Changelog

​

BGE-small embeddings (384-dim) and Qwen LLM default

​

Install and upgrade

pip install moorcheh-edge==0.2.3
# or
pip install -U moorcheh-edge

docker pull moorcheh/moorcheh-edge:0.2.3
# or
docker pull moorcheh/moorcheh-edge:latest

​

What’s new

• Default embedding: BAAI/bge-small-en-v1.5 (384 dimensions, ~67 MB download, ~200–300 MB RAM when warm). Replaces BAAI/bge-base-en-v1.5 (768d, ~210 MB).
• Default LLM: qwen2.5:0.5b-instruct (~400 MB) for /answer and voice RAG.
• Docs: limits, quickstart, UNO Q guide, and case study updated for BGE-small.

​

Upgrade on UNO Q

moorcheh-edge down
rm -rf ~/.moorcheh-edge/data/* ~/.moorcheh-edge/models
docker pull moorcheh/moorcheh-edge:0.2.3
pip install -U moorcheh-edge
moorcheh-edge up --with-llm -y
# re-upload catalog

​

Catalog sync (retail kiosk)

​

Install and upgrade

pip install moorcheh-edge==0.2.2
# or
pip install -U moorcheh-edge

docker pull moorcheh/moorcheh-edge:0.2.2
# or
docker pull moorcheh/moorcheh-edge:latest

​

What’s new

• GET /export - export all catalog chunk text from Moorcheh Edge (no vectors). See Export store.
• export_store() - Python client helper for the same endpoint. See export_store.
• Voice POST /catalog/document - chunk, embed on edge hardware, and upload to Moorcheh Edge. See Voice server.
• Retail kiosk sync - Admin Sync from UNO Q pulls GET /export and replaces the local SQLite catalog; document saves embed on the UNO Q. See Retail Kiosk.

​

Qwen LLM, anti-hallucination, and kiosk voice

​

Install and upgrade

pip install moorcheh-edge==0.2.1
# or
pip install -U moorcheh-edge

docker pull moorcheh/moorcheh-edge:0.2.1
# or
docker pull moorcheh/moorcheh-edge:latest

​

What’s new

• Default LLM - qwen2.5:0.5b-instruct (replaces llama3.2:1b-instruct-q4_K_M, ~400 MB pull on moorcheh-edge up --with-llm)
• No-context answers - when search returns zero passages, /answer and /answer/stream return I don’t have enough information to answer that question. without calling Ollama
• Voice streaming - /ask/stream with speak: true streams tokens to the UI immediately; TTS runs after the full answer is ready
• Thinking loop - ambient kiosk audio while RAG runs (voice cache-thinking, thinking_enabled on /ask/stream)
• Kiosk audio docs - voice cache-holding and voice cache-thinking documented under Retail Kiosk (not general CLI pages)
• Bluetooth - manual pairing guide; automated voice bluetooth CLI removed
• Docs - Qwen model, --with-llm, and no-context behavior updated across guides and API reference

​

Flat store, text mode, RAG answer, and client-side embeddings

​

Breaking changes

• Upload payload - use {"store_mode": "text"|"vector", "items": [...]} instead of {"vectors": [...]}
• Delete route - DELETE /items (was DELETE /delete)
• New route - DELETE /store to wipe all data (API, CLI, Python)
• Health fields - dimension, store_mode, embedding_model replace vector_dimension
• Text search - CLI --query-text and SDK search_text() embed locally; server accepts float vectors only

​

What’s new

• upload-documents CLI - embed text locally and upload to a text store
• clear-store CLI - wipe store with confirmation prompt (-y to skip)
• SDK - upload_documents(), upload_vectors(), search_text(), clear_store()
• Text store - BGE-base-en-v1.5, 768 dimensions, optional text on each item
• Vector store - dimensions 128, 256, 384, 768, 1024, 1536 on first upload
• POST /answer - search + prompt + Ollama LLM response (RAG)
• POST /answer/stream - same RAG flow, streamed as SSE (meta, token, done, error)
• moorcheh-edge answer - embed query locally, call /answer
• SDK - answer_text(), answer(); MoorchehEdgeApiClient.answer_stream() for streaming
• moorcheh-edge voice - STT (Whisper), TTS (Piper), local voice ask, and voice serve HTTP API on :8766 for edge kiosk hardware
• Docs: Retail Kiosk example - PC + Arduino UNO Q in-store demo (Admin/Customer UI, voice proxy, catalog upload)
• Ollama bootstrap - moorcheh-edge up installs Ollama on Linux and pulls llama3.2:1b-instruct-q4_K_M (~808 MB)
• --skip-ollama on up - search-only without LLM setup

​

Upgrade notes

moorcheh-edge clear-store -y
moorcheh-edge upload-documents --documents-file documents.json

​

CI, release pipeline, and multi-arch Docker

​

Install and upgrade

pip install moorcheh-edge==0.1.1
# or
pip install -U moorcheh-edge

​

What’s new

• CI pipeline - automated tests and checks on pull requests and pushes
• Publish pipeline - automated PyPI and Docker image releases for moorcheh-edge and moorcheh/moorcheh-edge
• Multi-arch Docker image - moorcheh/moorcheh-edge:0.1.1 and :latest ship linux/amd64 and linux/arm64

​

Initial release

​

Install

pip install moorcheh-edge
moorcheh-edge up
moorcheh-edge status

​

Runtime

• moorcheh-edge up / down / status - Docker container lifecycle
• Local data - ~/.moorcheh-edge/data (moorcheh_edge_store.json)
• Default API - http://localhost:8080