0.1.17

0.1.17: a failed reindex no longer turns memory off

The log was decisive: failing embeds took 5-10 s (COLD load after an
eviction) while a working one took 75 ms. Search was healthy; the
background reindex was dying — and its catch disabled the whole
memory even though the on-disk index was intact. Now only a MISSING
model (404) turns it off.

Also: keep_alive 30m on embeds (the model is 640 MB and warm does 80
chunks in 0.77 s, so the load window — exactly when chat evicts it —
nearly disappears), batches of 8 to bound the blast radius, and
buscar now THROWS instead of returning empty, because a silent []
made the journal answer "I have no record of that": lying about what
it remembers is worse than saying it can't reach memory. The
interview and web server swallow it; the ask tab says it in words.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

0.1.16

0.1.16: drop keep_alive:0 + cancel ghost reindexing

The unload was the friendly-fire weapon: it killed MID-BATCH the
embeds of any other instance (another Obsidian window, a background
reindex of an abandoned instance) — hence tasks processing while the
API returned 400 pointing at a dead port. And server.log proved
Ollama respawns the runner on its own with each new request (a
400->200 happened 54 ms apart with no unload): plain retries with
growing waits are enough. Plus Rag.cancelar(): instances the plugin
replaces release Ollama instead of embedding in the background
forever.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

0.1.15

0.1.15: serialized embeds + staged retry

Live Ollama server.log: background reindexing and the user's query
ran CONCURRENT embeds and their keep_alive:0 remediations killed each
other's runner (respawn every 20-40s, everything 400); also a request
arriving while a runner unloads gets the dead ref (scheduler race —
the 400 always points at the old port). Fix: per-instance embed
mutex, re-entrancy guard in refrescarRag, and a 3-step retry (plain
retry -> unload -> final attempt).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

0.1.14

0.1.14: recover from a dead embeddings runner

Seen live with bonsai: after VRAM eviction, Ollama keeps routing
/api/embed to a DEAD llama-server (400 "connection refused" to the
internal port) for minutes, so a plain retry didn't save memory
either. Before retrying, embedRobusto sends keep_alive:0
(best-effort) to unregister the runner so the retry spawns a fresh
one.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

0.1.13

0.1.13: resilient memory + per-reply speaker button

- Rag: a transient embedding failure (Ollama restarting, VRAM churn
  with bonsai) no longer turns long-term memory off — it retries once,
  and if reindexing still dies, partial progress is saved so the next
  pass resumes where it left off.
- The "memory is off" message now shows the actual cause.
- Speaker button on every assistant reply (interview and ask tabs):
  read it aloud again, or silence the one currently playing. Only
  shown when the local voice sidecar is up.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

0.1.12

0.1.12: cap num_ctx at 8192 and disable thinking for bonsai

Unbounded, Ollama reserves the model's maximum context window (bonsai
advertises 262k → 14.7 GB RAM, CPU spill and 7.8 tok/s, measured). With
num_ctx 8192 + think:false it loads in 4.3 s at ~58 tok/s using ~1 GB.
The reasoning-model regex now covers bonsai (1-bit qwen3.6, thinks by
default).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

0.1.11

0.1.11: real download progress bar, wrapping first-run card, self-hea…

0.1.10

0.1.10: recommended-model catalog with Bonsai 27B

First-run card and settings now offer gemma3:4b, gemma3:12b and the
experimental bonsai-27b (27B at 1-bit in 4.4 GB, needs Ollama 0.32.5+).
Each entry pins its own extraction model — Bonsai interviews but
extraction stays on gemma3:4b — and the download fetches both if needed.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

0.1.9

Memory without a terminal: embeddinggemma auto-downloads (0.1.9)

refrescarRag ensures the embeddings model before indexing: if Ollama
is up but the model is missing, the plugin pulls it itself (visible
Notices, double-pull guard). Also fixes a latent bug where a Rag
instance that self-deactivated (model absent at startup) was never
recreated — memory now activates as soon as the model arrives,
no Obsidian restart needed. Settings/README no longer ask for a
manual 'ollama pull'.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

0.1.8

STT on clean machines: canonical ffmpeg name + self-updating sidecar …