termgenerator_internal.cc

Go to the documentation of this file.

/* Copyright (C) 2007-2023 Olly Betts
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 2 of the License, or
 * (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301 USA
 */

#include <config.h>

#include "termgenerator_internal.h"

#include "api/omenquireinternal.h"
#include "api/queryinternal.h"

#include <xapian/document.h>
#include <xapian/queryparser.h>
#include <xapian/stem.h>
#include <xapian/unicode.h>

#include "stringutils.h"

#include <algorithm>
#include <cmath>
#include <deque>
#include <limits>
#include <list>
#include <string>
#include <unordered_map>
#include <vector>

#include "word-breaker.h"

using namespace std;

namespace Xapian {

static inline bool
U_isupper(unsigned ch)
{
    return ch < 128 && C_isupper(static_cast<unsigned char>(ch));
}

static inline unsigned
check_wordchar(unsigned ch)
{
    if (Unicode::is_wordchar(ch)) return Unicode::tolower(ch);
    return 0;
}

static inline bool
should_stem(const std::string & term)
{
    const unsigned int SHOULD_STEM_MASK =
        (1 << Unicode::LOWERCASE_LETTER) |
        (1 << Unicode::TITLECASE_LETTER) |
        (1 << Unicode::MODIFIER_LETTER) |
        (1 << Unicode::OTHER_LETTER);
    Utf8Iterator u(term);
    return ((SHOULD_STEM_MASK >> Unicode::get_category(*u)) & 1);
}

static const unsigned UNICODE_IGNORE = numeric_limits<unsigned>::max();

static inline unsigned
check_infix(unsigned ch)
{
    if (ch == '\'' || ch == '&' || ch == 0xb7 || ch == 0x5f4 || ch == 0x2027) {
        // Unicode includes all these except '&' in its word boundary rules,
        // as well as 0x2019 (which we handle below) and ':' (for Swedish
        // apparently, but we ignore this for now as it's problematic in
        // real world cases).
        return ch;
    }
    // 0x2019 is Unicode apostrophe and single closing quote.
    // 0x201b is Unicode single opening quote with the tail rising.
    if (ch == 0x2019 || ch == 0x201b) return '\'';
    if (ch >= 0x200b && (ch <= 0x200d || ch == 0x2060 || ch == 0xfeff))
        return UNICODE_IGNORE;
    return 0;
}

static inline unsigned
check_infix_digit(unsigned ch)
{
    // This list of characters comes from Unicode's word identifying algorithm.
    switch (ch) {
        case ',':
        case '.':
        case ';':
        case 0x037e: // GREEK QUESTION MARK
        case 0x0589: // ARMENIAN FULL STOP
        case 0x060D: // ARABIC DATE SEPARATOR
        case 0x07F8: // NKO COMMA
        case 0x2044: // FRACTION SLASH
        case 0xFE10: // PRESENTATION FORM FOR VERTICAL COMMA
        case 0xFE13: // PRESENTATION FORM FOR VERTICAL COLON
        case 0xFE14: // PRESENTATION FORM FOR VERTICAL SEMICOLON
            return ch;
    }
    if (ch >= 0x200b && (ch <= 0x200d || ch == 0x2060 || ch == 0xfeff))
        return UNICODE_IGNORE;
    return 0;
}

static inline bool
is_digit(unsigned ch) {
    return (Unicode::get_category(ch) == Unicode::DECIMAL_DIGIT_NUMBER);
}

static inline unsigned
check_suffix(unsigned ch)
{
    if (ch == '+' || ch == '#') return ch;
    // FIXME: what about '-'?
    return 0;
}

template<typename ACTION>
static void
parse_terms(Utf8Iterator itor, bool try_word_break, bool with_positions,
            ACTION action)
{
    while (true) {
        // Advance to the start of the next term.
        unsigned ch;
        while (true) {
            if (itor == Utf8Iterator()) return;
            ch = check_wordchar(*itor);
            if (ch) break;
            ++itor;
        }

        string term;
        // Look for initials separated by '.' (e.g. P.T.O., U.N.C.L.E).
        // Don't worry if there's a trailing '.' or not.
        if (U_isupper(*itor)) {
            const Utf8Iterator end;
            Utf8Iterator p = itor;
            do {
                Unicode::append_utf8(term, Unicode::tolower(*p++));
            } while (p != end && *p == '.' && ++p != end && U_isupper(*p));
            // One letter does not make an acronym!  If we handled a single
            // uppercase letter here, we wouldn't catch M&S below.
            if (term.size() > 1) {
                // Check there's not a (lower case) letter or digit
                // immediately after it.
                if (p == end || !Unicode::is_wordchar(*p)) {
                    itor = p;
                    goto endofterm;
                }
            }
            term.resize(0);
        }

        while (true) {
            if (try_word_break &&
                is_unbroken_script(*itor) &&
                Unicode::is_wordchar(*itor)) {
                NgramIterator tk(itor);
                while (tk != NgramIterator()) {
                    const string& ngram = *tk;
                    if (!action(ngram, with_positions && tk.unigram(),
                                tk.get_utf8iterator()))
                        return;
                    ++tk;
                }
                // Update itor to end of CJK text span.
                itor = tk.get_utf8iterator();
                while (true) {
                    if (itor == Utf8Iterator()) return;
                    ch = check_wordchar(*itor);
                    if (ch) break;
                    ++itor;
                }
                continue;
            }
            unsigned prevch;
            do {
                Unicode::append_utf8(term, ch);
                prevch = ch;
                if (++itor == Utf8Iterator() ||
                    (try_word_break && is_unbroken_script(*itor)))
                    goto endofterm;
                ch = check_wordchar(*itor);
            } while (ch);

            Utf8Iterator next(itor);
            ++next;
            if (next == Utf8Iterator()) break;
            unsigned nextch = check_wordchar(*next);
            if (!nextch) break;
            unsigned infix_ch = *itor;
            if (is_digit(prevch) && is_digit(*next)) {
                infix_ch = check_infix_digit(infix_ch);
            } else {
                // Handle things like '&' in AT&T, apostrophes, etc.
                infix_ch = check_infix(infix_ch);
            }
            if (!infix_ch) break;
            if (infix_ch != UNICODE_IGNORE)
                Unicode::append_utf8(term, infix_ch);
            ch = nextch;
            itor = next;
        }

        {
            size_t len = term.size();
            unsigned count = 0;
            while ((ch = check_suffix(*itor))) {
                if (++count > 3) {
                    term.resize(len);
                    break;
                }
                Unicode::append_utf8(term, ch);
                if (++itor == Utf8Iterator()) goto endofterm;
            }
            // Don't index fish+chips as fish+ chips.
            if (Unicode::is_wordchar(*itor))
                term.resize(len);
        }

endofterm:
        if (!action(term, with_positions, itor))
            return;
    }
}

void
TermGenerator::Internal::index_text(Utf8Iterator itor, termcount wdf_inc,
                                    const string & prefix, bool with_positions)
{
    bool try_word_break = (flags & FLAG_NGRAMS) || is_ngram_enabled();

    stop_strategy current_stop_mode;
    if (!stopper.get()) {
        current_stop_mode = TermGenerator::STOP_NONE;
    } else {
        current_stop_mode = stop_mode;
    }

    parse_terms(itor, try_word_break, with_positions,
        [=
#if __cplusplus >= 201907L
// C++20 no longer supports implicit `this` in lambdas but older C++ versions
// don't allow `this` here.
        , this
#endif
        ](const string & term, bool positional, const Utf8Iterator &) {
            if (term.size() > max_word_length) return true;

            if (current_stop_mode == TermGenerator::STOP_ALL &&
                (*stopper)(term)) {
                return true;
            }

            if (strategy == TermGenerator::STEM_SOME ||
                strategy == TermGenerator::STEM_NONE ||
                strategy == TermGenerator::STEM_SOME_FULL_POS) {
                if (positional) {
                    doc.add_posting(prefix + term, ++cur_pos, wdf_inc);
                } else {
                    doc.add_term(prefix + term, wdf_inc);
                }
            }

            // MSVC seems to need "this->" on member variables in this
            // situation.
            if ((this->flags & FLAG_SPELLING) && prefix.empty())
                db.add_spelling(term);

            if (strategy == TermGenerator::STEM_NONE || stemmer.is_none())
                return true;

            if (strategy == TermGenerator::STEM_SOME ||
                strategy == TermGenerator::STEM_SOME_FULL_POS) {
                if (current_stop_mode == TermGenerator::STOP_STEMMED &&
                    (*stopper)(term))
                    return true;

                // Note, this uses the lowercased term, but that's OK as we
                // only want to avoid stemming terms starting with a digit.
                if (!should_stem(term)) return true;
            }

            // Add stemmed form without positional information.
            const string& stem = stemmer(term);
            if (rare(stem.empty())) return true;
            string stemmed_term;
            if (strategy != TermGenerator::STEM_ALL) {
                stemmed_term += "Z";
            }
            stemmed_term += prefix;
            stemmed_term += stem;
            if (strategy != TermGenerator::STEM_SOME && with_positions) {
                if (strategy != TermGenerator::STEM_SOME_FULL_POS) ++cur_pos;
                doc.add_posting(stemmed_term, cur_pos, wdf_inc);
            } else {
                doc.add_term(stemmed_term, wdf_inc);
            }
            return true;
        });
}

struct Sniplet {
    double* relevance;

    size_t term_end;

    size_t highlight;

    Sniplet(double* r, size_t t, size_t h)
        : relevance(r), term_end(t), highlight(h) { }
};

class SnipPipe {
    deque<Sniplet> pipe;
    deque<Sniplet> best_pipe;

    // Requested length for snippet.
    size_t length;

    // Position in text of start of current pipe contents.
    size_t begin = 0;

    // Rolling sum of the current pipe contents.
    double sum = 0;

    size_t phrase_len = 0;

  public:
    size_t best_begin = 0;

    size_t best_end = 0;

    double best_sum = 0;

    // Add one to length to allow for inter-word space.
    // FIXME: We ought to correctly allow for multiple spaces.
    explicit SnipPipe(size_t length_) : length(length_ + 1) { }

    bool pump(double* r, size_t t, size_t h, unsigned flags);

    void done();

    bool drain(const string & input,
               const string & hi_start,
               const string & hi_end,
               const string & omit,
               string & output);
};

#define DECAY 2.0

inline bool
SnipPipe::pump(double* r, size_t t, size_t h, unsigned flags)
{
    if (h > 1) {
        if (pipe.size() >= h - 1) {
            // The final term of a phrase is entering the window.  Peg the
            // phrase's relevance onto the first term of the phrase, so it'll
            // be removed from `sum` when the phrase starts to leave the
            // window.
            auto & phrase_start = pipe[pipe.size() - (h - 1)];
            if (phrase_start.relevance) {
                *phrase_start.relevance *= DECAY;
                sum -= *phrase_start.relevance;
            }
            sum += *r;
            phrase_start.relevance = r;
            phrase_start.highlight = h;
            *r /= DECAY;
        }
        r = NULL;
        h = 0;
    }
    pipe.emplace_back(r, t, h);
    if (r) {
        sum += *r;
        *r /= DECAY;
    }

    // If necessary, discard words from the start of the pipe until it has the
    // desired length.
    // FIXME: Also shrink the window past words with relevance < 0?
    while (t - begin > length /* || pipe.front().relevance < 0.0 */) {
        const Sniplet& word = pipe.front();
        if (word.relevance) {
            *word.relevance *= DECAY;
            sum -= *word.relevance;
        }
        begin = word.term_end;
        if (best_end >= begin)
            best_pipe.push_back(word);
        pipe.pop_front();
        // E.g. can happen if the current term is longer than the requested
        // length!
        if (rare(pipe.empty())) break;
    }

    // Using > here doesn't work well, as we don't extend a snippet over terms
    // with 0 weight.
    if (sum >= best_sum) {
        // Discard any part of `best_pipe` which is before `begin`.
        if (begin >= best_end) {
            best_pipe.clear();
        } else {
            while (!best_pipe.empty() &&
                   best_pipe.front().term_end <= begin) {
                best_pipe.pop_front();
            }
        }
        best_sum = sum;
        best_begin = begin;
        best_end = t;
    } else if ((flags & Xapian::MSet::SNIPPET_EXHAUSTIVE) == 0) {
        if (best_sum > 0 && best_end < begin) {
            // We found something, and we aren't still looking near it.
            // FIXME: Benchmark this and adjust if necessary.
            return false;
        }
    }
    return true;
}

inline void
SnipPipe::done()
{
    // Discard any part of `pipe` which is after `best_end`.
    if (begin >= best_end) {
        pipe.clear();
    } else {
        // We should never empty the pipe (as that case should be handled
        // above).
        while (rare(!pipe.empty()) &&
               pipe.back().term_end > best_end) {
            pipe.pop_back();
        }
    }
}

// Check if a non-word character is should be included at the start of the
// snippet.  We want to include certain leading non-word characters, but not
// others.
static inline bool
snippet_check_leading_nonwordchar(unsigned ch) {
    if (Unicode::is_currency(ch) ||
        Unicode::get_category(ch) == Unicode::OPEN_PUNCTUATION ||
        Unicode::get_category(ch) == Unicode::INITIAL_QUOTE_PUNCTUATION) {
        return true;
    }
    switch (ch) {
        case '"':
        case '#':
        case '%':
        case '&':
        case '\'':
        case '+':
        case '-':
        case '/':
        case '<':
        case '@':
        case '\\':
        case '`':
        case '~':
        case 0x00A1: // INVERTED EXCLAMATION MARK
        case 0x00A7: // SECTION SIGN
        case 0x00BF: // INVERTED QUESTION MARK
            return true;
    }
    return false;
}

// Check if a non-word character is should be included at the end of the
// snippet.  We want to include certain trailing non-word characters, but not
// others.
static inline bool
snippet_check_trailing_nonwordchar(unsigned ch) {
    if (Unicode::is_currency(ch) ||
        Unicode::get_category(ch) == Unicode::CLOSE_PUNCTUATION ||
        Unicode::get_category(ch) == Unicode::FINAL_QUOTE_PUNCTUATION) {
        return true;
    }
    switch (ch) {
        case '"':
        case '%':
        case '\'':
        case '+':
        case '-':
        case '/':
        case '>':
        case '@':
        case '\\':
        case '`':
        case '~':
            return true;
    }
    return false;
}

static inline void
append_escaping_xml(const char* p, const char* end, string& output)
{
    while (p != end) {
        char ch = *p++;
        switch (ch) {
            case '&':
                output += "&amp;";
                break;
            case '<':
                output += "&lt;";
                break;
            case '>':
                output += "&gt;";
                break;
            default:
                output += ch;
        }
    }
}

inline bool
SnipPipe::drain(const string & input,
                const string & hi_start,
                const string & hi_end,
                const string & omit,
                string & output)
{
    if (best_pipe.empty() && !pipe.empty()) {
        swap(best_pipe, pipe);
    }

    if (best_pipe.empty()) {
        size_t tail_len = input.size() - best_end;
        if (tail_len == 0) return false;

        // See if this is the end of a sentence.
        // FIXME: This is quite simplistic - look at the Unicode rules:
        // https://unicode.org/reports/tr29/#Sentence_Boundaries
        bool sentence_end = false;
        Utf8Iterator i(input.data() + best_end, tail_len);
        while (i != Utf8Iterator()) {
            unsigned ch = *i;
            if (sentence_end && Unicode::is_whitespace(ch)) break;

            // Allow "...", "!!", "!?!", etc...
            sentence_end = (ch == '.' || ch == '?' || ch == '!');

            if (Unicode::is_wordchar(ch)) break;
            ++i;
        }

        if (sentence_end) {
            // Include end of sentence punctuation.
            append_escaping_xml(input.data() + best_end, i.raw(), output);
            return false;
        }

        // Include trailing punctuation which includes meaning or context.
        i.assign(input.data() + best_end, tail_len);
        int trailing_punc = 0;
        while (i != Utf8Iterator() && snippet_check_trailing_nonwordchar(*i)) {
            // But limit how much trailing punctuation we include.
            if (++trailing_punc > 4) {
                trailing_punc = 0;
                break;
            }
            ++i;
        }
        if (trailing_punc) {
            append_escaping_xml(input.data() + best_end, i.raw(), output);
            if (i == Utf8Iterator()) return false;
        }

        // Append "..." or equivalent as this doesn't seem to be the start
        // of a sentence.
        output += omit;

        return false;
    }

    const Sniplet & word = best_pipe.front();

    if (output.empty()) {
        // Start of the snippet.
        enum { NO, PUNC, YES } sentence_boundary = (best_begin == 0) ? YES : NO;

        Utf8Iterator i(input.data() + best_begin, word.term_end - best_begin);
        while (i != Utf8Iterator()) {
            unsigned ch = *i;
            switch (sentence_boundary) {
                case NO:
                    if (ch == '.' || ch == '?' || ch == '!') {
                        sentence_boundary = PUNC;
                    }
                    break;
                case PUNC:
                    if (Unicode::is_whitespace(ch)) {
                        sentence_boundary = YES;
                    } else if (ch == '.' || ch == '?' || ch == '!') {
                        // Allow "...", "!!", "!?!", etc...
                    } else {
                        sentence_boundary = NO;
                    }
                    break;
                case YES:
                    break;
            }

            // Start the snippet at the start of the first word, but include
            // certain punctuation too.
            if (Unicode::is_wordchar(ch)) {
                // But limit how much leading punctuation we include.
                size_t word_begin = i.raw() - input.data();
                if (word_begin - best_begin > 4) {
                    best_begin = word_begin;
                }
                break;
            }
            ++i;
            if (!snippet_check_leading_nonwordchar(ch)) {
                best_begin = i.raw() - input.data();
            }
        }

        // Add "..." or equivalent if this doesn't seem to be the start of a
        // sentence.
        if (sentence_boundary != YES) {
            output += omit;
        }
    }

    if (word.highlight) {
        // Don't include inter-word characters in the highlight.
        Utf8Iterator i(input.data() + best_begin, input.size() - best_begin);
        while (i != Utf8Iterator()) {
            unsigned ch = *i;
            if (Unicode::is_wordchar(ch)) {
                append_escaping_xml(input.data() + best_begin, i.raw(), output);
                best_begin = i.raw() - input.data();
                break;
            }
            ++i;
        }
    }

    if (!phrase_len) {
        phrase_len = word.highlight;
        if (phrase_len) output += hi_start;
    }

    const char* p = input.data();
    append_escaping_xml(p + best_begin, p + word.term_end, output);
    best_begin = word.term_end;

    if (phrase_len && --phrase_len == 0) output += hi_end;

    best_pipe.pop_front();
    return true;
}

static void
check_query(const Xapian::Query & query,
            list<vector<string>> & exact_phrases,
            unordered_map<string, double> & loose_terms,
            list<string> & wildcards,
            size_t & longest_phrase)
{
    // FIXME: OP_NEAR, non-tight OP_PHRASE, OP_PHRASE with non-term subqueries
    size_t n_subqs = query.get_num_subqueries();
    Xapian::Query::op op = query.get_type();
    if (op == query.LEAF_TERM) {
        const Xapian::Internal::QueryTerm & qt =
            *static_cast<const Xapian::Internal::QueryTerm *>(query.internal.get());
        loose_terms.insert(make_pair(qt.get_term(), 0));
    } else if (op == query.OP_WILDCARD) {
        const Xapian::Internal::QueryWildcard & qw =
            *static_cast<const Xapian::Internal::QueryWildcard *>(query.internal.get());
        wildcards.push_back(qw.get_pattern());
    } else if (op == query.OP_PHRASE) {
        const Xapian::Internal::QueryPhrase & phrase =
            *static_cast<const Xapian::Internal::QueryPhrase *>(query.internal.get());
        if (phrase.get_window() == n_subqs) {
            // Tight phrase.
            for (size_t i = 0; i != n_subqs; ++i) {
                if (query.get_subquery(i).get_type() != query.LEAF_TERM)
                    goto non_term_subquery;
            }

            // Tight phrase of terms.
            exact_phrases.push_back(vector<string>());
            vector<string> & terms = exact_phrases.back();
            terms.reserve(n_subqs);
            for (size_t i = 0; i != n_subqs; ++i) {
                Xapian::Query q = query.get_subquery(i);
                const Xapian::Internal::QueryTerm & qt =
                    *static_cast<const Xapian::Internal::QueryTerm *>(q.internal.get());
                terms.push_back(qt.get_term());
            }
            if (n_subqs > longest_phrase) longest_phrase = n_subqs;
            return;
        }
    }
non_term_subquery:
    for (size_t i = 0; i != n_subqs; ++i)
        check_query(query.get_subquery(i), exact_phrases, loose_terms,
                    wildcards, longest_phrase);
}

static double*
check_term(unordered_map<string, double> & loose_terms,
           const Xapian::Weight::Internal * stats,
           const string & term,
           double max_tw)
{
    auto it = loose_terms.find(term);
    if (it == loose_terms.end()) return NULL;

    if (it->second == 0.0) {
        double relevance;
        if (!stats->get_termweight(term, relevance)) {
            // FIXME: Assert?
            loose_terms.erase(it);
            return NULL;
        }

        it->second = relevance + max_tw;
    }
    return &it->second;
}

string
MSet::Internal::snippet(const string & text,
                        size_t length,
                        const Xapian::Stem & stemmer,
                        unsigned flags,
                        const string & hi_start,
                        const string & hi_end,
                        const string & omit) const
{
    if (hi_start.empty() && hi_end.empty() && text.size() <= length) {
        // Too easy!
        return text;
    }

    bool try_word_break = (flags & MSet::SNIPPET_CJK_NGRAM);
    if (!try_word_break) {
        try_word_break = true;
    }

    size_t term_start = 0;
    double min_tw = 0, max_tw = 0;
    if (stats) stats->get_max_termweight(min_tw, max_tw);
    if (max_tw == 0.0) {
        max_tw = 1.0;
    } else {
        // Scale up by (1 + 1/64) so that highlighting works better for terms
        // with termweight 0 (which happens for terms not in the database, and
        // also with some weighting schemes for terms which occur in almost all
        // documents.
        max_tw *= 1.015625;
    }

    Xapian::Query query;
    if (enquire.get()) {
        query = enquire->query;
    }
    SnipPipe snip(length);

    list<vector<string>> exact_phrases;
    unordered_map<string, double> loose_terms;
    list<string> wildcards;
    size_t longest_phrase = 0;
    check_query(query, exact_phrases, loose_terms,
                wildcards, longest_phrase);

    vector<double> exact_phrases_relevance;
    exact_phrases_relevance.reserve(exact_phrases.size());
    for (auto&& terms : exact_phrases) {
        // FIXME: What relevance to use?
        exact_phrases_relevance.push_back(max_tw * terms.size());
    }

    vector<double> wildcards_relevance;
    wildcards_relevance.reserve(exact_phrases.size());
    for (auto&& pattern : wildcards) {
        // FIXME: What relevance to use?
        (void)pattern;
        wildcards_relevance.push_back(max_tw + min_tw);
    }

    // Background relevance is the same for a given MSet, so cache it
    // between calls to MSet::snippet() on the same object.
    unordered_map<string, double>& background = snippet_bg_relevance;

    vector<string> phrase;
    if (longest_phrase) phrase.resize(longest_phrase - 1);
    size_t phrase_next = 0;
    bool matchfound = false;
    parse_terms(Utf8Iterator(text), try_word_break, true,
        [&](const string & term, bool positional, const Utf8Iterator & it) {
            // FIXME: Don't hardcode this here.
            const size_t max_word_length = 64;

            if (!positional) return true;
            if (term.size() > max_word_length) return true;

            // We get segments with any "inter-word" characters in front of
            // each word, e.g.:
            // [The][ cat][ sat][ on][ the][ mat]
            size_t term_end = text.size() - it.left();

            double* relevance = 0;
            size_t highlight = 0;
            if (stats) {
                size_t i = 0;
                for (auto&& terms : exact_phrases) {
                    if (term == terms.back()) {
                        size_t n = terms.size() - 1;
                        bool match = true;
                        while (n--) {
                            if (terms[n] != phrase[(n + phrase_next) % (longest_phrase - 1)]) {
                                match = false;
                                break;
                            }
                        }
                        if (match) {
                            // FIXME: Sort phrases, highest score first!
                            relevance = &exact_phrases_relevance[i];
                            highlight = terms.size();
                            goto relevance_done;
                        }
                    }
                    ++i;
                }

                relevance = check_term(loose_terms, stats, term, max_tw);
                if (relevance) {
                    // Matched unstemmed term.
                    highlight = 1;
                    goto relevance_done;
                }

                string stem = "Z";
                stem += stemmer(term);
                relevance = check_term(loose_terms, stats, stem, max_tw);
                if (relevance) {
                    // Matched stemmed term.
                    highlight = 1;
                    goto relevance_done;
                }

                // Check wildcards.
                // FIXME: Sort wildcards, shortest pattern first or something?
                i = 0;
                for (auto&& pattern : wildcards) {
                    if (startswith(term, pattern)) {
                        relevance = &wildcards_relevance[i];
                        highlight = 1;
                        goto relevance_done;
                    }
                    ++i;
                }

                if (flags & Xapian::MSet::SNIPPET_BACKGROUND_MODEL) {
                    // Background document model.
                    auto bgit = background.find(term);
                    if (bgit == background.end()) bgit = background.find(stem);
                    if (bgit == background.end()) {
                        Xapian::doccount tf = enquire->db.get_termfreq(term);
                        if (!tf) {
                            tf = enquire->db.get_termfreq(stem);
                        } else {
                            stem = term;
                        }
                        double r = 0.0;
                        if (tf) {
                            // Add one to avoid log(0) when a term indexes all
                            // documents.
                            Xapian::doccount num_docs = stats->collection_size + 1;
                            r = max_tw * log((num_docs - tf) / double(tf));
                            r /= (length + 1) * log(double(num_docs));
#if 0
                            if (r <= 0) {
                                Utf8Iterator i(text.data() + term_start, text.data() + term_end);
                                while (i != Utf8Iterator()) {
                                    if (Unicode::get_category(*i++) == Unicode::UPPERCASE_LETTER) {
                                        r = max_tw * 0.05;
                                    }
                                }
                            }
#endif
                        }
                        bgit = background.emplace(make_pair(stem, r)).first;
                    }
                    relevance = &bgit->second;
                }
            } else {
#if 0
                // In the absence of weight information, assume longer terms
                // are more relevant, and that unstemmed matches are a bit more
                // relevant than stemmed matches.
                if (queryterms.find(term) != queryterms.end()) {
                    relevance = term.size() * 3;
                } else {
                    string stem = "Z";
                    stem += stemmer(term);
                    if (queryterms.find(stem) != queryterms.end()) {
                        relevance = term.size() * 2;
                    }
                }
#endif
            }

            // FIXME: Allow Enquire without a DB set or an empty MSet() to be
            // used if you don't want the collection model?

#if 0
            // FIXME: Punctuation should somehow be included in the model, but this
            // approach is problematic - we don't want the first word of a sentence
            // to be favoured when it's at the end of the window.

            // Give first word in each sentence a relevance boost.
            if (term_start == 0) {
                relevance += 10;
            } else {
                for (size_t i = term_start; i + term.size() < term_end; ++i) {
                    if (text[i] == '.' && Unicode::is_whitespace(text[i + 1])) {
                        relevance += 10;
                        break;
                    }
                }
            }
#endif

relevance_done:
            if (longest_phrase) {
                phrase[phrase_next] = term;
                phrase_next = (phrase_next + 1) % (longest_phrase - 1);
            }

            if (highlight) matchfound = true;

            if (!snip.pump(relevance, term_end, highlight, flags)) return false;

            term_start = term_end;
            return true;
        });

    snip.done();

    // Put together the snippet.
    string result;
    if (matchfound || (flags & SNIPPET_EMPTY_WITHOUT_MATCH) == 0) {
        while (snip.drain(text, hi_start, hi_end, omit, result)) { }
    }

    return result;
}

}