Skip to content

Switchboard 0.3.0

Choose a tag to compare

View all tags
@aivinay aivinay released this 02 Jul 14:12
v0.3.0
71c48ee
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Highlights: Switchboard 0.3.0 adds 2026 local model packs and switchboard models --recommend, disabled-by-default confidence escalation, Arch-Router judge support, optional Headroom compression, a premium quota ledger, and the UI dashboard/routing chips.

Migration note: legacy Ollama chat profiles removed in this release fall back to defaults; migrate with switchboard models --recommend --apply or remap preferences.local_model_roles.

[0.3.0] - 2026-07-02

Added

  • Added the 2026 Ollama model pack, hardware-aware switchboard models --recommend, and explicit local role mappings for general, coding, and
    reasoning routes.
  • Added a sub-12 GiB floor tier for switchboard models --recommend so tiny
    machines get llama3.2:3b instead of heavier 16 GiB recommendations.
  • Added embedding-model-aware learned routing, dispatcher, sensitivity, and
    semantic-memory embedding calls, including task prefixes and fail-closed
    weight metadata checks.
  • Added disabled-by-default local answer-confidence checks with one-hop premium
    escalation for low-confidence non-sensitive Ollama answers.
  • Added first-class Arch-Router LLM judge support via router_llm_model, with
    policy-format prompting and fail-closed parsing.
  • Added optional headroom-ai compression support behind the headroom extra
    and preferences.compression_engine: "headroom", while keeping the
    dependency-free heuristic as the default.
  • Added a local, estimate-only premium quota ledger with user-declared soft
    budgets, quota-aware premium routing, switchboard quota, and /api/quota.
  • Added the dependency-free web UI upgrade: dynamic backend picker, routing
    chips, private-mode indicator, quota meters, and a savings dashboard fed by
    recorded metrics.

Changed

  • Made the packaged switchboard/config/ tree canonical, with make sync-config and drift tests keeping the root config/ copy aligned for
    development and Docker mounts.
  • Updated switchboard doctor to point at local model recommendations and
    report the GLM 4.7 Flash Ollama version requirement.

Removed

  • Removed legacy Ollama chat profiles replaced by the 2026 local model pack:
    ollama/qwen3:8b, ollama/qwen2.5-coder:7b,
    ollama/deepseek-r1:8b, ollama/gemma3:12b,
    ollama/qwen2.5-coder:14b, ollama/deepseek-r1:14b, and
    ollama/mistral-small3.2:24b. Existing role mappings that reference these
    profiles fall back to defaults; migrate with
    switchboard models --recommend --apply or remap
    preferences.local_model_roles.

Fixed

  • Made switchboard route previews honor private mode when Ollama is
    unavailable: sensitive prompts now preview as local-only/refused instead of
    recommending a subscription fallback.
  • Targeted mypy at Python 3.12 syntax so NumPy 2.1+ stubs using PEP 695 parse
    correctly, and annotated pre-existing training variables that this exposes.
  • Corrected Headroom installation docs to use the switchboard-local package
    name.
  • Corrected the real-backend smoke expectation for deterministic time-tool
    prompts, which are grounded and then formatted locally through Ollama.
Assets 2
Loading