2.0.0 (2026-04-24)
⚠ BREAKING CHANGES
- breaking: rename search modes to :vector/:keyword, evaluation UX overhaul
- dashboard: the Ask page URL shape changed from /arcana/ask (with a client-side mode toggle) to /arcana/ask/<sub_tab>. The old URL still lands on the Advanced sub-tab by default, so existing bookmarks keep working.
- module names changed. Custom behaviour implementers need to update their @behaviour declarations from Arcana.Agent.* to Arcana.Pipeline.. Code that used Arcana.Agent. still works via the deprecated facade.
Features
- breaking: rename search modes to :vector/:keyword, evaluation UX overhaul (f1c59da)
- config: global config defaults and unified opts merging (11880e1)
- dashboard: ChunkResults component, Loop grounding via LLMJudge (b3dba78)
- dashboard: compact Loop settings and simpler collection opts (daa4cd8)
- dashboard: markdown answers, live trace polish, role-aware temps (7836d37)
- dashboard: Pipeline UX overhaul (eb999a0)
- dashboard: rerank top_k, expanded Pipeline internals card (87dac70)
- dashboard: sub-tab polish and live tracing for Ask (958f7ce)
- dashboard: thousand separators and tabular-nums on stats (501a2f8)
- dashboard: windowed pagination with prev/next and ellipsis (aa3eba5)
- evaluation: pluggable retriever and answer correctness scoring (972375e)
- graph: community summary improvements and config extraction (3967270)
- graph: GraphRAG Local Search alignment (27ed3f4)
- graph: pluggable entity matcher for graph-enhanced search (2b56205)
- grounder: LLMJudge with atomic claim decomposition (d8f0525)
- loop: Arcana.Loop (agentic RAG) with grounding and tool-call attribution (b36c9b9)
- loop: custom tool callbacks, humanize guides (5c525a4)
- loop: per-call collection selection via dynamic search tool (afc9724)
- loop: per-tool-call telemetry for live tracing (aca51c2)
- loop: record synthesis fallback in tool_history (980efb7)
- loop: role-specific temperatures and customizable answer prompts (e3c2558)
- search: local cross-encoder reranker via Bumblebee (387e910)
Bug Fixes
- answer_metrics: strip markdown code fences from LLM judge (972375e)
- graph: pass entity metadata to Entity.changeset when upserting it (#57) (3786969)
- grounding: drop hardcoded EXLA, defer to global Nx.Defn defaults (cf46ec7)
- grounding: per-chunk multi-pass NLI scoring, attach pipeline.ground telemetry (29bd21f)
- grounding: top-K budget-aware concat with per-chunk-max fallback, sharpen rewriter prompt (ba705e2)
- pipeline: gate + self-correct use Arcana.LLM protocol, skip_retrieval honored by prep steps (b4bf007)
- pipeline: rerank forwards all opts to reranker module (2a3214a)
Miscellaneous
- deps-dev: bump credo from 1.7.17 to 1.7.18 (#69) (18ada82)
- deps-dev: bump floki from 0.38.0 to 0.38.1 (#62) (f386462)
- deps: bump googleapis/release-please-action from 4 to 5 (#71) (c5e51bb)
- deps: bump igniter from 0.7.2 to 0.7.3 (#54) (1ac1dd9)
- deps: bump igniter from 0.7.3 to 0.7.6 (#61) (d467e43)
- deps: bump mdex from 0.12.1 to 0.12.2 (#72) (19c797e)
- deps: bump phoenix_live_view from 1.1.25 to 1.1.26 (#53) (41b6724)
- deps: bump phoenix_live_view from 1.1.26 to 1.1.27 (#59) (29a3665)
- deps: bump req_llm from 1.6.0 to 1.7.1 (#60) (2acdfe5)
- deps: bump req_llm from 1.7.1 to 1.8.0 (#63) (d5a4197)
- deps: bump req_llm from 1.9.0 to 1.10.0 (#70) (fe2c0da)
- deps: update lockfile to latest compatible versions (b1ae28b)
- fix credo --strict findings to unblock CI (bb9d1b6)
- test: use Tools alias to satisfy credo AliasUsage (80ff3d3)
Documentation
- livebook: add Loop section with multi-collection examples (6aa4db2)
- readme: align architecture diagram to uniform 83-char width (d812dc5)
Code Refactoring
Contributors
behaviour