 3]);
    };

    for (size_t step = 0; step < response_maxlen; ++step){
        NextSet candidates;
        bool found = false;
        std::string context_tok;

        // Dynamic Backoff N-gram lookup
        int current_n = std::min(static_cast<int>(context.size()), n_gram_size);
        while (current_n > 0) {
            std::vector<StrPtr> search_ctx(context.end() - current_n, context.end());
            auto opt = kb.lookup_ngram(search_ctx);
            if (opt && !opt->empty()) {
                candidates = *opt;
                found = true;
                context_tok = *search_ctx.back();
                break;
            }
            current_n--;
        }

        // Fallback to checking earlier prompt tokens if sequence stalled
        if (!found) {
            for (ssize_t p = static_cast<ssize_t>(prompt_ptrs.size()) - 1; p >= 0; --p){
                std::vector<StrPtr> search_ctx = { prompt_ptrs[(size_t)p] };
                auto opt = kb.lookup_ngram(search_ctx);
                if (opt && !opt->empty()){
                    candidates = *opt;
                    found = true;
                    context_tok = *prompt_ptrs[(size_t)p];
                    break;
                }
            }
        }

        if (!found || candidates.empty()) break;

        if (candidates.size() == 1){
            std::string only = *candidates[0];
            std::string only_key = normalize_dictionary_key(only);
            if (recent_counts[only_key.empty() ? only : only_key] > 0) break;

            resp.push_back(only);
            StrPtr ptr = kb.interner.intern(only);
            resp_ptrs.push_back(ptr);
            context.push_back(ptr);
            recent_counts[only_key.empty() ? only : only_key] += 1;
            std::cout << only << ' ' << std::flush;
            continue;
        }

        std::string chosen = best_candidate_by_similarity(
            candidates, prompt_ptrs, resp_ptrs, kb.def_index, kb.interner,
            recent_counts, repeat_penalty, context_tok
        );

        if (chosen.empty()) break;
        if (would_create_2_cycle(chosen)) break;

        resp.push_back(chosen);
        StrPtr chosen_ptr = kb.interner.intern(chosen);
        resp_ptrs.push_back(chosen_ptr);
        context.push_back(chosen_ptr);

        std::string chosen_key = normalize_dictionary_key(chosen);
        recent_counts[chosen_key.empty() ? chosen : chosen_key] += 1;

        std::cout << chosen << ' ' << std::flush;
    }

    return resp;
}

static void learn_from_file(KnowledgeBase &kb, const std::string &fname, int n_gram_size){
    std::ifstream ifs(fname);
    if (!ifs) return;
    std::vector<std::string> tokens;
    std::string tok;
    while (ifs >> tok) tokens.push_back(tok);
    learn_tokens_ngram(kb, tokens, n_gram_size);
}

static void learn_files_parallel(KnowledgeBase &kb, const std::vector<std::string> &files, int n_gram_size){
    #pragma omp parallel for schedule(dynamic)
    for (ptrdiff_t i=0;i<static_cast<ptrdiff_t>(files.size());++i){
        learn_from_file(kb, files[(size_t)i], n_gram_size);
    }
}

static constexpr std::uint64_t KB_MAGIC   = 0x434850434B535641ULL;
static constexpr std::uint64_t KB_VERSION = 2ULL;

static void write_u64(std::ostream &os, std::uint64_t v){
    os.write(reinterpret_cast<const char*>(&v), sizeof(v));
    if(!os) throw std::runtime_error("write_u64 failed");
}

static std::uint64_t read_u64(std::istream &is){
    std::uint64_t v = 0;
    is.read(reinterpret_cast<char*>(&v), sizeof(v));
    if(!is) throw std::runtime_error("read_u64 failed");
    return v;
}

static void write_string(std::ostream &os, const std::string &s){
    write_u64(os, static_cast<std::uint64_t>(s.size()));
    if (!s.empty()){
        os.write(s.data(), static_cast<std::streamsize>(s.size()));
        if(!os) throw std::runtime_error("write_string failed");
    }
}

static std::string read_string(std::istream &is){
    std::uint64_t n = read_u64(is);
    if (n > (1ULL << 30)) throw std::runtime_error("save file is corrupted: string too large");

    std::string s;
    s.resize(static_cast<size_t>(n));

    if (n != 0){
        is.read(&s[0], static_cast<std::streamsize>(n));
        if(!is) throw std::runtime_error("read_string failed");
    }
    return s;
}

static void save_kb_binary(const KnowledgeBase &kb, const std::string &fname){
    const std::string temp = fname + ".tmp";

    {
        std::ofstream ofs(temp.c_str(), std::ios::binary | std::ios::trunc);
        if (!ofs) throw std::runtime_error("cannot open temp save file");

        std::vector<std::string> pool;
        pool.reserve(kb.interner.pool.size());
        for (const auto &s : kb.interner.pool) pool.push_back(s);

        std::sort(pool.begin(), pool.end());

        std::unordered_map<std::string, std::uint64_t> id;
        id.reserve(pool.size());
        for (std::uint64_t i = 0; i < static_cast<std::uint64_t>(pool.size()); ++i){
            id.emplace(pool[(size_t)i], i);
        }

        write_u64(ofs, KB_MAGIC);
        write_u64(ofs, KB_VERSION);
        write_u64(ofs, static_cast<std::uint64_t>(kb.def_depth));

        write_u64(ofs, static_cast<std::uint64_t>(pool.size()));
        for (const auto &s : pool) write_string(ofs, s);

        write_u64(ofs, static_cast<std::uint64_t>(kb.next.size()));
        for (const auto &pr : kb.next){
            write_u64(ofs, static_cast<std::uint64_t>(pr.first.size()));
            for (StrPtr ctx_tok : pr.first){
                write_u64(ofs, id.at(*ctx_tok));
            }
            write_u64(ofs, static_cast<std::uint64_t>(pr.second.size()));
            for (StrPtr nxt : pr.second){
                write_u64(ofs, id.at(*nxt));
            }
        }

        write_u64(ofs, static_cast<std::uint64_t>(kb.def_index.size()));
        for (const auto &pr : kb.def_index){
            write_u64(ofs, id.at(*pr.first));
            write_u64(ofs, static_cast<std::uint64_t>(pr.second.size()));
            for (StrPtr tok : pr.second){
                write_u64(ofs, id.at(*tok));
            }
        }

        ofs.flush();
        if (!ofs) throw std::runtime_error("failed while writing temp save file");
    }

    std::remove(fname.c_str());
    if (std::rename(temp.c_str(), fname.c_str()) != 0){
        std::remove(temp.c_str());
        throw std::runtime_error("failed to commit save file");
    }
}

static void load_kb_binary(KnowledgeBase &kb, const std::string &fname, int cli_def_depth){
    std::ifstream ifs(fname, std::ios::binary);
    if (!ifs) throw std::runtime_error("cannot open load file");

    const std::uint64_t magic = read_u64(ifs);
    if (magic != KB_MAGIC) throw std::runtime_error("bad save file magic");

    const std::uint64_t version = read_u64(ifs);
    if (version != KB_VERSION) throw std::runtime_error("unsupported save file version");

    const std::uint64_t file_def_depth = read_u64(ifs);

    const std::uint64_t N = read_u64(ifs);
    if (N > (1ULL << 26)) throw std::runtime_error("save file is corrupted: pool too large");

    std::vector<std::string> strings;
    strings.reserve(static_cast<size_t>(N));

    for (std::uint64_t i = 0; i < N; ++i){
        strings.push_back(read_string(ifs));
    }

    kb.interner.pool.clear();
    kb.interner.pool.reserve(static_cast<size_t>(N));

    std::vector<StrPtr> ptrs;
    ptrs.reserve(static_cast<size_t>(N));
    for (const auto &s : strings){
        ptrs.push_back(kb.interner.intern(s));
    }

    // Rebuild next
    const std::uint64_t E = read_u64(ifs);
    if (E > (1ULL << 26)) throw std::runtime_error("save file is corrupted: graph too large");

    {
        std::lock_guard<std::mutex> lk(kb.m);
        kb.next.clear();
        kb.next.reserve(static_cast<size_t>(E));
    }

    for (std::uint64_t i = 0; i < E; ++i){
        const std::uint64_t ctx_size = read_u64(ifs);
        if (ctx_size > (1ULL << 16)) throw std::runtime_error("save file is corrupted: n-gram context too large");

        std::vector<StrPtr> ctx;
        ctx.reserve(static_cast<size_t>(ctx_size));
        for(std::uint64_t c = 0; c < ctx_size; ++c) {
            const std::uint64_t c_idx = read_u64(ifs);
            if (c_idx >= ptrs.size()) throw std::runtime_error("save file is corrupted: bad context key");
            ctx.push_back(ptrs[(size_t)c_idx]);
        }

        const std::uint64_t M = read_u64(ifs);
        if (M > (1ULL << 26)) throw std::runtime_error("save file is corrupted: graph degree too large");

        NextSet vec;
        vec.reserve(static_cast<size_t>(M));

        for (std::uint64_t j = 0; j < M; ++j){
            const std::uint64_t v_idx = read_u64(ifs);
            if (v_idx >= ptrs.size()) throw std::runtime_error("save file is corrupted: bad graph value");
            vec.push_back(ptrs[(size_t)v_idx]);
        }

        {
            std::lock_guard<std::mutex> lk(kb.m);
            kb.next.emplace(std::move(ctx), std::move(vec));
        }
    }

    // Rebuild def_index from file
    const std::uint64_t K = read_u64(ifs);
    if (K > (1ULL << 26)) throw std::runtime_error("save file is corrupted: def_index too large");

    {
        std::lock_guard<std::mutex> lk(kb.def_m);
        kb.def_index.clear();
        kb.def_index.reserve(static_cast<size_t>(K));
        kb.def_depth = static_cast<int>(file_def_depth);
    }

    for (std::uint64_t i = 0; i < K; ++i){
        const std::uint64_t key_idx = read_u64(ifs);
        const std::uint64_t M = read_u64(ifs);

        if (key_idx >= ptrs.size()) throw std::runtime_error("save file is corrupted: bad def key");
        if (M > (1ULL << 26)) throw std::runtime_error("save file is corrupted: def list too large");

        std::vector<StrPtr> toks;
        toks.reserve(static_cast<size_t>(M));

        for (std::uint64_t j = 0; j < M; ++j){
            const std::uint64_t v_idx = read_u64(ifs);
            if (v_idx >= ptrs.size()) throw std::runtime_error("save file is corrupted: bad def value");
            toks.push_back(ptrs[(size_t)v_idx]);
        }

        {
            std::lock_guard<std::mutex> lk(kb.def_m);
            kb.def_index.emplace(ptrs[(size_t)key_idx], std::move(toks));
        }
    }

    if (cli_def_depth != static_cast<int>(file_def_depth)){
        kb.set_def_depth(cli_def_depth);

        std::vector<StrPtr> targets;
        targets.reserve(ptrs.size() + kb.next.size() * 2);

        std::unordered_set<StrPtr, PtrHash, PtrEq> seen;
        seen.reserve(ptrs.size() + kb.next.size() * 2);

        for (StrPtr p : ptrs){
            if (seen.insert(p).second) targets.push_back(p);
        }

        {
            std::lock_guard<std::mutex> lk(kb.m);
            for (const auto &pr : kb.next){
                for (StrPtr ctx_element : pr.first) {
                    if (seen.insert(ctx_element).second) {
                        targets.push_back(ctx_element);
                    }
                }

                for (StrPtr v : pr.second){
                    if (seen.insert(v).second) {
                        targets.push_back(v);
                    }
                }
            }
        }

        #pragma omp parallel for schedule(dynamic)
        for (ptrdiff_t i = 0; i < static_cast<ptrdiff_t>(targets.size()); ++i){
            kb.ensure_def_for_interned(targets[(size_t)i]);
        }
    }
}

static void print_kb_info(const std::string &fname) {
    std::ifstream ifs(fname, std::ios::binary);
    if (!ifs) {
        std::cerr << "Error: cannot open knowledge-base file " << fname << "\n";
        return;
    }

    std::cout << "Reckoning knowledge-base "<< fname << " info.\n";

    try {
        const std::uint64_t magic = read_u64(ifs);
        if (magic != KB_MAGIC) throw std::runtime_error("bad save file magic");

        const std::uint64_t version = read_u64(ifs);
        if (version != KB_VERSION) throw std::runtime_error("unsupported save file version");

        const std::uint64_t file_def_depth = read_u64(ifs);

        const std::uint64_t N = read_u64(ifs);
        if (N > (1ULL << 26)) throw std::runtime_error("save file is corrupted: pool too large");

        // Fast-skip string pool to reach the n-gram graph data
        for (std::uint64_t i = 0; i < N; ++i){
            std::uint64_t len = read_u64(ifs);
            if (len > (1ULL << 30)) throw std::runtime_error("save file is corrupted: string too large");
            if (len > 0) ifs.seekg(static_cast<std::streamoff>(len), std::ios_base::cur);
        }

        const std::uint64_t E = read_u64(ifs);
        if (E > (1ULL << 26)) throw std::runtime_error("save file is corrupted: graph too large");

        std::uint64_t max_ngram = 0;
        for (std::uint64_t i = 0; i < E; ++i){
            const std::uint64_t ctx_size = read_u64(ifs);
            if (ctx_size > (1ULL << 16)) throw std::runtime_error("save file is corrupted: n-gram context too large");

            // The maximum context size stored represents the n-gram size
            if (ctx_size > max_ngram) max_ngram = ctx_size;

            // Skip the context keys
            ifs.seekg(static_cast<std::streamoff>(ctx_size * sizeof(std::uint64_t)), std::ios_base::cur);

            const std::uint64_t M = read_u64(ifs);
            if (M > (1ULL << 26)) throw std::runtime_error("save file is corrupted: graph degree too large");

            // Skip the next-nodes values
            ifs.seekg(static_cast<std::streamoff>(M * sizeof(std::uint64_t)), std::ios_base::cur);
        }

        std::cout << fname << " dictionary depth: " << file_def_depth << "\n";
        std::cout << fname << " n-gram max size: " << max_ngram << "\n";

    } catch (const std::exception &e) {
        std::cerr << "Error reading knowledge-base info: " << e.what() << "\n";
    }
}

static void print_commands(const char *p){
    std::cout << p << " command-line interface options:\n";
    std::cout << "  --response-max-length N     Set maximum number of tokens in a response.\n";
    std::cout << "  --save-kb FILE              Save the knowledge-base to a binary file.\n";
    std::cout << "  --load-kb FILE              Load a previously saved knowledge-base from a binary file.\n";
    std::cout << "  --info-kb FILE              Show the dictionary-depth and n-gram max size of a saved knowledge-base.\n";
    std::cout << "  --dictionary-depth D        Set depth of dictionary-definition expansion used during learning.\n";
    std::cout << "  --n-gram-max-size N         Set maximum size of the n-gram where N is the size.\n";
    std::cout << "  --learn f1 f2 ...           Learn from one or more text files to update the knowledge-base.\n";
    std::cout << "  --repeat-penalty P          Set penalty for repeated tokens when constructing response (higher values reduce repetition).\n";
    std::cout << "  --activate-agi              Activate the Artificial General Intelligence (AGI) features.\n";
    std::cout << "  --help                      Show " << p << " command-line interface options.\n";
}

int main(int argc, char **argv){
    size_t response_maxlen = 500;
    std::string savefile;
    std::string load_txt;
    std::string load_kb;
    int def_depth = 3;
    int n_gram_size = 3;
    double repeat_penalty = 0.7; // default λ
    std::vector<std::string> learn_files;

    for (int i=1;i<argc;++i){
        std::string a = argv[i];
        if (a=="--help"){ print_commands(argv[0]); return 0; }
        if (a=="--response-max-length" && i+1<argc){ response_maxlen = std::stoul(argv[++i]); continue; }
        if (a=="--save-kb" && i+1<argc){ savefile = argv[++i]; continue; }
        if (a=="--load-kb" && i+1<argc){ load_kb = argv[++i]; continue; }
        if (a=="--info-kb" && i+1<argc){ print_kb_info(argv[++i]); return 0; }
        if (a=="--dictionary-depth" && i+1<argc){ def_depth = std::stoi(argv[++i]); continue; }
        if (a=="--n-gram-max-size" && i+1<argc){ n_gram_size = std::max(1, std::stoi(argv[++i])); continue; }
        if (a=="--repeat-penalty" && i+1<argc){ repeat_penalty = std::stod(argv[++i]); continue; }
        if (a=="--activate-agi"){
                std::cerr << "AGI subscription fee is $1,000 CAD per month. To pay the fee, contact Caleb Nwokocha via email cnwokocha@proton.me\n";
                return 0;
        }
        if (a=="--learn"){
            while(i+1<argc && argv[i+1][0] != '-') learn_files.push_back(argv[++i]);
            continue;
        }
        learn_files.push_back(a);
    }

    KnowledgeBase kb;

    global_dictionary_entries = parse_dictionary_json();
    build_def_tokens_cache();
    kb.set_def_depth(def_depth);

    if (!load_kb.empty()){
        try { std::cerr << "Loading knowledge base: " << load_kb << "\n";
        load_kb_binary(kb, load_kb, def_depth); std::cerr << "Loaded knowledge base: " << load_kb << "\n"; }
        catch (const std::exception &e){ std::cerr << "Error: " << e.what() << "\n"; }
    }

    if (!learn_files.empty()){
        std::cerr << "Learning from file/s (" << learn_files.size() << ") using threads=" << omp_get_max_threads() << "\n";
        learn_files_parallel(kb, learn_files, n_gram_size);
    }

    std::string line;
    std::cout << "Ready. Enter prompts.\n";
    while (std::cout << "> " , std::getline(std::cin, line)){
        if (line.empty()){ std::cout << "\n"; continue; }
        auto prompt_toks = tokenize_whitespace(line);
        learn_tokens_ngram(kb, prompt_toks, n_gram_size);
        auto resp = construct_response(kb, prompt_toks, response_maxlen, repeat_penalty, n_gram_size);
        std::cout << "\n";
        if (!resp.empty()){learn_tokens_ngram(kb, resp, n_gram_size);}
        if (!savefile.empty()){
            try { std::cerr << "Saving knowledge base: " << savefile << "\n";
            save_kb_binary(kb, savefile); std::cerr << "Saved knowledge base: " << savefile << "\n"; }
            catch (const std::exception &e){ std::cerr << "Error: " << e.what() << "\n"; }
        }
    }

    return 0;
}