Examples with BytesReader org.apache.lucene.util.fst.FST.BytesReader used on opensource projects

Example 1 with BytesReader

use of org.apache.lucene.util.fst.FST.BytesReader in project lucene-solr by apache.

the class FreeTextSuggester method lookup.

/** Retrieve suggestions. */
public List<LookupResult> lookup(final CharSequence key, Set<BytesRef> contexts, int num) throws IOException {
    if (contexts != null) {
        throw new IllegalArgumentException("this suggester doesn't support contexts");
    }
    if (fst == null) {
        throw new IllegalStateException("Lookup not supported at this time");
    }
    try (TokenStream ts = queryAnalyzer.tokenStream("", key.toString())) {
        TermToBytesRefAttribute termBytesAtt = ts.addAttribute(TermToBytesRefAttribute.class);
        OffsetAttribute offsetAtt = ts.addAttribute(OffsetAttribute.class);
        PositionLengthAttribute posLenAtt = ts.addAttribute(PositionLengthAttribute.class);
        PositionIncrementAttribute posIncAtt = ts.addAttribute(PositionIncrementAttribute.class);
        ts.reset();
        BytesRefBuilder[] lastTokens = new BytesRefBuilder[grams];
        //System.out.println("lookup: key='" + key + "'");
        // Run full analysis, but save only the
        // last 1gram, last 2gram, etc.:
        int maxEndOffset = -1;
        boolean sawRealToken = false;
        while (ts.incrementToken()) {
            BytesRef tokenBytes = termBytesAtt.getBytesRef();
            sawRealToken |= tokenBytes.length > 0;
            // TODO: this is somewhat iffy; today, ShingleFilter
            // sets posLen to the gram count; maybe we should make
            // a separate dedicated att for this?
            int gramCount = posLenAtt.getPositionLength();
            assert gramCount <= grams;
            // Safety: make sure the recalculated count "agrees":
            if (countGrams(tokenBytes) != gramCount) {
                throw new IllegalArgumentException("tokens must not contain separator byte; got token=" + tokenBytes + " but gramCount=" + gramCount + " does not match recalculated count=" + countGrams(tokenBytes));
            }
            maxEndOffset = Math.max(maxEndOffset, offsetAtt.endOffset());
            BytesRefBuilder b = new BytesRefBuilder();
            b.append(tokenBytes);
            lastTokens[gramCount - 1] = b;
        }
        ts.end();
        if (!sawRealToken) {
            throw new IllegalArgumentException("no tokens produced by analyzer, or the only tokens were empty strings");
        }
        // Carefully fill last tokens with _ tokens;
        // ShingleFilter appraently won't emit "only hole"
        // tokens:
        int endPosInc = posIncAtt.getPositionIncrement();
        // Note this will also be true if input is the empty
        // string (in which case we saw no tokens and
        // maxEndOffset is still -1), which in fact works out OK
        // because we fill the unigram with an empty BytesRef
        // below:
        boolean lastTokenEnded = offsetAtt.endOffset() > maxEndOffset || endPosInc > 0;
        if (lastTokenEnded) {
            // starting with "foo":
            for (int i = grams - 1; i > 0; i--) {
                BytesRefBuilder token = lastTokens[i - 1];
                if (token == null) {
                    continue;
                }
                token.append(separator);
                lastTokens[i] = token;
            }
            lastTokens[0] = new BytesRefBuilder();
        }
        Arc<Long> arc = new Arc<>();
        BytesReader bytesReader = fst.getBytesReader();
        // Try highest order models first, and if they return
        // results, return that; else, fallback:
        double backoff = 1.0;
        List<LookupResult> results = new ArrayList<>(num);
        // We only add a given suffix once, from the highest
        // order model that saw it; for subsequent lower order
        // models we skip it:
        final Set<BytesRef> seen = new HashSet<>();
        for (int gram = grams - 1; gram >= 0; gram--) {
            BytesRefBuilder token = lastTokens[gram];
            // Don't make unigram predictions from empty string:
            if (token == null || (token.length() == 0 && key.length() > 0)) {
                //System.out.println("  gram=" + gram + ": skip: not enough input");
                continue;
            }
            if (endPosInc > 0 && gram <= endPosInc) {
                //System.out.println("  break: only holes now");
                break;
            }
            //System.out.println("try " + (gram+1) + " gram token=" + token.utf8ToString());
            // TODO: we could add fuzziness here
            // match the prefix portion exactly
            //Pair<Long,BytesRef> prefixOutput = null;
            Long prefixOutput = null;
            try {
                prefixOutput = lookupPrefix(fst, bytesReader, token.get(), arc);
            } catch (IOException bogus) {
                throw new RuntimeException(bogus);
            }
            if (prefixOutput == null) {
                // This model never saw this prefix, e.g. the
                // trigram model never saw context "purple mushroom"
                backoff *= ALPHA;
                continue;
            }
            // TODO: we could do this division at build time, and
            // bake it into the FST?
            // Denominator for computing scores from current
            // model's predictions:
            long contextCount = totTokens;
            BytesRef lastTokenFragment = null;
            for (int i = token.length() - 1; i >= 0; i--) {
                if (token.byteAt(i) == separator) {
                    BytesRef context = new BytesRef(token.bytes(), 0, i);
                    Long output = Util.get(fst, Util.toIntsRef(context, new IntsRefBuilder()));
                    assert output != null;
                    contextCount = decodeWeight(output);
                    lastTokenFragment = new BytesRef(token.bytes(), i + 1, token.length() - i - 1);
                    break;
                }
            }
            final BytesRefBuilder finalLastToken = new BytesRefBuilder();
            if (lastTokenFragment == null) {
                finalLastToken.copyBytes(token.get());
            } else {
                finalLastToken.copyBytes(lastTokenFragment);
            }
            CharsRefBuilder spare = new CharsRefBuilder();
            // complete top-N
            TopResults<Long> completions = null;
            try {
                // Because we store multiple models in one FST
                // (1gram, 2gram, 3gram), we must restrict the
                // search so that it only considers the current
                // model.  For highest order model, this is not
                // necessary since all completions in the FST
                // must be from this model, but for lower order
                // models we have to filter out the higher order
                // ones:
                // Must do num+seen.size() for queue depth because we may
                // reject up to seen.size() paths in acceptResult():
                Util.TopNSearcher<Long> searcher = new Util.TopNSearcher<Long>(fst, num, num + seen.size(), weightComparator) {

                    BytesRefBuilder scratchBytes = new BytesRefBuilder();

                    @Override
                    protected void addIfCompetitive(Util.FSTPath<Long> path) {
                        if (path.arc.label != separator) {
                            //System.out.println("    keep path: " + Util.toBytesRef(path.input, new BytesRef()).utf8ToString() + "; " + path + "; arc=" + path.arc);
                            super.addIfCompetitive(path);
                        } else {
                        //System.out.println("    prevent path: " + Util.toBytesRef(path.input, new BytesRef()).utf8ToString() + "; " + path + "; arc=" + path.arc);
                        }
                    }

                    @Override
                    protected boolean acceptResult(IntsRef input, Long output) {
                        Util.toBytesRef(input, scratchBytes);
                        finalLastToken.grow(finalLastToken.length() + scratchBytes.length());
                        int lenSav = finalLastToken.length();
                        finalLastToken.append(scratchBytes);
                        //System.out.println("    accept? input='" + scratchBytes.utf8ToString() + "'; lastToken='" + finalLastToken.utf8ToString() + "'; return " + (seen.contains(finalLastToken) == false));
                        boolean ret = seen.contains(finalLastToken.get()) == false;
                        finalLastToken.setLength(lenSav);
                        return ret;
                    }
                };
                // since this search is initialized with a single start node 
                // it is okay to start with an empty input path here
                searcher.addStartPaths(arc, prefixOutput, true, new IntsRefBuilder());
                completions = searcher.search();
                assert completions.isComplete;
            } catch (IOException bogus) {
                throw new RuntimeException(bogus);
            }
            int prefixLength = token.length();
            BytesRefBuilder suffix = new BytesRefBuilder();
            nextCompletion: for (Result<Long> completion : completions) {
                token.setLength(prefixLength);
                // append suffix
                Util.toBytesRef(completion.input, suffix);
                token.append(suffix);
                //System.out.println("    completion " + token.utf8ToString());
                // Skip this path if a higher-order model already
                // saw/predicted its last token:
                BytesRef lastToken = token.get();
                for (int i = token.length() - 1; i >= 0; i--) {
                    if (token.byteAt(i) == separator) {
                        assert token.length() - i - 1 > 0;
                        lastToken = new BytesRef(token.bytes(), i + 1, token.length() - i - 1);
                        break;
                    }
                }
                if (seen.contains(lastToken)) {
                    //System.out.println("      skip dup " + lastToken.utf8ToString());
                    continue nextCompletion;
                }
                seen.add(BytesRef.deepCopyOf(lastToken));
                spare.copyUTF8Bytes(token.get());
                LookupResult result = new LookupResult(spare.toString(), (long) (Long.MAX_VALUE * backoff * ((double) decodeWeight(completion.output)) / contextCount));
                results.add(result);
                assert results.size() == seen.size();
            //System.out.println("  add result=" + result);
            }
            backoff *= ALPHA;
        }
        Collections.sort(results, new Comparator<LookupResult>() {

            @Override
            public int compare(LookupResult a, LookupResult b) {
                if (a.value > b.value) {
                    return -1;
                } else if (a.value < b.value) {
                    return 1;
                } else {
                    // Tie break by UTF16 sort order:
                    return ((String) a.key).compareTo((String) b.key);
                }
            }
        });
        if (results.size() > num) {
            results.subList(num, results.size()).clear();
        }
        return results;
    }
}

Example 2 with BytesReader

use of org.apache.lucene.util.fst.FST.BytesReader in project lucene-solr by apache.

the class AnalyzingSuggester method lookup.

@Override
public List<LookupResult> lookup(final CharSequence key, Set<BytesRef> contexts, boolean onlyMorePopular, int num) {
    assert num > 0;
    if (onlyMorePopular) {
        throw new IllegalArgumentException("this suggester only works with onlyMorePopular=false");
    }
    if (contexts != null) {
        throw new IllegalArgumentException("this suggester doesn't support contexts");
    }
    if (fst == null) {
        return Collections.emptyList();
    }
    //System.out.println("lookup key=" + key + " num=" + num);
    for (int i = 0; i < key.length(); i++) {
        if (key.charAt(i) == 0x1E) {
            throw new IllegalArgumentException("lookup key cannot contain HOLE character U+001E; this character is reserved");
        }
        if (key.charAt(i) == 0x1F) {
            throw new IllegalArgumentException("lookup key cannot contain unit separator character U+001F; this character is reserved");
        }
    }
    final BytesRef utf8Key = new BytesRef(key);
    try {
        Automaton lookupAutomaton = toLookupAutomaton(key);
        final CharsRefBuilder spare = new CharsRefBuilder();
        //System.out.println("  now intersect exactFirst=" + exactFirst);
        // Intersect automaton w/ suggest wFST and get all
        // prefix starting nodes & their outputs:
        //final PathIntersector intersector = getPathIntersector(lookupAutomaton, fst);
        //System.out.println("  prefixPaths: " + prefixPaths.size());
        BytesReader bytesReader = fst.getBytesReader();
        FST.Arc<Pair<Long, BytesRef>> scratchArc = new FST.Arc<>();
        final List<LookupResult> results = new ArrayList<>();
        List<FSTUtil.Path<Pair<Long, BytesRef>>> prefixPaths = FSTUtil.intersectPrefixPaths(convertAutomaton(lookupAutomaton), fst);
        if (exactFirst) {
            int count = 0;
            for (FSTUtil.Path<Pair<Long, BytesRef>> path : prefixPaths) {
                if (fst.findTargetArc(END_BYTE, path.fstNode, scratchArc, bytesReader) != null) {
                    // This node has END_BYTE arc leaving, meaning it's an
                    // "exact" match:
                    count++;
                }
            }
            // Searcher just to find the single exact only
            // match, if present:
            Util.TopNSearcher<Pair<Long, BytesRef>> searcher;
            searcher = new Util.TopNSearcher<>(fst, count * maxSurfaceFormsPerAnalyzedForm, count * maxSurfaceFormsPerAnalyzedForm, weightComparator);
            // ...:
            for (FSTUtil.Path<Pair<Long, BytesRef>> path : prefixPaths) {
                if (fst.findTargetArc(END_BYTE, path.fstNode, scratchArc, bytesReader) != null) {
                    // This node has END_BYTE arc leaving, meaning it's an
                    // "exact" match:
                    searcher.addStartPaths(scratchArc, fst.outputs.add(path.output, scratchArc.output), false, path.input);
                }
            }
            TopResults<Pair<Long, BytesRef>> completions = searcher.search();
            assert completions.isComplete;
            // maxSurfaceFormsPerAnalyzedForm:
            for (Result<Pair<Long, BytesRef>> completion : completions) {
                BytesRef output2 = completion.output.output2;
                if (sameSurfaceForm(utf8Key, output2)) {
                    results.add(getLookupResult(completion.output.output1, output2, spare));
                    break;
                }
            }
            if (results.size() == num) {
                // That was quick:
                return results;
            }
        }
        Util.TopNSearcher<Pair<Long, BytesRef>> searcher;
        searcher = new Util.TopNSearcher<Pair<Long, BytesRef>>(fst, num - results.size(), num * maxAnalyzedPathsForOneInput, weightComparator) {

            private final Set<BytesRef> seen = new HashSet<>();

            @Override
            protected boolean acceptResult(IntsRef input, Pair<Long, BytesRef> output) {
                // can get duplicate surface forms:
                if (seen.contains(output.output2)) {
                    return false;
                }
                seen.add(output.output2);
                if (!exactFirst) {
                    return true;
                } else {
                    // create duplicate results:
                    if (sameSurfaceForm(utf8Key, output.output2)) {
                        // have already found it in the first search:
                        assert results.size() == 1;
                        return false;
                    } else {
                        return true;
                    }
                }
            }
        };
        prefixPaths = getFullPrefixPaths(prefixPaths, lookupAutomaton, fst);
        for (FSTUtil.Path<Pair<Long, BytesRef>> path : prefixPaths) {
            searcher.addStartPaths(path.fstNode, path.output, true, path.input);
        }
        TopResults<Pair<Long, BytesRef>> completions = searcher.search();
        assert completions.isComplete;
        for (Result<Pair<Long, BytesRef>> completion : completions) {
            LookupResult result = getLookupResult(completion.output.output1, completion.output.output2, spare);
            // TODO: for fuzzy case would be nice to return
            // how many edits were required
            //System.out.println("    result=" + result);
            results.add(result);
            if (results.size() == num) {
                // produce one extra path
                break;
            }
        }
        return results;
    } catch (IOException bogus) {
        throw new RuntimeException(bogus);
    }
}

Example 3 with BytesReader

use of org.apache.lucene.util.fst.FST.BytesReader in project lucene-solr by apache.

the class TestFSTs method testExpandedCloseToRoot.

/**
   * Test state expansion (array format) on close-to-root states. Creates
   * synthetic input that has one expanded state on each level.
   *
   * @see <a href="https://issues.apache.org/jira/browse/LUCENE-2933">LUCENE-2933</a>
   */
public void testExpandedCloseToRoot() throws Exception {
    class SyntheticData {

        FST<Object> compile(String[] lines) throws IOException {
            final NoOutputs outputs = NoOutputs.getSingleton();
            final Object nothing = outputs.getNoOutput();
            final Builder<Object> b = new Builder<>(FST.INPUT_TYPE.BYTE1, outputs);
            int line = 0;
            final BytesRefBuilder term = new BytesRefBuilder();
            final IntsRefBuilder scratchIntsRef = new IntsRefBuilder();
            while (line < lines.length) {
                String w = lines[line++];
                if (w == null) {
                    break;
                }
                term.copyChars(w);
                b.add(Util.toIntsRef(term.get(), scratchIntsRef), nothing);
            }
            return b.finish();
        }

        void generate(ArrayList<String> out, StringBuilder b, char from, char to, int depth) {
            if (depth == 0 || from == to) {
                String seq = b.toString() + "_" + out.size() + "_end";
                out.add(seq);
            } else {
                for (char c = from; c <= to; c++) {
                    b.append(c);
                    generate(out, b, from, c == to ? to : from, depth - 1);
                    b.deleteCharAt(b.length() - 1);
                }
            }
        }

        public int verifyStateAndBelow(FST<Object> fst, Arc<Object> arc, int depth) throws IOException {
            if (FST.targetHasArcs(arc)) {
                int childCount = 0;
                BytesReader fstReader = fst.getBytesReader();
                for (arc = fst.readFirstTargetArc(arc, arc, fstReader); ; arc = fst.readNextArc(arc, fstReader), childCount++) {
                    boolean expanded = fst.isExpandedTarget(arc, fstReader);
                    int children = verifyStateAndBelow(fst, new FST.Arc<>().copyFrom(arc), depth + 1);
                    assertEquals(expanded, (depth <= FST.FIXED_ARRAY_SHALLOW_DISTANCE && children >= FST.FIXED_ARRAY_NUM_ARCS_SHALLOW) || children >= FST.FIXED_ARRAY_NUM_ARCS_DEEP);
                    if (arc.isLast())
                        break;
                }
                return childCount;
            }
            return 0;
        }
    }
    // Sanity check.
    assertTrue(FST.FIXED_ARRAY_NUM_ARCS_SHALLOW < FST.FIXED_ARRAY_NUM_ARCS_DEEP);
    assertTrue(FST.FIXED_ARRAY_SHALLOW_DISTANCE >= 0);
    SyntheticData s = new SyntheticData();
    ArrayList<String> out = new ArrayList<>();
    StringBuilder b = new StringBuilder();
    s.generate(out, b, 'a', 'i', 10);
    String[] input = out.toArray(new String[out.size()]);
    Arrays.sort(input);
    FST<Object> fst = s.compile(input);
    FST.Arc<Object> arc = fst.getFirstArc(new FST.Arc<>());
    s.verifyStateAndBelow(fst, arc, 1);
}

Example 4 with BytesReader

use of org.apache.lucene.util.fst.FST.BytesReader in project lucene-solr by apache.

the class TestFSTs method testShortestPathsRandom.

public void testShortestPathsRandom() throws Exception {
    final Random random = random();
    int numWords = atLeast(1000);
    final TreeMap<String, Long> slowCompletor = new TreeMap<>();
    final TreeSet<String> allPrefixes = new TreeSet<>();
    final PositiveIntOutputs outputs = PositiveIntOutputs.getSingleton();
    final Builder<Long> builder = new Builder<>(FST.INPUT_TYPE.BYTE1, outputs);
    final IntsRefBuilder scratch = new IntsRefBuilder();
    for (int i = 0; i < numWords; i++) {
        String s;
        while (true) {
            s = TestUtil.randomSimpleString(random);
            if (!slowCompletor.containsKey(s)) {
                break;
            }
        }
        for (int j = 1; j < s.length(); j++) {
            allPrefixes.add(s.substring(0, j));
        }
        // weights 1..100
        int weight = TestUtil.nextInt(random, 1, 100);
        slowCompletor.put(s, (long) weight);
    }
    for (Map.Entry<String, Long> e : slowCompletor.entrySet()) {
        //System.out.println("add: " + e);
        builder.add(Util.toIntsRef(new BytesRef(e.getKey()), scratch), e.getValue());
    }
    final FST<Long> fst = builder.finish();
    //System.out.println("SAVE out.dot");
    //Writer w = new OutputStreamWriter(new FileOutputStream("out.dot"));
    //Util.toDot(fst, w, false, false);
    //w.close();
    BytesReader reader = fst.getBytesReader();
    //System.out.println("testing: " + allPrefixes.size() + " prefixes");
    for (String prefix : allPrefixes) {
        // 1. run prefix against fst, then complete by value
        //System.out.println("TEST: " + prefix);
        long prefixOutput = 0;
        FST.Arc<Long> arc = fst.getFirstArc(new FST.Arc<Long>());
        for (int idx = 0; idx < prefix.length(); idx++) {
            if (fst.findTargetArc((int) prefix.charAt(idx), arc, arc, reader) == null) {
                fail();
            }
            prefixOutput += arc.output;
        }
        final int topN = TestUtil.nextInt(random, 1, 10);
        Util.TopResults<Long> r = Util.shortestPaths(fst, arc, fst.outputs.getNoOutput(), minLongComparator, topN, true);
        assertTrue(r.isComplete);
        // 2. go thru whole treemap (slowCompletor) and check it's actually the best suggestion
        final List<Result<Long>> matches = new ArrayList<>();
        // TODO: could be faster... but it's slowCompletor for a reason
        for (Map.Entry<String, Long> e : slowCompletor.entrySet()) {
            if (e.getKey().startsWith(prefix)) {
                //System.out.println("  consider " + e.getKey());
                matches.add(new Result<>(Util.toIntsRef(new BytesRef(e.getKey().substring(prefix.length())), new IntsRefBuilder()), e.getValue() - prefixOutput));
            }
        }
        assertTrue(matches.size() > 0);
        Collections.sort(matches, new TieBreakByInputComparator<>(minLongComparator));
        if (matches.size() > topN) {
            matches.subList(topN, matches.size()).clear();
        }
        assertEquals(matches.size(), r.topN.size());
        for (int hit = 0; hit < r.topN.size(); hit++) {
            //System.out.println("  check hit " + hit);
            assertEquals(matches.get(hit).input, r.topN.get(hit).input);
            assertEquals(matches.get(hit).output, r.topN.get(hit).output);
        }
    }
}

Example 5 with BytesReader

use of org.apache.lucene.util.fst.FST.BytesReader in project lucene-solr by apache.

the class Util method get.

/** Looks up the output for this input, or null if the
   *  input is not accepted. */
public static <T> T get(FST<T> fst, IntsRef input) throws IOException {
    // TODO: would be nice not to alloc this on every lookup
    final FST.Arc<T> arc = fst.getFirstArc(new FST.Arc<T>());
    final BytesReader fstReader = fst.getBytesReader();
    // Accumulate output as we go
    T output = fst.outputs.getNoOutput();
    for (int i = 0; i < input.length; i++) {
        if (fst.findTargetArc(input.ints[input.offset + i], arc, arc, fstReader) == null) {
            return null;
        }
        output = fst.outputs.add(output, arc.output);
    }
    if (arc.isFinal()) {
        return fst.outputs.add(output, arc.nextFinalOutput);
    } else {
        return null;
    }
}