Dashboard › opencode-lore › Distillation
Distillation
ID: ac5357a9-3ad2-4051-a99e-7ae66d6cfaf7
Generation: 0
Tokens: 595
R_compression: 7.559
C_norm: 0.000
Archived: No
Created: 2026-09-17 10:21:14
Source IDs:
["lore_tm_v1_gD2pkgn3_VoXw0MPdMsbIiR8hRbZyU3hhPETMaDzcsw","lore_tm_v1_61JT_LSdg6zTwDZQxMzUwuKeHutwDgFZ22ZwJ1UabzA"]
Observations
Date: Sep 17, 2026
- 🔴 (09:49) User stated the adaptive embedding cap never exceeds the Nomic v1.5 maximum sequence length.
- 🔴 (09:49) User stated a fixed memory limit never improves, so the embedding-cap logic never re-probes under that condition.
- 🔴 (09:49) User stated persisted free-memory baseline handling can cause a memory-rich reboot to never re-probe back up to a cap that the WASM heap has.
- 🔴 (09:49) User reiterated embedding worker admission is never below 1: the primary worker always runs.
- 🔴 (09:49) User stated the ×0.7 OOM backoff never fires for an uncatchable WASM out-of-memory event because the WASM path self-limited beforehand.
- 🟡 (09:49)
packages/core/src/embedding-cap.ts defines MIN_EMBED_TOKENS = 256.
- 🟡 (09:49) Embedding-cap source documentation says each OOM lowers the cap by multiplying it by
0.7, then respawns the worker on a fresh heap.
- 🟡 (09:49) Embedding memory sizing documentation records a measured baseline of approximately
665 MB, subtracting the model footprint from free memory before sizing transient allocation and relying on backoff to correct residual error.
- 🟡 (09:49) Pool admission computes how many workers fit in
freeBytes, using Math.floor(free / PER_WORKER_MEM_BUDGET_BYTES), while preserving the primary-worker floor of 1.
- 🟡 (09:49) Source documentation warns that inside a memory-capped container,
os.freemem() reports host memory rather than container memory; the associated adjustment is monotonic and can only reduce the relevant bound.
- 🟡 (09:49) Current-headroom selection uses
Math.min(limit, runtime) after reconciling Linux/container and runtime-available memory signals.
- 🟡 (09:49) The embedding-cap model clamps an attention-memory estimate computed as
Math.sqrt(budget / EMBED_ATTENTION_BYTES_PER_TOKEN_SQ).
- 🟡 (09:49) Source documentation states effective embedding context plateaus around
~2048 tokens, so capping around ~4962 tokens trades little useful context while reducing memory risk.
- 🟡 (09:49) Persisted-cap reconciliation uses
ratio <= 1 + EMBED_CAP_TRUST_BAND as a trust-band condition and guards against stale caps learned before a newer ceiling existed, exemplified by an old cap of 7000.
- 🟡 (09:49) Upward re-probing is monotonic: it never steps down; its result cannot exceed the current ceiling even if passed an above-ceiling
cap.