YantrikDB

Honors the bundled-embedder distribution narrative. Diagnosed by yantrikdb-hermes-agent v0.2 plugin smoke-test (swarm msg 1c8c78a8 → response f5c82ccf). NO Rust code changed; manifest-only release.

The contradiction this fixes

v0.7.4's YantrikDB.with_default() docstring promised:

Without any external sentence-transformers install or first-run model download.

But pyproject.toml listed sentence-transformers>=3.0.0 + numpy>=1.26.0 as hard dependencies. So a default pip install yantrikdb pulled torch 2.11.0, transformers 5.8.0, sentence-transformers 5.4.1, scipy, scikit-learn — easily 2 GB of disk and 90 seconds of install time — even though the documented Just-Works path uses the bundled potion-base-2M embedder running entirely in Rust.

Audit before fix

• numpy: imported NOWHERE in src/yantrikdb/. Got dragged in as a transitive of sentence-transformers, listed as a direct dep without checking. Dropped entirely.
• sentence-transformers: imported in 5 files but ALL inside function bodies (lazy imports). Only executes on yantrikdb-server startup, yantrikdb CLI with --model, MCP server startup, or eval scripts. Default from yantrikdb import YantrikDB never touches it. Moved to [sentence-transformers] extras.

Net pip install experience after v0.7.6

| Command | Pulls | Disk | Install time |
|---------|-------|------|--------------|
| pip install yantrikdb | wheel + uuid-utils + click | ~10 MB | ~2s |
| pip install yantrikdb[sentence-transformers] | + torch + transformers + scipy + sklearn | ~2 GB | ~90s |
| pip install yantrikdb[all] | sentence-transformers + server + sync + langchain + openai + agent | ~2 GB | ~90s |

The default install is now what the bundled-embedder narrative claims: a ~10 MB wheel that finishes installing in seconds and gives you working record_text() / recall_text() against a Rust-native potion-base-2M embedder out of the box. Users who BYO sentence-transformers opt into the heavy path explicitly.

Compatibility

Pure-additive for the [all] extra path — existing pip install yantrikdb[all] users keep the old behavior. Only breaking change is for users who relied on pip install yantrikdb to also install sentence-transformers; they need to add [sentence-transformers] to the spec.

Empirical baseline (Hermes plugin smoke-test, v0.7.4 wheel, Windows 3.13)

Numbers Hermes captured pre-fix that motivated the change:

• with_default() cold start: 77.5 ms
• record_text: 18.88 ms each (embed + write)
• recall_text: 2.24 ms with semantic top-hit on dim=64

Compare to HTTP path against a homelab cluster: 13–24 ms p50 per op over LAN. Embedded is 8–10× faster on writes, ~10× faster on recall, with no leader-failover or token-replication concerns. v0.7.6 makes that experience start with a 2-second pip install.

🤖 Generated with Claude Code