Skip to content

Ragbot 4.0: current dependencies, runtimes and models - #158

Merged
rajivpant merged 1 commit into
mainfrom
chore/latest-dependencies
Nov 6, 2026
Merged

rajivpant merged 1 commit into
mainfrom
chore/latest-dependencies

Conversation

@rajivpant

Copy link
Copy Markdown
Collaborator

Moves every external dependency, runtime and LLM to its latest release. Superseded models, runtimes and SDK paths are removed rather than kept. The full entry is in CHANGELOG.md under 4.0.0.

Models. engines.yaml now carries only current models, verified against provider docs on 2026-11-06. Sources are in docs/model-catalog-2026-11-06.yaml.

  • OpenAI: GPT-6 Astra, GPT-6.1 Sol (default), GPT-6.1 Terra, GPT-6 Luna. max_input_tokens is corrected to the documented 922K.
  • Anthropic: Claude Fable 5.1, Opus 5.5, Sonnet 5.5 (default), Haiku 5.5.
  • Google: Gemini 4 Argon, 3.8 Flash (default), 3.5 Flash-Lite. The previous default, gemini-3-flash-preview, and gemini-3.1-flash-lite-preview no longer exist.
  • Ollama: Gemma 4 (E4B, 12B, 26B, 31B), Qwen3.8 27B, Qwen3.8 Flash-Next (preview), Muse Glimmer 30B, Nemotron 3.5 Super, Mistral Medium 3.5. Llama 4, Qwen3.6 and the retired DeepSeek V3.2 and Mistral Large 3 tags are removed.

All 11 cloud IDs appear in each provider's list-models endpoint for the owner's keys. A live completion was not possible: all three cloud accounts returned billing errors. The local Gemma 4 models answered live.

Request handling.

  • Anthropic: effort goes in output_config, and sampling parameters are never sent (API 2026-10-21).
  • Gemini: thinking is set by level, never by budget.
  • OpenAI: max_completion_tokens.
  • The Claude-version heuristics, and the legacy paths they served, are deleted.
  • Model defaults in code (chat API, consolidator, cross-workspace synthesis, eval runner) now come from engines.yaml.

Fixed. local_only and downgrade-to-local routing classified every configured Ollama model as non-local, and suggested fallbacks that did not exist. Local models now come from engines.yaml, and Ollama :cloud tags are never treated as local.

Runtimes and images.

  • Python 3.14 in CI and the backend image. All 175 packages have amd64 and arm64 wheels, so the image no longer needs gcc. Python 3.15 still lacks wheels for PyYAML, fastuuid and orjson.
  • Node 26 LTS in CI and the web image, which drops its separate npm install.
  • pgvector pg18 on a new volume, plus scripts/migrate-postgres-16-to-18.sh. The script was tested on a copy of a real pg16 database: every table matched.
  • Jaeger v2.

Packages.

  • Python: every floor raised to the latest release; google-generativeai (end of life) removed, which frees protobuf 7; appdirs → platformdirs. Seven transitive packages stay held back by caps in their dependents' latest releases; they're listed in the triage doc.
  • Web: everything on its latest release, including jsdom 31. TypeScript stays on 6.0 until typescript-eslint accepts 7. npm audit: 0 vulnerabilities.

Verification.

  • Clean Python 3.14 install: pip check clean, 1085 passed / 13 skipped.
  • Web: npm ci, lint, 27 tests, tsc, next build.
  • Full docker compose stack built and run on the new images, with Jaeger v2 receiving traces and the API serving the new catalog.

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@rajivpant
rajivpant merged commit 2803935 into main Nov 6, 2026
2 checks passed
@rajivpant
rajivpant deleted the chore/latest-dependencies branch November 6, 2026 16:00
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant