[ferhimedamine]

Summary

Chart 0.12.4 for Dakera server v0.12.4 (Dashboard stays 0.5.2): the version bump, the server's new settings as chart values, new resource defaults, a checkpoint mode for the rollback Job, and the upgrade docs.

Merge order: after the engine's 0.12.4 release has published ghcr.io/dakera-ai/dakera:0.12.4, together with dakera-deploy. Merging publishes the chart. This hand-written release PR sets 0.12.4, so the sync-on-release workflow opens no duplicate.

Changes

Version

• version / appVersion 0.12.3 -> 0.12.4. The image tag follows appVersion.
• artifacthub images and changes annotations.
• CHANGELOG 0.12.4 section.
• README version tables, install examples and "Upgrading from chart 0.12.3".

Server 0.12.4 settings (dakera.config.*)

• Each is rendered only when set, so an image up to 0.12.3 is never handed an unknown name, and each is checked by values.schema.json:
  • contactEmail (DAKERA_CONTACT_EMAIL, must be an e-mail address)
  • defaultEntityTypes
  • glinerModel: one of the five models
  • entityTagging, standingInstructions, forgetRequests
  • recallImportanceWeighted, recallIncludeAssociated, rerankHead
  • autopilotConsolidationMode
• The chart refuses a glinerModel that does not read the embedding model's languages, as the server refuses it on a fresh install. "Multilingual" means features.multilingual, or DAKERA_MODEL set to bge-m3 / multilingual-minilm in extraEnv.
• values.yaml describes each setting and the value that restores 0.12.3's behaviour. It also lists the other defaults that moved (normalisation, long-content windows, bi-temporal ingest, autopilot dedup) under extraEnv.

Sizing

• Server limits go from 2 CPUs / 4Gi to 4 CPUs / 6Gi: embedding and the reranker want at least 4 CPUs (the server warns below that, config_warnings: sizing), and the default models hold about 3.5 GiB.
• New features.multilingual.resources (8 CPUs / 8Gi), because bge-m3 wants 8 CPUs. It is chosen before multimodal.resources, so multilingual + multimodal keeps 8 CPUs; vision still wins (it excludes multilingual anyway).

Rollback

• New rollback.command:
  • downgrade: the default, unchanged, to v0.11.108;
  • checkpoint: dakera checkpoint as a one-off Job, required before going back from 0.12.4 to an earlier 0.12 chart (an image swap).
• New README section "Rolling back to an earlier v0.12".

NOTES

• On upgrade: the 0.12.4 changes (entity extraction on by default plus the background sweep, the GLiNER model in use, ranking and autopilot defaults) and how to go back.
• A warning when the data root is an emptyDir (dakera.persistence.enabled: false, the server's ephemeral_data_dir warning).
• The optional contact address.

Models

• The image (about 2.65 GB) ships GLiNER and the forget/instruction encoder.
• Downloads come from huggingface.co/DakeraAI first; HF_ENDPOINT serves both sources.
• models.pull: [configured] does not include GLiNER (noted in the docs).

validate-chart.sh gains checks for:

• the new settings: left out by default, rendered when set;
• GLiNER pairs that are accepted and refused, and the other refused values (contact address, modes, entity labels, rerank head, rollback command);
• the 4 CPU / 6Gi defaults and 8 CPUs with multilingual (also beside multimodal);
• the checkpoint Job.

Testing

• values.yaml and Chart.yaml parse.
• The defaults validate against values.schema.json (Python jsonschema, draft 7). 14 hand cases (good and bad e-mail, GLiNER names, modes, labels, rerank head, rollback command) are accepted or refused as intended.
• Template blocks are balanced in every edited template.
• Not run locally: helm lint, helm template and scripts/validate-chart.sh. Helm is not installed on the host I worked from, and downloading it was not allowed there. Chart CI runs scripts/validate-chart.sh (helm 3.14.4) on this PR: please check that run before marking it ready.

Update (2026-11-04): Dashboard 0.5.3

• dashboard.image.tag 0.5.3 (the dashboard released with Dakera 0.12.4; Chart.yaml images, READMEs, CHANGELOG).
• New dashboard.telemetry (default true): false renders DAKERA_TELEMETRY=off in the dashboard pod. Schema: boolean. validate-chart.sh checks the image, nothing rendered by default, the false case, and refuses a non-boolean.
• Validated (2026-11-04) with Helm 3.17.3 (alpine/helm container): helm lint clean, scripts/validate-chart.sh passed, including the new cases (dashboard image 0.5.3, telemetry rendered only when off, a non-boolean refused). The chart CI job itself is still refused by GitHub (billing lock).

🤖 Generated with Claude Code