feature: add MMR support

Author: densumesh

frontends/search/src/components/ResultsPage.tsx

@@ -251,6 +251,7 @@ const ResultsPage = (props: ResultsPageProps) => {
       if (!dataset) return;
 
       let sort_by;
+      let mmr;
 
       if (isSortBySearchType(props.search.debounced.sort_by)) {
         props.search.debounced.sort_by.rerank_type != ""
@@ -262,6 +263,12 @@ const ResultsPage = (props: ResultsPageProps) => {
           : (sort_by = undefined);
       }
 
+      if (!props.search.debounced.mmr.use_mmr) {
+        mmr = undefined;
+      } else {
+        mmr = props.search.debounced.mmr;
+      }
+
       const query =
         props.search.debounced.multiQueries.length > 0
           ? props.search.debounced.multiQueries
@@ -280,6 +287,7 @@ const ResultsPage = (props: ResultsPageProps) => {
         score_threshold: props.search.debounced.scoreThreshold,
         sort_options: {
           sort_by: sort_by,
+          mmr: mmr,
         },
         slim_chunks: props.search.debounced.slimChunks ?? false,
         page_size: props.search.debounced.pageSize ?? 10,

frontends/search/src/components/SearchForm.tsx

@@ -1193,6 +1193,46 @@ const SearchForm = (props: {
                             }}
                           />
                         </div>
+                        <div class="flex items-center justify-between space-x-2 p-1">
+                          <label>Use MMR:</label>
+                          <input
+                            class="h-4 w-4"
+                            type="checkbox"
+                            checked={tempSearchValues().mmr.use_mmr}
+                            onChange={(e) => {
+                              setTempSearchValues((prev) => {
+                                return {
+                                  ...prev,
+                                  mmr: {
+                                    ...prev.mmr,
+                                    use_mmr: e.target.checked,
+                                  },
+                                };
+                              });
+                            }}
+                          />
+                        </div>
+                        <div class="flex items-center justify-between space-x-2 p-1">
+                          <label>MMR Lambda:</label>
+                          <input
+                            class="w-16 rounded border border-neutral-400 p-0.5 text-black"
+                            type="number"
+                            value={tempSearchValues().mmr.mmr_lambda}
+                            onChange={(e) => {
+                              setTempSearchValues((prev) => {
+                                return {
+                                  ...prev,
+                                  mmr: {
+                                    ...prev.mmr,
+                                    mmr_lambda: parseFloat(
+                                      e.currentTarget.value,
+                                    ),
+                                  },
+                                };
+                              });
+                            }}
+                          />
+                        </div>
                         <div class="px-1 font-bold">Search Refinement</div>
                         <div class="flex items-center justify-between space-x-2 p-1">
                           <label>Use Quote Negated Words:</label>

frontends/search/src/hooks/useSearch.ts

@@ -68,6 +68,10 @@ export interface SearchOptions {
   prioritize_domain_specifc_words: boolean | null;
   disableOnWords: string[];
   sort_by: SortByField | SortBySearchType;
+  mmr: {
+    use_mmr: boolean;
+    mmr_lambda?: number;
+  };
   pageSize: number;
   getTotalPages: boolean;
   highlightResults: boolean;
@@ -98,6 +102,9 @@ const initalState: SearchOptions = {
   sort_by: {
     field: "",
   },
+  mmr: {
+    use_mmr: false,
+  },
   pageSize: 10,
   getTotalPages: true,
   correctTypos: false,
@@ -146,6 +153,7 @@ const fromStateToParams = (state: SearchOptions): Params => {
     oneTypoWordRangeMax: state.oneTypoWordRangeMax?.toString() ?? "6",
     twoTypoWordRangeMin: state.twoTypoWordRangeMin.toString(),
     twoTypoWordRangeMax: state.twoTypoWordRangeMax?.toString() ?? "",
+    mmr: JSON.stringify(state.mmr),
     prioritize_domain_specifc_words:
       state.prioritize_domain_specifc_words?.toString() ?? "",
     disableOnWords: state.disableOnWords.join(","),
@@ -189,6 +197,11 @@ const fromParamsToState = (
       initalState.sort_by,
     pageSize: parseInt(params.pageSize ?? "10"),
     getTotalPages: (params.getTotalPages ?? "true") === "true",
+    mmr:
+      (JSON.parse(params.mmr ?? "{}") as {
+        use_mmr: boolean;
+        mmr_lambda?: number;
+      }) ?? initalState.mmr,
     correctTypos: (params.correctTypos ?? "false") === "true",
     oneTypoWordRangeMin: parseInt(params.oneTypoWordRangeMin ?? "4"),
     oneTypoWordRangeMax: parseIntOrNull(params.oneTypoWordRangeMax),

server/src/data/models.rs

@@ -3520,6 +3520,7 @@ impl ApiKeyRequestParams {
             new_message_content: payload.new_message_content,
             topic_id: payload.topic_id,
             user_id: payload.user_id,
+            sort_options: payload.sort_options,
             highlight_options: self.highlight_options.or(payload.highlight_options),
             search_type: self.search_type.or(payload.search_type),
             use_group_search: payload.use_group_search,
@@ -6670,6 +6671,17 @@ pub struct SortOptions {
     pub use_weights: Option<bool>,
     /// Tag weights is a JSON object which can be used to boost the ranking of chunks with certain tags. This is useful for when you want to be able to bias towards chunks with a certain tag on the fly. The keys are the tag names and the values are the weights.
     pub tag_weights: Option<HashMap<String, f32>>,
+    /// Set use_mmr to true to use the Maximal Marginal Relevance algorithm to rerank the results. If not specified, this defaults to false.
+    pub mmr: Option<MmrOptions>,
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone, ToSchema, Default)]
+/// MMR Options lets you specify different methods to rerank the chunks in the result set. If not specified, this defaults to the score of the chunks.
+pub struct MmrOptions {
+    /// Set use_mmr to true to use the Maximal Marginal Relevance algorithm to rerank the results.
+    pub use_mmr: bool,
+    /// Set mmr_lambda to a value between 0.0 and 1.0 to control the tradeoff between relevance and diversity. Closer to 1.0 will give more diverse results, closer to 0.0 will give more relevant results. If not specified, this defaults to 0.5.
+    pub mmr_lambda: Option<f32>,
 }
 
 #[derive(Serialize, Deserialize, Debug, Clone, ToSchema, Default)]
@@ -6790,6 +6802,9 @@ fn extract_sort_highlight_options(
     if let Some(value) = other.remove("tag_weights") {
         sort_options.tag_weights = serde_json::from_value(value).ok();
     }
+    if let Some(value) = other.remove("mmr") {
+        sort_options.mmr = serde_json::from_value(value).ok();
+    }
 
     // Extract highlight options
     if let Some(value) = other.remove("highlight_results") {
@@ -6818,6 +6833,7 @@ fn extract_sort_highlight_options(
         && sort_options.location_bias.is_none()
         && sort_options.use_weights.is_none()
         && sort_options.tag_weights.is_none()
+        && sort_options.mmr.is_none()
     {
         None
     } else {
@@ -7143,6 +7159,7 @@ impl<'de> Deserialize<'de> for CreateMessageReqPayload {
             pub search_type: Option<SearchMethod>,
             pub concat_user_messages_query: Option<bool>,
             pub search_query: Option<String>,
+            pub sort_options: Option<SortOptions>,
             pub page_size: Option<u64>,
             pub filters: Option<ChunkFilter>,
             pub score_threshold: Option<f32>,
@@ -7172,6 +7189,7 @@ impl<'de> Deserialize<'de> for CreateMessageReqPayload {
             new_message_content: helper.new_message_content,
             topic_id: helper.topic_id,
             highlight_options,
+            sort_options: helper.sort_options,
             search_type: helper.search_type,
             use_group_search: helper.use_group_search,
             concat_user_messages_query: helper.concat_user_messages_query,
@@ -7198,6 +7216,8 @@ impl<'de> Deserialize<'de> for RegenerateMessageReqPayload {
             pub highlight_options: Option<HighlightOptions>,
             pub search_type: Option<SearchMethod>,
             pub concat_user_messages_query: Option<bool>,
+            pub sort_options: Option<SortOptions>,
+
             pub search_query: Option<String>,
             pub page_size: Option<u64>,
             pub filters: Option<ChunkFilter>,
@@ -7227,6 +7247,7 @@ impl<'de> Deserialize<'de> for RegenerateMessageReqPayload {
         Ok(RegenerateMessageReqPayload {
             topic_id: helper.topic_id,
             highlight_options,
+            sort_options: helper.sort_options,
             search_type: helper.search_type,
             concat_user_messages_query: helper.concat_user_messages_query,
             search_query: helper.search_query,
@@ -7254,6 +7275,8 @@ impl<'de> Deserialize<'de> for EditMessageReqPayload {
             pub new_message_content: String,
             pub highlight_options: Option<HighlightOptions>,
             pub search_type: Option<SearchMethod>,
+            pub sort_options: Option<SortOptions>,
+
             pub use_group_search: Option<bool>,
             pub concat_user_messages_query: Option<bool>,
             pub search_query: Option<String>,
@@ -7284,6 +7307,7 @@ impl<'de> Deserialize<'de> for EditMessageReqPayload {
         Ok(EditMessageReqPayload {
             topic_id: helper.topic_id,
             message_sort_order: helper.message_sort_order,
+            sort_options: helper.sort_options,
             new_message_content: helper.new_message_content,
             highlight_options,
             search_type: helper.search_type,

server/src/handlers/message_handler.rs

@@ -6,7 +6,7 @@ use crate::{
     data::models::{
         self, ChunkMetadata, ChunkMetadataStringTagSet, ChunkMetadataTypes, ContextOptions,
         DatasetAndOrgWithSubAndPlan, DatasetConfiguration, HighlightOptions, LLMOptions, Pool,
-        QdrantChunkMetadata, RedisPool, SearchMethod, SuggestType,
+        QdrantChunkMetadata, RedisPool, SearchMethod, SortOptions, SuggestType,
     },
     errors::ServiceError,
     get_env,
@@ -98,6 +98,8 @@ pub struct CreateMessageReqPayload {
     pub search_query: Option<String>,
     /// Page size is the number of chunks to fetch during RAG. If 0, then no search will be performed. If specified, this will override the N retrievals to include in the dataset configuration. Default is None.
     pub page_size: Option<u64>,
+    /// Sort Options lets you specify different methods to rerank the chunks in the result set. If not specified, this defaults to the score of the chunks.
+    pub sort_options: Option<SortOptions>,
     /// Filters is a JSON object which can be used to filter chunks. This is useful for when you want to filter chunks by arbitrary metadata. Unlike with tag filtering, there is a performance hit for filtering on metadata.
     pub filters: Option<ChunkFilter>,
     /// Set score_threshold to a float to filter out chunks with a score below the threshold. This threshold applies before weight and bias modifications. If not specified, this defaults to 0.0.
@@ -349,6 +351,8 @@ pub struct RegenerateMessageReqPayload {
     pub search_query: Option<String>,
     /// Page size is the number of chunks to fetch during RAG. If 0, then no search will be performed. If specified, this will override the N retrievals to include in the dataset configuration. Default is None.
     pub page_size: Option<u64>,
+    /// Sort Options lets you specify different methods to rerank the chunks in the result set. If not specified, this defaults to the score of the chunks.
+    pub sort_options: Option<SortOptions>,
     /// Filters is a JSON object which can be used to filter chunks. This is useful for when you want to filter chunks by arbitrary metadata. Unlike with tag filtering, there is a performance hit for filtering on metadata.
     pub filters: Option<ChunkFilter>,
     /// Set score_threshold to a float to filter out chunks with a score below the threshold. This threshold applies before weight and bias modifications. If not specified, this defaults to 0.0.
@@ -381,6 +385,8 @@ pub struct EditMessageReqPayload {
     pub concat_user_messages_query: Option<bool>,
     /// Query is the search query. This can be any string. The search_query will be used to create a dense embedding vector and/or sparse vector which will be used to find the result set. If not specified, will default to the last user message or HyDE if HyDE is enabled in the dataset configuration. Default is None.
     pub search_query: Option<String>,
+    /// Sort Options lets you specify different methods to rerank the chunks in the result set. If not specified, this defaults to the score of the chunks.
+    pub sort_options: Option<SortOptions>,
     /// Page size is the number of chunks to fetch during RAG. If 0, then no search will be performed. If specified, this will override the N retrievals to include in the dataset configuration. Default is None.
     pub page_size: Option<u64>,
     /// Filters is a JSON object which can be used to filter chunks. This is useful for when you want to filter chunks by arbitrary metadata. Unlike with tag filtering, there is a performance hit for filtering on metadata.
@@ -404,6 +410,7 @@ impl From<EditMessageReqPayload> for CreateMessageReqPayload {
             topic_id: data.topic_id,
             highlight_options: data.highlight_options,
             search_type: data.search_type,
+            sort_options: data.sort_options,
             use_group_search: data.use_group_search,
             concat_user_messages_query: data.concat_user_messages_query,
             search_query: data.search_query,
@@ -426,6 +433,7 @@ impl From<RegenerateMessageReqPayload> for CreateMessageReqPayload {
             highlight_options: data.highlight_options,
             search_type: data.search_type,
             use_group_search: data.use_group_search,
+            sort_options: data.sort_options,
             concat_user_messages_query: data.concat_user_messages_query,
             search_query: data.search_query,
             page_size: data.page_size,

server/src/lib.rs

@@ -495,6 +495,7 @@ impl Modify for SecurityAddon {
             data::models::OrganizationUsageCount,
             data::models::Dataset,
             data::models::DatasetAndUsage,
+            data::models::MmrOptions,
             data::models::DatasetUsageCount,
             data::models::DatasetDTO,
             data::models::DatasetUsageCount,

server/src/operators/chunk_operator.rs

@@ -138,7 +138,7 @@ pub struct ChunkMetadataWithQdrantId {
     pub qdrant_id: uuid::Uuid,
 }
 
-pub async fn get_chunk_metadatas_and_collided_chunks_from_point_ids_query(
+pub async fn get_chunk_metadatas_from_point_ids_query(
     point_ids: Vec<uuid::Uuid>,
     pool: web::Data<Pool>,
 ) -> Result<Vec<ChunkMetadataTypes>, ServiceError> {

server/src/operators/message_operator.rs

@@ -348,6 +348,7 @@ pub async fn get_rag_chunks_query(
                     .page_size
                     .unwrap_or(n_retrievals_to_include.try_into().unwrap_or(8)),
             ),
+            sort_options: create_message_req_payload.sort_options,
             highlight_options: create_message_req_payload.highlight_options,
             filters: create_message_req_payload.filters,
             group_size: Some(1),
@@ -453,6 +454,7 @@ pub async fn get_rag_chunks_query(
             search_type: search_type.clone(),
             query: QueryTypes::Single(query.clone()),
             score_threshold: create_message_req_payload.score_threshold,
+            sort_options: create_message_req_payload.sort_options,
             page_size: Some(
                 create_message_req_payload
                     .page_size