Skip to content

Ragbot 4.0.0

Choose a tag to compare

@rajivpant rajivpant released this 06 Nov 19:12
· 1 commit to main since this release
ae804c4

Ragbot moves every dependency, runtime and model to its latest release. Backward
compatibility with superseded models, runtimes and SDK versions is not kept.

Changed

  • Models. engines.yaml carries only current models, verified against provider
    documentation on 2026-11-06 (docs/model-catalog-2026-11-06.yaml):

    • OpenAI: GPT-6 Astra, GPT-6.1 Sol (default), GPT-6.1 Terra, GPT-6 Luna.
    • 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.
    • Local (Ollama): Gemma 4 (E4B, 12B, 26B, 31B), Qwen3.8 27B, the Qwen3.8 Flash-Next
      preview, Muse Glimmer 30B, Nemotron 3.5 Super, Mistral Medium 3.5.

    Every cloud model has a strict request contract. Superseded models are removed,
    including gemini-3.1-flash-lite-preview, which Google shut down on 2026-05-25, and
    three local tags Ollama no longer serves locally.

  • Embedding model. RAG embeds with Microsoft's harrier-oss-v1-0.6b (MIT licence,
    1,024 dimensions, 32K-token context) in place of all-MiniLM-L6-v2 (384 dimensions,
    256 tokens). engines.yaml's embedding: block is now the only place the model is set,
    with a pinned Hugging Face revision; the EMBEDDING_MODEL environment override is gone.

    • Queries carry the model's retrieval instruction (web_search_query); documents carry
      none.
    • Indexing embeds chunks in batches instead of one model call per chunk.
    • Migration 0004_embedding_dimensions.sql moves both vector columns to 1,024
      dimensions. It clears the chunk index and entity embeddings (entities are kept), so
      every workspace needs ragbot index --workspace <name> --force after upgrading.
    • The vector store reads its schema dimension and refuses to index or search with
      vectors of another size; ragbot db init and ragbot db status report it.
    • The API container keeps downloaded models in a ragbot-hf-cache volume.
    • The model needs about 4.2 GB while loading in float32. Docker Desktop's default 4 GB
      is not enough; allow it at least 8 GB.
  • MCP SDK 2.x (#157). The client primitives, tasks, transports, OAuth callback, proxy
    and Ragbot's MCP server are ported to mcp 2.3.0.

  • Runtimes and images.

    • Python 3.14 in CI and the backend image, which no longer installs a compiler.
    • The backend image installs CPU-only torch from PyTorch's index. PyPI's Linux wheels are
      the CUDA builds, which added 4.2 GB of GPU libraries to an image with no GPU to use;
      the build now also runs pip check.
    • Node 26 (active LTS) in CI and the web image.
    • PostgreSQL 18 with pgvector 0.8.7 on a new volume;
      scripts/migrate-postgres-16-to-18.sh copies an existing database across.
    • Jaeger v2.
  • Python packages. Every floor is raised to the latest release; openai 3.x is allowed now
    that LiteLLM 1.108+ accepts it; google-generativeai (end of life) is removed, freeing
    protobuf 7; appdirs is replaced by platformdirs.

  • Web packages. Every package is on its latest release (Next.js 16.4.5, React 19.3.1,
    jsdom 31). TypeScript stays on 6.0 because typescript-eslint does not yet accept 7.

  • Request handling.

    • Anthropic requests never carry sampling parameters, and effort travels in
      output_config.
    • Gemini thinking is set by level, never by budget.
    • OpenAI requests use max_completion_tokens.

    The Claude version heuristics for superseded models are gone.

  • Defaults come from engines.yaml. The chat API default, the memory consolidator, cross-workspace synthesis and the eval runner no longer name models in code.

Fixed

  • A reasoning model that spent its whole output budget thinking returned an empty reply with
    no error. Every backend now raises OutputBudgetExhaustedError, which names the budget
    and says to raise max_tokens or lower the thinking effort; a partial reply is still
    returned.
  • A streamed chat that failed simply ended: the streaming thread dropped its exception.
    chat_stream now re-raises it as ChatError, which the web UI shows.
  • local_only and downgrade-to-local workspace policies recognised no configured local
    model and suggested models that did not exist. Local models are now those engines.yaml
    serves through Ollama, and Ollama :cloud tags are never treated as local.
  • The agent loop read an attribute mcp 2.x no longer has, which would have passed MCP tool
    errors through as successful output.
  • The MCP server's HTTP transport resolves bearer tokens from the request it serves,
    starts its session manager through the ASGI lifespan, and fails closed.
  • The OAuth callback page escaped nothing in its error message.
  • npm audit reports 0 vulnerabilities, down from 22 (1 critical), and CI now runs
    next build.

Security

  • Next.js advisories covering 16.0.0–16.4.0 (two unauthenticated remote code execution
    issues among them) are resolved. Transitive advisories (sharp, brace-expansion,
    minimatch, postcss-selector-parser, vite, zod, undici) are cleared by overrides and
    lockfile updates.
  • 13 CVEs in filelock 3.32.7 are resolved: filelock 4.2.9 installs. Every stable
    LiteLLM release from 1.104.0 through 1.108.1 caps it below 4, so the LiteLLM floor is
    the 1.109.0 release candidate, the first that accepts it; pip moves to 1.109.0 on its
    own once that ships. pip-audit reports no known vulnerabilities. See
    docs/security/dependabot-triage-2026-11.md.