Similarity ranking for Retrieval-Augmented Generation

Meta

✨ What is SymRank?

SymRank is a blazing-fast Python library for top-k cosine similarity ranking, designed for vector search, retrieval-augmented generation (RAG), and embedding-based matching.

Built with a Rust + SIMD backend, it offers the speed of native code with the ease of Python.

🚀 Why SymRank?

⚡ Fast: SIMD-accelerated cosine scoring with adaptive parallelism

🧠 Smart: Automatically selects serial or parallel mode based on workload

🔢 Top-K optimized: Efficient inlined heap selection (no full sort overhead)

🐍 Pythonic: Easy-to-use Python API

🦀 Powered by Rust: Safe, high-performance core engine

📉 Memory Efficient: Supports batching for speed and to reduce memory footprint

Below are single-query cosine similarity benchmarks comparing SymRank to NumPy and scikit-learn across realistic re-ranking candidate sizes.

Candidates (N)	SymRank matrix (ms)	NumPy normalized (ms)	sklearn (ms)	Fastest	SymRank Speedup
20	0.006	0.027	0.210	SymRank	4.50x
50	0.017	0.050	0.266	SymRank	2.92x
100	0.020	0.086	0.390	SymRank	4.18x
500	0.169	0.393	1.843	SymRank	2.32x
1,000	0.170	0.669	3.588	SymRank	3.95x
5,000	0.748	5.261	32.196	SymRank	7.03x
10,000	1.976	13.938	42.514	SymRank	7.05x

• Cosine similarity top k (k=5), embedding dimension 1536, float32.
• NumPy baseline uses candidates normalized once and query normalized per call.
• sklearn uses sklearn.metrics.pairwise.cosine_similarity.
• Times are mean milliseconds per query on Windows.

Real-world benchmark (Hugging Face embeddings)

Performance on 10,000 real OpenAI-style embeddings streamed from the Hugging Face Hub.

Method	Mean time (ms)	Speedup vs SymRank
SymRank matrix	1.800	1.0x
SymRank list	22.753	12.64x
NumPy normalized	44.665	24.81x
sklearn	42.709	23.73x

• Dataset: Qdrant/dbpedia-entities-openai3-text-embedding-3-large-1536-1M
• k=5, float32, Windows
• Mean time per query

📦 Installation

You can install SymRank with 'uv' or alternatively using 'pip'.

Recommended (with uv)

uv pip install symrank

Alternatively (using pip)

pip install symrank

🧪 Usage

SymRank provides two APIs optimized for different workflows.

---

Option 1: cosine_similarity_matrix (recommended for performance)

Best when:

• Candidate embeddings are already stored as a single 2D NumPy array
• Performance matters (about 10 to 14x faster for N=1,000 to 10,000 versus the list API)
• Running many queries against the same candidate set

import numpy as np
from symrank import cosine_similarity_matrix

# Example data (dimension = 4 for readability)
query = np.array([1.0, 0.0, 0.0, 0.0], dtype=np.float32)

candidate_matrix = np.array(
    [
        [1.0, 0.0, 0.0, 0.0],  # identical to query
        [0.0, 1.0, 0.0, 0.0],  # orthogonal
        [0.5, 0.5, 0.0, 0.0],  # partially aligned
        [0.2, 0.1, 0.0, 0.0],  # weakly aligned
    ],
    dtype=np.float32,
)

ids = ["doc_a", "doc_b", "doc_c", "doc_d"]

results = cosine_similarity_matrix(query, candidate_matrix, ids, k=3)
print(results)

Output:

[
  {"id": "doc_a", "score": 1.0},
  {"id": "doc_d", "score": 0.8944272},
  {"id": "doc_c", "score": 0.70710677}
]

Notes:

• Scores are cosine similarity (range -1 to 1, higher = more similar)
• Results are sorted by descending similarity

Typical production usage (1536-dimensional embeddings):

import numpy as np
from symrank import cosine_similarity_matrix

D = 1536
N = 10_000

query = np.random.rand(D).astype(np.float32)
candidate_matrix = np.random.rand(N, D).astype(np.float32)
ids = [f"doc_{i}" for i in range(N)]

top5 = cosine_similarity_matrix(query, candidate_matrix, ids, k=5)
for result in top5:
    print(f"{result['id']}: {result['score']:.4f}")

Optional batching for memory control:

# Process 10k candidates in batches of 2000
results = cosine_similarity_matrix(
    query, candidate_matrix, ids, k=5, batch_size=2000
)

---

Option 2: cosine_similarity (flexible and convenient)

Best when:

• Candidates come from mixed or streaming sources
• Vectors are naturally represented as (id, vector) pairs
• Simplicity is more important than maximum throughput

Basic example using Python lists:

import symrank as sr

query = [0.1, 0.2, 0.3, 0.4]
candidates = [
    ("doc_1", [0.1, 0.2, 0.3, 0.5]),
    ("doc_2", [0.9, 0.1, 0.2, 0.1]),
    ("doc_3", [0.0, 0.0, 0.0, 1.0]),
]

results = sr.cosine_similarity(query, candidates, k=2)
print(results)

Output:

[
  {"id": "doc_1", "score": 0.9939991235733032},
  {"id": "doc_3", "score": 0.7302967309951782}
]

Basic example using NumPy arrays:

import symrank as sr
import numpy as np

query = np.array([0.1, 0.2, 0.3, 0.4], dtype=np.float32)
candidates = [
    ("doc_1", np.array([0.1, 0.2, 0.3, 0.5], dtype=np.float32)),
    ("doc_2", np.array([0.9, 0.1, 0.2, 0.1], dtype=np.float32)),
    ("doc_3", np.array([0.0, 0.0, 0.0, 1.0], dtype=np.float32)),
]

results = sr.cosine_similarity(query, candidates, k=2)
print(results)

Output:

[
  {"id": "doc_1", "score": 0.9939991235733032},
  {"id": "doc_3", "score": 0.7302967309951782}
]

Optional batching:

results = sr.cosine_similarity(query, candidates, k=5, batch_size=1000)

---

Performance Comparison

Dataset Size	Option 1 (matrix)	Option 2 (list)	Speedup
N=100	0.02 ms	0.06 ms	3.3x
N=1,000	0.18 ms	2.28 ms	12.7x
N=10,000	1.50 ms	19.66 ms	13.1x

Benchmark: 1536-dimensional embeddings, k=5, Python 3.14, Windows. Benchmark includes Python-side overhead for each API.

---

Quick Decision Guide

Use cosine_similarity_matrix if:

• ✅ You have a pre-built NumPy matrix of candidates
• ✅ Performance is critical
• ✅ Processing many queries against the same corpus

Use cosine_similarity if:

• ✅ Building candidates on-the-fly
• ✅ Mixed vector input types (lists or NumPy arrays)
• ✅ Flexibility > raw speed

Both functions return the same format: a list of dicts sorted by descending similarity score.

🧩 API: cosine_similarity(...)

cosine_similarity(
    query_vector,              # List[float] or np.ndarray
    candidate_vectors,         # List[Tuple[str, List[float] or np.ndarray]]
    k=5,                       # Number of top results to return
    batch_size=None            # Optional: set for memory-efficient batching
)

'cosine_similarity(...)' Parameters

Parameter	Type	Default	Description
query_vector	list[float] or np.ndarray	required	The query vector you want to compare against the candidate vectors.
candidate_vectors	list[tuple[str, list[float] or np.ndarray]]	required	List of (id, vector) pairs. Each vector can be a list or NumPy array.
k	int	5	Number of top results to return, sorted by descending similarity.
batch_size	int or None	None	Optional batch size to reduce memory usage. If None, uses SIMD directly.

Returns

List of dictionaries with id and score (cosine similarity), sorted by descending similarity:

[{"id": "doc_42", "score": 0.8763}, {"id": "doc_17", "score": 0.8451}, ...]

📄 License

This project is licensed under the Apache License 2.0.