Achilles — Admin: Knowledge Store

Storage state

graph metrics

achilles.local/admin/knowledge-store /admin/knowledge-store · Owner / Admin

Knowledge store

All sources ▾ 1

1.24M Entities graph nodes · entities

4.81M Chunks chunks · vector + lexical

918k Edges edges · entity_edge

3.1k Unresolved edges 2 refs · entity_ref

38 GB Vector volume pgvector · HNSW

1 A per-source filter on the counts — each source's contribution; the whole graph by default.
2 Edge refs whose second node has not arrived yet — closed by edge materialization. A growing tail signals an under-collected graph, not an error.

Graph curation → Knowledge Store

curation run · re-embedding in progress

Curation run 1

running · 62% model change 2 started 8 min ago Cancel run 3

Re-embedding chunks AI — 8.2k of 13.1k 4 Embedding 3 Large → Embedding 3 Small

Search keeps working during the recompute. 5

Curation schedule

When to run the platform pass heavy curation is kept to quiet hours · organization time (Europe/Moscow) 6

Frequency

Daily ▾

Time

04:00 ▾

↳ every day at 04:00 · next run — today at 04:00

run · scheduled (destructive window)

running · 41% scheduled Cancel

edges ✓ merge ● decay retention v2 Destructive window (duplicate merge) · sync of affected sources is queued 7

idle · last run

succeeded today 04:00 · in 6 min

edges +1.2k · merged 84 · demoted 320 next scheduled — tomorrow 04:00 8

1 Kicks off a one-off run outside the schedule. There is one run per platform — while the current one is going, the button is disabled (curation_runs).
2 The trigger chip — scheduled, manual, or an embedding-model change.
3 Cancelling a re-embedding is safe: until the atomic swap the old index is used, while the new vectors are built in a separate column alongside (refresh). The run is expensive, so it is behind confirmation.
4 A shared activity indicator: which step is running now and how far it has gone. It is referenced by lane coordination.
5 Re-embedding is triggered by a change of embedding model. The mechanics — embedding refresh.
6 A single platform schedule, not per-source: frequency and time. Stored in platform_settings.
7 The run's steps are shown as chips. Only the destructive window (duplicate merge and retention) pauses writes from affected sources.
8 When idle — the outcome of the last run and the time of the next scheduled one.

Retention policy v2

What to keep, archive or delete by entity type and age — deferred to v2. The mechanics — Knowledge Store · Lifecycle.

retention by type · v2 sketch

Entity type	Keep	Archive after	Delete after
Chat messages	forever ▾	1 year ▾	—
Documents	forever ▾	—	—
Tasks / tickets	forever ▾	2 years ▾	5 years ▾

Sketch — the set of types and the age units settle during v2 design.

Backup & restore

Destination, cadence and retention of snapshots; restore starts from here too. The mechanics — Knowledge Store · Lifecycle.

backups · schedule · restore

Destination and schedule

External storage S3-compatible (incl. MinIO) · write-only key, empty = IAM role

s3://acme-achilles-backups ▾

••••a3f7

Cadence when to take a full database dump

Daily ▾

02:00 ▾

Snapshots to keep older ones rotate out as new ones accrue

14 ▾

Latest snapshots

ready today 02:00 · 12.4 GB Restore 1

ready yesterday 02:00 · 12.3 GB Restore

1 A restore overwrites the current database with the snapshot — irreversible, so it is behind type-to-confirm. For the duration of the restore the platform goes into maintenance.