Merge #4796

4796: Generate vectors in dumps r=dureuill a=dureuill

# Pull Request

## What does this PR do?
1. Add an Index::embeddings method to compute the embeddings of a document
2. Write generated vectors in dumps
3. Remove generated vectors when importing dumps
4. Cherry pick the `ACTIONS_ALLOW_USE_UNSECURE_NODE_VERSION` workaround so that the older CI can still build

## Manual Tests (TODO)

- [ ] Import a dump from a v1.8.3 into a v1.8.4 successfully
- [x] Import a dump from a v1.8.4 into a v1.8.4 successfully
- [x] Import a dump from a v1.8.4 into a v1.9.0 successfully
  - [x] generated vectors are not regenerated
  - [x] user provided vectors are still available
  - [x] generated vectors still have the correct value
  - [x] updating a document with generated vectors attempts to regenerate

Co-authored-by: Louis Dureuil <louis@meilisearch.com>

.github/workflows/flaky-tests.yml

@@ -1,4 +1,6 @@
 name: Look for flaky tests
+env:
+  ACTIONS_ALLOW_USE_UNSECURE_NODE_VERSION: true
 on:
   workflow_dispatch:
   schedule:

.github/workflows/fuzzer-indexing.yml

@@ -1,5 +1,6 @@
 name: Run the indexing fuzzer
-
+env:
+  ACTIONS_ALLOW_USE_UNSECURE_NODE_VERSION: true
 on:
   push:
     branches:

.github/workflows/publish-apt-brew-pkg.yml

@@ -15,6 +15,8 @@ jobs:
 
   debian:
     name: Publish debian packagge
+    env:
+      ACTIONS_ALLOW_USE_UNSECURE_NODE_VERSION: true
     runs-on: ubuntu-latest
     needs: check-version
     container:

.github/workflows/publish-binaries.yml

@@ -35,6 +35,8 @@ jobs:
   publish-linux:
     name: Publish binary for Linux
     runs-on: ubuntu-latest
+    env:
+      ACTIONS_ALLOW_USE_UNSECURE_NODE_VERSION: true
     needs: check-version
     container:
       # Use ubuntu-18.04 to compile with glibc 2.27
@@ -132,6 +134,8 @@ jobs:
     name: Publish binary for aarch64
     runs-on: ubuntu-latest
     needs: check-version
+    env:
+      ACTIONS_ALLOW_USE_UNSECURE_NODE_VERSION: true
     container:
       # Use ubuntu-18.04 to compile with glibc 2.27
       image: ubuntu:18.04

.github/workflows/test-suite.yml

@@ -21,6 +21,8 @@ jobs:
   test-linux:
     name: Tests on ubuntu-18.04
     runs-on: ubuntu-latest
+    env:
+      ACTIONS_ALLOW_USE_UNSECURE_NODE_VERSION: true
     container:
       # Use ubuntu-18.04 to compile with glibc 2.27, which are the production expectations
       image: ubuntu:18.04
@@ -77,6 +79,8 @@ jobs:
   test-all-features:
     name: Tests almost all features
     runs-on: ubuntu-latest
+    env:
+      ACTIONS_ALLOW_USE_UNSECURE_NODE_VERSION: true
     container:
       # Use ubuntu-18.04 to compile with glibc 2.27, which are the production expectations
       image: ubuntu:18.04
@@ -100,6 +104,8 @@ jobs:
 
   test-disabled-tokenization:
     name: Test disabled tokenization
+    env:
+      ACTIONS_ALLOW_USE_UNSECURE_NODE_VERSION: true
     runs-on: ubuntu-latest
     container:
       image: ubuntu:18.04
@@ -127,6 +133,8 @@ jobs:
   # We run tests in debug also, to make sure that the debug_assertions are hit
   test-debug:
     name: Run tests in debug
+    env:
+      ACTIONS_ALLOW_USE_UNSECURE_NODE_VERSION: true
     runs-on: ubuntu-latest
     container:
       # Use ubuntu-18.04 to compile with glibc 2.27, which are the production expectations

index-scheduler/src/batch.rs

@@ -914,8 +914,34 @@ impl IndexScheduler {
                         if self.must_stop_processing.get() {
                             return Err(Error::AbortedTask);
                         }
-                        let (_id, doc) = ret?;
-                        let document = milli::obkv_to_json(&all_fields, &fields_ids_map, doc)?;
+                        let (id, doc) = ret?;
+                        let mut document = milli::obkv_to_json(&all_fields, &fields_ids_map, doc)?;
+
+                        'inject_vectors: {
+                            let embeddings = index.embeddings(&rtxn, id)?;
+
+                            if embeddings.is_empty() {
+                                break 'inject_vectors;
+                            }
+
+                            let vectors = document
+                                .entry("_vectors".to_owned())
+                                .or_insert(serde_json::Value::Object(Default::default()));
+
+                            let serde_json::Value::Object(vectors) = vectors else {
+                                break 'inject_vectors;
+                            };
+
+                            for (embedder_name, embeddings) in embeddings {
+                                vectors.entry(embedder_name).or_insert_with(|| {
+                                    serde_json::json!({
+                                        "embeddings": embeddings,
+                                        "regenerate": true
+                                    })
+                                });
+                            }
+                        }
+
                         index_dumper.push_document(&document)?;
                     }
 

meilisearch/src/lib.rs

@@ -419,7 +419,41 @@ fn import_dump(
         let file = tempfile::tempfile()?;
         let mut builder = DocumentsBatchBuilder::new(BufWriter::new(file));
         for document in index_reader.documents()? {
-            builder.append_json_object(&document?)?;
+            let mut document = document?;
+
+            'remove_injected_vectors: {
+                let Some(vectors) = document.get_mut("_vectors") else {
+                    break 'remove_injected_vectors;
+                };
+
+                let Some(vectors) = vectors.as_object_mut() else { break 'remove_injected_vectors };
+
+                vectors.retain(|_embedder, embedding_object| {
+                    // don't touch values that aren't objects
+                    let Some(embedding_object) = embedding_object.as_object() else {
+                        return true;
+                    };
+
+                    let mut has_regenerate_true = false;
+                    for (field, value) in embedding_object {
+                        match (field.as_str(), value) {
+                            // detected regenerate : true
+                            // if we don't have any superfluous field, we'll remove the entire entry
+                            ("regenerate", serde_json::Value::Bool(true)) => {
+                                has_regenerate_true = true;
+                            }
+                            // ignore embeddings
+                            ("embeddings", _) => continue,
+                            // any other field: immediately retain the entry
+                            _ => return true,
+                        }
+                    }
+                    // retain the entry unless it has regenerate: true
+                    !has_regenerate_true
+                })
+            }
+
+            builder.append_json_object(&document)?;
         }
 
         // This flush the content of the batch builder.

milli/src/index.rs

@@ -22,7 +22,7 @@ use crate::heed_codec::{
 };
 use crate::order_by_map::OrderByMap;
 use crate::proximity::ProximityPrecision;
-use crate::vector::EmbeddingConfig;
+use crate::vector::{Embedding, EmbeddingConfig};
 use crate::{
     default_criteria, CboRoaringBitmapCodec, Criterion, DocumentId, ExternalDocumentsIds,
     FacetDistribution, FieldDistribution, FieldId, FieldIdWordCountCodec, GeoPoint, ObkvCodec,
@@ -1516,6 +1516,42 @@ impl Index {
             .unwrap_or_default())
     }
 
+    pub fn embeddings(
+        &self,
+        rtxn: &RoTxn<'_>,
+        docid: DocumentId,
+    ) -> Result<BTreeMap<String, Vec<Embedding>>> {
+        let mut res = BTreeMap::new();
+        for row in self.embedder_category_id.iter(rtxn)? {
+            let (embedder_name, embedder_id) = row?;
+            let embedder_id = (embedder_id as u16) << 8;
+            let mut embeddings = Vec::new();
+            'vectors: for i in 0..=u8::MAX {
+                let reader = arroy::Reader::open(rtxn, embedder_id | (i as u16), self.vector_arroy)
+                    .map(Some)
+                    .or_else(|e| match e {
+                        arroy::Error::MissingMetadata => Ok(None),
+                        e => Err(e),
+                    })
+                    .transpose();
+
+                let Some(reader) = reader else {
+                    break 'vectors;
+                };
+
+                let embedding = reader?.item_vector(rtxn, docid)?;
+                if let Some(embedding) = embedding {
+                    embeddings.push(embedding)
+                } else {
+                    break 'vectors;
+                }
+            }
+
+            res.insert(embedder_name.to_owned(), embeddings);
+        }
+        Ok(res)
+    }
+
     pub(crate) fn put_search_cutoff(&self, wtxn: &mut RwTxn<'_>, cutoff: u64) -> heed::Result<()> {
         self.main.remap_types::<Str, BEU64>().put(wtxn, main_key::SEARCH_CUTOFF, &cutoff)
     }