Skip to content

Weekly research scan — 2026-03-29 #7

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
max4c merged 1 commit into from
Mar 30, 2026
Merged

Weekly research scan — 2026-03-29 #7

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
115 changes: 115 additions & 0 deletions research/weekly-scan-2026-03-29.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,115 @@
# Weekly Research Scan — 2026-03-29

Repos monitored: **OpenOats**, **QMD**, **AppFlowy**
Period: March 22–29, 2026

---

## 1. OpenOats (yazinsai/OpenOats)

> Native macOS (Swift/SwiftUI) meeting note-taker with on-device WhisperKit transcription, local knowledge-base search via embeddings, and LLM-powered suggestions. 2.1k stars, MIT.

### Activity: Very High (33 commits, 13 merged PRs, 13 issues closed)

| PR | What shipped | Why it matters |
|----|-------------|----------------|
| #228 | **Sidecast** — multi-persona AI sidebar with intensity control, dedup, evidence gating | Novel UX for surfacing multiple AI perspectives simultaneously |
| #208 | **Real-time suggestion engine** — replaced 5-stage serial pipeline with 3-layer concurrent architecture (pre-fetch cache → local heuristic gate → streaming LLM). Sub-2s latency | Gold-standard pattern for "AI while you write" |
| #200 | **Granola import** — paginated fetch, speaker mapping, tag-based duplicate detection, Keychain credentials | Clean data-migration pattern with idempotent re-import |
| #210 | **Tabbed settings** — 16-section scroll → 5 native macOS tabs | Good SwiftUI settings reference as feature count grows |
| #224 | Model download progress (speed, size, ETA) | Polish detail for on-device model UX |
| #223 | Auto-discover Ollama models in settings | Pluggable local-LLM provider pattern |
| #205 | Webhook on meeting end | Lightweight extensibility without a plugin system |
| #191 | Auto-stop recording when meeting app exits | Smart lifecycle management |

### Patterns worth adapting for Bugbook

| Pattern | Detail | Effort | Impact |
|---------|--------|--------|--------|
| **Markdown chunking + incremental embedding** | Chunk by heading (80–500 words), prepend header breadcrumb, hash-based change detection, cache embeddings locally. Only re-embed changed files. | Med | High |
| **Pluggable LLM provider** | Common interface over OpenRouter (cloud) and Ollama (local). "Bring your own model" is table-stakes. | Med | High |
| **3-layer suggestion architecture** | Pre-fetch KB cache → local heuristic gate → streaming LLM synthesis. Avoids unnecessary LLM calls. | High | High |
| **Tag-based import dedup** | `source:"granola"` + `granola:{noteId}` tags for idempotent re-import. Works for any import source (Apple Notes, Bear, Obsidian). | Low | Med |
| **Tabbed settings (SwiftUI)** | Native `TabView` with 5 categories. Copy-paste-ready pattern. | Low | Med |

---

## 2. QMD (tobi/qmd)

> On-device hybrid search engine for markdown files. Combines BM25 (SQLite FTS5), vector similarity (sqlite-vec), and LLM reranking — all local via node-llama-cpp with GGUF models. 17.2k stars, MIT. By Tobi Lütke.

### Activity: Very High (10 merged PRs, 6+ new issues, active RFCs)

| PR | What shipped | Why it matters |
|----|-------------|----------------|
| #455 | **50× FTS5 speedup** — CTE wrapper prevents query planner from abandoning the FTS index when collection filters are present (19.8s → 0.4s) | Critical SQLite gotcha; directly applicable |
| #449 | **AST-aware chunking** via tree-sitter for code files | Respects code structure during chunking |
| #463 | Fix hyphenated token handling in FTS5 lex queries | Edge-case fix worth knowing about |
| #462 | Fix BM25 field weights to include all 3 FTS columns | Correct weighting = better ranking |
| #478 | Add `rerank` parameter to MCP query tool | MCP integration getting richer |
| #456 | Handle `vec0 OR REPLACE` limitation | sqlite-vec compatibility fix |

**Notable open PRs:** #484 Symbol extraction (Phase 2), #480 Remote OpenAI-compatible embeddings, #470 `qmd bench` for search quality benchmarks, #469 REST search endpoints.

**Notable issues:** #481 Feature request for knowledge-graph support, #483 Vector search unstable on Apple Silicon.

### Patterns worth adapting for Bugbook

| Pattern | Detail | Effort | Impact |
|---------|--------|--------|--------|
| **CTE trick for FTS5 + filters** | Wrap FTS MATCH in a CTE, then JOIN/filter. Without this, SQLite abandons the FTS index. 50× difference. | **Low** | **High** |
| **Hybrid search pipeline** | BM25 → vector similarity → LLM rerank. Each stage is optional. Clean `searchFTS() → searchVec() → rerank()` pipeline. | Med | High |
| **Scored break-point chunking** | Headings scored 100/90/80…, code fences, horizontal rules, paragraph boundaries. `findBestCutoff()` picks optimal splits within token budget (900 tokens, 15% overlap). | Med | High |
| **Content-addressable storage** | SHA256 hash as key, deduplicates content, trivial change detection. | **Low** | Med |
| **MCP server exposure** | Expose search as MCP tools (`query`, `get`, `multi_get`, `list_collections`). Makes the app accessible to Claude Desktop, Cursor, etc. | Med | High |
| **Collection context metadata** | Per-folder semantic hints (e.g., "these are engineering meeting notes") injected into search/embedding without modifying notes. | **Low** | Med |
| **LLM query expansion (HyDE)** | Generate hypothetical answer, search for documents similar to it. Also generates lex/vec query variants. | Med | Med |
| **Virtual path URIs** | `qmd://collection/path` — stable, location-independent document references. Enables deep linking and stable backlinks. | **Low** | Med |
| **Index health monitoring** | `{needsEmbedding, totalDocs, daysStale}` — simple signal for "your search is out of date". | **Low** | Low–Med |

---

## 3. AppFlowy (AppFlowy-IO/AppFlowy)

> Open-source AI-powered collaborative workspace (Flutter + Rust), 68.9k stars. Notion alternative emphasizing data privacy and local-first architecture.

### Activity: Moderate (v0.11.5 released March 26)

**v0.11.5 highlights:**
- AI Meeting Transcript — paste YouTube link → transcript
- Database bulk edit optimization
- Mobile: database embedded views support Feed and List views
- Fixes: shortcut conflicts, plus-menu insertion, filter evaluation

**Notable new issue:** #8599 — CLI Interface for Local AI Agent Integration & Headless Operations (proposes using the existing Rust `flowy-core` event dispatch to enable CLI/agent CRUD on workspaces).

### Patterns worth adapting for Bugbook

| Pattern | Detail | Effort | Impact |
|---------|--------|--------|--------|
| **Slash command menu** | "/" opens a filtered list of block types, AI actions, and formatting. Redesigned in v0.11.4 for discoverability. Straightforward in SwiftUI (filtered list overlay). | **Low** | **High** |
| **CRDT-based local-first sync** | Yrs (Yjs Rust port) wrapped in a `Collab` abstraction with plugin system: disk persistence, cloud sync, snapshot as separate plugins. Dual storage: SQLite for metadata, CollabKVDB for content. | Med | High |
| **In-editor AI text actions** | Select text → write/edit/translate/improve. Pre-built + custom prompts loaded from database pages (prompts-as-data). | **Low** | Med |
| **RAG search over workspace** | Natural language queries returning results with source links + AI-generated overview. | Med | High |
| **Event-dispatch / FFI architecture** | Flutter ↔ Rust via Protobuf event dispatch. Enables the proposed CLI (#8599) — same backend serves GUI and CLI. Swift equivalent via UniFFI. | Med | High |
| **Block-component-builder plugin system** | 22+ content types registered via builders. Toolbar items, shortcuts, slash commands, context menus are all separate extension points. | High | High |
| **Notification center** | In-app hub for mentions, reminders, backlink activity. Good for surfacing stale notes and pending tasks. | Low | Med |
| **Custom AI prompts as data** | Users define reusable AI workflows as database pages, not code. Clever for power users. | Low | Med |

---

## Top 3 This Week

The three highest-impact, lowest-effort items to act on first:

### 1. FTS5 CTE Query Pattern (from QMD #455)
**Effort: Low | Impact: High**
Wrap any FTS5 MATCH query in a CTE before applying additional WHERE filters. Without this, SQLite's query planner silently abandons the full-text index, causing 50× slowdowns. This is a one-line SQL restructure that applies immediately to Bugbook's existing search layer. No new dependencies.

### 2. Slash Command Menu (from AppFlowy v0.11.4)
**Effort: Low | Impact: High**
A "/" command menu is the standard interaction pattern for block-based editors. In SwiftUI, it's a filtered list overlay triggered by a single character. It provides a discoverable entry point for block insertion, formatting, and (eventually) AI actions. Implement now; extend later.

### 3. Content-Addressable Storage + Incremental Embedding Cache (from QMD + OpenOats)
**Effort: Low | Impact: High**
Both QMD and OpenOats independently converged on the same pattern: SHA256-hash documents, only re-process changed content. QMD uses it for dedup and change detection in the search index; OpenOats uses it for incremental embedding. This is the foundation for any local semantic search — get it in early so the embedding/indexing pipeline scales from day one.
Loading