Tagvico AI v2.0.0

Tagvico AI v2.0.0 is the first stable release of the self-hosted, reviewable
AI filing workflow for Paperless-ngx.

Highlights

Queue durable metadata suggestions for review or enable validated automatic
writes per installation.
Apply or reject structured diffs and restore the first metadata snapshot,
including exact tag, correspondent, document-type, date, language, custom
field, and owner values.
Constrain generated tags with controlled groups, per-document caps, and an
exception queue.
Use Ollama, Ollama Cloud, OpenAI, Anthropic, OpenRouter, OpenCode Go, Azure
OpenAI, GitHub Copilot, OpenAI-compatible gateways, or experimental ChatGPT
subscription access.
Recover OCR-poor documents, interrupted work, retries, and terminal failures
from durable queues.
Inspect processing history, token usage, estimated cost, runner state, and
provider health from the web interface.
Optionally share a coarse installation heartbeat. Analytics are off by
default, the exact payload is locally previewable, identifiers rotate, and
no official collector endpoint is embedded.

Stop Tagvico and back up the complete tagvico_ai_data volume.
Replace deprecated pre-Tagvico environment variables with their documented
TAGVICO_* equivalents.
Pin ghcr.io/arturict/tagvico-ai:2.0.0 in Compose.
Start the container and verify both /health and /api/health.
Sign in, test the Paperless and model-provider connections, and process a
representative non-sensitive document in Review first.
Enable Automatic mode only after reviewing the resulting metadata.

Do not run two Tagvico versions against the same data volume. Restore the
volume backup before rolling back if a newer version migrated the database.

The v2 Docker, persistent-volume, canonical environment-variable, migration,
setup, review, history, and restore contracts are stable.
ChatGPT subscription access remains experimental and account-specific.
Model catalogs, pricing, quotas, and entitlements remain provider-controlled.
Compatible and subscription-backed adapters may report provider health as
unknown; use the explicit connection tests in Settings before processing.