diff --git a/src/rime/dict/user_dictionary.cc b/src/rime/dict/user_dictionary.cc
index 9b57a69b0..f2105c45b 100644
--- a/src/rime/dict/user_dictionary.cc
+++ b/src/rime/dict/user_dictionary.cc
@@ -325,6 +325,21 @@ an<UserDictEntryCollector> UserDictionary::Lookup(
   for (auto& v : state.query_result) {
     v.second.Sort();
   }
+  auto entries_with_word_completion =
+      state.query_result.find(state.predict_word_from_depth);
+  if (entries_with_word_completion != state.query_result.end()) {
+    auto& entries = entries_with_word_completion->second;
+    // if the top candidate is predictive match,
+    if (!entries.empty() && entries.front()->IsPredictiveMatch()) {
+      auto found =
+          std::find_if(entries.begin(), entries.end(),
+                       [](const auto& e) { return e->IsExactMatch(); });
+      if (found != entries.end()) {
+        // move the first exact match candidate to top.
+        std::rotate(entries.begin(), found, found + 1);
+      }
+    }
+  }
   return collect(&state.query_result);
 }
 
diff --git a/src/rime/dict/vocabulary.cc b/src/rime/dict/vocabulary.cc
index 1328c3c4f..b2d76d008 100644
--- a/src/rime/dict/vocabulary.cc
+++ b/src/rime/dict/vocabulary.cc
@@ -57,10 +57,6 @@ string Code::ToString() const {
   return stream.str();
 }
 
-inline ShortDictEntry DictEntry::ToShort() const {
-  return {text, code, weight};
-}
-
 bool ShortDictEntry::operator<(const ShortDictEntry& other) const {
   // Sort different entries sharing the same code by weight desc.
   if (weight != other.weight)
diff --git a/src/rime/dict/vocabulary.h b/src/rime/dict/vocabulary.h
index 0dc196ec4..ca7b397be 100644
--- a/src/rime/dict/vocabulary.h
+++ b/src/rime/dict/vocabulary.h
@@ -55,7 +55,13 @@ struct DictEntry {
   int matching_code_size = 0;
 
   DictEntry() = default;
-  ShortDictEntry ToShort() const;
+  ShortDictEntry ToShort() const { return {text, code, weight}; }
+  bool IsExactMatch() const {
+    return matching_code_size == 0 || matching_code_size == code.size();
+  }
+  bool IsPredictiveMatch() const {
+    return matching_code_size != 0 && matching_code_size < code.size();
+  }
   bool operator<(const DictEntry& other) const;
 };
 
diff --git a/src/rime/gear/contextual_translation.cc b/src/rime/gear/contextual_translation.cc
index 604d96cce..b037749c7 100644
--- a/src/rime/gear/contextual_translation.cc
+++ b/src/rime/gear/contextual_translation.cc
@@ -18,7 +18,8 @@ bool ContextualTranslation::Replenish() {
     DLOG(INFO) << cand->text() << " cache/queue: " << cache_.size() << "/"
                << queue.size();
     if (cand->type() == "phrase" || cand->type() == "user_phrase" ||
-        cand->type() == "table" || cand->type() == "user_table") {
+        cand->type() == "table" || cand->type() == "user_table" ||
+        cand->type() == "completion") {
       if (end_pos != cand->end() || last_type != cand->type()) {
         end_pos = cand->end();
         last_type = cand->type();
diff --git a/src/rime/gear/script_translator.cc b/src/rime/gear/script_translator.cc
index 176f9461d..853b8fc33 100644
--- a/src/rime/gear/script_translator.cc
+++ b/src/rime/gear/script_translator.cc
@@ -130,7 +130,7 @@ class ScriptTranslation : public Translation {
  protected:
   bool CheckEmpty();
   bool IsNormalSpelling() const;
-  void PrepareCandidate();
+  bool PrepareCandidate();
   template <class QueryResult>
   void EnrollEntries(map<int, DictEntryList>& entries_by_end_pos,
                      const an<QueryResult>& query_result);
@@ -147,11 +147,19 @@ class ScriptTranslation : public Translation {
   an<Sentence> sentence_;
 
   an<Phrase> candidate_ = nullptr;
+  size_t candidate_index_ = 0;
+  enum CandidateSource {
+    kUninitialized,
+    kUserPhrase,
+    kSysPhrase,
+    kSentence,
+  };
+  CandidateSource candidate_source_ = kUninitialized;
 
   DictEntryCollector::reverse_iterator phrase_iter_;
   UserDictEntryCollector::reverse_iterator user_phrase_iter_;
 
-  size_t max_corrections_ = 4;
+  const size_t max_corrections_ = 4;
   size_t correction_count_ = 0;
 
   bool enable_correction_;
@@ -342,9 +350,10 @@ string ScriptSyllabifier::GetOriginalSpelling(const Phrase& cand) const {
   return string();
 }
 
-static bool is_exact_match_phrase(const an<DictEntry>& entry) {
-  return entry && entry->matching_code_size == 0 ||
-         entry->matching_code_size == entry->code.size();
+template <class Ptr, class Iter>
+static bool has_exact_match_phrase(Ptr ptr, Iter iter, size_t consumed) {
+  return ptr && iter->first == consumed && !iter->second.exhausted() &&
+         iter->second.Peek()->IsExactMatch();
 }
 
 // ScriptTranslation implementation
@@ -372,15 +381,10 @@ bool ScriptTranslation::Evaluate(Dictionary* dict, UserDictionary* user_dict) {
     user_phrase_iter_ = user_phrase_->rbegin();
 
   // make sentences when there is no exact-matching phrase candidate
-  bool has_exact_match_phrase =
-      phrase_ && phrase_iter_->first == consumed &&
-      is_exact_match_phrase(phrase_iter_->second.Peek());
-  bool has_exact_match_user_phrase =
-      user_phrase_ && user_phrase_iter_->first == consumed &&
-      is_exact_match_phrase(user_phrase_iter_->second.Peek());
   bool has_at_least_two_syllables = syllable_graph.edges.size() >= 2;
-  if (!has_exact_match_phrase && !has_exact_match_user_phrase &&
-      has_at_least_two_syllables) {
+  if (has_at_least_two_syllables &&
+      !has_exact_match_phrase(phrase_, phrase_iter_, consumed) &&
+      !has_exact_match_phrase(user_phrase_, user_phrase_iter_, consumed)) {
     sentence_ = MakeSentence(dict, user_dict);
   }
 
@@ -393,43 +397,42 @@ bool ScriptTranslation::Next() {
     is_correction = false;
     if (exhausted())
       return false;
-    if (sentence_) {
-      sentence_.reset();
-      return !CheckEmpty();
+    if (candidate_source_ == kUninitialized) {
+      PrepareCandidate();  // to determine candidate_source_
     }
-    int user_phrase_code_length = 0;
-    if (user_phrase_ && user_phrase_iter_ != user_phrase_->rend()) {
-      user_phrase_code_length = user_phrase_iter_->first;
-    }
-    int phrase_code_length = 0;
-    if (phrase_ && phrase_iter_ != phrase_->rend()) {
-      phrase_code_length = phrase_iter_->first;
-    }
-    if (user_phrase_code_length > 0 &&
-        user_phrase_code_length >= phrase_code_length) {
-      UserDictEntryIterator& uter(user_phrase_iter_->second);
-      if (!uter.Next()) {
-        ++user_phrase_iter_;
-      }
-    } else if (phrase_code_length > 0) {
-      DictEntryIterator& iter(phrase_iter_->second);
-      if (!iter.Next()) {
-        ++phrase_iter_;
-      }
+    switch (candidate_source_) {
+      case kUninitialized:
+        break;
+      case kSentence:
+        sentence_.reset();
+        break;
+      case kUserPhrase: {
+        UserDictEntryIterator& uter(user_phrase_iter_->second);
+        if (!uter.Next()) {
+          ++user_phrase_iter_;
+        }
+      } break;
+      case kSysPhrase: {
+        DictEntryIterator& iter(phrase_iter_->second);
+        if (!iter.Next()) {
+          ++phrase_iter_;
+        }
+      } break;
     }
+    candidate_.reset();
+    candidate_source_ = kUninitialized;
     if (enable_correction_) {
-      PrepareCandidate();
-      if (!candidate_) {
+      // populate next candidate and skip it if it's a correction beyond max
+      // numbers.
+      if (!PrepareCandidate()) {
         break;
       }
-      is_correction = syllabifier_->IsCandidateCorrection(*candidate_);
     }
-  } while (  // limit the number of correction candidates
-      enable_correction_ && is_correction &&
-      correction_count_ > max_corrections_);
-  if (is_correction) {
-    ++correction_count_;
-  }
+  } while (enable_correction_ &&
+           syllabifier_->IsCandidateCorrection(*candidate_) &&
+           // limit the number of correction candidates
+           ++correction_count_ > max_corrections_);
+  ++candidate_index_;
   return !CheckEmpty();
 }
 
@@ -440,8 +443,7 @@ bool ScriptTranslation::IsNormalSpelling() const {
 }
 
 an<Candidate> ScriptTranslation::Peek() {
-  PrepareCandidate();
-  if (!candidate_) {
+  if (candidate_source_ == kUninitialized && !PrepareCandidate()) {
     return nullptr;
   }
   if (candidate_->preedit().empty()) {
@@ -458,14 +460,29 @@ an<Candidate> ScriptTranslation::Peek() {
   return candidate_;
 }
 
-void ScriptTranslation::PrepareCandidate() {
+static bool always_true() {
+  return true;
+}
+
+template <typename T>
+inline static bool prefer_user_phrase(
+    T user_phrase_weight,
+    T sys_phrase_weight,
+    function<bool()> compare_on_tie = always_true) {
+  return user_phrase_weight > sys_phrase_weight ||
+         (user_phrase_weight == sys_phrase_weight && compare_on_tie());
+}
+
+bool ScriptTranslation::PrepareCandidate() {
   if (exhausted()) {
+    candidate_source_ = kUninitialized;
     candidate_ = nullptr;
-    return;
+    return false;
   }
   if (sentence_) {
+    candidate_source_ = kSentence;
     candidate_ = sentence_;
-    return;
+    return true;
   }
   size_t user_phrase_code_length = 0;
   if (user_phrase_ && user_phrase_iter_ != user_phrase_->rend()) {
@@ -475,28 +492,44 @@ void ScriptTranslation::PrepareCandidate() {
   if (phrase_ && phrase_iter_ != phrase_->rend()) {
     phrase_code_length = phrase_iter_->first;
   }
-  an<Phrase> cand;
   if (user_phrase_code_length > 0 &&
-      user_phrase_code_length >= phrase_code_length) {
+      prefer_user_phrase(user_phrase_code_length, phrase_code_length, [this]() {
+        const int kNumExactMatchOnTop = 1;
+        size_t full_code_length = end_of_input_ - start_;
+        return candidate_index_ >= kNumExactMatchOnTop ||
+               prefer_user_phrase(
+                   has_exact_match_phrase(user_phrase_, user_phrase_iter_,
+                                          full_code_length),
+                   has_exact_match_phrase(phrase_, phrase_iter_,
+                                          full_code_length));
+      })) {
     UserDictEntryIterator& uter = user_phrase_iter_->second;
     const auto& entry = uter.Peek();
     DLOG(INFO) << "user phrase '" << entry->text
                << "', code length: " << user_phrase_code_length;
-    cand = New<Phrase>(translator_->language(), "user_phrase", start_,
-                       start_ + user_phrase_code_length, entry);
-    cand->set_quality(std::exp(entry->weight) + translator_->initial_quality() +
-                      (IsNormalSpelling() ? 0.5 : -0.5));
+    candidate_source_ = kUserPhrase;
+    candidate_ =
+        New<Phrase>(translator_->language(),
+                    entry->IsPredictiveMatch() ? "completion" : "user_phrase",
+                    start_, start_ + user_phrase_code_length, entry);
+    candidate_->set_quality(std::exp(entry->weight) +
+                            translator_->initial_quality() +
+                            (IsNormalSpelling() ? 0.5 : -0.5));
   } else if (phrase_code_length > 0) {
     DictEntryIterator& iter = phrase_iter_->second;
     const auto& entry = iter.Peek();
     DLOG(INFO) << "phrase '" << entry->text
                << "', code length: " << phrase_code_length;
-    cand = New<Phrase>(translator_->language(), "phrase", start_,
-                       start_ + phrase_code_length, entry);
-    cand->set_quality(std::exp(entry->weight) + translator_->initial_quality() +
-                      (IsNormalSpelling() ? 0 : -1));
+    candidate_source_ = kSysPhrase;
+    candidate_ =
+        New<Phrase>(translator_->language(),
+                    entry->IsPredictiveMatch() ? "completion" : "phrase",
+                    start_, start_ + phrase_code_length, entry);
+    candidate_->set_quality(std::exp(entry->weight) +
+                            translator_->initial_quality() +
+                            (IsNormalSpelling() ? 0 : -1));
   }
-  candidate_ = cand;
+  return true;
 }
 
 bool ScriptTranslation::CheckEmpty() {