What's new

Release notes for the Community Edition, one entry perGitHub release. Since v2.1.0, every release also publishes the savvinaai/savvina-backend and savvinaai/savvina-frontend images to Docker Hub under the same tag.

git pull
docker compose pull
docker compose up -d --no-build

• v2.1.1September 22, 2026

  v2.1.1: Setup fixes

  • Appending LOCAL_UID and LOCAL_GID to .env on WSL and Linux no longer corrupts the file
  • The sample PostgreSQL database seeds its dates from CURRENT_DATE, so questions about this month return rows whichever day you start it
  • README: clearer instructions for the sample databases
• v2.1.0September 22, 2026

  v2.1.0: Pre-built images on Docker Hub

  • Every release now publishes multi-arch images (linux/amd64 and linux/arm64) to Docker Hub as savvinaai/savvina-backend and savvinaai/savvina-frontend — docker compose pull replaces the local build
  • SAVVINA_IMAGE_TAG pins a release; latest tracks the newest one
  • TLS moved out of the frontend container: the stack serves plain HTTP on APP_PORT and a reverse proxy of your choice terminates HTTPS, with BEHIND_TLS_PROXY and TRUSTED_LB_CIDR telling the backend and nginx it is there
  • Review findings closed and the custom-provider service list moved into its own module
• v2.0.0September 20, 2026

  v2.0.0: API keys in the UI only, settings rework

  • Breaking: LLM provider API keys are no longer read from environment variables. Enter them under Settings → LLM Providers; they are stored encrypted
  • GitHub Models provider removed
  • Settings page restructured: cache TTL is read live, cache and pruning toggles are staged until Save, and saved settings write through to the running app
  • Frontend dev dependencies updated to clear all nine npm audit advisories
  • Docs corrected where they disagreed with the code, and the unused provider config schema deleted
• v1.0.6August 27, 2026

  v1.0.6: Semantic model save fix

  • Adding a business metric no longer rejects the whole semantic model save with a 422
  • Frontend semantic types synced with the backend: segments, notes, table grain and domain, and the column intelligence fields are now visible
  • Thumbs-down feedback sends a semantic correction into a new Suggestions tab
  • sqlparse 0.6.0, failing closed on its new SQLParseError
  • CI: semantic-release restored under GitPython 3.1.60
• v1.0.5August 12, 2026

  v1.0.5: Version reporting, dependency patches

  • The app reports its real version
  • Security patches for cryptography, react-router and js-yaml
  • Node version aligned across the Compose files; the unused esbuild dependency dropped
  • Chart view no longer needs a container ref workaround
  • Docs: every endpoint documented, TRUSTED_PROXIES explained, frontend test suite described, and two inaccurate claims in the Docker guide corrected
• v1.0.4July 22, 2026

  v1.0.4: First-run setup unblocked

  • First-run setup no longer stalls
  • DATABASE_URL is derived from APP_DB_PASSWORD, so the bundled database needs one value in .env instead of two
• v1.0.3July 22, 2026

  v1.0.3: Dependency patches

  • Security patches for axios, brace-expansion and json-repair
  • README polished with a tagline and badges
• v1.0.2June 16, 2026

  v1.0.2: Semantic generation races closed

  • Concurrent batch generation of the semantic model no longer races
  • Security patches for form-data, js-yaml, ws, starlette, cryptography and python-multipart
• v1.0.1June 15, 2026

  v1.0.1: Faster Docker setup

  • ENCRYPTION_KEY and JWT_SECRET_KEY are generated on first boot instead of by hand
  • HF_TOKEN avoids anonymous rate limits when the build downloads the embedding model
  • Security patches for Vite and React
  • Contributor License Agreement check added to pull requests
• v1.0.0June 1, 2026

  v1.0.0: Initial release

  • Natural language to SQL with multi-turn conversation history
  • Semantic model layer — auto-generate business context from your schema, with display names, value mappings, metrics, and segments
  • Support for 8 LLM providers: Claude, OpenAI, Groq, Google Gemini, Cerebras, Mistral, Ollama (local), and any OpenAI-compatible endpoint
  • PostgreSQL and MySQL / MariaDB adapters with read-only enforcement and dialect-specific SQL validation
  • Three execution modes: Auto-execute, Review First (human-in-the-loop), Generate Only
  • Two-level query cache — exact match and semantic similarity (cosine ≥ 0.87) with thumbs-up/thumbs-down training
  • Per-connection privacy controls: exclude schemas, tables, and columns from the LLM prompt; auto-detect sensitive column patterns
  • Charts and BI: bar, line, area, pie, scatter, combo, KPI cards, and gauges with client-side filtering
  • PDF report builder — combine multiple query results into a single exportable report
  • Self-hosted Docker Compose stack — single command to start, no external dependencies, runs on a single Linux host
  • Business Source License 1.1 — free for non-commercial use, converts to Apache 2.0 on 2030-06-01