💻 code-embed-v2

CodeEmbed v2 is a lightweight, state-of-the-art multilingual code embedding model (≤131M parameters) developed by Divakant Pandey. Specialized for semantic code search, code-to-code retrieval, text-to-SQL mapping, and repository navigation across 6 major programming languages with an extended 8,192 token context window and native Matryoshka Representation Learning (MRL).

---

🚀 Highlights & Functional Capabilities

• Multilingual Support: High-accuracy semantic search across Python, JavaScript, Go, Java, Ruby, and PHP.
• Extended Context: Full 8,192 token context window allowing whole-file, multi-function, and long class embeddings without aggressive truncation.
• Matryoshka Representation Learning (MRL): Embeddings can be dynamically truncated from 1536 down to 768, 384, 256, 128, or 64 dimensions with negligible loss in accuracy, enabling significant vector database memory and latency savings.
• Compact & Efficient: 131M parameters (**484 MB** safetensors footprint) optimized for low-latency CPU inference and fast GPU batch indexing.
• Universal Compatibility: Works out of the box with sentence-transformers, transformers, and PyTorch.

---

📊 Benchmark Performance

CodeEmbed v2 was thoroughly evaluated across 18 official benchmark tasks comprising the primary code retrieval suites on the Massive Text Embedding Benchmark (MTEB):

🏆 Benchmark Suite Macro-Averages (meanTask)

Benchmark Suite	Total Tasks	Metric	Official Macro-Average	Coverage
MTEB(Code, v1)	12 / 12	NDCG@10	40.64%	100% (Complete)
CoIR (Code IR)	10 / 10	NDCG@10	37.29%	100% (Complete)
CoREB(v1)	6 / 6	NDCG@10	27.11%	100% (Complete)

📋 Detailed Task-by-Task Scorecard

Task	Category	NDCG@10	MRR@10	Evaluation Scope
CodeSearchNetRetrieval	Text-to-Code Search	77.78%	73.85%	6 Languages (Python, JS, Go, Java, Ruby, PHP)
COIRCodeSearchNetRetrieval	Multilingual Search	74.43%	71.37%	Full-scale 300k+ codebase corpus
CorebC2TReranking	Code-to-Text Reranking	66.74%	57.59%	Cross-lingual code description ranking
StackOverflowQA	Developer Q&A Search	55.30%	51.59%	Real-world developer query matching
CodeFeedbackST	Instruction Feedback	46.93%	42.51%	Single-turn instruction and code feedback
CodeSearchNetCCRetrieval	Code-to-Code Search	41.70%	37.84%	Cross-language code function matching
CodeEditSearchRetrieval	Edit / Patch Search	37.03%	32.83%	13 Git commit and patch subsets
CorebC2CReranking	Code-to-Code Reranking	36.87%	34.40%	Direct code snippet similarity ranking
CodeTransOceanContest	Code Translation	36.40%	31.50%	Cross-language algorithmic translation
CodeTransOceanDL	Framework Translation	34.39%	22.04%	Deep learning API cross-matching
SyntheticText2SQL	Text-to-SQL	32.77%	25.13%	Natural language to structured SQL queries
CodeFeedbackMT	Multi-Turn Feedback	31.88%	28.82%	Conversational multi-turn code context
CorebT2CReranking	Text-to-Code Reranking	26.44%	27.65%	Natural language to code reranking
CorebC2TRetrieval	Code-to-Text Search	22.32%	18.83%	Source code to documentation retrieval
CosQA	Web Query Code Search	15.99%	11.35%	Real-world noisy web search queries
CorebT2CRetrieval	Text-to-Code Search	7.45%	8.10%	Zero-shot query to code retrieval
AppsRetrieval	Algorithmic Problem Search	3.06%	2.41%	Hard algorithmic programming specifications
CorebC2CRetrieval	Code-to-Code Retrieval	2.87%	2.25%	Direct code-to-code corpus retrieval

---

💻 Quickstart Usage

1. Basic Semantic Code Search with sentence-transformers

from sentence_transformers import SentenceTransformer

# Load model directly from Hugging Face Hub
model = SentenceTransformer("divakantpandey/code-embed-v2")

queries = [
    "binary search implementation",
    "fetch json from api endpoint",
    "goroutine channel worker pool"
]

code_snippets = [
    "def binary_search(arr, target):\n    left, right = 0, len(arr) - 1\n    while left <= right:\n        mid = (left + right) // 2\n        if arr[mid] == target: return mid\n        elif arr[mid] < target: left = mid + 1\n        else: right = mid - 1\n    return -1",
    "async function fetchJSON(url) {\n    const response = await fetch(url);\n    return await response.json();\n}",
    "func worker(id int, jobs <-chan int, results chan<- int) {\n    for j := range jobs {\n        results <- j * 2\n    }\n}"
]

# Generate normalized embeddings
query_embeddings = model.encode(queries, normalize_embeddings=True)
code_embeddings = model.encode(code_snippets, normalize_embeddings=True)

# Compute cosine similarity
similarity_matrix = query_embeddings @ code_embeddings.T
print(similarity_matrix)

2. Matryoshka Dimension Truncation (Adaptive Dimensions)

CodeEmbed v2 natively supports Matryoshka Representation Learning. You can truncate embeddings to smaller dimensions (e.g., 768, 384, 256) for faster vector database indexing with minimal accuracy impact:

# Truncate to 768 dimensions (50% storage & compute reduction)
compact_query_embs = model.encode(queries, truncate_dim=768, normalize_embeddings=True)
compact_code_embs = model.encode(code_snippets, truncate_dim=768, normalize_embeddings=True)

similarity = compact_query_embs @ compact_code_embs.T

---

📜 License & Commercial Terms

This model and its associated weights are licensed under the Creative Commons Attribution-NonCommercial 4.0 International Public License (CC-BY-NC 4.0).

• Non-Commercial Use: Free for academic research, educational use, personal exploration, open-source evaluations, and non-commercial benchmarking with appropriate attribution.
• Attribution: Please attribute Divakant Pandey and reference the repository at https://github.com/divakantpandey/code-embed.
• Commercial Licensing: To license CodeEmbed v2 for commercial products, enterprise SaaS applications, closed-source systems, or proprietary developer tools, please contact:
  • Licensor / Maintainer: Divakant Pandey
  • Email: divakant01@gmail.com
  • GitHub: @divakantpandey