v0.4.0
Changes
- Unified note search across REST, gRPC CoreBridge, and MCP with rich hits (
note_title, excerpt, matched_by, block_id, score).
- Per-block Qdrant indexing with reindex on writes, purge on delete, and startup backfill when embeddings are enabled.
- Replaced Intelligence preview-apply note flow with an Ollama native tool-agent (
create_note, write_note_text, find_semantic_blocks, get_note_content).
- Limited Intelligence providers to Ollama multi-config; removed GitHub Models from the release surface.
- Updated the Web search dialog and agent panel for immediate tool execution, source notes, and unified search API.
- Restored normative
docs/SPEC.md and docs/DESIGN.md with a Perplexity-style spacing and token scale.
Scope
- Unified note search (REST, gRPC, MCP) with rich hits
- Per-block Qdrant indexing and startup backfill
- Ollama tool-agent (create / write / search / read) without preview-apply
- Ollama-only Intelligence providers
- Perplexity-style design tokens and spacing scale
- Restored
docs/SPEC.md and docs/DESIGN.md
Runtime requirements
docker compose up -d # Postgres, Redis, Qdrant, NATS (as needed)
ollama serve
ollama pull nomic-embed-text-v2-moe
# chat model with tool support, e.g. llama3.2 or similar
just run-full
Build verification
cd core && cargo check
cd web && bun run build
cd intelligence && python3 -m compileall src
Manual acceptance
Notes
Search
MCP
Agent
UI
Breaking changes
- Intelligence: removed
POST /chat/note-actions/apply and GitHub Models providers
- Core gRPC:
SearchNotes → FindSemanticBlocks, UpdateNote → WriteNoteText
- Web:
POST /search/semantic → POST /search/notes