ragworkspace-local-ai-starter-pack

RAGWorkspace Local AI Starter Pack v5

This repository distributes a version-pinned convenience ZIP for the RAGWorkspace llama.cpp installer. Version 5 contains one compact chat model, one text embedding model, and one text reranking model. RAGWorkspace packages the files but is not the author or owner of the included models.

All model credit, immutable upstream revisions, exact hashes, licenses, and original-model lineage are recorded below, in NOTICE.md, and in SOURCE-MANIFEST.json.

Download and integrity

During online installation, the verified archive and extracted model payload can temporarily coexist. RAGWorkspace calculates the exact free-space requirement from manifest.json and any resumable partial download.

RAGWorkspace checks the archive's exact byte length and SHA-256 before extraction, rejects unexpected or unsafe ZIP members, and verifies each GGUF's size, magic header, and SHA-256 while extracting it.

Included models

The exact per-model SHA-256 values are in manifest.json and SOURCE-MANIFEST.json.

Runtime roles

Qwen3.5 0.8B is the only included chat model and the declared default. Nomic v1.5 is registered only for embeddings, uses mean pooling, and requires search_document: for indexed text and search_query: for retrieval queries. BGE Reranker v2 M3 is registered as a distinct multilingual reranker with rank pooling; it is not a chat model or an embedder and is intended to load on demand when a retrieval result set needs reranking.

The release catalog can list additional optional downloads, but entries marked included_in_package: false are not members of this ZIP.

RAGWorkspace installation and live catalog

1. Install the managed llama.cpp runtime in RAGWorkspace.
2. Open Native llama.cpp settings.
3. If this release was just published, select Check Hugging Face for updates to bypass the validated one-hour catalog cache.
4. Select the model download/install action.
5. Keep RAGWorkspace open while it downloads, verifies, extracts, and registers the models.

The app reads the small catalog from this repository's main/manifest.json, but that manifest may reference package archives only inside this same repository at an immutable release tag or 40-character commit. If the live catalog cannot be fetched or validated, the app uses its bundled manifest or last validated cache. Downloads and upgrades are always user-initiated.

Control of this Hugging Face repository is the catalog's publisher-authentication boundary. SHA-256 verifies that the downloaded bytes match the catalog, but the catalog supplies that digest and is not independently signed. Release policy treats published vN tags as immutable. The app also pins the exact ID, version, and SHA shown at confirmation time, so a later catalog change requires the user to review and confirm again.

For offline installation, put manifest.json and the complete ZIP in a native_model_assets folder beside the RAGWorkspace installer or portable EXE, then select the model installation action.

Text-only scope

This package supports text/chat generation, text embedding, and text ranking. It intentionally omits multimodal projector files, so the pack alone does not enable image, audio, or video prompting.

License, provenance, and safety

The pinned source/original-model lineage records Apache-2.0 for all included weights. A complete Apache License 2.0 text is provided beside and inside the ZIP. Qwen3.5's distinct license file, including its 2026 Alibaba Cloud copyright notice, is preserved as LICENSE-QWEN3.5-APACHE-2.0.txt.

The model weights are redistributed byte-for-byte from the pinned GGUF sources; RAGWorkspace did not further quantize or fine-tune them. Collection-level changes are limited to assembling a deterministic ZIP64 archive, adding legal and provenance material, and declaring runtime roles and selection policy.

Model output and rankings can be inaccurate, offensive, unsafe, or unlawful. Users remain responsible for evaluating results, respecting third-party rights, and complying with applicable law and each model's license. This summary is informational, does not replace the license or upstream model cards, and is not legal advice.

Version 5 keeps the v4 chat and embedding choices while replacing its Qwen3 reranker with BGE Reranker v2 M3. Existing v1, v2, v2.1, v3, and v4 tags remain immutable historical releases.