YantrikDB

The v0.7.0 milestone bundles 10 commits since v0.6.7 into the engine release that closes the wedge fix structurally (Phase 4.3 = SQL writes off foreground) and ships a real out-of-the-box embedding story (bundled potion-base-2M static embedder, ~7.9MB).

yantrikdb-server can now bump its yantrikdb dep to "0.7" — the cluster API surface (issue #9 primitives + Phase 6 RYW) and the wedge-fix architecture (Phases 1-5 in v0.6.6, 4.3 in this release) are both production-shaped.

Highlights

Phase 4.3 — SQL writes off foreground

Foreground record() / record_with_rid() now do only fixed-cost SQL (memories INSERT + session UPDATE) before returning. The unbounded entity / memory_entities / claims / graph_index loops that previously scaled with extracted_entities.len() run on the materializer thread.

• Wedge primitive #2 closed at the engine layer
• Composes with yantrikdb-server v0.8.14's read pool for the full multi-process story
• Synchronous read-after-write of entity-graph queries shifts from instant to ms-scale; delta-recall stays synchronous

Bundled default embedder (Slices A + B)

pip install yantrikdb now works with real semantic embeddings out of the box. No pip install sentence-transformers. No first-run HuggingFace download. No ONNX, no candle, no GPU.

db = yantrikdb.YantrikDB.with_default("memory.db")
db.record_text("Alice is the engineering lead", ...)
results = db.recall_text("who leads the team?", top_k=3)

The bundled default is potion-base-2M — a model2vec distillation of bge-base-en-v1.5. Token lookup → mean-pool → L2-normalize, ~500× faster than transformer inference, ~10× smaller than MiniLM.

Empirical quality on yantrikdb-shaped memories (eval at scratch/eval_potion_2m.py):

| Embedder | R@5 | R@10 | MRR | Size |
|---|---|---|---|---|
| hash-trick (Slice A foundation) | 0.75 | 0.85 | 0.71 | 0 |
| potion-base-2M (default) | 0.90 | 0.90 | 0.78 | ~8 MB |
| all-MiniLM-L6-v2 (gold) | 0.95 | 1.00 | 0.90 | ~80 MB |

Slim deployments: yantrikdb = { version = "0.7", default-features = false } drops the embedder + its deps.

Compactor cycle 1s → 250ms

Smaller per-cycle pauses for sustained-load p99. Same total clone work, 4× more cycles. Idle deployments still pay near-zero CPU — should_compact() short-circuits when the delta is empty. Options 3/4 (LSM-style CoW, adaptive cadence) deferred to v0.7.x pending Phase 7 soak empirical signal.

Engine pressure surface (cross-stack rule)

New public getters: db.delta_max(), db.delta_len(), db.cold_len(). yantrikdb-server's tick loop uses these to scale its enrichment_pause_threshold proportionally to engine capacity. CONCURRENCY.md documents the cross-stack rule: engine pressure suppresses enrichment, NEVER decay (memory aging is wall-clock-dependent; pausing creates timeline divergence on the same logical clock = correctness bug).

Architecture diagrams

README.md gains two Mermaid diagrams covering the decoupled write path and cluster RYW.

What's in this release (commits 11aa937..45e11e2 on main)

9182526 release: v0.7.0
45e11e2 feat(embedder): Slice B — bundled potion-base-2M static embedder
dd94ace feat(embedder): bundled default embedder (Slice A foundation)
876cecb feat(record_with_rid): Phase 4.3 Commit C — flip cluster path
d289063 feat(record): Phase 4.3 Commit B — flip record() to enqueue
cf45e12 perf(compactor): COMPACTOR_INTERVAL 1s -> 250ms
a7c1f0a feat(materializer): Phase 4.3 dispatch arm
93126c3 docs(readme): Mermaid fix + crate version bump
99f943f docs(readme): architecture diagrams
9077e05 ops(wedge_repro): RSS + delta-pressure instrumentation
bcf1c78 feat(engine): public delta_max/delta_len/cold_len getters

Migration notes

• Existing dim=384 deployments using db.set_embedder(MiniLM) are unaffected — that path stays. Only YantrikDB::new(path, dim) auto-attach behavior changes.
• Use YantrikDB::with_default(path) to stay bundle-dim agnostic.
• Schema unchanged from v0.6.7 (still v25). No migration needed.
• Slim builds (--no-default-features) lose the bundled embedder; must set_embedder() explicitly.

Tests

1416 lib tests pass. Slim build (--no-default-features) clean.

Still pending (post-v0.7.0)

• Phase 7 long-duration soak — runs on Linux homelab; validates v0.7.0's read p99 fix empirically against the CT 168 production load profile.
• Compactor scaling Options 3/4 — LSM-style CoW or adaptive cadence; wait on soak data.
• Slice C (db.set_embedder_named("potion-base-8M") runtime download from yantrikos/yantrikdb-models) — additive feature, queued for v0.7.x.
• Per-namespace fairness scoring — blocked on Phase 4.x multi-namespace HNSW.

Anti-goals respected

• No ort / ONNX runtime (retired class of bugs from runtime version-mismatch hell).
• No candle dep — pure Rust static-embedding inference via model2vec-rs.
• No first-run model download — files committed to repo, baked via include_bytes!.
• No mandatory pip install sentence-transformers for record_text to work.