← Back tolarr; Back to Changelog →

Changelog

July 2026

📚

Multi knowledge bases

One Qdrant collection = one KB. Agents create, list, inspect, and delete with kb commands; ingest/query take -c NAME so vectors never mix.

🧪

Machin-native eval gates

eval cerise, eval hash, and eval synonyms live in the binary — no Python. CI blocks on hash + Cerise; synonyms stay a soft ONNX signal.

🍒

Cerise landing backtest

Real French FAQ corpus with curated service cards. Ranking covers doula, retraites, aquatique, contact, and photo sidecars.

🖼

Vision without LLMs

Phase 0 sidecars (photo.jpg.md) and Phase 1 folder walk + optional Tesseract OCR. No cloud vision, no VLM captioning.

🔤

hash-sha256-v2 embedder

Stopword drop + light plural stem for better French FAQ ranking. Re-ingest after upgrade. ONNX stays paused while the synonym gate holds.

🚀 Features

  • [18/07]feat: multi-KB, machin eval suites, vision Phase 0–1 (v0.6.5)
  • [17/07]feat: OCR policy for text/mixed/images-only plus CI smoke
  • [17/07]feat: PDF text-layer ingest and DOCX phase 3 structure
  • [17/07]feat: DOCX phase 2 — reading-order walk with rich image payloads
  • [17/07]feat: native DOCX ingest via ZIP/zlib and OOXML text scrape
  • [17/07]feat: md/csv/json ingest with config merge and chunk pipeline
  • [17/07]feat: initial machin-rag MVP — agent-first local RAG

🐛 Bug Fixes

  • [17/07]fix: OCR tessdata path and images-only always-OCR

📚 Documentation

  • [17/07]docs: changelog for OCR tessdata fix
  • [17/07]docs: changelog for OCR policy and CI smoke
  • [17/07]docs: changelog for PDF + DOCX phase 3
  • [17/07]docs: refresh July changelog for DOCX phase 2
  • [17/07]docs: add DOCX ingest to July changelog
  • [17/07]docs: refresh July changelog for structured ingest
  • [17/07]docs: remove csv-to-rag references from project copy
  • [17/07]docs: landing page + July 2026 dual-view changelog