4769: Federated search r=ManyTheFish a=dureuill

# Pull Request

## Related issue
Fixes #4747 

[Usage](https://meilisearch.notion.site/v1-10-federated-search-698dfe36ab6b4668b044f735fb40f0b2)

## What does this PR do?
- multi-search now allows a top-level federation object. When not `null`, the results of multi-search are modified to be a single list of results rather than a list of a list of results
- changed lifetimes around tokenizer et al. to be able to make hits one by one rather than using a vector
- adds `roaring` to Meilisearch itself. As the federated search happens at the Meilisearch level (reuses the search functions declared at the Meilisearch level + merge happens after the hits were created), `RoaringBitmap`s are needed to track the candidates: hits that were seen,  all candidates.
- Refactor `make_hits` to allow for an individual, optimized `make_hit` 
- Score details comparison no longer fail when sorting on different field names or target point (for geo)

Co-authored-by: Louis Dureuil <louis@meilisearch.com>
This commit is contained in:
meili-bors[bot] 2024-07-16 08:14:46 +00:00 committed by GitHub
commit 1582c7e788
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
16 changed files with 5985 additions and 216 deletions

1
Cargo.lock generated
View File

@ -3394,6 +3394,7 @@ dependencies = [
"rayon",
"regex",
"reqwest",
"roaring",
"rustls 0.21.12",
"rustls-pemfile 1.0.4",
"segment",

View File

@ -192,6 +192,7 @@ merge_with_error_impl_take_error_message!(ParseOffsetDateTimeError);
merge_with_error_impl_take_error_message!(ParseTaskKindError);
merge_with_error_impl_take_error_message!(ParseTaskStatusError);
merge_with_error_impl_take_error_message!(IndexUidFormatError);
merge_with_error_impl_take_error_message!(InvalidMultiSearchWeight);
merge_with_error_impl_take_error_message!(InvalidSearchSemanticRatio);
merge_with_error_impl_take_error_message!(InvalidSearchRankingScoreThreshold);
merge_with_error_impl_take_error_message!(InvalidSimilarRankingScoreThreshold);

View File

@ -238,6 +238,11 @@ InvalidIndexLimit , InvalidRequest , BAD_REQUEST ;
InvalidIndexOffset , InvalidRequest , BAD_REQUEST ;
InvalidIndexPrimaryKey , InvalidRequest , BAD_REQUEST ;
InvalidIndexUid , InvalidRequest , BAD_REQUEST ;
InvalidMultiSearchFederated , InvalidRequest , BAD_REQUEST ;
InvalidMultiSearchFederationOptions , InvalidRequest , BAD_REQUEST ;
InvalidMultiSearchQueryPagination , InvalidRequest , BAD_REQUEST ;
InvalidMultiSearchQueryRankingRules , InvalidRequest , BAD_REQUEST ;
InvalidMultiSearchWeight , InvalidRequest , BAD_REQUEST ;
InvalidSearchAttributesToSearchOn , InvalidRequest , BAD_REQUEST ;
InvalidSearchAttributesToCrop , InvalidRequest , BAD_REQUEST ;
InvalidSearchAttributesToHighlight , InvalidRequest , BAD_REQUEST ;
@ -512,6 +517,12 @@ impl fmt::Display for deserr_codes::InvalidSearchSemanticRatio {
}
}
impl fmt::Display for deserr_codes::InvalidMultiSearchWeight {
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
write!(f, "the value of `weight` is invalid, expected a positive float (>= 0.0).")
}
}
impl fmt::Display for deserr_codes::InvalidSimilarId {
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
write!(

View File

@ -102,6 +102,7 @@ tracing-subscriber = { version = "0.3.18", features = ["json"] }
tracing-trace = { version = "0.1.0", path = "../tracing-trace" }
tracing-actix-web = "0.7.11"
build-info = { version = "1.7.0", path = "../build-info" }
roaring = "0.10.2"
[dev-dependencies]
actix-rt = "2.10.0"

View File

@ -42,7 +42,7 @@ pub struct MultiSearchAggregator;
#[allow(dead_code)]
impl MultiSearchAggregator {
pub fn from_queries(_: &dyn Any, _: &dyn Any) -> Self {
pub fn from_federated_search(_: &dyn Any, _: &dyn Any) -> Self {
Self
}

View File

@ -34,8 +34,8 @@ use crate::routes::indexes::documents::{DocumentEditionByFunction, UpdateDocumen
use crate::routes::indexes::facet_search::FacetSearchQuery;
use crate::routes::{create_all_stats, Stats};
use crate::search::{
FacetSearchResult, MatchingStrategy, SearchQuery, SearchQueryWithIndex, SearchResult,
SimilarQuery, SimilarResult, DEFAULT_CROP_LENGTH, DEFAULT_CROP_MARKER,
FacetSearchResult, FederatedSearch, MatchingStrategy, SearchQuery, SearchQueryWithIndex,
SearchResult, SimilarQuery, SimilarResult, DEFAULT_CROP_LENGTH, DEFAULT_CROP_MARKER,
DEFAULT_HIGHLIGHT_POST_TAG, DEFAULT_HIGHLIGHT_PRE_TAG, DEFAULT_SEARCH_LIMIT,
DEFAULT_SEMANTIC_RATIO,
};
@ -1095,22 +1095,33 @@ pub struct MultiSearchAggregator {
show_ranking_score: bool,
show_ranking_score_details: bool,
// federation
use_federation: bool,
// context
user_agents: HashSet<String>,
}
impl MultiSearchAggregator {
pub fn from_queries(query: &[SearchQueryWithIndex], request: &HttpRequest) -> Self {
pub fn from_federated_search(
federated_search: &FederatedSearch,
request: &HttpRequest,
) -> Self {
let timestamp = Some(OffsetDateTime::now_utc());
let user_agents = extract_user_agents(request).into_iter().collect();
let distinct_indexes: HashSet<_> = query
let use_federation = federated_search.federation.is_some();
let distinct_indexes: HashSet<_> = federated_search
.queries
.iter()
.map(|query| {
let query = &query;
// make sure we get a compilation error if a field gets added to / removed from SearchQueryWithIndex
let SearchQueryWithIndex {
index_uid,
federation_options: _,
q: _,
vector: _,
offset: _,
@ -1142,8 +1153,10 @@ impl MultiSearchAggregator {
})
.collect();
let show_ranking_score = query.iter().any(|query| query.show_ranking_score);
let show_ranking_score_details = query.iter().any(|query| query.show_ranking_score_details);
let show_ranking_score =
federated_search.queries.iter().any(|query| query.show_ranking_score);
let show_ranking_score_details =
federated_search.queries.iter().any(|query| query.show_ranking_score_details);
Self {
timestamp,
@ -1151,10 +1164,11 @@ impl MultiSearchAggregator {
total_succeeded: 0,
total_distinct_index_count: distinct_indexes.len(),
total_single_index: if distinct_indexes.len() == 1 { 1 } else { 0 },
total_search_count: query.len(),
total_search_count: federated_search.queries.len(),
show_ranking_score,
show_ranking_score_details,
user_agents,
use_federation,
}
}
@ -1180,6 +1194,7 @@ impl MultiSearchAggregator {
let show_ranking_score_details =
this.show_ranking_score_details || other.show_ranking_score_details;
let mut user_agents = this.user_agents;
let use_federation = this.use_federation || other.use_federation;
for user_agent in other.user_agents.into_iter() {
user_agents.insert(user_agent);
@ -1196,6 +1211,7 @@ impl MultiSearchAggregator {
user_agents,
show_ranking_score,
show_ranking_score_details,
use_federation,
// do not add _ or ..Default::default() here
};
@ -1214,6 +1230,7 @@ impl MultiSearchAggregator {
user_agents,
show_ranking_score,
show_ranking_score_details,
use_federation,
} = self;
if total_received == 0 {
@ -1238,6 +1255,9 @@ impl MultiSearchAggregator {
"scoring": {
"show_ranking_score": show_ranking_score,
"show_ranking_score_details": show_ranking_score_details,
},
"federation": {
"use_federation": use_federation,
}
});

View File

@ -25,6 +25,10 @@ pub enum MeilisearchHttpError {
DocumentNotFound(String),
#[error("Sending an empty filter is forbidden.")]
EmptyFilter,
#[error("Using `federationOptions` is not allowed in a non-federated search.\n Hint: remove `federationOptions` from query #{0} or add `federation: {{}}` to the request.")]
FederationOptionsInNonFederatedRequest(usize),
#[error("Inside `.queries[{0}]`: Using pagination options is not allowed in federated queries.\n Hint: remove `{1}` from query #{0} or remove `federation: {{}}` from the request")]
PaginationInFederatedQuery(usize, &'static str),
#[error("Invalid syntax for the filter parameter: `expected {}, found: {1}`.", .0.join(", "))]
InvalidExpression(&'static [&'static str], Value),
#[error("A {0} payload is missing.")]
@ -86,6 +90,12 @@ impl ErrorCode for MeilisearchHttpError {
MeilisearchHttpError::DocumentFormat(e) => e.error_code(),
MeilisearchHttpError::Join(_) => Code::Internal,
MeilisearchHttpError::MissingSearchHybrid => Code::MissingSearchHybrid,
MeilisearchHttpError::FederationOptionsInNonFederatedRequest(_) => {
Code::InvalidMultiSearchFederationOptions
}
MeilisearchHttpError::PaginationInFederatedQuery(_, _) => {
Code::InvalidMultiSearchQueryPagination
}
}
}
}

View File

@ -10,12 +10,14 @@ use serde::Serialize;
use tracing::debug;
use crate::analytics::{Analytics, MultiSearchAggregator};
use crate::error::MeilisearchHttpError;
use crate::extractors::authentication::policies::ActionPolicy;
use crate::extractors::authentication::{AuthenticationError, GuardedData};
use crate::extractors::sequential_extractor::SeqHandler;
use crate::routes::indexes::search::search_kind;
use crate::search::{
add_search_rules, perform_search, RetrieveVectors, SearchQueryWithIndex, SearchResultWithIndex,
add_search_rules, perform_federated_search, perform_search, FederatedSearch, RetrieveVectors,
SearchQueryWithIndex, SearchResultWithIndex,
};
use crate::search_queue::SearchQueue;
@ -28,85 +30,44 @@ struct SearchResults {
results: Vec<SearchResultWithIndex>,
}
#[derive(Debug, deserr::Deserr)]
#[deserr(error = DeserrJsonError, rename_all = camelCase, deny_unknown_fields)]
pub struct SearchQueries {
queries: Vec<SearchQueryWithIndex>,
}
pub async fn multi_search_with_post(
index_scheduler: GuardedData<ActionPolicy<{ actions::SEARCH }>, Data<IndexScheduler>>,
search_queue: Data<SearchQueue>,
params: AwebJson<SearchQueries, DeserrJsonError>,
params: AwebJson<FederatedSearch, DeserrJsonError>,
req: HttpRequest,
analytics: web::Data<dyn Analytics>,
) -> Result<HttpResponse, ResponseError> {
let queries = params.into_inner().queries;
let mut multi_aggregate = MultiSearchAggregator::from_queries(&queries, &req);
let features = index_scheduler.features();
// Since we don't want to process half of the search requests and then get a permit refused
// we're going to get one permit for the whole duration of the multi-search request.
let _permit = search_queue.try_get_search_permit().await?;
// Explicitly expect a `(ResponseError, usize)` for the error type rather than `ResponseError` only,
// so that `?` doesn't work if it doesn't use `with_index`, ensuring that it is not forgotten in case of code
// changes.
let search_results: Result<_, (ResponseError, usize)> = async {
let mut search_results = Vec::with_capacity(queries.len());
for (query_index, (index_uid, mut query)) in
queries.into_iter().map(SearchQueryWithIndex::into_index_query).enumerate()
{
debug!(on_index = query_index, parameters = ?query, "Multi-search");
let federated_search = params.into_inner();
let mut multi_aggregate = MultiSearchAggregator::from_federated_search(&federated_search, &req);
let FederatedSearch { mut queries, federation } = federated_search;
let features = index_scheduler.features();
// regardless of federation, check authorization and apply search rules
let auth = 'check_authorization: {
for (query_index, federated_query) in queries.iter_mut().enumerate() {
let index_uid = federated_query.index_uid.as_str();
// Check index from API key
if !index_scheduler.filters().is_index_authorized(&index_uid) {
return Err(AuthenticationError::InvalidToken).with_index(query_index);
if !index_scheduler.filters().is_index_authorized(index_uid) {
break 'check_authorization Err(AuthenticationError::InvalidToken)
.with_index(query_index);
}
// Apply search rules from tenant token
if let Some(search_rules) = index_scheduler.filters().get_index_search_rules(&index_uid)
if let Some(search_rules) = index_scheduler.filters().get_index_search_rules(index_uid)
{
add_search_rules(&mut query.filter, search_rules);
add_search_rules(&mut federated_query.filter, search_rules);
}
let index = index_scheduler
.index(&index_uid)
.map_err(|err| {
let mut err = ResponseError::from(err);
// Patch the HTTP status code to 400 as it defaults to 404 for `index_not_found`, but
// here the resource not found is not part of the URL.
err.code = StatusCode::BAD_REQUEST;
err
})
.with_index(query_index)?;
let search_kind = search_kind(&query, index_scheduler.get_ref(), &index, features)
.with_index(query_index)?;
let retrieve_vector =
RetrieveVectors::new(query.retrieve_vectors, features).with_index(query_index)?;
let search_result = tokio::task::spawn_blocking(move || {
perform_search(&index, query, search_kind, retrieve_vector)
})
.await
.with_index(query_index)?;
search_results.push(SearchResultWithIndex {
index_uid: index_uid.into_inner(),
result: search_result.with_index(query_index)?,
});
}
Ok(search_results)
}
.await;
Ok(())
};
if search_results.is_ok() {
multi_aggregate.succeed();
}
analytics.post_multi_search(multi_aggregate);
let search_results = search_results.map_err(|(mut err, query_index)| {
auth.map_err(|(mut err, query_index)| {
// Add the query index that failed as context for the error message.
// We're doing it only here and not directly in the `WithIndex` trait so that the `with_index` function returns a different type
// of result and we can benefit from static typing.
@ -114,9 +75,95 @@ pub async fn multi_search_with_post(
err
})?;
debug!(returns = ?search_results, "Multi-search");
let response = match federation {
Some(federation) => {
let search_result = tokio::task::spawn_blocking(move || {
perform_federated_search(&index_scheduler, queries, federation, features)
})
.await;
Ok(HttpResponse::Ok().json(SearchResults { results: search_results }))
if let Ok(Ok(_)) = search_result {
multi_aggregate.succeed();
}
analytics.post_multi_search(multi_aggregate);
HttpResponse::Ok().json(search_result??)
}
None => {
// Explicitly expect a `(ResponseError, usize)` for the error type rather than `ResponseError` only,
// so that `?` doesn't work if it doesn't use `with_index`, ensuring that it is not forgotten in case of code
// changes.
let search_results: Result<_, (ResponseError, usize)> = async {
let mut search_results = Vec::with_capacity(queries.len());
for (query_index, (index_uid, query, federation_options)) in queries
.into_iter()
.map(SearchQueryWithIndex::into_index_query_federation)
.enumerate()
{
debug!(on_index = query_index, parameters = ?query, "Multi-search");
if federation_options.is_some() {
return Err((
MeilisearchHttpError::FederationOptionsInNonFederatedRequest(
query_index,
)
.into(),
query_index,
));
}
let index = index_scheduler
.index(&index_uid)
.map_err(|err| {
let mut err = ResponseError::from(err);
// Patch the HTTP status code to 400 as it defaults to 404 for `index_not_found`, but
// here the resource not found is not part of the URL.
err.code = StatusCode::BAD_REQUEST;
err
})
.with_index(query_index)?;
let search_kind =
search_kind(&query, index_scheduler.get_ref(), &index, features)
.with_index(query_index)?;
let retrieve_vector = RetrieveVectors::new(query.retrieve_vectors, features)
.with_index(query_index)?;
let search_result = tokio::task::spawn_blocking(move || {
perform_search(&index, query, search_kind, retrieve_vector)
})
.await
.with_index(query_index)?;
search_results.push(SearchResultWithIndex {
index_uid: index_uid.into_inner(),
result: search_result.with_index(query_index)?,
});
}
Ok(search_results)
}
.await;
if search_results.is_ok() {
multi_aggregate.succeed();
}
analytics.post_multi_search(multi_aggregate);
let search_results = search_results.map_err(|(mut err, query_index)| {
// Add the query index that failed as context for the error message.
// We're doing it only here and not directly in the `WithIndex` trait so that the `with_index` function returns a different type
// of result and we can benefit from static typing.
err.message = format!("Inside `.queries[{query_index}]`: {}", err.message);
err
})?;
debug!(returns = ?search_results, "Multi-search");
HttpResponse::Ok().json(SearchResults { results: search_results })
}
};
Ok(response)
}
/// Local `Result` extension trait to avoid `map_err` boilerplate.

View File

@ -0,0 +1,629 @@
use std::cmp::Ordering;
use std::collections::BTreeMap;
use std::fmt;
use std::iter::Zip;
use std::rc::Rc;
use std::str::FromStr as _;
use std::time::Duration;
use std::vec::{IntoIter, Vec};
use actix_http::StatusCode;
use index_scheduler::{IndexScheduler, RoFeatures};
use meilisearch_types::deserr::DeserrJsonError;
use meilisearch_types::error::deserr_codes::{
InvalidMultiSearchWeight, InvalidSearchLimit, InvalidSearchOffset,
};
use meilisearch_types::error::ResponseError;
use meilisearch_types::milli::score_details::{ScoreDetails, ScoreValue};
use meilisearch_types::milli::{self, DocumentId, TimeBudget};
use roaring::RoaringBitmap;
use serde::Serialize;
use super::ranking_rules::{self, RankingRules};
use super::{
prepare_search, AttributesFormat, HitMaker, HitsInfo, RetrieveVectors, SearchHit, SearchKind,
SearchQuery, SearchQueryWithIndex,
};
use crate::error::MeilisearchHttpError;
use crate::routes::indexes::search::search_kind;
pub const DEFAULT_FEDERATED_WEIGHT: f64 = 1.0;
#[derive(Debug, Default, Clone, Copy, PartialEq, deserr::Deserr)]
#[deserr(error = DeserrJsonError, rename_all = camelCase, deny_unknown_fields)]
pub struct FederationOptions {
#[deserr(default, error = DeserrJsonError<InvalidMultiSearchWeight>)]
pub weight: Weight,
}
#[derive(Debug, Clone, Copy, PartialEq, deserr::Deserr)]
#[deserr(try_from(f64) = TryFrom::try_from -> InvalidMultiSearchWeight)]
pub struct Weight(f64);
impl Default for Weight {
fn default() -> Self {
Weight(DEFAULT_FEDERATED_WEIGHT)
}
}
impl std::convert::TryFrom<f64> for Weight {
type Error = InvalidMultiSearchWeight;
fn try_from(f: f64) -> Result<Self, Self::Error> {
if f < 0.0 {
Err(InvalidMultiSearchWeight)
} else {
Ok(Weight(f))
}
}
}
impl std::ops::Deref for Weight {
type Target = f64;
fn deref(&self) -> &Self::Target {
&self.0
}
}
#[derive(Debug, deserr::Deserr)]
#[deserr(error = DeserrJsonError, rename_all = camelCase, deny_unknown_fields)]
pub struct Federation {
#[deserr(default = super::DEFAULT_SEARCH_LIMIT(), error = DeserrJsonError<InvalidSearchLimit>)]
pub limit: usize,
#[deserr(default = super::DEFAULT_SEARCH_OFFSET(), error = DeserrJsonError<InvalidSearchOffset>)]
pub offset: usize,
}
#[derive(Debug, deserr::Deserr)]
#[deserr(error = DeserrJsonError, rename_all = camelCase, deny_unknown_fields)]
pub struct FederatedSearch {
pub queries: Vec<SearchQueryWithIndex>,
#[deserr(default)]
pub federation: Option<Federation>,
}
#[derive(Serialize, Clone, PartialEq)]
#[serde(rename_all = "camelCase")]
pub struct FederatedSearchResult {
pub hits: Vec<SearchHit>,
pub processing_time_ms: u128,
#[serde(flatten)]
pub hits_info: HitsInfo,
#[serde(skip_serializing_if = "Option::is_none")]
pub semantic_hit_count: Option<u32>,
// These fields are only used for analytics purposes
#[serde(skip)]
pub degraded: bool,
#[serde(skip)]
pub used_negative_operator: bool,
}
impl fmt::Debug for FederatedSearchResult {
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
let FederatedSearchResult {
hits,
processing_time_ms,
hits_info,
semantic_hit_count,
degraded,
used_negative_operator,
} = self;
let mut debug = f.debug_struct("SearchResult");
// The most important thing when looking at a search result is the time it took to process
debug.field("processing_time_ms", &processing_time_ms);
debug.field("hits", &format!("[{} hits returned]", hits.len()));
debug.field("hits_info", &hits_info);
if *used_negative_operator {
debug.field("used_negative_operator", used_negative_operator);
}
if *degraded {
debug.field("degraded", degraded);
}
if let Some(semantic_hit_count) = semantic_hit_count {
debug.field("semantic_hit_count", &semantic_hit_count);
}
debug.finish()
}
}
struct WeightedScore<'a> {
details: &'a [ScoreDetails],
weight: f64,
}
impl<'a> WeightedScore<'a> {
pub fn new(details: &'a [ScoreDetails], weight: f64) -> Self {
Self { details, weight }
}
pub fn weighted_global_score(&self) -> f64 {
ScoreDetails::global_score(self.details.iter()) * self.weight
}
pub fn compare_weighted_global_scores(&self, other: &Self) -> Ordering {
self.weighted_global_score()
.partial_cmp(&other.weighted_global_score())
// both are numbers, possibly infinite
.unwrap()
}
pub fn compare(&self, other: &Self) -> Ordering {
let mut left_it = ScoreDetails::score_values(self.details.iter());
let mut right_it = ScoreDetails::score_values(other.details.iter());
loop {
let left = left_it.next();
let right = right_it.next();
match (left, right) {
(None, None) => return Ordering::Equal,
(None, Some(_)) => return Ordering::Less,
(Some(_), None) => return Ordering::Greater,
(Some(ScoreValue::Score(left)), Some(ScoreValue::Score(right))) => {
let left = left * self.weight;
let right = right * other.weight;
if (left - right).abs() <= f64::EPSILON {
continue;
}
return left.partial_cmp(&right).unwrap();
}
(Some(ScoreValue::Sort(left)), Some(ScoreValue::Sort(right))) => {
match left.partial_cmp(right) {
Some(Ordering::Equal) => continue,
Some(order) => return order,
None => return self.compare_weighted_global_scores(other),
}
}
(Some(ScoreValue::GeoSort(left)), Some(ScoreValue::GeoSort(right))) => {
match left.partial_cmp(right) {
Some(Ordering::Equal) => continue,
Some(order) => return order,
None => {
return self.compare_weighted_global_scores(other);
}
}
}
// not comparable details, use global
(Some(ScoreValue::Score(_)), Some(_))
| (Some(_), Some(ScoreValue::Score(_)))
| (Some(ScoreValue::GeoSort(_)), Some(ScoreValue::Sort(_)))
| (Some(ScoreValue::Sort(_)), Some(ScoreValue::GeoSort(_))) => {
let left_count = left_it.count();
let right_count = right_it.count();
// compare how many remaining groups of rules each side has.
// the group with the most remaining groups wins.
return left_count
.cmp(&right_count)
// breaks ties with the global ranking score
.then_with(|| self.compare_weighted_global_scores(other));
}
}
}
}
}
struct QueryByIndex {
query: SearchQuery,
federation_options: FederationOptions,
query_index: usize,
}
struct SearchResultByQuery<'a> {
documents_ids: Vec<DocumentId>,
document_scores: Vec<Vec<ScoreDetails>>,
federation_options: FederationOptions,
hit_maker: HitMaker<'a>,
query_index: usize,
}
struct SearchResultByQueryIter<'a> {
it: Zip<IntoIter<DocumentId>, IntoIter<Vec<ScoreDetails>>>,
federation_options: FederationOptions,
hit_maker: Rc<HitMaker<'a>>,
query_index: usize,
}
impl<'a> SearchResultByQueryIter<'a> {
fn new(
SearchResultByQuery {
documents_ids,
document_scores,
federation_options,
hit_maker,
query_index,
}: SearchResultByQuery<'a>,
) -> Self {
let it = documents_ids.into_iter().zip(document_scores);
Self { it, federation_options, hit_maker: Rc::new(hit_maker), query_index }
}
}
struct SearchResultByQueryIterItem<'a> {
docid: DocumentId,
score: Vec<ScoreDetails>,
federation_options: FederationOptions,
hit_maker: Rc<HitMaker<'a>>,
query_index: usize,
}
fn merge_index_local_results(
results_by_query: Vec<SearchResultByQuery<'_>>,
) -> impl Iterator<Item = SearchResultByQueryIterItem> + '_ {
itertools::kmerge_by(
results_by_query.into_iter().map(SearchResultByQueryIter::new),
|left: &SearchResultByQueryIterItem, right: &SearchResultByQueryIterItem| {
let left_score = WeightedScore::new(&left.score, *left.federation_options.weight);
let right_score = WeightedScore::new(&right.score, *right.federation_options.weight);
match left_score.compare(&right_score) {
// the biggest score goes first
Ordering::Greater => true,
// break ties using query index
Ordering::Equal => left.query_index < right.query_index,
Ordering::Less => false,
}
},
)
}
fn merge_index_global_results(
results_by_index: Vec<SearchResultByIndex>,
) -> impl Iterator<Item = SearchHitByIndex> {
itertools::kmerge_by(
results_by_index.into_iter().map(|result_by_index| result_by_index.hits.into_iter()),
|left: &SearchHitByIndex, right: &SearchHitByIndex| {
let left_score = WeightedScore::new(&left.score, *left.federation_options.weight);
let right_score = WeightedScore::new(&right.score, *right.federation_options.weight);
match left_score.compare(&right_score) {
// the biggest score goes first
Ordering::Greater => true,
// break ties using query index
Ordering::Equal => left.query_index < right.query_index,
Ordering::Less => false,
}
},
)
}
impl<'a> Iterator for SearchResultByQueryIter<'a> {
type Item = SearchResultByQueryIterItem<'a>;
fn next(&mut self) -> Option<Self::Item> {
let (docid, score) = self.it.next()?;
Some(SearchResultByQueryIterItem {
docid,
score,
federation_options: self.federation_options,
hit_maker: Rc::clone(&self.hit_maker),
query_index: self.query_index,
})
}
}
struct SearchHitByIndex {
hit: SearchHit,
score: Vec<ScoreDetails>,
federation_options: FederationOptions,
query_index: usize,
}
struct SearchResultByIndex {
hits: Vec<SearchHitByIndex>,
candidates: RoaringBitmap,
degraded: bool,
used_negative_operator: bool,
}
pub fn perform_federated_search(
index_scheduler: &IndexScheduler,
queries: Vec<SearchQueryWithIndex>,
federation: Federation,
features: RoFeatures,
) -> Result<FederatedSearchResult, ResponseError> {
let before_search = std::time::Instant::now();
// this implementation partition the queries by index to guarantee an important property:
// - all the queries to a particular index use the same read transaction.
// This is an important property, otherwise we cannot guarantee the self-consistency of the results.
// 1. partition queries by index
let mut queries_by_index: BTreeMap<String, Vec<QueryByIndex>> = Default::default();
for (query_index, federated_query) in queries.into_iter().enumerate() {
if let Some(pagination_field) = federated_query.has_pagination() {
return Err(MeilisearchHttpError::PaginationInFederatedQuery(
query_index,
pagination_field,
)
.into());
}
let (index_uid, query, federation_options) = federated_query.into_index_query_federation();
queries_by_index.entry(index_uid.into_inner()).or_default().push(QueryByIndex {
query,
federation_options: federation_options.unwrap_or_default(),
query_index,
})
}
// 2. perform queries, merge and make hits index by index
let required_hit_count = federation.limit + federation.offset;
// In step (2), semantic_hit_count will be set to Some(0) if any search kind uses semantic
// Then in step (3), we'll update its value if there is any semantic search
let mut semantic_hit_count = None;
let mut results_by_index = Vec::with_capacity(queries_by_index.len());
let mut previous_query_data: Option<(RankingRules, usize, String)> = None;
for (index_uid, queries) in queries_by_index {
let index = match index_scheduler.index(&index_uid) {
Ok(index) => index,
Err(err) => {
let mut err = ResponseError::from(err);
// Patch the HTTP status code to 400 as it defaults to 404 for `index_not_found`, but
// here the resource not found is not part of the URL.
err.code = StatusCode::BAD_REQUEST;
if let Some(query) = queries.first() {
err.message =
format!("Inside `.queries[{}]`: {}", query.query_index, err.message);
}
return Err(err);
}
};
// Important: this is the only transaction we'll use for this index during this federated search
let rtxn = index.read_txn()?;
let criteria = index.criteria(&rtxn)?;
// stuff we need for the hitmaker
let script_lang_map = index.script_language(&rtxn)?;
let dictionary = index.dictionary(&rtxn)?;
let dictionary: Option<Vec<_>> =
dictionary.as_ref().map(|x| x.iter().map(String::as_str).collect());
let separators = index.allowed_separators(&rtxn)?;
let separators: Option<Vec<_>> =
separators.as_ref().map(|x| x.iter().map(String::as_str).collect());
// each query gets its individual cutoff
let cutoff = index.search_cutoff(&rtxn)?;
let mut degraded = false;
let mut used_negative_operator = false;
let mut candidates = RoaringBitmap::new();
// 2.1. Compute all candidates for each query in the index
let mut results_by_query = Vec::with_capacity(queries.len());
for QueryByIndex { query, federation_options, query_index } in queries {
// use an immediately invoked lambda to capture the result without returning from the function
let res: Result<(), ResponseError> = (|| {
let search_kind = search_kind(&query, index_scheduler, &index, features)?;
let canonicalization_kind = match (&search_kind, &query.q) {
(SearchKind::SemanticOnly { .. }, _) => {
ranking_rules::CanonicalizationKind::Vector
}
(_, Some(q)) if !q.is_empty() => ranking_rules::CanonicalizationKind::Keyword,
_ => ranking_rules::CanonicalizationKind::Placeholder,
};
let sort = if let Some(sort) = &query.sort {
let sorts: Vec<_> =
match sort.iter().map(|s| milli::AscDesc::from_str(s)).collect() {
Ok(sorts) => sorts,
Err(asc_desc_error) => {
return Err(milli::Error::from(milli::SortError::from(
asc_desc_error,
))
.into())
}
};
Some(sorts)
} else {
None
};
let ranking_rules = ranking_rules::RankingRules::new(
criteria.clone(),
sort,
query.matching_strategy.into(),
canonicalization_kind,
);
if let Some((previous_ranking_rules, previous_query_index, previous_index_uid)) =
previous_query_data.take()
{
if let Err(error) = ranking_rules.is_compatible_with(&previous_ranking_rules) {
return Err(error.to_response_error(
&ranking_rules,
&previous_ranking_rules,
query_index,
previous_query_index,
&index_uid,
&previous_index_uid,
));
}
previous_query_data = if previous_ranking_rules.constraint_count()
> ranking_rules.constraint_count()
{
Some((previous_ranking_rules, previous_query_index, previous_index_uid))
} else {
Some((ranking_rules, query_index, index_uid.clone()))
};
} else {
previous_query_data = Some((ranking_rules, query_index, index_uid.clone()));
}
match search_kind {
SearchKind::KeywordOnly => {}
_ => semantic_hit_count = Some(0),
}
let retrieve_vectors = RetrieveVectors::new(query.retrieve_vectors, features)?;
let time_budget = match cutoff {
Some(cutoff) => TimeBudget::new(Duration::from_millis(cutoff)),
None => TimeBudget::default(),
};
let (mut search, _is_finite_pagination, _max_total_hits, _offset) =
prepare_search(&index, &rtxn, &query, &search_kind, time_budget)?;
search.scoring_strategy(milli::score_details::ScoringStrategy::Detailed);
search.offset(0);
search.limit(required_hit_count);
let (result, _semantic_hit_count) = super::search_from_kind(search_kind, search)?;
let format = AttributesFormat {
attributes_to_retrieve: query.attributes_to_retrieve,
retrieve_vectors,
attributes_to_highlight: query.attributes_to_highlight,
attributes_to_crop: query.attributes_to_crop,
crop_length: query.crop_length,
crop_marker: query.crop_marker,
highlight_pre_tag: query.highlight_pre_tag,
highlight_post_tag: query.highlight_post_tag,
show_matches_position: query.show_matches_position,
sort: query.sort,
show_ranking_score: query.show_ranking_score,
show_ranking_score_details: query.show_ranking_score_details,
};
let milli::SearchResult {
matching_words,
candidates: query_candidates,
documents_ids,
document_scores,
degraded: query_degraded,
used_negative_operator: query_used_negative_operator,
} = result;
candidates |= query_candidates;
degraded |= query_degraded;
used_negative_operator |= query_used_negative_operator;
let tokenizer = HitMaker::tokenizer(
&script_lang_map,
dictionary.as_deref(),
separators.as_deref(),
);
let formatter_builder = HitMaker::formatter_builder(matching_words, tokenizer);
let hit_maker = HitMaker::new(&index, &rtxn, format, formatter_builder)?;
results_by_query.push(SearchResultByQuery {
federation_options,
hit_maker,
query_index,
documents_ids,
document_scores,
});
Ok(())
})();
if let Err(mut error) = res {
error.message = format!("Inside `.queries[{query_index}]`: {}", error.message);
return Err(error);
}
}
// 2.2. merge inside index
let mut documents_seen = RoaringBitmap::new();
let merged_result: Result<Vec<_>, ResponseError> =
merge_index_local_results(results_by_query)
// skip documents we've already seen & mark that we saw the current document
.filter(|SearchResultByQueryIterItem { docid, .. }| documents_seen.insert(*docid))
.take(required_hit_count)
// 2.3 make hits
.map(
|SearchResultByQueryIterItem {
docid,
score,
federation_options,
hit_maker,
query_index,
}| {
let mut hit = hit_maker.make_hit(docid, &score)?;
let weighted_score =
ScoreDetails::global_score(score.iter()) * (*federation_options.weight);
let _federation = serde_json::json!(
{
"indexUid": index_uid,
"queriesPosition": query_index,
"weightedRankingScore": weighted_score,
}
);
hit.document.insert("_federation".to_string(), _federation);
Ok(SearchHitByIndex { hit, score, federation_options, query_index })
},
)
.collect();
let merged_result = merged_result?;
results_by_index.push(SearchResultByIndex {
hits: merged_result,
candidates,
degraded,
used_negative_operator,
});
}
// 3. merge hits and metadata across indexes
// 3.1 merge metadata
let (estimated_total_hits, degraded, used_negative_operator) = {
let mut estimated_total_hits = 0;
let mut degraded = false;
let mut used_negative_operator = false;
for SearchResultByIndex {
hits: _,
candidates,
degraded: degraded_by_index,
used_negative_operator: used_negative_operator_by_index,
} in &results_by_index
{
estimated_total_hits += candidates.len() as usize;
degraded |= *degraded_by_index;
used_negative_operator |= *used_negative_operator_by_index;
}
(estimated_total_hits, degraded, used_negative_operator)
};
// 3.2 merge hits
let merged_hits: Vec<_> = merge_index_global_results(results_by_index)
.skip(federation.offset)
.take(federation.limit)
.inspect(|hit| {
if let Some(semantic_hit_count) = &mut semantic_hit_count {
if hit.score.iter().any(|score| matches!(&score, ScoreDetails::Vector(_))) {
*semantic_hit_count += 1;
}
}
})
.map(|hit| hit.hit)
.collect();
let search_result = FederatedSearchResult {
hits: merged_hits,
processing_time_ms: before_search.elapsed().as_millis(),
hits_info: HitsInfo::OffsetLimit {
limit: federation.limit,
offset: federation.offset,
estimated_total_hits,
},
semantic_hit_count,
degraded,
used_negative_operator,
};
Ok(search_result)
}

View File

@ -1,6 +1,6 @@
use core::fmt;
use std::cmp::min;
use std::collections::{BTreeMap, BTreeSet, HashSet};
use std::collections::{BTreeMap, BTreeSet, HashMap, HashSet};
use std::str::FromStr;
use std::sync::Arc;
use std::time::{Duration, Instant};
@ -31,6 +31,11 @@ use serde_json::{json, Value};
use crate::error::MeilisearchHttpError;
mod federated;
pub use federated::{perform_federated_search, FederatedSearch, Federation, FederationOptions};
mod ranking_rules;
type MatchesPosition = BTreeMap<String, Vec<MatchBounds>>;
pub const DEFAULT_SEARCH_OFFSET: fn() -> usize = || 0;
@ -257,11 +262,13 @@ pub struct HybridQuery {
pub embedder: Option<String>,
}
#[derive(Clone)]
pub enum SearchKind {
KeywordOnly,
SemanticOnly { embedder_name: String, embedder: Arc<Embedder> },
Hybrid { embedder_name: String, embedder: Arc<Embedder>, semantic_ratio: f32 },
}
impl SearchKind {
pub(crate) fn semantic(
index_scheduler: &index_scheduler::IndexScheduler,
@ -358,7 +365,7 @@ impl SearchQuery {
}
}
/// A `SearchQuery` + an index UID.
/// A `SearchQuery` + an index UID and optional FederationOptions.
// This struct contains the fields of `SearchQuery` inline.
// This is because neither deserr nor serde support `flatten` when using `deny_unknown_fields.
// The `From<SearchQueryWithIndex>` implementation ensures both structs remain up to date.
@ -373,10 +380,10 @@ pub struct SearchQueryWithIndex {
pub vector: Option<Vec<f32>>,
#[deserr(default, error = DeserrJsonError<InvalidHybridQuery>)]
pub hybrid: Option<HybridQuery>,
#[deserr(default = DEFAULT_SEARCH_OFFSET(), error = DeserrJsonError<InvalidSearchOffset>)]
pub offset: usize,
#[deserr(default = DEFAULT_SEARCH_LIMIT(), error = DeserrJsonError<InvalidSearchLimit>)]
pub limit: usize,
#[deserr(default, error = DeserrJsonError<InvalidSearchOffset>)]
pub offset: Option<usize>,
#[deserr(default, error = DeserrJsonError<InvalidSearchLimit>)]
pub limit: Option<usize>,
#[deserr(default, error = DeserrJsonError<InvalidSearchPage>)]
pub page: Option<usize>,
#[deserr(default, error = DeserrJsonError<InvalidSearchHitsPerPage>)]
@ -417,12 +424,33 @@ pub struct SearchQueryWithIndex {
pub attributes_to_search_on: Option<Vec<String>>,
#[deserr(default, error = DeserrJsonError<InvalidSearchRankingScoreThreshold>, default)]
pub ranking_score_threshold: Option<RankingScoreThreshold>,
#[deserr(default)]
pub federation_options: Option<FederationOptions>,
}
impl SearchQueryWithIndex {
pub fn into_index_query(self) -> (IndexUid, SearchQuery) {
pub fn has_federation_options(&self) -> bool {
self.federation_options.is_some()
}
pub fn has_pagination(&self) -> Option<&'static str> {
if self.offset.is_some() {
Some("offset")
} else if self.limit.is_some() {
Some("limit")
} else if self.page.is_some() {
Some("page")
} else if self.hits_per_page.is_some() {
Some("hitsPerPage")
} else {
None
}
}
pub fn into_index_query_federation(self) -> (IndexUid, SearchQuery, Option<FederationOptions>) {
let SearchQueryWithIndex {
index_uid,
federation_options,
q,
vector,
offset,
@ -454,8 +482,8 @@ impl SearchQueryWithIndex {
SearchQuery {
q,
vector,
offset,
limit,
offset: offset.unwrap_or(DEFAULT_SEARCH_OFFSET()),
limit: limit.unwrap_or(DEFAULT_SEARCH_LIMIT()),
page,
hits_per_page,
attributes_to_retrieve,
@ -480,6 +508,7 @@ impl SearchQueryWithIndex {
// do not use ..Default::default() here,
// rather add any missing field from `SearchQuery` to `SearchQueryWithIndex`
},
federation_options,
)
}
}
@ -864,15 +893,7 @@ pub fn perform_search(
used_negative_operator,
},
semantic_hit_count,
) = match &search_kind {
SearchKind::KeywordOnly => (search.execute()?, None),
SearchKind::SemanticOnly { .. } => {
let results = search.execute()?;
let semantic_hit_count = results.document_scores.len() as u32;
(results, Some(semantic_hit_count))
}
SearchKind::Hybrid { semantic_ratio, .. } => search.execute_hybrid(*semantic_ratio)?,
};
) = search_from_kind(search_kind, search)?;
let SearchQuery {
q,
@ -919,8 +940,13 @@ pub fn perform_search(
show_ranking_score_details,
};
let documents =
make_hits(index, &rtxn, format, matching_words, documents_ids, document_scores)?;
let documents = make_hits(
index,
&rtxn,
format,
matching_words,
documents_ids.iter().copied().zip(document_scores.iter()),
)?;
let number_of_hits = min(candidates.len() as usize, max_total_hits);
let hits_info = if is_finite_pagination {
@ -988,6 +1014,22 @@ pub fn perform_search(
Ok(result)
}
pub fn search_from_kind(
search_kind: SearchKind,
search: milli::Search<'_>,
) -> Result<(milli::SearchResult, Option<u32>), MeilisearchHttpError> {
let (milli_result, semantic_hit_count) = match &search_kind {
SearchKind::KeywordOnly => (search.execute()?, None),
SearchKind::SemanticOnly { .. } => {
let results = search.execute()?;
let semantic_hit_count = results.document_scores.len() as u32;
(results, Some(semantic_hit_count))
}
SearchKind::Hybrid { semantic_ratio, .. } => search.execute_hybrid(*semantic_ratio)?,
};
Ok((milli_result, semantic_hit_count))
}
struct AttributesFormat {
attributes_to_retrieve: Option<BTreeSet<String>>,
retrieve_vectors: RetrieveVectors,
@ -1033,131 +1075,191 @@ impl RetrieveVectors {
}
}
fn make_hits(
index: &Index,
rtxn: &RoTxn<'_>,
format: AttributesFormat,
matching_words: milli::MatchingWords,
documents_ids: Vec<u32>,
document_scores: Vec<Vec<ScoreDetails>>,
) -> Result<Vec<SearchHit>, MeilisearchHttpError> {
let fields_ids_map = index.fields_ids_map(rtxn).unwrap();
let displayed_ids =
index.displayed_fields_ids(rtxn)?.map(|fields| fields.into_iter().collect::<BTreeSet<_>>());
struct HitMaker<'a> {
index: &'a Index,
rtxn: &'a RoTxn<'a>,
fields_ids_map: FieldsIdsMap,
displayed_ids: BTreeSet<FieldId>,
vectors_fid: Option<FieldId>,
retrieve_vectors: RetrieveVectors,
to_retrieve_ids: BTreeSet<FieldId>,
embedding_configs: Vec<milli::index::IndexEmbeddingConfig>,
formatter_builder: MatcherBuilder<'a>,
formatted_options: BTreeMap<FieldId, FormatOptions>,
show_ranking_score: bool,
show_ranking_score_details: bool,
sort: Option<Vec<String>>,
show_matches_position: bool,
}
let vectors_fid = fields_ids_map.id(milli::vector::parsed_vectors::RESERVED_VECTORS_FIELD_NAME);
impl<'a> HitMaker<'a> {
pub fn tokenizer<'b>(
script_lang_map: &'b HashMap<milli::tokenizer::Script, Vec<milli::tokenizer::Language>>,
dictionary: Option<&'b [&'b str]>,
separators: Option<&'b [&'b str]>,
) -> milli::tokenizer::Tokenizer<'b> {
let mut tokenizer_builder = TokenizerBuilder::default();
tokenizer_builder.create_char_map(true);
if !script_lang_map.is_empty() {
tokenizer_builder.allow_list(script_lang_map);
}
let vectors_is_hidden = match (&displayed_ids, vectors_fid) {
// displayed_ids is a wildcard, so `_vectors` can be displayed regardless of its fid
(None, _) => false,
// displayed_ids is a finite list, and `_vectors` cannot be part of it because it is not an existing field
(Some(_), None) => true,
// displayed_ids is a finit list, so hide if `_vectors` is not part of it
(Some(map), Some(vectors_fid)) => map.contains(&vectors_fid),
};
if let Some(separators) = separators {
tokenizer_builder.separators(separators);
}
let retrieve_vectors = if let RetrieveVectors::Retrieve = format.retrieve_vectors {
if vectors_is_hidden {
RetrieveVectors::Hide
if let Some(dictionary) = dictionary {
tokenizer_builder.words_dict(dictionary);
}
tokenizer_builder.into_tokenizer()
}
pub fn formatter_builder(
matching_words: milli::MatchingWords,
tokenizer: milli::tokenizer::Tokenizer<'_>,
) -> MatcherBuilder<'_> {
let formatter_builder = MatcherBuilder::new(matching_words, tokenizer);
formatter_builder
}
pub fn new(
index: &'a Index,
rtxn: &'a RoTxn<'a>,
format: AttributesFormat,
mut formatter_builder: MatcherBuilder<'a>,
) -> Result<Self, MeilisearchHttpError> {
formatter_builder.crop_marker(format.crop_marker);
formatter_builder.highlight_prefix(format.highlight_pre_tag);
formatter_builder.highlight_suffix(format.highlight_post_tag);
let fields_ids_map = index.fields_ids_map(rtxn)?;
let displayed_ids = index
.displayed_fields_ids(rtxn)?
.map(|fields| fields.into_iter().collect::<BTreeSet<_>>());
let vectors_fid =
fields_ids_map.id(milli::vector::parsed_vectors::RESERVED_VECTORS_FIELD_NAME);
let vectors_is_hidden = match (&displayed_ids, vectors_fid) {
// displayed_ids is a wildcard, so `_vectors` can be displayed regardless of its fid
(None, _) => false,
// displayed_ids is a finite list, and `_vectors` cannot be part of it because it is not an existing field
(Some(_), None) => true,
// displayed_ids is a finit list, so hide if `_vectors` is not part of it
(Some(map), Some(vectors_fid)) => map.contains(&vectors_fid),
};
let displayed_ids =
displayed_ids.unwrap_or_else(|| fields_ids_map.iter().map(|(id, _)| id).collect());
let retrieve_vectors = if let RetrieveVectors::Retrieve = format.retrieve_vectors {
if vectors_is_hidden {
RetrieveVectors::Hide
} else {
RetrieveVectors::Retrieve
}
} else {
RetrieveVectors::Retrieve
}
} else {
format.retrieve_vectors
};
format.retrieve_vectors
};
let displayed_ids =
displayed_ids.unwrap_or_else(|| fields_ids_map.iter().map(|(id, _)| id).collect());
let fids = |attrs: &BTreeSet<String>| {
let mut ids = BTreeSet::new();
for attr in attrs {
if attr == "*" {
ids.clone_from(&displayed_ids);
break;
let fids = |attrs: &BTreeSet<String>| {
let mut ids = BTreeSet::new();
for attr in attrs {
if attr == "*" {
ids.clone_from(&displayed_ids);
break;
}
if let Some(id) = fields_ids_map.id(attr) {
ids.insert(id);
}
}
ids
};
let to_retrieve_ids: BTreeSet<_> = format
.attributes_to_retrieve
.as_ref()
.map(fids)
.unwrap_or_else(|| displayed_ids.clone())
.intersection(&displayed_ids)
.cloned()
.collect();
if let Some(id) = fields_ids_map.id(attr) {
ids.insert(id);
}
}
ids
};
let to_retrieve_ids: BTreeSet<_> = format
.attributes_to_retrieve
.as_ref()
.map(fids)
.unwrap_or_else(|| displayed_ids.clone())
.intersection(&displayed_ids)
.cloned()
.collect();
let attr_to_highlight = format.attributes_to_highlight.unwrap_or_default();
let attr_to_crop = format.attributes_to_crop.unwrap_or_default();
let formatted_options = compute_formatted_options(
&attr_to_highlight,
&attr_to_crop,
format.crop_length,
&to_retrieve_ids,
&fields_ids_map,
&displayed_ids,
);
let attr_to_highlight = format.attributes_to_highlight.unwrap_or_default();
let attr_to_crop = format.attributes_to_crop.unwrap_or_default();
let formatted_options = compute_formatted_options(
&attr_to_highlight,
&attr_to_crop,
format.crop_length,
&to_retrieve_ids,
&fields_ids_map,
&displayed_ids,
);
let mut tokenizer_builder = TokenizerBuilder::default();
tokenizer_builder.create_char_map(true);
let script_lang_map = index.script_language(rtxn)?;
if !script_lang_map.is_empty() {
tokenizer_builder.allow_list(&script_lang_map);
let embedding_configs = index.embedding_configs(rtxn)?;
Ok(Self {
index,
rtxn,
fields_ids_map,
displayed_ids,
vectors_fid,
retrieve_vectors,
to_retrieve_ids,
embedding_configs,
formatter_builder,
formatted_options,
show_ranking_score: format.show_ranking_score,
show_ranking_score_details: format.show_ranking_score_details,
show_matches_position: format.show_matches_position,
sort: format.sort,
})
}
let separators = index.allowed_separators(rtxn)?;
let separators: Option<Vec<_>> =
separators.as_ref().map(|x| x.iter().map(String::as_str).collect());
if let Some(ref separators) = separators {
tokenizer_builder.separators(separators);
}
let dictionary = index.dictionary(rtxn)?;
let dictionary: Option<Vec<_>> =
dictionary.as_ref().map(|x| x.iter().map(String::as_str).collect());
if let Some(ref dictionary) = dictionary {
tokenizer_builder.words_dict(dictionary);
}
let mut formatter_builder = MatcherBuilder::new(matching_words, tokenizer_builder.build());
formatter_builder.crop_marker(format.crop_marker);
formatter_builder.highlight_prefix(format.highlight_pre_tag);
formatter_builder.highlight_suffix(format.highlight_post_tag);
let mut documents = Vec::new();
let embedding_configs = index.embedding_configs(rtxn)?;
let documents_iter = index.documents(rtxn, documents_ids)?;
for ((id, obkv), score) in documents_iter.into_iter().zip(document_scores.into_iter()) {
pub fn make_hit(
&self,
id: u32,
score: &[ScoreDetails],
) -> Result<SearchHit, MeilisearchHttpError> {
let (_, obkv) =
self.index.iter_documents(self.rtxn, std::iter::once(id))?.next().unwrap()?;
// First generate a document with all the displayed fields
let displayed_document = make_document(&displayed_ids, &fields_ids_map, obkv)?;
let displayed_document = make_document(&self.displayed_ids, &self.fields_ids_map, obkv)?;
let add_vectors_fid =
vectors_fid.filter(|_fid| retrieve_vectors == RetrieveVectors::Retrieve);
self.vectors_fid.filter(|_fid| self.retrieve_vectors == RetrieveVectors::Retrieve);
// select the attributes to retrieve
let attributes_to_retrieve = to_retrieve_ids
let attributes_to_retrieve = self
.to_retrieve_ids
.iter()
// skip the vectors_fid if RetrieveVectors::Hide
.filter(|fid| match vectors_fid {
.filter(|fid| match self.vectors_fid {
Some(vectors_fid) => {
!(retrieve_vectors == RetrieveVectors::Hide && **fid == vectors_fid)
!(self.retrieve_vectors == RetrieveVectors::Hide && **fid == vectors_fid)
}
None => true,
})
// need to retrieve the existing `_vectors` field if the `RetrieveVectors::Retrieve`
.chain(add_vectors_fid.iter())
.map(|&fid| fields_ids_map.name(fid).expect("Missing field name"));
.map(|&fid| self.fields_ids_map.name(fid).expect("Missing field name"));
let mut document =
permissive_json_pointer::select_values(&displayed_document, attributes_to_retrieve);
if retrieve_vectors == RetrieveVectors::Retrieve {
if self.retrieve_vectors == RetrieveVectors::Retrieve {
// Clippy is wrong
#[allow(clippy::manual_unwrap_or_default)]
let mut vectors = match document.remove("_vectors") {
Some(Value::Object(map)) => map,
_ => Default::default(),
};
for (name, vector) in index.embeddings(rtxn, id)? {
let user_provided = embedding_configs
for (name, vector) in self.index.embeddings(self.rtxn, id)? {
let user_provided = self
.embedding_configs
.iter()
.find(|conf| conf.name == name)
.is_some_and(|conf| conf.user_provided.contains(id));
@ -1170,21 +1272,21 @@ fn make_hits(
let (matches_position, formatted) = format_fields(
&displayed_document,
&fields_ids_map,
&formatter_builder,
&formatted_options,
format.show_matches_position,
&displayed_ids,
&self.fields_ids_map,
&self.formatter_builder,
&self.formatted_options,
self.show_matches_position,
&self.displayed_ids,
)?;
if let Some(sort) = format.sort.as_ref() {
if let Some(sort) = self.sort.as_ref() {
insert_geo_distance(sort, &mut document);
}
let ranking_score =
format.show_ranking_score.then(|| ScoreDetails::global_score(score.iter()));
self.show_ranking_score.then(|| ScoreDetails::global_score(score.iter()));
let ranking_score_details =
format.show_ranking_score_details.then(|| ScoreDetails::to_json_map(score.iter()));
self.show_ranking_score_details.then(|| ScoreDetails::to_json_map(score.iter()));
let hit = SearchHit {
document,
@ -1193,7 +1295,38 @@ fn make_hits(
ranking_score_details,
ranking_score,
};
documents.push(hit);
Ok(hit)
}
}
fn make_hits<'a>(
index: &Index,
rtxn: &RoTxn<'_>,
format: AttributesFormat,
matching_words: milli::MatchingWords,
documents_ids_scores: impl Iterator<Item = (u32, &'a Vec<ScoreDetails>)> + 'a,
) -> Result<Vec<SearchHit>, MeilisearchHttpError> {
let mut documents = Vec::new();
let script_lang_map = index.script_language(rtxn)?;
let dictionary = index.dictionary(rtxn)?;
let dictionary: Option<Vec<_>> =
dictionary.as_ref().map(|x| x.iter().map(String::as_str).collect());
let separators = index.allowed_separators(rtxn)?;
let separators: Option<Vec<_>> =
separators.as_ref().map(|x| x.iter().map(String::as_str).collect());
let tokenizer =
HitMaker::tokenizer(&script_lang_map, dictionary.as_deref(), separators.as_deref());
let formatter_builder = HitMaker::formatter_builder(matching_words, tokenizer);
let hit_maker = HitMaker::new(index, rtxn, format, formatter_builder)?;
for (id, score) in documents_ids_scores {
documents.push(hit_maker.make_hit(id, score)?);
}
Ok(documents)
}
@ -1309,7 +1442,13 @@ pub fn perform_similar(
show_ranking_score_details,
};
let hits = make_hits(index, &rtxn, format, Default::default(), documents_ids, document_scores)?;
let hits = make_hits(
index,
&rtxn,
format,
Default::default(),
documents_ids.iter().copied().zip(document_scores.iter()),
)?;
let max_total_hits = index
.pagination_max_total_hits(&rtxn)
@ -1482,10 +1621,10 @@ fn make_document(
Ok(document)
}
fn format_fields<'a>(
fn format_fields(
document: &Document,
field_ids_map: &FieldsIdsMap,
builder: &'a MatcherBuilder<'a>,
builder: &MatcherBuilder<'_>,
formatted_options: &BTreeMap<FieldId, FormatOptions>,
compute_matches: bool,
displayable_ids: &BTreeSet<FieldId>,
@ -1540,9 +1679,9 @@ fn format_fields<'a>(
Ok((matches_position, document))
}
fn format_value<'a>(
fn format_value(
value: Value,
builder: &'a MatcherBuilder<'a>,
builder: &MatcherBuilder<'_>,
format_options: Option<FormatOptions>,
infos: &mut Vec<MatchBounds>,
compute_matches: bool,

View File

@ -0,0 +1,823 @@
use std::collections::HashMap;
use std::fmt::Write;
use itertools::Itertools as _;
use meilisearch_types::error::{Code, ResponseError};
use meilisearch_types::milli::{AscDesc, Criterion, Member, TermsMatchingStrategy};
pub struct RankingRules {
canonical_criteria: Vec<Criterion>,
canonical_sort: Option<Vec<AscDesc>>,
canonicalization_actions: Vec<CanonicalizationAction>,
source_criteria: Vec<Criterion>,
source_sort: Option<Vec<AscDesc>>,
}
pub enum CanonicalizationAction {
PrependedWords {
prepended_index: RankingRuleSource,
},
RemovedDuplicate {
earlier_occurrence: RankingRuleSource,
removed_occurrence: RankingRuleSource,
},
RemovedWords {
reason: RemoveWords,
removed_occurrence: RankingRuleSource,
},
RemovedPlaceholder {
removed_occurrence: RankingRuleSource,
},
TruncatedVector {
vector_rule: RankingRuleSource,
truncated_from: RankingRuleSource,
},
RemovedVector {
vector_rule: RankingRuleSource,
removed_occurrence: RankingRuleSource,
},
RemovedSort {
removed_occurrence: RankingRuleSource,
},
}
pub enum RemoveWords {
WasPrepended,
MatchingStrategyAll,
}
impl std::fmt::Display for RemoveWords {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
let reason = match self {
RemoveWords::WasPrepended => "it was previously prepended",
RemoveWords::MatchingStrategyAll => "`query.matchingWords` is set to `all`",
};
f.write_str(reason)
}
}
pub enum CanonicalizationKind {
Placeholder,
Keyword,
Vector,
}
pub struct CompatibilityError {
previous: RankingRule,
current: RankingRule,
}
impl CompatibilityError {
pub(crate) fn to_response_error(
&self,
ranking_rules: &RankingRules,
previous_ranking_rules: &RankingRules,
query_index: usize,
previous_query_index: usize,
index_uid: &str,
previous_index_uid: &str,
) -> meilisearch_types::error::ResponseError {
let rule = self.current.as_string(
&ranking_rules.canonical_criteria,
&ranking_rules.canonical_sort,
query_index,
index_uid,
);
let previous_rule = self.previous.as_string(
&previous_ranking_rules.canonical_criteria,
&previous_ranking_rules.canonical_sort,
previous_query_index,
previous_index_uid,
);
let canonicalization_actions = ranking_rules.canonicalization_notes();
let previous_canonicalization_actions = previous_ranking_rules.canonicalization_notes();
let mut msg = String::new();
let reason = self.reason();
let _ = writeln!(
&mut msg,
"The results of queries #{previous_query_index} and #{query_index} are incompatible: "
);
let _ = writeln!(&mut msg, " 1. {previous_rule}");
let _ = writeln!(&mut msg, " 2. {rule}");
let _ = writeln!(&mut msg, " - {reason}");
if !previous_canonicalization_actions.is_empty() {
let _ = write!(&mut msg, " - note: The ranking rules of query #{previous_query_index} were modified during canonicalization:\n{previous_canonicalization_actions}");
}
if !canonicalization_actions.is_empty() {
let _ = write!(&mut msg, " - note: The ranking rules of query #{query_index} were modified during canonicalization:\n{canonicalization_actions}");
}
ResponseError::from_msg(msg, Code::InvalidMultiSearchQueryRankingRules)
}
pub fn reason(&self) -> &'static str {
match (self.previous.kind, self.current.kind) {
(RankingRuleKind::Relevancy, RankingRuleKind::AscendingSort)
| (RankingRuleKind::Relevancy, RankingRuleKind::DescendingSort)
| (RankingRuleKind::AscendingSort, RankingRuleKind::Relevancy)
| (RankingRuleKind::DescendingSort, RankingRuleKind::Relevancy) => {
"cannot compare a relevancy rule with a sort rule"
}
(RankingRuleKind::Relevancy, RankingRuleKind::AscendingGeoSort)
| (RankingRuleKind::Relevancy, RankingRuleKind::DescendingGeoSort)
| (RankingRuleKind::AscendingGeoSort, RankingRuleKind::Relevancy)
| (RankingRuleKind::DescendingGeoSort, RankingRuleKind::Relevancy) => {
"cannot compare a relevancy rule with a geosort rule"
}
(RankingRuleKind::AscendingSort, RankingRuleKind::DescendingSort)
| (RankingRuleKind::DescendingSort, RankingRuleKind::AscendingSort) => {
"cannot compare two sort rules in opposite directions"
}
(RankingRuleKind::AscendingSort, RankingRuleKind::AscendingGeoSort)
| (RankingRuleKind::AscendingSort, RankingRuleKind::DescendingGeoSort)
| (RankingRuleKind::DescendingSort, RankingRuleKind::AscendingGeoSort)
| (RankingRuleKind::DescendingSort, RankingRuleKind::DescendingGeoSort)
| (RankingRuleKind::AscendingGeoSort, RankingRuleKind::AscendingSort)
| (RankingRuleKind::AscendingGeoSort, RankingRuleKind::DescendingSort)
| (RankingRuleKind::DescendingGeoSort, RankingRuleKind::AscendingSort)
| (RankingRuleKind::DescendingGeoSort, RankingRuleKind::DescendingSort) => {
"cannot compare a sort rule with a geosort rule"
}
(RankingRuleKind::AscendingGeoSort, RankingRuleKind::DescendingGeoSort)
| (RankingRuleKind::DescendingGeoSort, RankingRuleKind::AscendingGeoSort) => {
"cannot compare two geosort rules in opposite directions"
}
(RankingRuleKind::Relevancy, RankingRuleKind::Relevancy)
| (RankingRuleKind::AscendingSort, RankingRuleKind::AscendingSort)
| (RankingRuleKind::DescendingSort, RankingRuleKind::DescendingSort)
| (RankingRuleKind::AscendingGeoSort, RankingRuleKind::AscendingGeoSort)
| (RankingRuleKind::DescendingGeoSort, RankingRuleKind::DescendingGeoSort) => {
"internal error, comparison should be possible"
}
}
}
}
impl RankingRules {
pub fn new(
criteria: Vec<Criterion>,
sort: Option<Vec<AscDesc>>,
terms_matching_strategy: TermsMatchingStrategy,
canonicalization_kind: CanonicalizationKind,
) -> Self {
let (canonical_criteria, canonical_sort, canonicalization_actions) =
Self::canonicalize(&criteria, &sort, terms_matching_strategy, canonicalization_kind);
Self {
canonical_criteria,
canonical_sort,
canonicalization_actions,
source_criteria: criteria,
source_sort: sort,
}
}
fn canonicalize(
criteria: &[Criterion],
sort: &Option<Vec<AscDesc>>,
terms_matching_strategy: TermsMatchingStrategy,
canonicalization_kind: CanonicalizationKind,
) -> (Vec<Criterion>, Option<Vec<AscDesc>>, Vec<CanonicalizationAction>) {
match canonicalization_kind {
CanonicalizationKind::Placeholder => Self::canonicalize_placeholder(criteria, sort),
CanonicalizationKind::Keyword => {
Self::canonicalize_keyword(criteria, sort, terms_matching_strategy)
}
CanonicalizationKind::Vector => Self::canonicalize_vector(criteria, sort),
}
}
fn canonicalize_placeholder(
criteria: &[Criterion],
sort_query: &Option<Vec<AscDesc>>,
) -> (Vec<Criterion>, Option<Vec<AscDesc>>, Vec<CanonicalizationAction>) {
let mut sort = None;
let mut sorted_fields = HashMap::new();
let mut canonicalization_actions = Vec::new();
let mut canonical_criteria = Vec::new();
let mut canonical_sort = None;
for (criterion_index, criterion) in criteria.iter().enumerate() {
match criterion.clone() {
Criterion::Words
| Criterion::Typo
| Criterion::Proximity
| Criterion::Attribute
| Criterion::Exactness => {
canonicalization_actions.push(CanonicalizationAction::RemovedPlaceholder {
removed_occurrence: RankingRuleSource::Criterion(criterion_index),
})
}
Criterion::Sort => {
if let Some(previous_index) = sort {
canonicalization_actions.push(CanonicalizationAction::RemovedDuplicate {
earlier_occurrence: RankingRuleSource::Criterion(previous_index),
removed_occurrence: RankingRuleSource::Criterion(criterion_index),
});
} else if let Some(sort_query) = sort_query {
sort = Some(criterion_index);
canonical_criteria.push(criterion.clone());
canonical_sort = Some(canonicalize_sort(
&mut sorted_fields,
sort_query.as_slice(),
criterion_index,
&mut canonicalization_actions,
));
} else {
canonicalization_actions.push(CanonicalizationAction::RemovedSort {
removed_occurrence: RankingRuleSource::Criterion(criterion_index),
})
}
}
Criterion::Asc(s) | Criterion::Desc(s) => match sorted_fields.entry(s) {
std::collections::hash_map::Entry::Occupied(entry) => canonicalization_actions
.push(CanonicalizationAction::RemovedDuplicate {
earlier_occurrence: *entry.get(),
removed_occurrence: RankingRuleSource::Criterion(criterion_index),
}),
std::collections::hash_map::Entry::Vacant(entry) => {
entry.insert(RankingRuleSource::Criterion(criterion_index));
canonical_criteria.push(criterion.clone())
}
},
}
}
(canonical_criteria, canonical_sort, canonicalization_actions)
}
fn canonicalize_vector(
criteria: &[Criterion],
sort_query: &Option<Vec<AscDesc>>,
) -> (Vec<Criterion>, Option<Vec<AscDesc>>, Vec<CanonicalizationAction>) {
let mut sort = None;
let mut sorted_fields = HashMap::new();
let mut canonicalization_actions = Vec::new();
let mut canonical_criteria = Vec::new();
let mut canonical_sort = None;
let mut vector = None;
'criteria: for (criterion_index, criterion) in criteria.iter().enumerate() {
match criterion.clone() {
Criterion::Words
| Criterion::Typo
| Criterion::Proximity
| Criterion::Attribute
| Criterion::Exactness => match vector {
Some(previous_occurrence) => {
if sorted_fields.is_empty() {
canonicalization_actions.push(CanonicalizationAction::RemovedVector {
vector_rule: RankingRuleSource::Criterion(previous_occurrence),
removed_occurrence: RankingRuleSource::Criterion(criterion_index),
});
} else {
canonicalization_actions.push(
CanonicalizationAction::TruncatedVector {
vector_rule: RankingRuleSource::Criterion(previous_occurrence),
truncated_from: RankingRuleSource::Criterion(criterion_index),
},
);
break 'criteria;
}
}
None => {
canonical_criteria.push(criterion.clone());
vector = Some(criterion_index);
}
},
Criterion::Sort => {
if let Some(previous_index) = sort {
canonicalization_actions.push(CanonicalizationAction::RemovedDuplicate {
earlier_occurrence: RankingRuleSource::Criterion(previous_index),
removed_occurrence: RankingRuleSource::Criterion(criterion_index),
});
} else if let Some(sort_query) = sort_query {
sort = Some(criterion_index);
canonical_criteria.push(criterion.clone());
canonical_sort = Some(canonicalize_sort(
&mut sorted_fields,
sort_query.as_slice(),
criterion_index,
&mut canonicalization_actions,
));
} else {
canonicalization_actions.push(CanonicalizationAction::RemovedSort {
removed_occurrence: RankingRuleSource::Criterion(criterion_index),
})
}
}
Criterion::Asc(s) | Criterion::Desc(s) => match sorted_fields.entry(s) {
std::collections::hash_map::Entry::Occupied(entry) => canonicalization_actions
.push(CanonicalizationAction::RemovedDuplicate {
earlier_occurrence: *entry.get(),
removed_occurrence: RankingRuleSource::Criterion(criterion_index),
}),
std::collections::hash_map::Entry::Vacant(entry) => {
entry.insert(RankingRuleSource::Criterion(criterion_index));
canonical_criteria.push(criterion.clone())
}
},
}
}
(canonical_criteria, canonical_sort, canonicalization_actions)
}
fn canonicalize_keyword(
criteria: &[Criterion],
sort_query: &Option<Vec<AscDesc>>,
terms_matching_strategy: TermsMatchingStrategy,
) -> (Vec<Criterion>, Option<Vec<AscDesc>>, Vec<CanonicalizationAction>) {
let mut words = None;
let mut typo = None;
let mut proximity = None;
let mut sort = None;
let mut attribute = None;
let mut exactness = None;
let mut sorted_fields = HashMap::new();
let mut canonical_criteria = Vec::new();
let mut canonical_sort = None;
let mut canonicalization_actions = Vec::new();
for (criterion_index, criterion) in criteria.iter().enumerate() {
let criterion = criterion.clone();
match criterion.clone() {
Criterion::Words => {
if let TermsMatchingStrategy::All = terms_matching_strategy {
canonicalization_actions.push(CanonicalizationAction::RemovedWords {
reason: RemoveWords::MatchingStrategyAll,
removed_occurrence: RankingRuleSource::Criterion(criterion_index),
});
continue;
}
if let Some(maybe_previous_index) = words {
if let Some(previous_index) = maybe_previous_index {
canonicalization_actions.push(
CanonicalizationAction::RemovedDuplicate {
earlier_occurrence: RankingRuleSource::Criterion(
previous_index,
),
removed_occurrence: RankingRuleSource::Criterion(
criterion_index,
),
},
);
continue;
}
canonicalization_actions.push(CanonicalizationAction::RemovedWords {
reason: RemoveWords::WasPrepended,
removed_occurrence: RankingRuleSource::Criterion(criterion_index),
})
}
words = Some(Some(criterion_index));
canonical_criteria.push(criterion);
}
Criterion::Typo => {
canonicalize_criterion(
criterion,
criterion_index,
terms_matching_strategy,
&mut words,
&mut canonicalization_actions,
&mut canonical_criteria,
&mut typo,
);
}
Criterion::Proximity => {
canonicalize_criterion(
criterion,
criterion_index,
terms_matching_strategy,
&mut words,
&mut canonicalization_actions,
&mut canonical_criteria,
&mut proximity,
);
}
Criterion::Attribute => {
canonicalize_criterion(
criterion,
criterion_index,
terms_matching_strategy,
&mut words,
&mut canonicalization_actions,
&mut canonical_criteria,
&mut attribute,
);
}
Criterion::Exactness => {
canonicalize_criterion(
criterion,
criterion_index,
terms_matching_strategy,
&mut words,
&mut canonicalization_actions,
&mut canonical_criteria,
&mut exactness,
);
}
Criterion::Sort => {
if let Some(previous_index) = sort {
canonicalization_actions.push(CanonicalizationAction::RemovedDuplicate {
earlier_occurrence: RankingRuleSource::Criterion(previous_index),
removed_occurrence: RankingRuleSource::Criterion(criterion_index),
});
} else if let Some(sort_query) = sort_query {
sort = Some(criterion_index);
canonical_criteria.push(criterion);
canonical_sort = Some(canonicalize_sort(
&mut sorted_fields,
sort_query.as_slice(),
criterion_index,
&mut canonicalization_actions,
));
} else {
canonicalization_actions.push(CanonicalizationAction::RemovedSort {
removed_occurrence: RankingRuleSource::Criterion(criterion_index),
})
}
}
Criterion::Asc(s) | Criterion::Desc(s) => match sorted_fields.entry(s) {
std::collections::hash_map::Entry::Occupied(entry) => canonicalization_actions
.push(CanonicalizationAction::RemovedDuplicate {
earlier_occurrence: *entry.get(),
removed_occurrence: RankingRuleSource::Criterion(criterion_index),
}),
std::collections::hash_map::Entry::Vacant(entry) => {
entry.insert(RankingRuleSource::Criterion(criterion_index));
canonical_criteria.push(criterion)
}
},
}
}
(canonical_criteria, canonical_sort, canonicalization_actions)
}
pub fn is_compatible_with(&self, previous: &Self) -> Result<(), CompatibilityError> {
for (current, previous) in self.coalesce_iterator().zip(previous.coalesce_iterator()) {
if current.kind != previous.kind {
return Err(CompatibilityError { current, previous });
}
}
Ok(())
}
pub fn constraint_count(&self) -> usize {
self.coalesce_iterator().count()
}
fn coalesce_iterator(&self) -> impl Iterator<Item = RankingRule> + '_ {
self.canonical_criteria
.iter()
.enumerate()
.flat_map(|(criterion_index, criterion)| {
RankingRule::from_criterion(criterion_index, criterion, &self.canonical_sort)
})
.coalesce(
|previous @ RankingRule { source: previous_source, kind: previous_kind },
current @ RankingRule { source, kind }| {
match (previous_kind, kind) {
(RankingRuleKind::Relevancy, RankingRuleKind::Relevancy) => {
let merged_source = match (previous_source, source) {
(
RankingRuleSource::Criterion(previous),
RankingRuleSource::Criterion(current),
) => RankingRuleSource::CoalescedCriteria(previous, current),
(
RankingRuleSource::CoalescedCriteria(begin, _end),
RankingRuleSource::Criterion(current),
) => RankingRuleSource::CoalescedCriteria(begin, current),
(_previous, current) => current,
};
Ok(RankingRule { source: merged_source, kind })
}
_ => Err((previous, current)),
}
},
)
}
fn canonicalization_notes(&self) -> String {
use CanonicalizationAction::*;
let mut notes = String::new();
for (index, action) in self.canonicalization_actions.iter().enumerate() {
let index = index + 1;
let _ = match action {
PrependedWords { prepended_index } => writeln!(
&mut notes,
" {index}. Prepended rule `words` before first relevancy rule `{}` at position {}",
prepended_index.rule_name(&self.source_criteria, &self.source_sort),
prepended_index.rule_position()
),
RemovedDuplicate { earlier_occurrence, removed_occurrence } => writeln!(
&mut notes,
" {index}. Removed duplicate rule `{}` at position {} as it already appears at position {}",
earlier_occurrence.rule_name(&self.source_criteria, &self.source_sort),
removed_occurrence.rule_position(),
earlier_occurrence.rule_position(),
),
RemovedWords { reason, removed_occurrence } => writeln!(
&mut notes,
" {index}. Removed rule `words` at position {} because {reason}",
removed_occurrence.rule_position()
),
RemovedPlaceholder { removed_occurrence } => writeln!(
&mut notes,
" {index}. Removed relevancy rule `{}` at position {} because the query is a placeholder search (`q`: \"\")",
removed_occurrence.rule_name(&self.source_criteria, &self.source_sort),
removed_occurrence.rule_position()
),
TruncatedVector { vector_rule, truncated_from } => writeln!(
&mut notes,
" {index}. Truncated relevancy rule `{}` at position {} and later rules because the query is a vector search and `vector` was inserted at position {}",
truncated_from.rule_name(&self.source_criteria, &self.source_sort),
truncated_from.rule_position(),
vector_rule.rule_position(),
),
RemovedVector { vector_rule, removed_occurrence } => writeln!(
&mut notes,
" {index}. Removed relevancy rule `{}` at position {} because the query is a vector search and `vector` was already inserted at position {}",
removed_occurrence.rule_name(&self.source_criteria, &self.source_sort),
removed_occurrence.rule_position(),
vector_rule.rule_position(),
),
RemovedSort { removed_occurrence } => writeln!(
&mut notes,
" {index}. Removed rule `sort` at position {} because `query.sort` is empty",
removed_occurrence.rule_position()
),
};
}
notes
}
}
fn canonicalize_sort(
sorted_fields: &mut HashMap<String, RankingRuleSource>,
sort_query: &[AscDesc],
criterion_index: usize,
canonicalization_actions: &mut Vec<CanonicalizationAction>,
) -> Vec<AscDesc> {
let mut geo_sorted = None;
let mut canonical_sort = Vec::new();
for (sort_index, asc_desc) in sort_query.iter().enumerate() {
let source = RankingRuleSource::Sort { criterion_index, sort_index };
let asc_desc = asc_desc.clone();
match asc_desc.clone() {
AscDesc::Asc(Member::Field(s)) | AscDesc::Desc(Member::Field(s)) => {
match sorted_fields.entry(s) {
std::collections::hash_map::Entry::Occupied(entry) => canonicalization_actions
.push(CanonicalizationAction::RemovedDuplicate {
earlier_occurrence: *entry.get(),
removed_occurrence: source,
}),
std::collections::hash_map::Entry::Vacant(entry) => {
entry.insert(source);
canonical_sort.push(asc_desc);
}
}
}
AscDesc::Asc(Member::Geo(_)) | AscDesc::Desc(Member::Geo(_)) => match geo_sorted {
Some(earlier_sort_index) => {
canonicalization_actions.push(CanonicalizationAction::RemovedDuplicate {
earlier_occurrence: RankingRuleSource::Sort {
criterion_index,
sort_index: earlier_sort_index,
},
removed_occurrence: source,
})
}
None => {
geo_sorted = Some(sort_index);
canonical_sort.push(asc_desc);
}
},
}
}
canonical_sort
}
fn canonicalize_criterion(
criterion: Criterion,
criterion_index: usize,
terms_matching_strategy: TermsMatchingStrategy,
words: &mut Option<Option<usize>>,
canonicalization_actions: &mut Vec<CanonicalizationAction>,
canonical_criteria: &mut Vec<Criterion>,
rule: &mut Option<usize>,
) {
*words = match (terms_matching_strategy, words.take()) {
(TermsMatchingStrategy::All, words) => words,
(_, None) => {
// inject words
canonicalization_actions.push(CanonicalizationAction::PrependedWords {
prepended_index: RankingRuleSource::Criterion(criterion_index),
});
canonical_criteria.push(Criterion::Words);
Some(None)
}
(_, words) => words,
};
if let Some(previous_index) = *rule {
canonicalization_actions.push(CanonicalizationAction::RemovedDuplicate {
earlier_occurrence: RankingRuleSource::Criterion(previous_index),
removed_occurrence: RankingRuleSource::Criterion(criterion_index),
});
} else {
*rule = Some(criterion_index);
canonical_criteria.push(criterion)
}
}
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum RankingRuleKind {
Relevancy,
AscendingSort,
DescendingSort,
AscendingGeoSort,
DescendingGeoSort,
}
#[derive(Debug, Clone, Copy)]
pub struct RankingRule {
source: RankingRuleSource,
kind: RankingRuleKind,
}
#[derive(Debug, Clone, Copy)]
pub enum RankingRuleSource {
Criterion(usize),
CoalescedCriteria(usize, usize),
Sort { criterion_index: usize, sort_index: usize },
}
impl RankingRuleSource {
fn rule_name(&self, criteria: &[Criterion], sort: &Option<Vec<AscDesc>>) -> String {
match self {
RankingRuleSource::Criterion(criterion_index) => criteria
.get(*criterion_index)
.map(|c| c.to_string())
.unwrap_or_else(|| "unknown".into()),
RankingRuleSource::CoalescedCriteria(begin, end) => {
let rules: Vec<_> = criteria
.get(*begin..=*end)
.iter()
.flat_map(|c| c.iter())
.map(|c| c.to_string())
.collect();
rules.join(", ")
}
RankingRuleSource::Sort { criterion_index: _, sort_index } => {
match sort.as_deref().and_then(|sort| sort.get(*sort_index)) {
Some(sort) => match sort {
AscDesc::Asc(Member::Field(field_name)) => format!("{field_name}:asc"),
AscDesc::Desc(Member::Field(field_name)) => {
format!("{field_name}:desc")
}
AscDesc::Asc(Member::Geo(_)) => "_geo(..):asc".to_string(),
AscDesc::Desc(Member::Geo(_)) => "_geo(..):desc".to_string(),
},
None => "unknown".into(),
}
}
}
}
fn rule_position(&self) -> String {
match self {
RankingRuleSource::Criterion(criterion_index) => {
format!("#{criterion_index} in ranking rules")
}
RankingRuleSource::CoalescedCriteria(begin, end) => {
format!("#{begin} to #{end} in ranking rules")
}
RankingRuleSource::Sort { criterion_index, sort_index } => format!(
"#{sort_index} in `query.sort` (as `sort` is #{criterion_index} in ranking rules)"
),
}
}
}
impl RankingRule {
fn from_criterion<'a>(
criterion_index: usize,
criterion: &'a Criterion,
sort: &'a Option<Vec<AscDesc>>,
) -> impl Iterator<Item = Self> + 'a {
let kind = match criterion {
Criterion::Words
| Criterion::Typo
| Criterion::Proximity
| Criterion::Attribute
| Criterion::Exactness => RankingRuleKind::Relevancy,
Criterion::Asc(s) if s == "_geo" => RankingRuleKind::AscendingGeoSort,
Criterion::Asc(_) => RankingRuleKind::AscendingSort,
Criterion::Desc(s) if s == "_geo" => RankingRuleKind::DescendingGeoSort,
Criterion::Desc(_) => RankingRuleKind::DescendingSort,
Criterion::Sort => {
return either::Right(sort.iter().flatten().enumerate().map(
move |(rule_index, asc_desc)| {
Self::from_asc_desc(asc_desc, criterion_index, rule_index)
},
))
}
};
either::Left(std::iter::once(Self {
source: RankingRuleSource::Criterion(criterion_index),
kind,
}))
}
fn from_asc_desc(asc_desc: &AscDesc, sort_index: usize, rule_index_in_sort: usize) -> Self {
let kind = match asc_desc {
AscDesc::Asc(Member::Field(_)) => RankingRuleKind::AscendingSort,
AscDesc::Desc(Member::Field(_)) => RankingRuleKind::DescendingSort,
AscDesc::Asc(Member::Geo(_)) => RankingRuleKind::AscendingGeoSort,
AscDesc::Desc(Member::Geo(_)) => RankingRuleKind::DescendingGeoSort,
};
Self {
source: RankingRuleSource::Sort {
criterion_index: sort_index,
sort_index: rule_index_in_sort,
},
kind,
}
}
fn as_string(
&self,
canonical_criteria: &[Criterion],
canonical_sort: &Option<Vec<AscDesc>>,
query_index: usize,
index_uid: &str,
) -> String {
let kind = match self.kind {
RankingRuleKind::Relevancy => "relevancy",
RankingRuleKind::AscendingSort => "ascending sort",
RankingRuleKind::DescendingSort => "descending sort",
RankingRuleKind::AscendingGeoSort => "ascending geo sort",
RankingRuleKind::DescendingGeoSort => "descending geo sort",
};
let rules = self.fetch_from_source(canonical_criteria, canonical_sort);
let source = match self.source {
RankingRuleSource::Criterion(criterion_index) => format!("`queries[{query_index}]`, `{index_uid}.rankingRules[{criterion_index}]`"),
RankingRuleSource::CoalescedCriteria(begin, end) => format!("`queries[{query_index}]`, `{index_uid}.rankingRules[{begin}..={end}]`"),
RankingRuleSource::Sort { criterion_index, sort_index } => format!("`queries[{query_index}].sort[{sort_index}]`, `{index_uid}.rankingRules[{criterion_index}]`"),
};
format!("{source}: {kind} {rules}")
}
fn fetch_from_source(
&self,
canonical_criteria: &[Criterion],
canonical_sort: &Option<Vec<AscDesc>>,
) -> String {
let rule_name = match self.source {
RankingRuleSource::Criterion(index) => {
canonical_criteria.get(index).map(|criterion| criterion.to_string())
}
RankingRuleSource::CoalescedCriteria(begin, end) => {
let rules: Vec<String> = canonical_criteria
.get(begin..=end)
.into_iter()
.flat_map(|criteria| criteria.iter())
.map(|criterion| criterion.to_string())
.collect();
(!rules.is_empty()).then_some(rules.join(", "))
}
RankingRuleSource::Sort { criterion_index: _, sort_index } => canonical_sort
.as_deref()
.and_then(|canonical_sort| canonical_sort.get(sort_index))
.and_then(|asc_desc: &AscDesc| match asc_desc {
AscDesc::Asc(Member::Field(s)) | AscDesc::Desc(Member::Field(s)) => {
Some(format!("on field `{s}`"))
}
_ => None,
}),
};
let rule_name = rule_name.unwrap_or_else(|| "default".into());
format!("rule(s) {rule_name}")
}
}

View File

@ -310,6 +310,23 @@ macro_rules! compute_authorized_single_search {
tenant_token,
key_content
);
// federated
let (response, code) = server.multi_search(json!({"federation": {}, "queries" : [{"indexUid": "sales", "filter": $filter}]})).await;
assert_eq!(
200, code,
"{} using tenant_token: {:?} generated with parent_key: {:?}",
response, tenant_token, key_content
);
assert_eq!(
// same count as the search is federated over a single query
$expected_count,
response["hits"].as_array().unwrap().len(),
"{} using tenant_token: {:?} generated with parent_key: {:?}",
response,
tenant_token,
key_content
);
}
}
};
@ -375,6 +392,25 @@ macro_rules! compute_authorized_multiple_search {
tenant_token,
key_content
);
let (response, code) = server.multi_search(json!({"federation": {}, "queries" : [
{"indexUid": "sales", "filter": $filter1},
{"indexUid": "products", "filter": $filter2},
]})).await;
assert_eq!(
code, 200,
"{} using tenant_token: {:?} generated with parent_key: {:?}",
response, tenant_token, key_content
);
assert_eq!(
response["hits"].as_array().unwrap().len(),
// sum of counts as the search is federated across to queries in different indexes
$expected_count1 + $expected_count2,
"{} using tenant_token: {:?} generated with parent_key: {:?}",
response,
tenant_token,
key_content
);
}
}
};
@ -433,6 +469,24 @@ macro_rules! compute_forbidden_single_search {
"{} using tenant_token: {:?} generated with parent_key: {:?}",
response, tenant_token, key_content
);
let (mut response, code) = server.multi_search(json!({"federation": {}, "queries" : [{"indexUid": "sales"}]})).await;
if failed_query_index.is_none() && !response["message"].is_null() {
response["message"] = serde_json::json!(null);
}
assert_eq!(
response,
invalid_response(failed_query_index),
"{} using tenant_token: {:?} generated with parent_key: {:?}",
response,
tenant_token,
key_content
);
assert_eq!(
code, 403,
"{} using tenant_token: {:?} generated with parent_key: {:?}",
response, tenant_token, key_content
);
}
}
};
@ -494,6 +548,27 @@ macro_rules! compute_forbidden_multiple_search {
"{} using tenant_token: {:?} generated with parent_key: {:?}",
response, tenant_token, key_content
);
let (mut response, code) = server.multi_search(json!({"federation": {}, "queries" : [
{"indexUid": "sales"},
{"indexUid": "products"},
]})).await;
if failed_query_index.is_none() && !response["message"].is_null() {
response["message"] = serde_json::json!(null);
}
assert_eq!(
response,
invalid_response(failed_query_index),
"{} using tenant_token: {:?} generated with parent_key: {:?}",
response,
tenant_token,
key_content
);
assert_eq!(
code, 403,
"{} using tenant_token: {:?} generated with parent_key: {:?}",
response, tenant_token, key_content
);
}
}
};

View File

@ -132,6 +132,79 @@ static NESTED_DOCUMENTS: Lazy<Value> = Lazy::new(|| {
])
});
static FRUITS_DOCUMENTS: Lazy<Value> = Lazy::new(|| {
json!([
{
"name": "Exclusive sale: green apple",
"id": "green-apple-boosted",
"BOOST": true
},
{
"name": "Pear",
"id": "pear",
},
{
"name": "Red apple gala",
"id": "red-apple-gala",
},
{
"name": "Exclusive sale: Red Tomato",
"id": "red-tomatoes-boosted",
"BOOST": true
},
{
"name": "Exclusive sale: Red delicious apple",
"id": "red-delicious-boosted",
"BOOST": true,
}
])
});
static VECTOR_DOCUMENTS: Lazy<Value> = Lazy::new(|| {
json!([
{
"id": "A",
"description": "the dog barks at the cat",
"_vectors": {
// dimensions [canine, feline, young]
"animal": [0.9, 0.8, 0.05],
// dimensions [negative/positive, energy]
"sentiment": [-0.1, 0.55]
}
},
{
"id": "B",
"description": "the kitten scratched the beagle",
"_vectors": {
// dimensions [canine, feline, young]
"animal": [0.8, 0.9, 0.5],
// dimensions [negative/positive, energy]
"sentiment": [-0.2, 0.65]
}
},
{
"id": "C",
"description": "the dog had to stay alone today",
"_vectors": {
// dimensions [canine, feline, young]
"animal": [0.85, 0.02, 0.1],
// dimensions [negative/positive, energy]
"sentiment": [-1.0, 0.1]
}
},
{
"id": "D",
"description": "the little boy pets the puppy",
"_vectors": {
// dimensions [canine, feline, young]
"animal": [0.8, 0.09, 0.8],
// dimensions [negative/positive, energy]
"sentiment": [0.8, 0.3]
}
},
])
});
#[actix_rt::test]
async fn simple_placeholder_search() {
let server = Server::new().await;

File diff suppressed because it is too large Load Diff

View File

@ -425,9 +425,6 @@ pub struct Sort {
impl PartialOrd for Sort {
fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
if self.field_name != other.field_name {
return None;
}
if self.ascending != other.ascending {
return None;
}
@ -466,9 +463,6 @@ pub struct GeoSort {
impl PartialOrd for GeoSort {
fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
if self.target_point != other.target_point {
return None;
}
if self.ascending != other.ascending {
return None;
}

View File

@ -46,7 +46,7 @@ impl<'m> MatcherBuilder<'m> {
self
}
pub fn build<'t>(&'m self, text: &'t str) -> Matcher<'t, 'm> {
pub fn build<'t>(&self, text: &'t str) -> Matcher<'t, 'm, '_> {
let crop_marker = match &self.crop_marker {
Some(marker) => marker.as_str(),
None => DEFAULT_CROP_MARKER,
@ -105,19 +105,19 @@ pub struct MatchBounds {
pub length: usize,
}
/// Structure used to analize a string, compute words that match,
/// Structure used to analyze a string, compute words that match,
/// and format the source string, returning a highlighted and cropped sub-string.
pub struct Matcher<'t, 'm> {
pub struct Matcher<'t, 'tokenizer, 'b> {
text: &'t str,
matching_words: &'m MatchingWords,
tokenizer: &'m Tokenizer<'m>,
crop_marker: &'m str,
highlight_prefix: &'m str,
highlight_suffix: &'m str,
matching_words: &'b MatchingWords,
tokenizer: &'b Tokenizer<'tokenizer>,
crop_marker: &'b str,
highlight_prefix: &'b str,
highlight_suffix: &'b str,
matches: Option<(Vec<Token<'t>>, Vec<Match>)>,
}
impl<'t> Matcher<'t, '_> {
impl<'t, 'tokenizer> Matcher<'t, 'tokenizer, '_> {
/// Iterates over tokens and save any of them that matches the query.
fn compute_matches(&mut self) -> &mut Self {
/// some words are counted as matches only if they are close together and in the good order,