MeiliSearch/milli/src/search/facet/facet_distribution.rs

261 lines
9.5 KiB
Rust

use std::collections::{HashSet, BTreeMap};
use std::ops::Bound::Unbounded;
use std::{cmp, fmt};
use anyhow::Context;
use heed::BytesDecode;
use roaring::RoaringBitmap;
use crate::facet::{FacetType, FacetValue};
use crate::heed_codec::facet::{FacetValueStringCodec, FacetLevelValueF64Codec, FacetLevelValueI64Codec};
use crate::heed_codec::facet::{FieldDocIdFacetStringCodec, FieldDocIdFacetF64Codec, FieldDocIdFacetI64Codec};
use crate::search::facet::{FacetIter, FacetRange};
use crate::{Index, FieldId, DocumentId};
/// The default number of values by facets that will
/// be fetched from the key-value store.
const DEFAULT_VALUES_BY_FACET: usize = 100;
/// The hard limit in the number of values by facets that will be fetched from
/// the key-value store. Searching for more values could slow down the engine.
const MAX_VALUES_BY_FACET: usize = 1000;
/// Threshold on the number of candidates that will make
/// the system to choose between one algorithm or another.
const CANDIDATES_THRESHOLD: u64 = 1000;
pub struct FacetDistribution<'a> {
facets: Option<HashSet<String>>,
candidates: Option<RoaringBitmap>,
max_values_by_facet: usize,
rtxn: &'a heed::RoTxn<'a>,
index: &'a Index,
}
impl<'a> FacetDistribution<'a> {
pub fn new(rtxn: &'a heed::RoTxn, index: &'a Index) -> FacetDistribution<'a> {
FacetDistribution {
facets: None,
candidates: None,
max_values_by_facet: DEFAULT_VALUES_BY_FACET,
rtxn,
index,
}
}
pub fn facets<I: IntoIterator<Item=A>, A: AsRef<str>>(&mut self, names: I) -> &mut Self {
self.facets = Some(names.into_iter().map(|s| s.as_ref().to_string()).collect());
self
}
pub fn candidates(&mut self, candidates: RoaringBitmap) -> &mut Self {
self.candidates = Some(candidates);
self
}
pub fn max_values_by_facet(&mut self, max: usize) -> &mut Self {
self.max_values_by_facet = cmp::min(max, MAX_VALUES_BY_FACET);
self
}
/// There is a small amount of candidates OR we ask for facet string values so we
/// decide to iterate over the facet values of each one of them, one by one.
fn facet_values_from_documents(
&self,
field_id: FieldId,
facet_type: FacetType,
candidates: &RoaringBitmap,
) -> heed::Result<BTreeMap<FacetValue, u64>>
{
fn fetch_facet_values<'t, KC, K: 't>(
index: &Index,
rtxn: &'t heed::RoTxn,
field_id: FieldId,
candidates: &RoaringBitmap,
) -> heed::Result<BTreeMap<FacetValue, u64>>
where
KC: BytesDecode<'t, DItem = (FieldId, DocumentId, K)>,
K: Into<FacetValue>,
{
let mut facet_values = BTreeMap::new();
let mut key_buffer = vec![field_id];
for docid in candidates.into_iter().take(CANDIDATES_THRESHOLD as usize) {
key_buffer.truncate(1);
key_buffer.extend_from_slice(&docid.to_be_bytes());
let iter = index.field_id_docid_facet_values
.prefix_iter(rtxn, &key_buffer)?
.remap_key_type::<KC>();
for result in iter {
let ((_, _, value), ()) = result?;
*facet_values.entry(value.into()).or_insert(0) += 1;
}
}
Ok(facet_values)
}
let index = self.index;
let rtxn = self.rtxn;
match facet_type {
FacetType::String => {
fetch_facet_values::<FieldDocIdFacetStringCodec, _>(index, rtxn, field_id, candidates)
},
FacetType::Float => {
fetch_facet_values::<FieldDocIdFacetF64Codec, _>(index, rtxn, field_id, candidates)
},
FacetType::Integer => {
fetch_facet_values::<FieldDocIdFacetI64Codec, _>(index, rtxn, field_id, candidates)
},
}
}
/// There is too much documents, we use the facet levels to move throught
/// the facet values, to find the candidates and values associated.
fn facet_values_from_facet_levels(
&self,
field_id: FieldId,
facet_type: FacetType,
candidates: &RoaringBitmap,
) -> heed::Result<BTreeMap<FacetValue, u64>>
{
let iter = match facet_type {
FacetType::String => unreachable!(),
FacetType::Float => {
let iter = FacetIter::<f64, FacetLevelValueF64Codec>::new_non_reducing(
self.rtxn, self.index, field_id, candidates.clone(),
)?;
let iter = iter.map(|r| r.map(|(v, docids)| (FacetValue::from(v), docids)));
Box::new(iter) as Box::<dyn Iterator<Item=_>>
},
FacetType::Integer => {
let iter = FacetIter::<i64, FacetLevelValueI64Codec>::new_non_reducing(
self.rtxn, self.index, field_id, candidates.clone(),
)?;
Box::new(iter.map(|r| r.map(|(v, docids)| (FacetValue::from(v), docids))))
},
};
let mut facet_values = BTreeMap::new();
for result in iter {
let (value, mut docids) = result?;
docids.intersect_with(candidates);
if !docids.is_empty() {
facet_values.insert(value, docids.len());
}
if facet_values.len() == self.max_values_by_facet {
break;
}
}
Ok(facet_values)
}
/// Placeholder search, a.k.a. no candidates were specified. We iterate throught the
/// facet values one by one and iterate on the facet level 0 for numbers.
fn facet_values_from_raw_facet_database(
&self,
field_id: FieldId,
facet_type: FacetType,
) -> heed::Result<BTreeMap<FacetValue, u64>>
{
let db = self.index.facet_field_id_value_docids;
let level = 0;
let iter = match facet_type {
FacetType::String => {
let iter = db
.prefix_iter(self.rtxn, &[field_id])?
.remap_key_type::<FacetValueStringCodec>()
.map(|r| r.map(|((_, v), docids)| (FacetValue::from(v), docids)));
Box::new(iter) as Box::<dyn Iterator<Item=_>>
},
FacetType::Float => {
let db = db.remap_key_type::<FacetLevelValueF64Codec>();
let range = FacetRange::<f64, _>::new(
self.rtxn, db, field_id, level, Unbounded, Unbounded,
)?;
Box::new(range.map(|r| r.map(|((_, _, v, _), docids)| (FacetValue::from(v), docids))))
},
FacetType::Integer => {
let db = db.remap_key_type::<FacetLevelValueI64Codec>();
let range = FacetRange::<i64, _>::new(
self.rtxn, db, field_id, level, Unbounded, Unbounded,
)?;
Box::new(range.map(|r| r.map(|((_, _, v, _), docids)| (FacetValue::from(v), docids))))
},
};
let mut facet_values = BTreeMap::new();
for result in iter {
let (value, docids) = result?;
facet_values.insert(value, docids.len());
if facet_values.len() == self.max_values_by_facet {
break;
}
}
Ok(facet_values)
}
fn facet_values(
&self,
field_id: FieldId,
facet_type: FacetType,
) -> heed::Result<BTreeMap<FacetValue, u64>>
{
if let Some(candidates) = self.candidates.as_ref() {
// Classic search, candidates were specified, we must return facet values only related
// to those candidates. We also enter here for facet strings for performance reasons.
if candidates.len() <= CANDIDATES_THRESHOLD || facet_type == FacetType::String {
self.facet_values_from_documents(field_id, facet_type, candidates)
} else {
self.facet_values_from_facet_levels(field_id, facet_type, candidates)
}
} else {
self.facet_values_from_raw_facet_database(field_id, facet_type)
}
}
pub fn execute(&self) -> anyhow::Result<BTreeMap<String, BTreeMap<FacetValue, u64>>> {
let fields_ids_map = self.index.fields_ids_map(self.rtxn)?;
let faceted_fields = self.index.faceted_fields(self.rtxn)?;
let fields_ids: Vec<_> = match &self.facets {
Some(names) => names
.iter()
.filter_map(|n| faceted_fields.get(n).map(|t| (n.to_string(), *t)))
.collect(),
None => faceted_fields.into_iter().collect(),
};
let mut facets_values = BTreeMap::new();
for (name, ftype) in fields_ids {
let fid = fields_ids_map.id(&name).with_context(|| {
format!("missing field name {:?} from the fields id map", name)
})?;
let values = self.facet_values(fid, ftype)?;
facets_values.insert(name, values);
}
Ok(facets_values)
}
}
impl fmt::Debug for FacetDistribution<'_> {
fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
let FacetDistribution {
facets,
candidates,
max_values_by_facet,
rtxn: _,
index: _,
} = self;
f.debug_struct("FacetDistribution")
.field("facets", facets)
.field("candidates", candidates)
.field("max_values_by_facet", max_values_by_facet)
.finish()
}
}