Skip to content

ROADMAP.md

Latest commit

 

History

History
111 lines (92 loc) · 4.66 KB

ROADMAP.md

File metadata and controls

111 lines (92 loc) · 4.66 KB
Raw

Roadmap

v0.0.1 — Minimum scaffold (current)

  • Repository structure with uv-managed Python package
  • LLM provider abstraction with five adapters (Anthropic, OpenAI, OpenAI-compat, Google, Ollama) covering 20+ services
  • Plain .txt ingest adapter with speaker-label heuristics
  • Single-pass pipeline (ingest → compose prompts → LLM → markdown out)
  • SQLite library with terms / aliases / speakers / sessions
  • Markdown and DOCX exporters
  • CLI (clearscript run, clearscript providers, clearscript lib)
  • Bundled desensitized prompt library (system + 7 layers + 5 stage prompts)
  • Bilingual README, MIT license, GitHub issue/PR templates
  • CI (lint + tests on macOS / Linux / Windows × Py 3.11/3.12/3.13)

v0.1.0 — Full pipeline (target: end of v0.0.1 + 8 weeks)

Pipeline

  • Stage decomposition (each stage produces serializable artifact)
  • Pre-scan stage with structured JSON output
  • Context Briefing UI flow (web UI ships in v0.0.2)
  • Chunking by semantic boundary at speaker turns (v0.0.5, ~3.5k target tokens)
  • L3.5 sentence-level reasoning layer (in prompts/layers/)
  • Self-review stage
  • Batch-ask UI
  • Diff-aware Re-scan stage
  • Project re-run (clearscript projects rerun <slug> + UI button, v0.0.11)
  • Project compare (/api/projects/{slug}/compare?with=... + UI modal, v0.0.13)
  • Audit trail with full change provenance (change_log.json per project)

Ingest (12 formats)

  • .txt
  • .md with AI-summary detection (v0.0.4)
  • .docx (generic + Feishu Miaoji specific) (v0.0.4)
  • .srt / .vtt (v0.0.4)
  • .json (PLAUD, common ASR APIs) (v0.0.4)
  • .html (Feishu Miaoji web export)
  • .lrc
  • Tongyi Tingwu (Alibaba)
  • Tencent Meeting
  • Yuanbao
  • Typeless (with summary stripping)

Library

  • Mode A: library-into-prompt activation (v0.0.3, fixed to scan transcript itself v0.0.11)
  • Mode B: end-of-session harvest UI (v0.0.3 per-run, v0.0.14 persistent inbox)
  • Mode C: cross-chunk learning (v0.0.11)
  • Markdown view (lib export --md, v0.0.13)
  • CLI: lib search (v0.0.13), lib export (v0.0.12), lib import (v0.0.12)
  • Universal pack seeded at install (17 terms, 3 negatives, v0.0.9)
  • Health check (duplicates / low-conf / stale, v0.0.13)
  • Negatives CRUD (v0.0.13)
  • Bulk delete (v0.0.12)

Frontend

  • FastAPI server with SSE progress streaming (v0.0.8 + v0.0.9 token-level deltas)
  • Single-page web UI in Bauhaus design system (v0.0.2)
  • Library / Editor / Projects views (v0.0.3)
  • Real-time diff display (rerun compare modal, v0.0.13)
  • Cost estimator before each LLM call (v0.0.9 + pre-flight confirmation v0.0.14)

Distribution

  • uv tool install + pipx install from git (v0.0.11)
  • PyInstaller-packaged single-file executable for macOS / Windows / Linux
  • Homebrew formula
  • Docker image

Documentation

  • MkDocs Material site auto-deployed to GitHub Pages (live at chen17-sq.github.io/clearscript)
  • Per-provider setup guides
  • Per-workflow examples (VC ref, podcast cleanup, medical interview, etc.)

v0.2.0 — Quality and trust

  • Mode B: end-of-session harvest UI (v0.0.3 + persistent inbox v0.0.14)
  • Library health dashboard (duplicates / low-conf / stale, v0.0.13 + UI v0.0.14)
  • Privacy redact mode (local NER → mask before LLM → unmask after)
  • Postgres adapter as alternative to SQLite (for team / shared library)
  • OS keyring for API key storage
  • Local embedding model (BGE-small) for semantic library search
  • PDF / JSON / SRT export
  • Encrypted project export (*.zip.enc)

v0.3.0 — Ecosystem

  • Domain pack system + 5 official packs (vc / ai-infra / consumer / medical / podcast)
  • Plugin system for custom layers and ingest adapters
  • Tauri desktop app (auto-update, system tray, file associations, native menus)
  • Multi-language UI (en + zh-CN, with framework for more)
  • Library bulk operations (merge / split / retag / mass-deprecate)

v0.4+ — Polish and scale

  • Full-text search across all transcripts (FTS5)
  • Speaker fingerprinting across sessions
  • Project templates (one-click setup for common workflows)
  • Notion / Obsidian integration
  • Optional git-based collaboration / sync
  • Read-only mobile companion (PWA)

Design principles that constrain the roadmap

These are not negotiable:

  • No telemetry, ever
  • No mandatory network calls beyond user-configured LLM provider
  • No proprietary data formats
  • No assumed cloud services
  • No data collection or aggregation across users

Features that conflict with the above will not ship, regardless of utility.