feat: Add sparse vectors benchmark support for Qdrant (#114)

* feat: Add sparse vectors benchmark support in Qdrant

* fix: Self review

* feat: Add sparse dataset for CI benchmarks

* feat: Introduce SparseVector class

* feat: Disallow sparse vector dataset being run with non sparse vector engine configs

* feat: use different engine config to run sparse vector benchmarks

* fix: use different engine config to run sparse vector benchmarks

* feat: Optimize CI benchmarks workflow

* feat: Add 1M sparse dataset

* fix: remove scipy, read csr matrix manually (#117)

* fix: remove scipy, read csr matrix manually

* fix: Dataset query reader should have sparse_vector=None by default

* refactor: Changes based on feedback

* refactoring: refactor sparse vector support (#118)

* refactoring: refactor sparse vector support

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* feat: Use pydantic construct

* refactor: Update all engines to use Query and Record dataclasses (#116)

* refactor: Update all engines to use Query and Record dataclasses

* feat: Add ruff in pre-commit hooks

* fix: Type mismatches

* fix: Redis search client types and var names

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix: Type issues detected by linter

* fix: iter_batches func type

* refactor: knn_conditions should be class level constant

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix: Type issue

* fix: Allow python 3.8 since scipy is now removed

* fix: Add missing redis-m-16-ef-128 config

* fix: redis container port

* fix linter

---------

Co-authored-by: George <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: generall <[email protected]>

Author: 4 people

.github/workflows/continuous-benchmark.yaml

@@ -15,18 +15,29 @@ jobs:
       - uses: webfactory/[email protected]
         with:
           ssh-private-key: ${{ secrets.SSH_PRIVATE_KEY }}
+      - name: Setup CI
+        run: bash -x tools/setup_ci.sh
       - name: Benches
         run: |
-          export HCLOUD_TOKEN=${{ secrets.HCLOUD_TOKEN }}
-          export GCS_KEY=${{ secrets.GCS_KEY }}
-          export GCS_SECRET=${{ secrets.GCS_SECRET }}
-          export POSTGRES_PASSWORD=${{ secrets.POSTGRES_PASSWORD }}
-          export POSTGRES_HOST=${{ secrets.POSTGRES_HOST }}
+            export HCLOUD_TOKEN=${{ secrets.HCLOUD_TOKEN }}
+            export GCS_KEY=${{ secrets.GCS_KEY }}
+            export GCS_SECRET=${{ secrets.GCS_SECRET }}
+            export POSTGRES_PASSWORD=${{ secrets.POSTGRES_PASSWORD }}
+            export POSTGRES_HOST=${{ secrets.POSTGRES_HOST }}
 
-          # Benchmark the dev branch:
-          export QDRANT_VERSION=ghcr/dev
-          bash -x tools/run_ci.sh
+            declare -A DATASET_TO_ENGINE
+            DATASET_TO_ENGINE["laion-small-clip"]="qdrant-continuous-benchmark"
+            DATASET_TO_ENGINE["msmarco-sparse-1M"]="qdrant-sparse-vector"
 
-          # Benchmark the master branch:
-          export QDRANT_VERSION=docker/master
-          bash -x tools/run_ci.sh
+            for dataset in "${!DATASET_TO_ENGINE[@]}"; do
+              export ENGINE_NAME=${DATASET_TO_ENGINE[$dataset]}
+              export DATASETS=$dataset
+
+              # Benchmark the dev branch:
+              export QDRANT_VERSION=ghcr/dev
+              bash -x tools/run_ci.sh
+
+              # Benchmark the master branch:
+              export QDRANT_VERSION=docker/master
+              bash -x tools/run_ci.sh
+            done

.pre-commit-config.yaml

@@ -28,3 +28,10 @@ repos:
       - id: isort
         name: "Sort Imports"
         args: ["--profile", "black"]
+
+  - repo: https://github.com/astral-sh/ruff-pre-commit
+    rev: v0.3.5
+    hooks:
+      # Run the linter.
+      - id: ruff
+        args: [ --fix ]

benchmark/dataset.py

@@ -10,20 +10,28 @@
 from dataset_reader.ann_h5_reader import AnnH5Reader
 from dataset_reader.base_reader import BaseReader
 from dataset_reader.json_reader import JSONReader
+from dataset_reader.sparse_reader import SparseReader
 
 
 @dataclass
 class DatasetConfig:
-    vector_size: int
-    distance: str
     name: str
     type: str
     path: str
+
     link: Optional[str] = None
     schema: Optional[Dict[str, str]] = field(default_factory=dict)
+    # None in case of sparse vectors:
+    vector_size: Optional[int] = None
+    distance: Optional[str] = None
 
 
-READER_TYPE = {"h5": AnnH5Reader, "jsonl": JSONReader, "tar": AnnCompoundReader}
+READER_TYPE = {
+    "h5": AnnH5Reader,
+    "jsonl": JSONReader,
+    "tar": AnnCompoundReader,
+    "sparse": SparseReader,
+}
 
 
 class Dataset:

dataset_reader/ann_compound_reader.py

@@ -33,6 +33,7 @@ def read_queries(self) -> Iterator[Query]:
                     vector /= np.linalg.norm(vector)
                 yield Query(
                     vector=vector.tolist(),
+                    sparse_vector=None,
                     meta_conditions=row_json["conditions"],
                     expected_result=row_json["closest_ids"],
                     expected_scores=row_json["closest_scores"],

dataset_reader/ann_h5_reader.py

@@ -22,6 +22,7 @@ def read_queries(self) -> Iterator[Query]:
                 vector /= np.linalg.norm(vector)
             yield Query(
                 vector=vector.tolist(),
+                sparse_vector=None,
                 meta_conditions=None,
                 expected_result=expected_result.tolist(),
                 expected_scores=expected_scores.tolist(),
@@ -33,7 +34,9 @@ def read_data(self) -> Iterator[Record]:
         for idx, vector in enumerate(data["train"]):
             if self.normalize:
                 vector /= np.linalg.norm(vector)
-            yield Record(id=idx, vector=vector.tolist(), metadata=None)
+            yield Record(
+                id=idx, vector=vector.tolist(), sparse_vector=None, metadata=None
+            )
 
 
 if __name__ == "__main__":

dataset_reader/base_reader.py

@@ -2,16 +2,24 @@
 from typing import Iterator, List, Optional
 
 
+@dataclass
+class SparseVector:
+    indices: List[int]
+    values: List[float]
+
+
 @dataclass
 class Record:
     id: int
-    vector: List[float]
+    vector: Optional[List[float]]
+    sparse_vector: Optional[SparseVector]
     metadata: Optional[dict]
 
 
 @dataclass
 class Query:
-    vector: List[float]
+    vector: Optional[List[float]]
+    sparse_vector: Optional[SparseVector]
     meta_conditions: Optional[dict]
     expected_result: Optional[List[int]]
     expected_scores: Optional[List[float]] = None

dataset_reader/json_reader.py

@@ -58,13 +58,18 @@ def read_queries(self) -> Iterator[Query]:
         ):
             # ToDo: add meta_conditions
 
-            yield Query(vector=vector, meta_conditions=None, expected_result=neighbours)
+            yield Query(
+                vector=vector,
+                sparse_vector=None,
+                meta_conditions=None,
+                expected_result=neighbours,
+            )
 
     def read_data(self) -> Iterator[Record]:
         for idx, (vector, payload) in enumerate(
             zip(self.read_vectors(), self.read_payloads())
         ):
-            yield Record(id=idx, vector=vector, metadata=payload)
+            yield Record(id=idx, vector=vector, sparse_vector=None, metadata=payload)
 
 
 if __name__ == "__main__":

dataset_reader/sparse_reader.py

@@ -0,0 +1,100 @@
+import os
+from pathlib import Path
+from typing import Iterator, List, Tuple, Union
+
+import numpy as np
+
+from dataset_reader.base_reader import BaseReader, Query, Record, SparseVector
+
+
+def read_sparse_matrix_fields(
+    filename: Union[Path, str]
+) -> Tuple[np.array, np.array, np.array]:
+    """Read the fields of a CSR matrix without instantiating it"""
+
+    with open(filename, "rb") as f:
+        sizes = np.fromfile(f, dtype="int64", count=3)
+        n_row, n_col, n_non_zero = sizes
+        index_pointer = np.fromfile(f, dtype="int64", count=n_row + 1)
+        assert n_non_zero == index_pointer[-1]
+        columns = np.fromfile(f, dtype="int32", count=n_non_zero)
+        assert np.all(columns >= 0) and np.all(columns < n_col)
+        values = np.fromfile(f, dtype="float32", count=n_non_zero)
+        return values, columns, index_pointer
+
+
+def csr_to_sparse_vectors(
+    values: List[float], columns: List[int], index_pointer: List[int]
+) -> Iterator[SparseVector]:
+    num_rows = len(index_pointer) - 1
+
+    for i in range(num_rows):
+        start = index_pointer[i]
+        end = index_pointer[i + 1]
+        row_values, row_indices = [], []
+        for j in range(start, end):
+            row_values.append(values[j])
+            row_indices.append(columns[j])
+        yield SparseVector(indices=row_indices, values=row_values)
+
+
+def read_csr_matrix(filename: Union[Path, str]) -> Iterator[SparseVector]:
+    """Read a CSR matrix in spmat format"""
+    values, columns, index_pointer = read_sparse_matrix_fields(filename)
+    values = values.tolist()
+    columns = columns.tolist()
+    index_pointer = index_pointer.tolist()
+
+    yield from csr_to_sparse_vectors(values, columns, index_pointer)
+
+
+def knn_result_read(
+    filename: Union[Path, str]
+) -> Tuple[List[List[int]], List[List[float]]]:
+    n, d = map(int, np.fromfile(filename, dtype="uint32", count=2))
+    assert os.stat(filename).st_size == 8 + n * d * (4 + 4)
+    with open(filename, "rb") as f:
+        f.seek(4 + 4)
+        ids = np.fromfile(f, dtype="int32", count=n * d).reshape(n, d).tolist()
+        scores = np.fromfile(f, dtype="float32", count=n * d).reshape(n, d).tolist()
+    return ids, scores
+
+
+class SparseReader(BaseReader):
+    def __init__(self, path, normalize=False):
+        self.path = path
+        self.normalize = normalize
+
+    def read_queries(self) -> Iterator[Query]:
+        queries_path = self.path / "queries.csr"
+        X = read_csr_matrix(queries_path)
+
+        gt_path = self.path / "results.gt"
+        gt_indices, _ = knn_result_read(gt_path)
+
+        for i, sparse_vector in enumerate(X):
+            yield Query(
+                vector=None,
+                sparse_vector=sparse_vector,
+                meta_conditions=None,
+                expected_result=gt_indices[i],
+            )
+
+    def read_data(self) -> Iterator[Record]:
+        data_path = self.path / "data.csr"
+        X = read_csr_matrix(data_path)
+
+        for i, sparse_vector in enumerate(X):
+            yield Record(id=i, vector=None, sparse_vector=sparse_vector, metadata=None)
+
+
+if __name__ == "__main__":
+    vals = [1, 3, 2, 3, 6, 4, 5]
+    cols = [0, 2, 2, 1, 3, 0, 2]
+    pointers = [0, 2, 3, 5, 7]
+    vecs = [vec for vec in csr_to_sparse_vectors(vals, cols, pointers)]
+
+    assert vecs[0] == SparseVector(indices=[0, 2], values=[1, 3])
+    assert vecs[1] == SparseVector(indices=[2], values=[2])
+    assert vecs[2] == SparseVector(indices=[1, 3], values=[3, 6])
+    assert vecs[3] == SparseVector(indices=[0, 2], values=[4, 5])

datasets/datasets.json

@@ -66,6 +66,18 @@
     "path": "dbpedia-openai-1M-1536-angular/dbpedia_openai_1M",
     "link": "https://storage.googleapis.com/ann-filtered-benchmark/datasets/dbpedia_openai_1M.tgz"
   },
+  {
+    "name": "msmarco-sparse-100K",
+    "type": "sparse",
+    "path": "msmarco-sparse/100K",
+    "link": "https://storage.googleapis.com/ann-filtered-benchmark/datasets/msmacro-sparse-100K.tar.gz"
+  },
+  {
+    "name": "msmarco-sparse-1M",
+    "type": "sparse",
+    "path": "msmarco-sparse/1M",
+    "link": "https://storage.googleapis.com/ann-filtered-benchmark/datasets/msmacro-sparse-1M.tar.gz"
+  },
   {
     "name": "h-and-m-2048-angular-filters",
     "vector_size": 2048,

engine/base_client/__init__.py

@@ -6,3 +6,12 @@
 
 class IncompatibilityError(Exception):
     pass
+
+
+__all__ = [
+    "BaseClient",
+    "BaseConfigurator",
+    "BaseSearcher",
+    "BaseUploader",
+    "IncompatibilityError",
+]