Skip to content

Changelog

English · 繁體中文首頁 · full history on GitHub: CHANGELOG.md

繁中讀者

此頁為英文精簡變更摘要。完整逐條紀錄請見 GitHub 上的 CHANGELOG.md。

This page is a curated summary of recent releases for the docs site. For complete notes (including every fix and script-level detail), use the repository changelog linked above.

v1.9 — Comprehensive audit fixes (2026-05-23)

Security, correctness, and CI hardening after a second full audit:

  • Removed hardcoded API keys from test configs; minimized subprocess environments
  • Import-safe LLM / pipeline modules when config is missing
  • Fail-closed health checks; soft-fail single embedding windows; safer corrupt-JSON handling
  • Shared embedding client / circuit-breaker consistency; dead-code removal
  • Test harness + CI workflow (pytest, compileall, GitHub Actions)
  • Traditional Chinese README (README.zh-TW.md)

Also delivered multi-model SRT quality-check parallelism, consensus merge reports, and batch file-id ranges (see full changelog for operator knobs).

v1.6 — Recovery, idempotent writes & live validation (2026-05-16)

Pipeline reliability for long batch jobs:

  • Sidecar locks for status and phase-slot writes (no stale-inode lost updates)
  • Content-aware summarization resume (source_text_hash); block partial summaries before embedding
  • Strict embedding validation; LanceDB schema preflight
  • Idempotent merge-upsert on stable file_id / chunk_id
  • Live validation on sample file IDs with unique chunk rows in LanceDB

v1.5 — API key security, sessions & atomic writes (2026-05-14)

  • API keys stay in .env only (not committed config.json)
  • Shared requests.Session() connection pooling
  • Atomic save_status() / locked read-write paths
  • Smart Merge split into testable helper functions

v1.4 — Circuit breaker, JSON integrity & unified LLM client (2026-05-14)

  • Embedding path always goes through the circuit breaker
  • Checkpoint JSON array integrity for incremental summarization writes
  • Non-blocking lock retries with timeout (no permanent hang)
  • Shared llm_client.py for call / model helpers across scripts

v1.3 — Parallel quality check & evaluation models (2026-05-14)

  • Concurrent SRT quality windows (--concurrency / config)
  • Per-tool model lists for quality check and evaluation scripts
  • Dead-state cleanup in the file state machine

v1.2 — LLM quality evaluation & SRT checker (2026-05-13)

New operator tooling:

  • Sliding-window SRT quality review
  • Chunk-boundary and summary-fidelity evaluation
  • Manifest generator for data directories

v1.1 — Test pipeline & boundary diagnostics (2026-05-13)

  • Standalone chunk test runner and multi-config test suite
  • Boundary strength metadata and discarded-chunk diagnostics from Smart Merge

v1.0 — Production baseline (2026-05-13)

Core product path:

  • Smart Merge 3.0 semantic chunking
  • Four-phase pipeline: chunking → summarizing → embedding → LanceDB insert
  • Explicit file state machine + watchdog automation
  • Atomic checkpoints, chunk-level retries, model diagnostics, batch audit

Full changelog

Every release bullet, security note, and script change lives in the repo:

CHANGELOG.md on GitHub