Skip to content

v0.3.0 — Long-Context Map-Reduce

Choose a tag to compare

View all tags
@SYuan03 SYuan03 released this 01 Jun 15:37
· 3 commits to main since this release
v0.3.0
4c83b8e

Skill-Anything v0.3

Highlights

v0.3 makes Skill-Anything actually usable on long sources — books, multi-hour talks, large repos.

Previous versions silently degraded on anything past ~15K characters: the knowledge generator truncated input, quiz/flashcard generation hit hard per-call caps before reaching later chapters, and every LLM call was sequential with no caching. v0.3 introduces a section-aware, map-reduce pipeline with concurrency and a disk cache, so a 12-chapter book produces a study pack that actually covers every chapter.

New

  • Section-aware parsing across every source type

    • PDFs use the embedded outline (TOC) when available, with a page-range fallback
    • Text/Markdown splits by headings into structured sections
    • Web pages reuse heading detection for long articles
    • Video/audio bucket transcripts into 5-minute time sections with anchored timestamps
    • Repos produce one Section per selected file

  • Map-reduce knowledge generation

    • Per-section "map" calls produce local summary / key concepts / glossary / notes
    • A single "reduce" call synthesizes a global summary, cheat sheet, takeaways, and learning path from the map outputs
    • detailed_notes is assembled deterministically from per-section notes — no truncation
    • Two-tier model routing: SKILL_ANYTHING_MODEL_FAST for map, SKILL_ANYTHING_MODEL_SMART for reduce

  • Per-section quota allocation for quiz / flashcards / exercises

    • Largest-remainder weighting by section size, with a guaranteed minimum per section
    • Every chapter gets coverage even when the total budget is small

  • Concurrent LLM execution with a disk cache

    • ThreadPoolExecutor-based map runner with rich progress bars
    • Per-section failure isolation: a single failed section doesn't sink the whole pack
    • Cache key = sha256(prompt + model + version); second runs of the same source skip every cached call
    • Cache lives under ./output/.skill-anything/<slug>/ (gitignored)

  • New CLI flags for every source command

    • --concurrency / -c to control parallel LLM calls (default 6)
    • --no-cache to bypass the disk cache for a clean run

  • New environment variables

    • SKILL_ANYTHING_MODEL_FAST — fast/cheap tier for per-section map calls
    • SKILL_ANYTHING_MODEL_SMART — stronger tier for the global reduce call
    • Both fall back to SKILL_ANYTHING_MODEL when unset

Improved

  • Repo parser now uses a 240K-character budget split (docs 50% / manifest 15% / code 35%) instead of hard [:12]+[:10]+[:8] caps. Long repos surface meaningful code in addition to README/docs.
  • Practice exercises no longer concatenate every chunk into a single prompt — each section gets its own call, so long sources don't fail with context overruns.
  • Engine.write() no longer re-renders the concept map twice when --format all is used.
  • Quiz runner removed an incorrect substring fallback that matched A inside ABLE.

Case Study

A 12-chapter distributed-systems primer (~4.4KB / ~12 sections) processed end-to-end with --concurrency 6:

Output Count
Outline entries 13
Key concepts 15
Glossary terms 62
Quiz questions 30
Flashcards 40
Exercises 10
Takeaways 10

Chapter coverage verified by keyword matching across the generated quiz / flashcards / exercises: every chapter (1–12) is represented, including narrower topics like CRDTs, observability, and messaging semantics that v0.2 routinely dropped once the per-call cap was hit.

Example Commands

# Long PDF with full concurrency and caching
sa pdf book.pdf --format all --concurrency 8

# Same source again — second run hits the cache and finishes in seconds
sa pdf book.pdf --format all --concurrency 8

# Bypass cache for a clean run
sa pdf book.pdf --no-cache

# Two-tier model routing via env
export SKILL_ANYTHING_MODEL_FAST=gpt-4o-mini
export SKILL_ANYTHING_MODEL_SMART=gpt-4o
sa repo . --format all

Notes

  • The cache directory ./output/.skill-anything/ is gitignored by default.
  • --concurrency defaults to 6; raise it for IO-bound LLM endpoints with generous rate limits.
  • The map-reduce pipeline is fully backward-compatible: existing Parser.parse() callers still receive a flat list of chunks via a shim, and parse_sections() is the new canonical entry point.
  • All 75 tests pass (62 existing + 13 new v0.3 tests covering quota allocation, cache round-trip, failure isolation, concurrency caps, and per-chapter coverage).
Assets 2
Loading