edismax支持boost函数与score相乘作为,而dismax只能使用bf作用效果是相加,所以在处理多个维度排序时,score其实也应该是其中一个维度 ,用相加的方式处理调整麻烦。
而dismax的实现代码逻辑比较简单,看起来比较易理解,edismax是它的加强版,其实是改变了不少。。比如在以下:
先看看dismax的解析主要实现思路:
首先取出搜索字段名qf
将最终解析成一个BooleanQuery
先解析主mainQuery:
用户主要是搜索串的解析
altQuery解析处理,看是否使用用户定义的后备搜索串
PhraseQuery解析组装
再解析bq查询,主要是额外加分的查询,不会影响搜索结果数,只会影响排序
再则是bf解析,函数搜索最后会以加的方式作用于文档评分
看主要代码更清晰:
@Override
public Query parse() throws ParseException {
SolrParams solrParams = SolrParams.wrapDefaults(localParams, params);
queryFields = SolrPluginUtils.parseFieldBoosts(solrParams.getParams(DisMaxParams.QF));
if (0 == queryFields.size()) {
queryFields.put(req.getSchema().getDefaultSearchFieldName(), 1.0f);
}
/* the main query we will execute. we disable the coord because
* this query is an artificial construct
*/
BooleanQuery query = new BooleanQuery(true);
boolean notBlank = addMainQuery(query, solrParams);
if (!notBlank)
return null;
addBoostQuery(query, solrParams);
addBoostFunctions(query, solrParams);
return query;
}
edismax的主要实现思路跟dismax差不多,以下是一些主要差别之处:
edismax解析含有+,OR,NOT,-语法时,就会忽略掉使用MM。
以下是主要代码实现:
统计搜索串中+,OR ,NOT,-语法元个数
// defer escaping and only do if lucene parsing fails, or we need phrases
// parsing fails. Need to sloppy phrase queries anyway though.
List<Clause> clauses = null;
int numPluses = 0;
int numMinuses = 0;
int numOR = 0;
int numNOT = 0;
clauses = splitIntoClauses(userQuery, false);
for (Clause clause : clauses) {
if (clause.must == '+') numPluses++;
if (clause.must == '-') numMinuses++;
if (clause.isBareWord()) {
String s = clause.val;
if ("OR".equals(s)) {
numOR++;
} else if ("NOT".equals(s)) {
numNOT++;
} else if (lowercaseOperators && "or".equals(s)) {
numOR++;
}
}
}
/当搜索串里包含有+,OR ,NOT,-这四种时候,mm就会失效
boolean doMinMatched = (numOR + numNOT + numPluses + numMinuses) == 0;
if (parsedUserQuery != null && doMinMatched) {
String minShouldMatch = solrParams.get(DisMaxParams.MM, "100%");
if (parsedUserQuery instanceof BooleanQuery) {
SolrPluginUtils.setMinShouldMatch((BooleanQuery)parsedUserQuery, minShouldMatch);
}
}
短语查询,先找出普通的查询,原来就是短语查询的、或者属于“OR”,“AND”,“NOT”,’TO‘类型的都不要。由于edismax支持解析符合lucene语法的搜索串,所以不像dismax那样,只需要简单的将搜索串去掉\“,然后加个“”括起来就行
// find non-field clauses
List<Clause>normalClauses =new ArrayList<Clause>(clauses.size());
for (Clauseclause :clauses) {
if (clause.field !=null ||clause.isPhrase)continue;
// check for keywords "AND,OR,TO"
if (clause.isBareWord()) {
String s =clause.val.toString();
// avoid putting explict operators in the phrase query
if ("OR".equals(s) ||"AND".equals(s) ||"NOT".equals(s) || "TO".equals(s))continue;
}
normalClauses.add(clause);
}
// full phrase...
addShingledPhraseQueries(query, normalClauses, phraseFields, 0,
tiebreaker,pslop);
// shingles...
addShingledPhraseQueries(query, normalClauses, phraseFields2, 2,
tiebreaker,pslop);
addShingledPhraseQueries(query, normalClauses, phraseFields3, 3,
tiebreaker,pslop);
下面是dismax获取短语查询的作法:
protected Query getPhraseQuery(String userQuery, SolrPluginUtils.DisjunctionMaxQueryParser pp) throws ParseException {
String userPhraseQuery = userQuery.replace("\"", "");
return pp.parse("\"" + userPhraseQuery + "\"");
}
下面是edismax的作法:
private void addShingledPhraseQueries(final BooleanQuery mainQuery,
final List<Clause> clauses,
final Map<String,Float> fields,
int shingleSize,
final float tiebreaker,
final int slop)
throws ParseException {
if (null == fields || fields.isEmpty() ||
null == clauses || clauses.size() <= shingleSize )
return;
if (0 == shingleSize) shingleSize = clauses.size();
final int goat = shingleSize-1; // :TODO: better name for var?
StringBuilder userPhraseQuery = new StringBuilder();
for (int i=0; i < clauses.size() - goat; i++) {
userPhraseQuery.append('"');
for (int j=0; j <= goat; j++) {
userPhraseQuery.append(clauses.get(i + j).val);
userPhraseQuery.append(' ');
}
userPhraseQuery.append('"');
userPhraseQuery.append(' ');
}
ExtendedSolrQueryParser pp =
new ExtendedSolrQueryParser(this, IMPOSSIBLE_FIELD_NAME);
pp.addAlias(IMPOSSIBLE_FIELD_NAME, tiebreaker, fields);
pp.setPhraseSlop(slop);
pp.setRemoveStopFilter(true); // remove stop filter and keep stopwords
pp.makeDismax = true;
pp.minClauseSize = 2;
Query phrase = pp.parse(userPhraseQuery.toString());
if (phrase != null) {
mainQuery.add(phrase, BooleanClause.Occur.SHOULD);
}
}
edismax技术另一个重要的boost查询,
boost查询也是不会影响搜索结果数,但是影响排序,主要作用是将最后得分以相乘的方式作用于score,函数的解析跟bf差不多。
//
// create a boosted query (scores multiplied by boosts)
//
Query topQuery = query;
multBoosts = solrParams.getParams("boost");
if (multBoosts!=null && multBoosts.length>0) {
List<ValueSource> boosts = new ArrayList<ValueSource>();
for (String boostStr : multBoosts) {
if (boostStr==null || boostStr.length()==0) continue;
Query boost = subQuery(boostStr, FunctionQParserPlugin.NAME).getQuery();
ValueSource vs;
if (boost instanceof FunctionQuery) {
vs = ((FunctionQuery)boost).getValueSource();
} else {
vs = new QueryValueSource(boost, 1.0f);
}
boosts.add(vs);
}
if (boosts.size()>1) {
ValueSource prod = new ProductFloatFunction(boosts.toArray(new ValueSource[boosts.size()]));
topQuery = new BoostedQuery(query, prod);
} else if (boosts.size() == 1) {
topQuery = new BoostedQuery(query, boosts.get(0));
}
}
可以看到最后不是一个BooleanQuery,而是一个BoostedQuery。
它就是简单处理子查询的分值再与函数查询的分值相乘返回 :主要的score方法如下:
public float score() throws IOException {
float score = qWeight * scorer.score() * vals.floatVal(scorer.docID());
return score>Float.NEGATIVE_INFINITY ? score : -Float.MAX_VALUE;
}