syntax = "proto3";

package hermes;

// Search service
service SearchService {
  // Search for documents
  rpc Search(SearchRequest) returns (SearchResponse);

  // Get document by ID
  rpc GetDocument(GetDocumentRequest) returns (GetDocumentResponse);

  // Get index info
  rpc GetIndexInfo(GetIndexInfoRequest) returns (GetIndexInfoResponse);

  // Text statistics (document frequencies, corpus sizes, average lengths)
  // of the BM25 terms of a query over this backend's index. A broker that
  // scatters one index over several shards sums the responses and sends the
  // total back as `SearchRequest.text_stats`, so every shard scores with
  // the same IDF.
  rpc GetTextStats(GetTextStatsRequest) returns (GetTextStatsResponse);
}

// Index service
service IndexService {
  // Create a new index (supports both structured schema and SDL string)
  rpc CreateIndex(CreateIndexRequest) returns (CreateIndexResponse);

  // Add documents to index (streaming)
  rpc IndexDocuments(stream IndexDocumentRequest) returns (IndexDocumentsResponse);

  // Add documents in batch
  rpc BatchIndexDocuments(BatchIndexDocumentsRequest) returns (BatchIndexDocumentsResponse);

  // Stage whole-document deletions by exact primary key, including every chunk.
  rpc DeleteDocuments(DeleteDocumentsRequest) returns (DocumentMutationResponse);

  // Stage complete replacements (inserts if absent). Commit publishes them.
  rpc UpsertDocuments(UpsertDocumentsRequest) returns (DocumentMutationResponse);

  // Commit pending changes
  rpc Commit(CommitRequest) returns (CommitResponse);

  // Force merge segments
  rpc ForceMerge(ForceMergeRequest) returns (ForceMergeResponse);

  // Run bounded text, BMP, Seismic, and binary ANN layout maintenance
  rpc Reorder(ReorderRequest) returns (ReorderResponse);

  // Delete an index
  rpc DeleteIndex(DeleteIndexRequest) returns (DeleteIndexResponse);

  // List all indexes
  rpc ListIndexes(ListIndexesRequest) returns (ListIndexesResponse);

  // Retrain global vector codebooks and atomically rebuild every ANN segment
  rpc RetrainVectorIndex(RetrainVectorIndexRequest) returns (RetrainVectorIndexResponse);

  // Atomically replace one vector field's IVF/ScaNN algorithm or parameters
  rpc AlterVectorIndex(AlterVectorIndexRequest) returns (AlterVectorIndexResponse);
}

// Query types
message Query {
  oneof query {
    TermQuery term = 1;
    BooleanQuery boolean = 2;
    BoostQuery boost = 3;
    AllQuery all = 4;
    SparseVectorQuery sparse_vector = 5;
    DenseVectorQuery dense_vector = 6;
    MatchQuery match = 7;
    RangeQuery range = 8;
    PrefixQuery prefix = 9;
    BinaryDenseVectorQuery binary_dense_vector = 10;
    FusionQuery fusion = 11;  // Top-level only (cannot be nested)
    PhraseQuery phrase = 12;
  }
}

// Method for fusing sub-query result lists
enum FusionMethod {
  FUSION_RRF = 0;                      // Reciprocal Rank Fusion (rank-based, default)
  FUSION_NORMALIZED_WEIGHTED_SUM = 1;  // Min-max normalized weighted score sum
  // Return complete bounded nomination lists/union for coordinator-side fusion.
  FUSION_CANDIDATES = 2;
}

// Weighted sub-query for hybrid fusion
message WeightedQuery {
  Query query = 1;
  float weight = 2;  // Legacy RRF/NWS contribution; 0 or unset = 1.0
  // Required, unique feature identity when l1 or score_export is supplied.
  string name = 3;
  ScoreScope scope = 4;
  // Backfill this branch without using it to nominate candidates (L1/export only).
  bool score_only = 5;
}

// Union fusion of independently-executed sub-queries (e.g. sparse + dense
// hybrid retrieval). Unlike the reranker (which can only re-score documents
// the first-stage query found), fusion keeps documents found by ANY
// sub-query. Only valid at the top level of SearchRequest.query; composes
// with SearchRequest.reranker (the fused list becomes the L1 candidates).
message FusionQuery {
  repeated WeightedQuery queries = 1;
  FusionMethod method = 2;
  float rrf_k = 3;         // RRF rank constant; 0 = default 60
  // Combiner for fused per-chunk (ordinal) scores into a document score.
  // Fusion runs at chunk granularity: same-chunk hits across sub-queries
  // compound, and results carry per-chunk ordinal_scores.
  // Unset (0) = MAX (recommended; LogSumExp is unsuitable at RRF magnitudes).
  MultiValueCombiner combiner = 5;
  // Common document eligibility for every branch. Does not create an L1 feature.
  repeated Query filters = 6;
  // Independent per-branch nomination depth (L1/export only). 0 uses
  // candidate_limit. The complete union survives until L1 ranking/export.
  uint32 candidate_depth = 7;
}

// Scope is explicit: document features never become body ordinal zero.
enum ScoreScope {
  SCORE_SCOPE_UNSPECIFIED = 0;
  SCORE_SCOPE_DOCUMENT = 1;
  SCORE_SCOPE_CHUNK = 2;
}

// Fixed preprocessing trained offline, never derived from a shard/result page.
// The only L1 scoring contract: a compiled symbolic formula over branch scores.
message L1Ranking {
  reserved 1, 2, 3, 6;
  reserved "weights", "bias", "transforms", "rrf_weight";
  // Omitted=true: backfill only missing raw cells. False uses organic values.
  optional bool backfill = 4;
  // Raw defaults for missing formula variables; omitted variables use zero.
  map<string, double> missing_values = 5;
  // Required, nonempty expression over named raw scores and reserved `rrf`.
  // Runs before passage/document selection. Requires formula_v1.
  string formula = 7;
}

message ScoreExport {
  // Maximum raw passage rows per hit; 0 = all (65536). With l1, exported rows
  // are the best passages after complete scoring. Without l1, exceeding this
  // bound fails instead of silently omitting features the caller needs to rank.
  uint32 passages_per_document = 1;
  // Diagnostics: score every stored passage of nominated documents. Default
  // false scores the union of nominated passages only, plus document context.
  bool all_passages = 2;
  // Capability 4: score real body ordinals for document-only nominees.
  // Requires backfill and a chunk-scoped feature. Existing body nominations
  // retain their original passage union; no organic votes are fabricated.
  bool seed_document_passages = 3;
}

// How to combine scores for multi-valued documents
enum MultiValueCombiner {
  COMBINER_LOG_SUM_EXP = 0;  // Softmax-weighted smooth maximum (default); count-invariant
  COMBINER_MAX = 1;          // Take maximum score
  COMBINER_AVG = 2;          // Average all scores
  COMBINER_SUM = 3;          // Sum all scores
  COMBINER_WEIGHTED_TOP_K = 4;  // Weighted top-k with decay
}

// Sparse vector query for similarity search
// Either provide (indices, values) directly, or provide text for server-side tokenization
message SparseVectorQuery {
  string field = 1;
  repeated uint32 indices = 2;   // Pre-computed token indices
  repeated float values = 3;     // Pre-computed token values
  string text = 4;               // Raw text (tokenized server-side if tokenizer configured)
  MultiValueCombiner combiner = 5;  // How to combine scores for multi-value fields
  float heap_factor = 6; // BMP/MaxScore pruning factor (0 = schema default, 1 = exact).
  float combiner_temperature = 7;   // Temperature for LogSumExp (default: 1.5)
  uint32 combiner_top_k = 8;        // K for WeightedTopK (default: 5)
  float combiner_decay = 9;         // Decay for WeightedTopK (default: 0.7)
  float weight_threshold = 10;      // Min abs(weight) for query dims (0 = no filtering)
  uint32 max_query_dims = 11;       // Max query dimensions to process (0 = all)
  float pruning = 12;               // Fraction of query dims to keep (0 = no pruning, 0.1 = top 10%)
  optional uint32 lsp_gamma = 13; // BMP LSP gamma (unset = depth-derived, 0 = exhaustive).
  optional uint32 seismic_cut = 14; // Nomination query dimensions (1..64)
  optional float seismic_factor = 15; // Summary pruning factor (0..1)
  optional bool exhaustive = 16; // Exact shared-forward scan; false allows approximation
}

// Dense vector query for similarity search
message DenseVectorQuery {
  string field = 1;
  repeated float vector = 2;
  uint32 nprobe = 3;           // Number of clusters to probe (for IVF indexes)
  MultiValueCombiner combiner = 5;  // How to combine scores for multi-value fields
  float combiner_temperature = 6;   // Temperature for LogSumExp (default: 1.5)
  uint32 combiner_top_k = 7;        // K for WeightedTopK (default: 5)
  float combiner_decay = 8;         // Decay for WeightedTopK (default: 0.7)
}

// Binary dense vector query for Hamming distance search
message BinaryDenseVectorQuery {
  string field = 1;
  bytes vector = 2;                        // Packed-bit query vector (ceil(dim/8) bytes)
  MultiValueCombiner combiner = 3;         // How to combine scores for multi-value fields
  float combiner_temperature = 4;
  uint32 combiner_top_k = 5;
  float combiner_decay = 6;
}

message TermQuery {
  string field = 1;
  string term = 2;
  // Optional hint for the field's tokenizer (see MatchQuery.tokenizer_hint)
  string tokenizer_hint = 3;
}

message BooleanQuery {
  repeated Query must = 1;
  repeated Query should = 2;
  repeated Query must_not = 3;
}

message BoostQuery {
  Query query = 1;
  float boost = 2;
}

message AllQuery {}

// Range query on a fast field (inclusive bounds)
// Exactly one pair of bounds should be set (u64, i64, or f64)
message RangeQuery {
  string field = 1;
  optional uint64 min_u64 = 2;
  optional uint64 max_u64 = 3;
  optional int64 min_i64 = 4;
  optional int64 max_i64 = 5;
  optional double min_f64 = 6;
  optional double max_f64 = 7;
}

// Full-text match query - tokenizes text server-side and searches as OR of tokens
// Use this instead of TermQuery when searching with natural language text
message MatchQuery {
  string field = 1;
  string text = 2;
  // Optional hint passed to the field's tokenizer. Static tokenizers ignore
  // it; a `lex(by: languages, ...)` tokenizer
  // reads it as a comma-separated list of language codes ("ru,en") and stems
  // each query token with the first listed language of the token's script.
  // Empty = the tokenizer's default.
  string tokenizer_hint = 3;
  // Proximity rescoring (sequential dependence): the top candidates of the
  // BM25 pass gain a bonus for adjacent query terms occurring next to each
  // other (ordered windows) or within `proximity_window` positions
  // (unordered windows, half weight), scaled by this weight. 0 = off.
  // Approximate: the bonus applies to an over-fetched candidate pool.
  float proximity_weight = 4;
  // Unordered window size; 0 = 8.
  uint32 proximity_window = 5;
  // Approximate MaxScore: scale the pruning threshold by 1/heap_factor
  // (0/unset or 1 = exact, rank-safe; 0.8 prunes more aggressively at some
  // recall cost). Must be finite and in [0, 1]. This option applies to text queries.
  float heap_factor = 6;
  // Long queries: keep only the `max_terms` rarest (highest idf) tokens for
  // scoring; 0 = all tokens. Approximate.
  uint32 max_terms = 7;
}

// Phrase query - text is tokenized server-side with the field's tokenizer
// (so stemming matches indexing) and the terms must occur at the distances
// the tokenizer assigned (consecutively unless the tokenizer dropped stop
// words between them, whose positions are kept as gaps) or within `slop`
// positions of that. Requires the field to be indexed with token positions
// (`indexed<token_position>` or `indexed<positions>`); without positions the
// engine degrades to a MUST of the terms. Scored with BM25 over the phrase
// frequency (occurrences of the whole phrase) and the summed idf of its terms.
// A phrase whose words are all stop words is rejected standalone and dropped
// (with a server warning) when it is a clause of a BooleanQuery.
message PhraseQuery {
  string field = 1;
  string text = 2;
  uint32 slop = 3;
  string tokenizer_hint = 4;  // See MatchQuery.tokenizer_hint
}

// Prefix query - matches all documents containing any term starting with prefix
// Score is always 1.0 (filter-style). Efficient as a MUST clause in BooleanQuery.
message PrefixQuery {
  string field = 1;
  string prefix = 2;
}

// L2 reranker: rerank L1 candidates by exact vector distance
// For dense vectors: set `vector` (f32). For binary dense vectors: set `binary_vector` (bytes).
message Reranker {
  string field = 1;                    // Vector field (dense or binary dense)
  repeated float vector = 2;           // Query vector (f32, for dense fields)
  MultiValueCombiner combiner = 4;
  float combiner_temperature = 5;
  uint32 combiner_top_k = 6;
  float combiner_decay = 7;
  uint32 matryoshka_dims = 8;          // Matryoshka pre-filter dims (0 = disabled)
  bytes binary_vector = 9;             // Query vector (packed bits, for binary dense fields)
  float rrf_k = 10;                    // Reciprocal Rank Fusion k (0 = disabled, typical: 60)
}

// Search request/response
message SearchRequest {
  string index_name = 1;
  Query query = 2;
  uint32 limit = 3;
  uint32 offset = 4;
  repeated string fields_to_load = 5;
  Reranker reranker = 6;               // Optional L2 reranker
  // First-stage candidate pool. 0 uses offset + limit; explicit values must
  // be between offset + limit and the 2x ceiling.
  uint32 candidate_limit = 7;
  // Anytime mode: wall-clock budget in milliseconds for the scoring phase.
  // Text (BM25) MaxScore executors stop at the deadline and return the best
  // results found so far; `SearchResponse.truncated` reports that it fired.
  // 0 = unbounded (exact top-k).
  uint64 time_budget_ms = 8;
  // Cross-shard text statistics for BM25 (see `GetTextStats`). When set they
  // replace the backend's own statistics for the listed fields and terms;
  // unset = the backend aggregates over its own segments.
  TextStats text_stats = 9;
  // Requires named, scoped fusion branches. The formula may reference the RRF feature.
  L1Ranking l1 = 10;
  // With no l1, exports the complete candidate union for external inference.
  // Export-only requests require offset=0 and a limit covering the whole union.
  ScoreExport score_export = 11;
  // Return RRF diagnostics alongside the requested ranking score. Requires
  // top-level fusion and complete nomination; does not change ranking.
  bool include_rrf_scores = 12;
  // Preserve bounded per-shard nomination and selection traces. Default false;
  // does not rerun subqueries or change retrieval/ranking budgets.
  bool tracing = 13;
}

message GetTextStatsRequest {
  string index_name = 1;
  // The query whose BM25 terms are collected (after this backend's own
  // tokenization); sparse, dense and filter clauses contribute nothing.
  Query query = 2;
}

message GetTextStatsResponse {
  TextStats stats = 1;
}

// BM25 statistics of a set of text terms.
message TextStats {
  // Documents in the index (or the sum over shards).
  uint64 total_docs = 1;
  repeated TextFieldStats fields = 2;
}

message TextFieldStats {
  string field = 1;
  // Scoring units of the field: documents for a plain field, chunks for a
  // chunked one.
  uint64 corpus_size = 2;
  // Average length of a scoring unit in tokens.
  float avg_len = 3;
  repeated TermDocFreq terms = 4;
}

message TermDocFreq {
  // Index-time token bytes.
  bytes term = 1;
  uint64 doc_freq = 2;
}

// Unique document address: segment + local doc_id
message DocAddress {
  string segment_id = 1;          // Segment ID (hex, 32 chars)
  uint32 doc_id = 2;              // Segment-local document ID
}

message SearchHit {
  DocAddress address = 1;
  float score = 2;
  map<string, FieldValueList> fields = 3;
  repeated OrdinalScore ordinal_scores = 4;  // Per-ordinal scores for multi-value fields
  CandidateScores candidate_scores = 5; // Present only when score_export is requested.
  optional float rrf_score = 6; // Present when include_rrf_scores is requested.
  repeated RrfContribution rrf_contributions = 7;
}

// One organic nomination vote. Document context has no ordinal; chunk votes
// are combined per ordinal before the document's fusion combiner is applied.
message RrfContribution {
  uint32 query_index = 1;
  string query_name = 2; // Empty for unnamed legacy branches.
  uint32 rank = 3; // One-based rank in the complete merged nomination list.
  float score = 4; // Branch weight / (k + rank).
  optional uint32 ordinal = 5;
}

message CandidateScores {
  // Branch name -> raw score. Absent key means unavailable indexed data;
  // a valid nonmatch is explicitly 0. Negative dense scores are preserved.
  map<string, float> document = 1;
  repeated PassageScores passages = 2;
  // Before limiting exported rows; L1 selected from all of these passages.
  uint32 scored_passages = 3;
}
message PassageScores {
  uint32 ordinal = 1;
  map<string, float> scores = 2;
  // Final formula score when l1 is present, otherwise no ranking score.
  optional float l1_score = 3;
}

// List of field values — supports multi-value fields (e.g. multiple URIs per document)
message FieldValueList {
  repeated FieldValue values = 1;
}

// Score contribution from a specific ordinal in a multi-valued field
message OrdinalScore {
  uint32 ordinal = 1;  // Which value in the multi-valued field (0-indexed)
  float score = 2;     // Score contribution from this ordinal
}

message FieldValue {
  oneof value {
    string text = 1;
    uint64 u64 = 2;
    int64 i64 = 3;
    double f64 = 4;
    bytes bytes_value = 5;
    SparseVector sparse_vector = 6;
    DenseVector dense_vector = 7;
    string json_value = 8;  // JSON serialized as string
    bytes binary_dense_vector = 9;  // Packed-bit binary vector
  }
}

// Sparse vector with term indices and weights
message SparseVector {
  repeated uint32 indices = 1;
  repeated float values = 2;
}

// Dense vector (float32 values)
message DenseVector {
  repeated float values = 1;
}

message SearchResponse {
  repeated SearchHit hits = 1;
  uint64 total_hits = 2;
  uint64 took_ms = 3;
  SearchTimings timings = 4;
  // True when `time_budget_ms` expired before every candidate was scored;
  // the hits are then the best-so-far, not the exact top-k.
  bool truncated = 5;
  // Feature contracts: "formula_v1" or "feature_export_v2".
  string ranking_method = 6;
  // Present for FUSION_CANDIDATES. Hits hold hydrated union entries once;
  // these lists preserve independent branch scores and nominated ordinals.
  repeated FusionCandidateList fusion_candidates = 7;
  SearchTrace trace = 8; // Present only when tracing is requested.
  // Capability 4 acknowledgment, true only when every backend honored the
  // requested document-only passage seeding policy. Old adapters default false.
  bool seeded_document_passages = 9;
}

message SearchTrace {
  repeated ShardSearchTrace shards = 1;
}

message ShardSearchTrace {
  string shard_id = 1; // Populated by the broker; empty on direct server calls.
  string backend_id = 2;
  string index_name = 3;
  repeated QueryTrace queries = 4;
  repeated FusionCandidate selected = 5; // Shard results before broker selection.
  string ranking_method = 6;
  bool truncated = 7;
  repeated Query filters = 8; // Common fusion filters applied during nomination.
}

message QueryTrace {
  uint32 query_index = 1;
  string query_name = 2;
  Query query = 3; // Original branch/root expression, including nested clauses.
  ScoreScope scope = 4;
  bool score_only = 5;
  uint32 candidate_depth = 6;
  uint32 total_seen = 7;
  // All organic nominations within the requested depth, before fusion/L1/L2.
  repeated FusionCandidate candidates = 8;
}

message FusionCandidateList {
  uint32 query_index = 1;
  repeated FusionCandidate candidates = 2;
}

message FusionCandidate {
  DocAddress address = 1;
  float score = 2;
  repeated OrdinalScore ordinal_scores = 3;
}

// Detailed timing breakdown for search phases (all values in microseconds)
message SearchTimings {
  uint64 search_us = 1;       // L1 retrieval (query scoring across segments)
  uint64 rerank_us = 2;       // L2 reranking (dense vector rescoring)
  uint64 load_us = 3;         // Document field loading from store
  uint64 total_us = 4;        // Wall-clock total (includes overhead)
  uint64 candidate_scoring_us = 5; // Missing-only feature backfill and formula selection.
}

// Get document request/response
message GetDocumentRequest {
  string index_name = 1;
  DocAddress address = 2;
}

message GetDocumentResponse {
  map<string, FieldValueList> fields = 1;
}

// Index info request/response
message GetIndexInfoRequest {
  string index_name = 1;
}

message GetIndexInfoResponse {
  string index_name = 1;
  uint32 num_docs = 2;
  uint32 num_segments = 3;
  // Schema in SDL format
  string schema = 4;
  // Memory usage breakdown (if available)
  MemoryStats memory_stats = 5;
  // Per-field vector statistics
  repeated VectorFieldStats vector_stats = 6;
  // Per text field: how it is tokenized, so clients pick query forms
  // without parsing the schema SDL.
  repeated TextFieldInfo text_fields = 7;
  uint32 candidate_scoring_version = 8; // 3 requires symbolic L1 formula; preserves organic scores, optional backfill and raw defaults.
  repeated string unprepared_candidate_fields = 9; // Fields requiring lookup preparation/migration.
  uint64 physical_num_docs = 10; // Includes tombstones; num_docs counts live rows.
  uint64 num_deleted_docs = 11;
  double deleted_ratio = 12; // num_deleted_docs / physical_num_docs; zero for an empty index.
}

// Tokenization facts of one text field.
message TextFieldInfo {
  string field = 1;
  // The field uses a `lex(...)` tokenizer.
  bool lexical = 2;
  // A `lex(by: <field>)` tokenizer: the query's `tokenizer_hint` selects
  // the stemmer (empty when the field has no hint field).
  string hint_field = 3;
  // `lex(variants: true)`: the written word is indexed next to its stem and
  // folded forms, so one clause per field suffices (phrases match the
  // written form, match queries the stem).
  bool variants = 4;
  // Token positions are indexed, so phrase queries verify adjacency.
  bool positions = 5;
  // `indexed<chunked>`: every value is its own BM25 unit with an ordinal.
  bool chunked = 6;
}

// Per-field vector statistics (dense or sparse)
message VectorFieldStats {
  string field_name = 1;
  string vector_type = 2;  // "dense" or "sparse"
  uint64 total_vectors = 3;
  uint32 dimension = 4;    // dim for dense, num_dimensions for sparse
  float avg_terms_per_vector = 5;  // sparse only: avg postings per vector (0 for dense)
}

// Memory usage statistics
message MemoryStats {
  // Total estimated memory usage in bytes
  uint64 total_bytes = 1;
  // Indexing buffer memory (pending documents not yet flushed)
  IndexingBufferStats indexing_buffer = 2;
  // Segment reader memory (loaded for search)
  SegmentReaderStats segment_reader = 3;
}

// Indexing buffer memory breakdown
message IndexingBufferStats {
  uint64 total_bytes = 1;
  uint64 postings_bytes = 2;
  uint64 sparse_vectors_bytes = 3;
  uint64 dense_vectors_bytes = 4;
  uint64 interner_bytes = 5;
  uint64 position_index_bytes = 6;
  uint32 pending_docs = 7;
  uint32 unique_terms = 8;
}

// Segment-reader heap memory (search structures and decompressed caches).
// File-backed mappings and pinned residency are intentionally excluded:
// mapped file extent is not equivalent to resident memory.
message SegmentReaderStats {
  uint64 total_bytes = 1;
  uint64 term_dict_cache_bytes = 2;
  uint64 store_cache_bytes = 3;
  uint64 sparse_index_bytes = 4;
  uint64 dense_index_bytes = 5;
  uint32 num_segments_loaded = 6;
}

// Create index request/response
message CreateIndexRequest {
  string index_name = 1;
  // Schema definition
  string schema = 2;
}

message CreateIndexResponse {
  bool success = 1;
}

// Field entry for multi-value field support
message FieldEntry {
  string name = 1;
  FieldValue value = 2;
}

// Named document for batch indexing
// Uses repeated FieldEntry to support multi-value fields (same name, multiple values)
message NamedDocument {
  repeated FieldEntry fields = 1;
}

// Batch index documents request
message BatchIndexDocumentsRequest {
  string index_name = 1;
  repeated NamedDocument documents = 2;
}

message BatchIndexDocumentsResponse {
  uint32 indexed_count = 1;
  uint32 error_count = 2;
  repeated DocumentError errors = 3;
}

// Requests are bounded before conversion or writer admission. Both require a
// primary-key schema and an explicit Commit. No cross-partition transaction.
message DeleteDocumentsRequest {
  string index_name = 1;
  repeated string primary_keys = 2; // <=100,000 keys, <=8 MiB total key bytes
}

message UpsertDocumentsRequest {
  string index_name = 1;
  repeated NamedDocument documents = 2; // <=1,000 documents, <=32 MiB encoded request (<=200 MiB for one document)
}

message DocumentMutationResponse {
  // Admitted operations, not affected rows. A missing deletion is accepted.
  // Every input is accepted or has one error; Commit publishes accepted work.
  uint32 accepted_count = 1;
  repeated DocumentError errors = 2;
}

// Per-document error detail (e.g. duplicate primary key)
message DocumentError {
  uint32 index = 1;   // 0-based index in the request batch
  string error = 2;   // Human-readable error message
}

// Index document request/response
message IndexDocumentRequest {
  string index_name = 1;
  repeated FieldEntry fields = 2;
}

message IndexDocumentsResponse {
  uint32 indexed_count = 1;
  repeated DocumentError errors = 2;
}

// Commit request/response
message CommitRequest {
  string index_name = 1;
}

message CommitResponse {
  bool success = 1;
  uint32 num_docs = 2;
}

// Force merge request/response
message ForceMergeRequest {
  string index_name = 1;
  // Expensive physical removal of deleted rows, once per final output.
  // Omitted/false copies encoded data and carries deletion masks unchanged.
  bool compact = 2;
}

message ForceMergeResponse {
  bool success = 1;
  uint32 num_segments = 2;
}

// Delete index request/response
message DeleteIndexRequest {
  string index_name = 1;
}

message DeleteIndexResponse {
  bool success = 1;
}

// List indexes request/response
message ListIndexesRequest {}

message ListIndexesResponse {
  repeated string index_names = 1;
}

// Retrain vector index request/response
message RetrainVectorIndexRequest {
  string index_name = 1;
}

message RetrainVectorIndexResponse {
  bool success = 1;
}

message AlterVectorIndexRequest {
  string index_name = 1;
  string field_name = 2;
  // SDL field type/options, for example:
  // dense_vector<768, f16> [indexed<scann, tree_levels: 2, nprobe: 1024>]
  string field_schema = 3;
}

enum VectorIndexAlterState {
  VECTOR_INDEX_ALTER_STATE_UNSPECIFIED = 0;
  VECTOR_INDEX_ALTER_STATE_BUILT = 1;
  VECTOR_INDEX_ALTER_STATE_DEFERRED_FLAT = 2;
  VECTOR_INDEX_ALTER_STATE_PARAMETERS_ONLY = 3;
}

message AlterVectorIndexResponse {
  uint64 publication_generation = 1;
  VectorIndexAlterState state = 2;
}

// Reorder request/response
message ReorderRequest {
  string index_name = 1;
}

message ReorderResponse {
  bool success = 1;
  uint32 num_segments = 2;
}