e5-base-v2-code-search (v9-200k)

A fine-tuned code search embedding model based on intfloat/e5-base-v2 (110M parameters). Trained with call-graph false-negative filtering on 200K balanced pairs across 9 programming languages.

Built for cqs — code intelligence and RAG for AI agents.

Key Results

Eval	Metric	Score
Pipeline (55 confusable functions, enriched)	R@1	94.5%
Pipeline	MRR	0.966
Raw code embedding (no enrichment)	R@1	70.9%
CodeSearchNet (6 languages)	NDCG@10	0.615

The 94.5% pipeline score ties BGE-large (335M) at 1/3 the parameter count. The 70.9% raw R@1 exceeds BGE-large (61.8%).

Training

• Base model: intfloat/e5-base-v2 (110M params, 768 dimensions)
• Data: 200K balanced pairs (22,222 per language × 9 languages) from cqs-indexed Stack repos
• Key technique: Call-graph false-negative filtering — uses code structure (caller/callee relationships) to exclude structurally related functions from contrastive negatives. Zero API cost (SQLite lookup).
• Loss: CachedGISTEmbedLoss + MatryoshkaLoss (768/384/192/128 dims)
• LoRA: rank 16, alpha 32, targets: query, key, value, dense
• Epochs: 1 (more epochs degrades enrichment compatibility)
• Dataset: jamie8johnson/cqs-code-search-200k

The 89.1% Basin

Six independent perturbations from this configuration (more data, less data, FAISS hard negatives, more epochs, contrastive query augmentation) all produce exactly -5.4pp pipeline regression to 89.1%. The 94.5% result appears to occupy a narrow peak in the loss landscape around ~22K examples per language with CG filtering.

Usage with cqs

# Default model in cqs v1.9.0+
cqs init && cqs index

# Or specify explicitly
export CQS_EMBEDDING_MODEL=e5-base
cqs index

Usage with sentence-transformers

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("jamie8johnson/e5-base-v2-code-search")
query_emb = model.encode("query: find functions that validate email addresses")
code_emb = model.encode("passage: def validate_email(addr): ...")

Languages

Go, Java, JavaScript, PHP, Python, Ruby, Rust, TypeScript, C++

ONNX

Includes model.onnx for inference with ONNX Runtime (used by cqs for local GPU/CPU inference).

Citation

Paper in preparation. See research log for methodology.