Proto commits in google/mozc

These commits are when the Protocol Buffers files have changed: (only the last 100 relevant commits are shown)

2026-06-10

Add ProtobufUtil for sanitizing string fields in protobuf messages. The `SanitizeMessageStrings` function uses protobuf reflection to traverse a given message. It applies a provided `absl::FunctionRef` to all string fields found, including those in nested and repeated messages. If the sanitizer returns a new string, the field is updated; otherwise, it remains unchanged. PiperOrigin-RevId: 929669400

The documentation is generated from this commit.

2026-06-01

Refactor: Use named fields for candidate view text styles This change refactors the RendererStyle proto to use individual named fields (shortcut_style, gap1_style, candidate_style, description_style) instead of a repeated field for candidate view text styles. This improves readability and type safety. The code accessing these styles in Mac and Win32 renderers has been updated accordingly. #codehealth PiperOrigin-RevId: 924578705

Enables allow exact match mode by default. Removed relevant flags. Enabled on prod 100%. PiperOrigin-RevId: 924539162

2026-05-27

Refactor: Use TextFormat to initialize RendererStyle This change replaces the programmatic initialization of the RendererStyle proto with parsing from a human-readable text proto string. This makes the default style definition easier to read and maintain. PiperOrigin-RevId: 922002571

2026-05-26

Improve next word prediction performance by optimizing state expansion. Profile revealed that Predictor::Next spent a significant amount of time in NextStates and shrinking the results. This change introduces NextTopStates to fetch only the top N states directly, avoiding allocation and sorting of all states. A feature flag `ngram_use_next_top_states` is added to select the implementation for online experiments. Benchmark results show ~55% reduction in latency. PiperOrigin-RevId: 921251452

2026-05-19

Disables context-aware candidate mixing features because it didn't improve the quality. PiperOrigin-RevId: 917596171

2026-05-08

Enable partial matches in user history prediction. This feature is already launched. Note that this feature is also enabled on desktop. PiperOrigin-RevId: 912343877

2026-04-24

Delete CalculatePrefixPenaltyLegacy. Migrated to the new prefix penalty scorer. PiperOrigin-RevId: 904840249

2026-04-21

enables cache inner segment mode by default. This feature has already been launched in production. PiperOrigin-RevId: 902943135

2026-03-30

Suggests full-sentence entries in exact-match case. This mode is enabled when both `cache_inner_segment` and `allow_exact_match` are true. PiperOrigin-RevId: 891449677

2026-03-19

Added the Mendel flag to disable legacy rewriters. The plan is to eventually migrate these personalization and collocation features to UserHistoryPredictor and supplemental model PiperOrigin-RevId: 886064572

2026-03-16

Enable partial-match functionality in the user history predictor - Uses SuffixDecoder to compute the remaining suffix cost and comparing the default cost against the combined prefix and suffix costs. - The decoder will then adjust behavior based on this comparison, either promoting partial matches, ranking them lower, or suppressing them entirely. PiperOrigin-RevId: 884316795

2026-03-13

The new RelatimeDecoder::DecodeSuffix method computes a more precise penalty for partial candidates. This new mode is guarded by a flag. We've also introduced a new Mendel parameter to control the additional penalty for demoting partial candidates over full candidates. PiperOrigin-RevId: 883007961

2026-03-02

Add the is_a11y_talkback_enabled option to Request. Some conversion behaviors should be changed when TalkBack is enabled. e.g. The candidate window should be always visible to propagate the candidate words to TalkBack. PiperOrigin-RevId: 877163530

2026-02-20

Override the top history suggestion with dictionary suggestion when the post correction is reasonably high enough. This promotion is triggered when the top candidate is not a proper noun. PiperOrigin-RevId: 872865405

2026-02-06

Adds API to insert new history entry from the client. PiperOrigin-RevId: 866296300

2026-01-26

Hook search key on mobile software keyboard. This information is going to be used in partial-revert features. Decoder wants to know whether corrections and conversion have been fully completed. PiperOrigin-RevId: 860963825

2026-01-19

Add a new session command to request NextWordPrediction PiperOrigin-RevId: 857965692

2026-01-15

Introduced DELETABLE attribute. Whether the candidate is deletable is not determined by user history flag. We would like to allow the decoder to control the deletable behavior. PiperOrigin-RevId: 856540231

2026-01-14

Replaced custom error codes in UserDictionary with absl::Status for better consistency. Merged all custom error code into single ExtendedErrorCode and store it as absl::raw_code. PiperOrigin-RevId: 856093474

Second trial: Removed user_history_predict_space_prefix flag.predict_space_prefix is already launched. PiperOrigin-RevId: 856054760

2026-01-13

Rename use_auto_ime_turn_off -> auto_switch_composition_mode in Mozc Config API. This Mozc feature does context-based "automatic switching of composition mode", not "automatically turn IME off". The current naming seems inaccurate & potentially misleading. PiperOrigin-RevId: 855530664

Deprecates UserDictionarySession. Migrated to simpler UserDictioanryImporter API. PiperOrigin-RevId: 855517196

Automated g4 rollback of changelist 855504417. *** Reason for rollback *** Autorollback has determined with high confidence that [cl/855504417](http://cl/855504417) has broken the TAP targets listed below. Autorollback is designed to address the situations where inaccuracy is most common, but you could still get unlucky. [cl/855504417](http://cl/855504417) has been rolled back. *** Original change description *** Removed user_history_predict_space_prefix flag.predict_space_prefix is already launched. *** PiperOrigin-RevId: 855511418

Removed user_history_predict_space_prefix flag.predict_space_prefix is already launched. PiperOrigin-RevId: 855504417

2026-01-06

Second trial of ImportUserDictionary migration. Keeps the original UserDictionarySession as chromeos is still using this legacy API. Uses the new and simple ImportUserDictionary API to import a TSV dictionary fed from the client. We will replace the current UserDictionarySession with this simple API. The new user-dictionary-import API is thread-safe and executed asynchronously on the C++ decoder. PiperOrigin-RevId: 852766629

2025-12-28

Automated g4 rollback of changelist 849473003. *** Reason for rollback *** Autorollback has determined with high confidence that [cl/849473003](http://cl/849473003) has broken the TAP targets listed below. Autorollback is designed to address the situations where inaccuracy is most common, but you could still get unlucky. *** Original change description *** Uses the new and simple ImportUserDictionary API to import a TSV dictionary fed from the client. We will replace the current UserDictionarySession with this simple API. The new user-dictionary-import API is thread-safe and executed asynchronously on the C++ decoder. *** PiperOrigin-RevId: 849704005

2025-12-27

Uses the new and simple ImportUserDictionary API to import a TSV dictionary fed from the client. We will replace the current UserDictionarySession with this simple API. The new user-dictionary-import API is thread-safe and executed asynchronously on the C++ decoder. PiperOrigin-RevId: 849473003

2025-12-19

Fix 17-year-old misnomers in PunctuationMethod options of Mozc Config. These longstanding misnomers have existed [since Feb 2009](http://cl/9971834). It's since "spread" across multiple components, incl. cross-language and/or cross-repo ones. This CL fixes the root causes in Mozc Config where everything originates, and their direct callsites. These aren't persisted to disks in stringly manner. Similar misnomers in cross-language and/or cross-repo components will necessarily be handled separately in follow-up CLs in their respective repos, while considering cross-repo inter-dependencies to avoid breakages induced by version skews, as long as they aren't persisted to user disks in stringly manner. PiperOrigin-RevId: 846472067

2025-12-16

Introduced fst_decoder_nwp_penalty. Special penalty/bonus value added to NWP candidate of FST decoder. PiperOrigin-RevId: 845259394

2025-11-25

Rename SWITCH_INPUT_MODE command to SWITCH_COMPOSITION_MODE in Mozc API. This helps improve terminology consistency & reduce confusion in Mozc API. The command operates on a CompositionMode param, whereas "input mode" is the typical user-facing label (e.g. on CrOS) for the unrelated concept known internally as PreeditMethod (romaji vs kana). Related terminology at layers immediately adjacent to the Mozc API, on both client & Mozc-internal sides, is adapted accordingly. PiperOrigin-RevId: 836610651

Introduced helper function. fst::PupulateResults to manages the triggering conditions of FstDecoder, and merges/mixes the English decoder's result with the current results. Converts the FstDecoder's score to Mozc cost using linear conversion. PiperOrigin-RevId: 836531914

2025-11-19

Cleanup the flag. user_history_suppress_min_length. The A/B test was neutral, so we can safely disable the length-based filtering on user-history prediction. This changes makes the code simpler and cleaner. PiperOrigin-RevId: 834107787

2025-11-15

Deprecates demote partial suggestion feature. We could not see a positive impact. In order to simplify the code, remove this functionality. PiperOrigin-RevId: 832675362

2025-11-12

Implement particle skipping in N-gram prediction. This CL introduces a feature to the N-gram based Next Word Prediction (NWP) to optionally skip common Japanese particles (e.g., "は", "を", "が", "に") during candidate generation. This improves prediction quality by offering more fluent and relevant candidates that don't strictly follow the preceding particle. PiperOrigin-RevId: 831215606

2025-11-04

Added General, simple and non-recursive RichText message for display_value v2. PiperOrigin-RevId: 827927562

2025-10-29

Caches the inner segment boundary info in the user history predictor. When this model is enabled, the full sentence with the inner segment boundary is cached. The coverage of user history prediction is improved in exact-match cases, while the final goal of this mode is to unify the storage of personalization features, PiperOrigin-RevId: 825375430

2025-10-24

Support NWP containing space as prefix. Since the history-segments are cleared with space, the history information is heuristically restored from the preceding text. PiperOrigin-RevId: 823552309

2025-10-22

Improved bi-gram zero query suggestion ranking based on generation time. Updated the `next_entry_fps()` storage so that `next_entry_fps()` can work as an LRU cache. The last item is the latest. PiperOrigin-RevId: 822490770

Added display_value_capability to Request. The client can declare whether it supports display_value and decoration level. Server can change the behavior or functionality depending on this level. PiperOrigin-RevId: 822354807

2025-10-20

Add support for Emoji 17.0. * This is a preparation to update the Emoji data to 17.0. #codehealth PiperOrigin-RevId: 821522004

2025-10-17

Removed the Mendel flag user_history_cache_full_sentence. This flag is now only enabled on canary and has never been pushed to prod. We found that this model makes the suggestion pretty aggressive, which is inconsistent with the fact that users prefer exact match results. We will merge this flag to the upcoming model to remember inner segment boundary information. PiperOrigin-RevId: 820500823

2025-10-16

Introduced "display_value" annotation. `display_value` is the value displayed on candidate window. We can encode meta-information in the display_value. PiperOrigin-RevId: 820215818

2025-10-14

Add experiment parameters for SymbolRewriter. * symbol_rewriter_candidate_position + Position of the first symbol rewriter candidate (0-origin). * symbol_rewriter_promotion_size + Number of candidates to be promoted. For example, if symbol_rewriter_candidate_position == 0, and symbol_rewriter_promotion_size == 5, the first 5 candidates are symbols. PiperOrigin-RevId: 818988399

2025-10-07

Resubmit the previous CL after fixing the build error. Cloned from CL 815593311 by 'g4 patch'. Removed the flags and code to configure the history cache size and lifetime. Uses 10k for all devices. PiperOrigin-RevId: 816204712

2025-10-06

Automated g4 rollback of changelist 815593311. *** Reason for rollback *** Autorollback has determined with high confidence that cl/815593311 has broken the TAP targets listed below. Autorollback is designed to address the situations where inaccuracy is most common, but you could still get unlucky. *** Original change description *** Removed the flags and code to configure the history cache size and lifetime. Uses 10k for all devices. *** PiperOrigin-RevId: 815600970

Removed the flags and code to configure the history cache size and lifetime. Uses 10k for all devices. PiperOrigin-RevId: 815593311

2025-10-03

Stores the inner segment boundary information in user history. This change has no quality impact. In the short term, we want to correctly implement the matching conditions for history deletion using boundary information. In the long term, we aim to integrate the converter's training module into the user history. PiperOrigin-RevId: 814619726

2025-10-02

Add a flag to cache the full sentence in user history learning. The desktop already caches full sentences. We would like to unify and simplify the implementation. PiperOrigin-RevId: 814128517

2025-09-29

Introduced a flag to control the minimum value length to suppress the candidate. We will have a rollout experiment to disable this length-based filtering logic. Eventually, we'll deprecate the filtering logic to simplify the implementation. PiperOrigin-RevId: 812607502

2025-09-24

Allows to run post corrections without inner segment boundary constraint. This mode is enabled when the value is hiragana. e.g., "靴は|いた" -> "靴|履いた" PiperOrigin-RevId: 810709916

2025-09-22

Add a new parameter for the number of rescoring targets for decoding. This CL is a preparation to use a different number of rescoring targets for decoding and next word prediction. We may want to use a different number to control the latency as the LSTM is called more times in decoding than in NWP. PiperOrigin-RevId: 809944298

2025-09-17

Demote short partial suggestions. Extremely short partial candidates occasionally appear at the top few candidates. We would like to demote them so that they do not appear at the top N candidates, where N is the parameter. (demote_partial_candidate_top_n). We also introduce the maximum length of partial candidates to be demoted. PiperOrigin-RevId: 807944144

2025-09-05

Remove user_history_partial_revert_mode and enables partial revert feature by default. PiperOrigin-RevId: 803333457

2025-09-02

Uses 64bit fingerprint to store the chain links. We also deprecated NextEntry message to save the serialized storage. PiperOrigin-RevId: 802066636

2025-08-29

Automated g4 rollback of changelist 800367114. *** Reason for rollback *** The joined prediction feature is already disabled on mobile, so we cannot measure the impact via experiments. We need to discuss whether desktop specific features are necessary. *** Original change description *** Added flag to disable joined (bigram) prediction from user history. When user types "東京" and "大学" sequentially, the current decoder predicts "東京大学" from "とう". At least in the Mobile environment, longer prediction is not preferred. We would like to confirm whether this prediction is useful through the experiments. *** PiperOrigin-RevId: 800726449

2025-08-28

Added flag to disable joined (bigram) prediction from user history. When user types "東京" and "大学" sequentially, the current decoder predicts "東京大学" from "とう". At least in the Mobile environment, longer prediction is not preferred. We would like to confirm whether this prediction is useful through the experiments. PiperOrigin-RevId: 800367114

2025-07-22

Remove apply_single_inner_segment_boundary flag. The behavior is unified to the case where the flag is true. PiperOrigin-RevId: 785677783

2025-07-17

Supports partial revert feature. Heuristically determines the number of characters actually removed from backspace keys, instead of reverting all characters. The context information must be populated from the client. PiperOrigin-RevId: 784080634

2025-07-16

Support different interpolation weights for decoding and NWP. Base scores of decoding and NWP are very different. In decoding, scores are mostly assigned by the underlying class language model and it includes the score of reading model. Taking interpolation between these base scores and rescoring LM's scores is thus reasonable. However, in NWP, many components provide results with rule-based scores just for ordering (e.g., 1, 10, 20, ... etc.). Therefore, it is often meaningless to take interpolation with such scores. By supporting different weights, one can control the behavior in a more flexible manner. This CL renames the proto field name but this is safe, as it is not referenced yet even in canary. PiperOrigin-RevId: 783650572

Use the LSTM beam search in next word prediction. PiperOrigin-RevId: 783638134

2025-07-14

Deprecates conversion_freq. Uses suggestion_freq only. suggestion_freq and conversion_freq were managed separately, but the distinction between them has become ambiguous, and there's no longer a good reason to differentiate their behavior. Supports the move operation in LURCache::Insert to directly transfer the user history storage to LRU cache. PiperOrigin-RevId: 782760366

2025-07-10

Apply LSTM generation method in the supplemental next word prediction. PiperOrigin-RevId: 781391940

2025-07-09

Remove findability oriented order feature. The experiment result was not good and we decided not to launch the feature. PiperOrigin-RevId: 780907417

2025-07-03

Support penalty for <unk> token. LM may give high scores for <unk>, especially when the vocabulary size is small. This CL adds a workaround to penalize <unk>. This CL also adds a flag to control the usage of softmax for rescoring. If the model is trained with self normalization, one may skip it. PiperOrigin-RevId: 778726029

Add a new rescoring feature. The new rescoring feature has two target selection modes: 1) simple selection by top N and 2) selection based on key length. Decoding and next word prediciton can use different modes. This CL introduces the rescoring by the LMTM LM. PiperOrigin-RevId: 778716506

2025-06-25

Deprecates the separation of DesktopPredictor and MobilePredictor. - engine_type is deprecated as it is used to switch Desktop/Mobile predictor. - We only have one Predictor and switch the logic based on mixed_conversion flag. PiperOrigin-RevId: 775495629

2025-06-19

Removed unused Mendel flag bigram_nwp_filtering_mode PiperOrigin-RevId: 773301232

Removed unused flag mobile_history_prediction_size: History size is now determined by the character coverage automatically, so this flag is no longer used. PiperOrigin-RevId: 773269540

2025-06-16

Added Mendel parameters to control the size and lifetime of user history entries. We will experiment with the lifetime and size of history data to find optimal values. If the effectiveness of the lifetime parameter cannot be verified, we will use only the size constraint to simplify the code. PiperOrigin-RevId: 771890898

2025-05-30

Remove EntryType of UserHistoryPredictor entry type was used to store the cleanup event as a history entry. However this feature was originally designed for sync feature and no longer used now. PiperOrigin-RevId: 765027613

2025-04-24

Deprecated user_history_prediction_filter_redundant_candidates_mode and user_history_prediction_aggressive_bigram. These mode already launched to prod. PiperOrigin-RevId: 750827634

2025-04-09

Add a feature flag to filter NWP suffix candidates based on their transition cost PiperOrigin-RevId: 745503591

2025-04-03

Add a comment about the alternative with the obsolete `mode`. Note, this `mode` was obsoleted by cl/16872841. #codehealth PiperOrigin-RevId: 743375728

2025-03-31

Add a clarification comment for id in commands.proto. #codehealth PiperOrigin-RevId: 742164769

Add 'is_incognito_mode' field to Request proto message. This is an additional option to Config::incognito_mode. Clients can use Request::is_incognito_mode instead of Config::incognito_mode so that the clients can skip calling SET_CONFIG commands to the server. #codehealth PiperOrigin-RevId: 742109093

2025-03-21

Update the comment about HIGHLIGHT_CANDIDATE #codehealth PiperOrigin-RevId: 739082082

2025-03-11

Clean up the experimental flag: realtime conversion candidate checker PiperOrigin-RevId: 735604374

Clean up the experimental flag: user_segment_history_rewriter_replace_proper_noun PiperOrigin-RevId: 735595946

2025-03-10

Deprecated enabled and auto_registered fields of UserDictioanryStorage. These fields is no longer used. PiperOrigin-RevId: 735289024

2025-03-09

Completely removed spelling correction features from Mozc. PiperOrigin-RevId: 735104457

2025-03-07

Remvoed dictionary sync features. They are no longer used and mantained. PiperOrigin-RevId: 734589342

2025-03-04

Add the experimental flag to filter NWP bigram entries. PiperOrigin-RevId: 733138670

2025-03-03

Add is_handwriting to Request PiperOrigin-RevId: 732786172

2025-02-25

Add a experimental flag to control the bugfix in cl/694010741 PiperOrigin-RevId: 730762929

2025-02-13

- Introduced Options::incognito_mode to make an on-the-fly incognito mode request. - Added ConversionRequest::incognito_mode() - Stop having incognito_config_ in EngineConverter. PiperOrigin-RevId: 726384302

2025-02-06

Removed usage_stats feature. UsageStats is no longer used. PiperOrigin-RevId: 723855520

2024-12-19

Deprecated install_path. install_path is no longer used after MDD. PiperOrigin-RevId: 707856864

2024-11-28

Fix typos. PiperOrigin-RevId: 700945320

Reranks suffix-dictionary-based zero query suggestions with neutral sentence classifier model. PiperOrigin-RevId: 700894055

2024-11-11

Add experimental param to relax the condition to remember candidates for the segment including proper noun words. Example: 1. Submit "崇が" for "たかしが". - The candidate may not have personal name POS, but some others should hae personal name POS. 2. Input "たかしの" 3. "崇の" will be promoted with this feature. PiperOrigin-RevId: 695144738

2024-11-05

Add logics to support Emoji 16.0. * This is a preparation to update the Emoji data to 16.0. #codehealth PiperOrigin-RevId: 693255328

Add a feature flag to change the cost offset for handwriting conversion candidates PiperOrigin-RevId: 693200996

Rename `java_outer_classname` of `candidate_window.proto` to `ProtoCandidateWindow`. PiperOrigin-RevId: 692825449

2024-10-30

Clean up the experimental param: user_history_prediction_min_selected_ratio PiperOrigin-RevId: 691275706

2024-10-23

Rename `candidates.proto` to `candidate_window.proto` for better readability. PiperOrigin-RevId: 688832889

2024-10-22

Rename `SessionState.candidates` to `SessionState.candidate_window` for better readability. PiperOrigin-RevId: 688396251

Rename `CandidateWindow.subcandidates` to `CandidateWindow.sub_candidate_window` for better readability. PiperOrigin-RevId: 688383891

Deprecates the T13n promotion based on the base LM score. PiperOrigin-RevId: 688376533

2024-10-21

Rename `Output.candidates` to `Output.candidate_window` for better readability. PiperOrigin-RevId: 688146644