diff --git a/raptor/src/lib.rs b/raptor/src/lib.rs
index 2786d0ca8..85ad4628e 100644
--- a/raptor/src/lib.rs
+++ b/raptor/src/lib.rs
@@ -5,12 +5,7 @@ pub mod metadata;
 pub mod vec_read_only;
 pub mod automaton;
 
-pub use self::metadata::{
-    Metadata, MetadataBuilder,
-    // Stream, StreamBuilder,
-    // Union, OpBuilder,
-    // IndexedValues,
-};
+pub use self::metadata::{Metadata, MetadataBuilder};
 pub use self::rank::RankedStream;
 
 pub type DocumentId = u64;
@@ -23,7 +18,6 @@ pub type DocumentId = u64;
 #[derive(Debug, Copy, Clone, Eq, PartialEq, PartialOrd, Ord, Hash)]
 #[repr(C)]
 pub struct DocIndex {
-
     /// The document identifier where the word was found.
     pub document: DocumentId,
 
@@ -50,7 +44,6 @@ pub struct DocIndex {
 // TODO do data oriented programming ? very arrays ?
 #[derive(Debug, Copy, Clone, Eq, PartialEq, PartialOrd, Ord, Hash)]
 pub struct Match {
-
     /// The word index in the query sentence.
     /// Same as the `attribute_index` but for the query words.
     ///
diff --git a/raptor/src/metadata/difference.rs b/raptor/src/metadata/difference.rs
new file mode 100644
index 000000000..50c648459
--- /dev/null
+++ b/raptor/src/metadata/difference.rs
@@ -0,0 +1,126 @@
+use fst::{Streamer, Automaton};
+use crate::metadata::ops::{self, IndexedDocIndexes};
+use crate::metadata::{stream_ops, Metadata};
+
+fn union_with_automatons<'a, A>(metas: &'a [Metadata], autos: Vec<A>) -> ops::Union
+where A: 'a + Automaton + Clone,
+{
+    let mut op = ops::OpBuilder::with_automatons(autos);
+    for metadata in metas {
+        op.push(metadata);
+    }
+    op.union()
+}
+
+pub struct Difference<'f> {
+    inner: stream_ops::Difference<'f>,
+}
+
+impl<'f> Difference<'f> {
+    pub fn new<A>(positives: &'f [Metadata], negatives: &'f [Metadata], automatons: Vec<A>) -> Self
+    where A: 'f + Automaton + Clone
+    {
+        let positives = union_with_automatons(positives, automatons.clone());
+        let negatives = union_with_automatons(negatives, automatons);
+
+        let mut builder = stream_ops::OpBuilder::new();
+        builder.push(positives);
+        builder.push(negatives);
+
+        Difference { inner: builder.difference() }
+    }
+}
+
+impl<'a, 'f> Streamer<'a> for Difference<'f> {
+    type Item = (&'a [u8], &'a [IndexedDocIndexes]);
+
+    fn next(&'a mut self) -> Option<Self::Item> {
+        self.inner.next()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use fst::automaton::AlwaysMatch;
+    use crate::metadata::{Metadata, MetadataBuilder};
+    use crate::vec_read_only::VecReadOnly;
+    use crate::DocIndex;
+
+    fn construct_metadata(documents: Vec<(String, DocIndex)>) -> Metadata {
+        let mapw = Vec::new();
+        let indexesw = Vec::new();
+
+        let mut builder = MetadataBuilder::new(mapw, indexesw);
+
+        for (string, doc_index) in documents {
+            builder.insert(string, doc_index);
+        }
+
+        let (map, indexes) = builder.into_inner().unwrap();
+        Metadata::from_bytes(map, indexes).unwrap()
+    }
+
+    #[test]
+    fn empty() {
+        let positive_metas = construct_metadata(vec![
+            ("chameau".into(), DocIndex{ document: 12, attribute: 1, attribute_index: 22 }),
+            ("chameau".into(), DocIndex{ document: 31, attribute: 0, attribute_index: 1 }),
+        ]);
+
+        let negative_metas = construct_metadata(vec![
+            ("chameau".into(), DocIndex{ document: 12, attribute: 1, attribute_index: 22 }),
+            ("chameau".into(), DocIndex{ document: 31, attribute: 0, attribute_index: 1 }),
+        ]);
+
+        let positives = &[positive_metas];
+        let negatives = &[negative_metas];
+        let mut diff = Difference::new(positives, negatives, vec![AlwaysMatch]);
+
+        assert_eq!(diff.next(), None);
+    }
+
+    #[test]
+    fn one_positive() {
+        let di1 = DocIndex{ document: 12, attribute: 1, attribute_index: 22 };
+        let di2 = DocIndex{ document: 31, attribute: 0, attribute_index: 1 };
+
+        let positive_metas = construct_metadata(vec![
+            ("chameau".into(), di1),
+            ("chameau".into(), di2),
+        ]);
+
+        let negative_metas = construct_metadata(vec![
+            ("chameau".into(), di1),
+        ]);
+
+        let positives = &[positive_metas];
+        let negatives = &[negative_metas];
+        let mut diff = Difference::new(positives, negatives, vec![AlwaysMatch]);
+
+        let idi = IndexedDocIndexes{ index: 0, doc_indexes: VecReadOnly::new(vec![di2]) };
+        assert_eq!(diff.next(), Some(("chameau".as_bytes(), &[idi][..])));
+        assert_eq!(diff.next(), None);
+    }
+
+    #[test]
+    fn more_negative_than_positive() {
+        let di1 = DocIndex{ document: 12, attribute: 1, attribute_index: 22 };
+        let di2 = DocIndex{ document: 31, attribute: 0, attribute_index: 1 };
+
+        let positive_metas = construct_metadata(vec![
+            ("chameau".into(), di1),
+        ]);
+
+        let negative_metas = construct_metadata(vec![
+            ("chameau".into(), di1),
+            ("chameau".into(), di2),
+        ]);
+
+        let positives = &[positive_metas];
+        let negatives = &[negative_metas];
+        let mut diff = Difference::new(positives, negatives, vec![AlwaysMatch]);
+
+        assert_eq!(diff.next(), None);
+    }
+}
diff --git a/raptor/src/metadata/mod.rs b/raptor/src/metadata/mod.rs
index 7c63d2d53..9e594d49b 100644
--- a/raptor/src/metadata/mod.rs
+++ b/raptor/src/metadata/mod.rs
@@ -1,6 +1,8 @@
-mod ops_indexed_value;
 pub mod ops;
+pub mod stream_ops;
 pub mod doc_indexes;
+pub mod difference;
+pub mod ops_indexed_value;
 
 use fst::{Map, MapBuilder};
 use std::error::Error;
diff --git a/raptor/src/metadata/stream_ops.rs b/raptor/src/metadata/stream_ops.rs
new file mode 100644
index 000000000..230a54b07
--- /dev/null
+++ b/raptor/src/metadata/stream_ops.rs
@@ -0,0 +1,309 @@
+use std::rc::Rc;
+use std::collections::{BinaryHeap, HashMap, BTreeMap};
+use std::cmp;
+use fst::{IntoStreamer, Streamer};
+use sdset::multi::OpBuilder as SdOpBuilder;
+use sdset::{SetOperation, Set};
+use crate::metadata::ops::IndexedDocIndexes;
+use crate::vec_read_only::VecReadOnly;
+use crate::DocIndex;
+
+type BoxedStream<'f> = Box<for<'a> Streamer<'a, Item=(&'a [u8], &'a [IndexedDocIndexes])> + 'f>;
+
+pub struct OpBuilder<'f> {
+    streams: Vec<BoxedStream<'f>>,
+}
+
+impl<'f> OpBuilder<'f> {
+    pub fn new() -> Self {
+        Self { streams: Vec::new() }
+    }
+
+    /// Push a stream of `IndexedDocIndexes`.
+    ///
+    /// # Warning
+    ///
+    /// You must ensure yourself that the automatons are
+    /// all the same in the same order for each stream you push.
+    pub fn push<I, S>(&mut self, stream: I)
+    where
+        I: for<'a> IntoStreamer<'a, Into=S, Item=(&'a [u8], &'a [IndexedDocIndexes])>,
+        S: 'f + for<'a> Streamer<'a, Item=(&'a [u8], &'a [IndexedDocIndexes])>,
+    {
+        self.streams.push(Box::new(stream.into_stream()));
+    }
+
+    pub fn union(self) -> Union<'f> {
+        Union {
+            heap: StreamHeap::new(self.streams),
+            outs: Vec::new(),
+            cur_slot: None,
+        }
+    }
+
+    pub fn intersection(self) -> Intersection<'f> {
+        Intersection {
+            heap: StreamHeap::new(self.streams),
+            outs: Vec::new(),
+            cur_slot: None,
+        }
+    }
+
+    pub fn difference(self) -> Difference<'f> {
+        Difference {
+            heap: StreamHeap::new(self.streams),
+            outs: Vec::new(),
+            cur_slot: None,
+        }
+    }
+
+    pub fn symmetric_difference(self) -> SymmetricDifference<'f> {
+        SymmetricDifference {
+            heap: StreamHeap::new(self.streams),
+            outs: Vec::new(),
+            cur_slot: None,
+        }
+    }
+}
+
+// FIXME reuse it from metadata::ops
+struct SlotIndexedDocIndexes {
+    aut_index: usize,
+    start: usize,
+    len: usize,
+}
+
+macro_rules! logical_operation {
+    (struct $name:ident, $operation:ident) => {
+
+pub struct $name<'f> {
+    heap: StreamHeap<'f>,
+    outs: Vec<IndexedDocIndexes>,
+    cur_slot: Option<Slot>,
+}
+
+impl<'a, 'f> Streamer<'a> for $name<'f> {
+    type Item = (&'a [u8], &'a [IndexedDocIndexes]);
+
+    // The Metadata could be types as "key-values present" and "key-values possibly not present"
+    // in other words Metadata that "needs" to have key-values and other that doesn't needs.
+    //
+    // We could probably allow the user to define in Metadata some Document
+    // that needs to be deleted and only declare the DocumentId, and not every DocIndex of each words.
+    fn next(&'a mut self) -> Option<Self::Item> {
+        if let Some(slot) = self.cur_slot.take() {
+            self.heap.refill(slot);
+        }
+        let slot = match self.heap.pop() {
+            None => return None,
+            Some(slot) => {
+                self.cur_slot = Some(slot);
+                self.cur_slot.as_mut().unwrap()
+            }
+        };
+
+        self.outs.clear();
+
+        // retrieve all the doc_indexes of all the streams,
+        // store them in an HashMap which the key is
+        // the aut_index (associated with the state that is ignored),
+        // the doc_indexes must be stored in another BTreeMap which the key
+        // is the rdr_index.
+        //
+        // This will permit us to do set operations on readers (using the rdr_index)
+        // the BTreeMap will gives the rdr_index in order and the final result
+        // will be aggregated in a Vec of IndexedDocIndexes which the aut_index and state
+        // are the key of the first HashMap
+
+        // TODO use the fnv Hasher!
+
+        let mut builders = HashMap::new();
+        let iv = slot.indexed_value();
+        let builder = builders.entry(iv.index).or_insert_with(BTreeMap::new);
+        builder.insert(slot.rdr_index, iv.doc_indexes);
+
+        while let Some(mut slot) = self.heap.pop_if_equal(slot.input()) {
+            let iv = slot.indexed_value();
+            let builder = builders.entry(iv.index).or_insert_with(BTreeMap::new);
+            builder.insert(slot.rdr_index, iv.doc_indexes);
+
+            self.heap.refill(slot);
+        }
+
+        // now that we have accumulated all the doc_indexes like so:
+        // HashMap<(aut_index, state*), BtreeMap<rdr_index, doc_indexes>>
+        // we will be able to retrieve, for each aut_index, the doc_indexes
+        // that are needed to do the set operation
+
+        let mut doc_indexes = Vec::new();
+        let mut doc_indexes_slots = Vec::with_capacity(builders.len());
+        for (aut_index, values) in builders {
+
+            let sets = values.iter().map(|(_, v)| Set::new_unchecked(v.as_slice())).collect();
+            let builder = SdOpBuilder::from_vec(sets);
+
+            let start = doc_indexes.len();
+            builder.$operation().extend_vec(&mut doc_indexes);
+            let len = doc_indexes.len() - start;
+            if len == 0 { continue }
+
+            let slot = SlotIndexedDocIndexes {
+                aut_index: aut_index,
+                start: start,
+                len: len,
+            };
+            doc_indexes_slots.push(slot);
+        }
+
+        let read_only = VecReadOnly::new(doc_indexes);
+        self.outs.reserve(doc_indexes_slots.len());
+        for slot in doc_indexes_slots {
+            let indexes = IndexedDocIndexes {
+                index: slot.aut_index,
+                doc_indexes: read_only.range(slot.start, slot.len),
+            };
+            self.outs.push(indexes);
+        }
+
+        if self.outs.is_empty() { return None }
+        Some((slot.input(), &self.outs))
+    }
+}
+}}
+
+logical_operation!(struct Union, union);
+logical_operation!(struct Intersection, intersection);
+logical_operation!(struct Difference, difference);
+logical_operation!(struct SymmetricDifference, symmetric_difference);
+
+struct StreamHeap<'f> {
+    rdrs: Vec<BoxedStream<'f>>,
+    heap: BinaryHeap<Slot>,
+}
+
+impl<'f> StreamHeap<'f> {
+    fn new(streams: Vec<BoxedStream<'f>>) -> StreamHeap<'f> {
+        let mut heap = StreamHeap {
+            rdrs: streams,
+            heap: BinaryHeap::new(),
+        };
+        for i in 0..heap.rdrs.len() {
+            heap.refill(Slot::new(i));
+        }
+        heap
+    }
+
+    fn pop(&mut self) -> Option<Slot> {
+        self.heap.pop()
+    }
+
+    fn peek_is_duplicate(&self, key: &[u8]) -> bool {
+        self.heap.peek().map(|s| s.input() == key).unwrap_or(false)
+    }
+
+    fn pop_if_equal(&mut self, key: &[u8]) -> Option<Slot> {
+        if self.peek_is_duplicate(key) {
+            self.pop()
+        } else {
+            None
+        }
+    }
+
+    fn pop_if_le(&mut self, key: &[u8]) -> Option<Slot> {
+        if self.heap.peek().map(|s| s.input() <= key).unwrap_or(false) {
+            self.pop()
+        } else {
+            None
+        }
+    }
+
+    fn num_slots(&self) -> usize {
+        self.rdrs.len()
+    }
+
+    fn refill(&mut self, mut slot: Slot) {
+        if let Some((input, outputs)) = self.rdrs[slot.rdr_index].next() {
+            slot.set_input(input);
+            for output in outputs {
+                slot.set_aut_index(output.index);
+                slot.set_output(output.doc_indexes.clone());
+                self.heap.push(slot.clone());
+            }
+        }
+    }
+}
+
+#[derive(Debug, Clone)]
+struct Slot {
+    rdr_index: usize,
+    aut_index: usize,
+    input: Rc<Vec<u8>>,
+    output: Option<VecReadOnly<DocIndex>>,
+}
+
+impl PartialEq for Slot {
+    fn eq(&self, other: &Self) -> bool {
+        (&self.input, self.rdr_index, self.aut_index)
+        .eq(&(&other.input, other.rdr_index, other.aut_index))
+    }
+}
+
+impl Eq for Slot { }
+
+impl PartialOrd for Slot {
+    fn partial_cmp(&self, other: &Self) -> Option<cmp::Ordering> {
+        (&self.input, self.rdr_index, self.aut_index)
+        .partial_cmp(&(&other.input, other.rdr_index, other.aut_index))
+        .map(|ord| ord.reverse())
+    }
+}
+
+impl Ord for Slot {
+    fn cmp(&self, other: &Self) -> cmp::Ordering {
+        self.partial_cmp(other).unwrap()
+    }
+}
+
+impl Slot {
+    fn new(rdr_index: usize) -> Self {
+        Slot {
+            rdr_index: rdr_index,
+            aut_index: 0,
+            input: Rc::new(Vec::with_capacity(64)),
+            output: None,
+        }
+    }
+
+    fn indexed_value(&mut self) -> IndexedDocIndexes {
+        IndexedDocIndexes {
+            index: self.aut_index,
+            doc_indexes: self.output.take().unwrap(),
+        }
+    }
+
+    fn input(&self) -> &[u8] {
+        &self.input
+    }
+
+    fn set_input(&mut self, input: &[u8]) {
+        if *self.input != input {
+            let inner = Rc::make_mut(&mut self.input);
+            inner.clear();
+            inner.extend(input);
+        }
+    }
+
+    fn set_aut_index(&mut self, aut_index: usize) {
+        self.aut_index = aut_index;
+    }
+
+    fn set_output(&mut self, output: VecReadOnly<DocIndex>) {
+        self.output = Some(output);
+    }
+}
+
+#[derive(Copy, Clone, Debug, Eq, Hash, Ord, PartialEq, PartialOrd)]
+pub struct IndexedValueWithState {
+    pub index: usize,
+    pub value: u64,
+}