prx 0.6.2

Praxis — agent-native Unix tools. Single binary replacing grep, cat, find, sed, diff for AI coding agents.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49

#!/usr/bin/env python3
"""Distill a code-specific embedding model into Model2Vec format for prx.

Usage:
    pip install model2vec
    python scripts/distill_model.py

Output:
    models/codesage-m2v-256/ directory with:
    - model.safetensors (the weights, ~8 MB)
    - tokenizer.json (the tokenizer)

These can be loaded by prx's DenseIndex via include_bytes! or file path.
"""

from model2vec import distill

MODELS = {
    "standard": {
        "source": "codesage/codesage-base-v2",
        "pca_dims": 256,
        "output": "models/codesage-m2v-256",
    },
    "large": {
        "source": "jinaai/jina-embeddings-v3",
        "pca_dims": 512,
        "output": "models/jina-code-m2v-512",
    },
}

if __name__ == "__main__":
    for tier, config in MODELS.items():
        print(f"Distilling {tier} tier: {config['source']} -> {config['output']}")
        print(f"  PCA dims: {config['pca_dims']}")

        model = distill(
            config["source"],
            pca_dims=config["pca_dims"],
        )

        model.save_pretrained(config["output"])
        print(f"  Saved to {config['output']}/")
        print()

    print("Done. Next steps:")
    print("  1. Check model sizes (standard ~8 MB, large ~30 MB)")
    print("  2. Run: prx bench-ndcg --model-path models/codesage-m2v-256/")
    print("  3. Run: prx bench-ndcg --model-path models/jina-code-m2v-512/")
    print("  4. Compare NDCG@10 against builtin baseline")