vinyteles/mobilerag-plus-assets

MobileRAG+ assets for the CEMIG mobile assistant

Files the CEMIG mobile app downloads to run MobileRAG+ on the phone over the benchmark_mobile corpus: retrieval with an EcoVector index, file-neighbor expansion, and Selective Content Reduction (SCR) with precomputed window vectors. The answer model (Energy-Mobile, LiteRT) is downloaded separately from Fazzioni/Energy-Mobile.

mobilerag-plus.json lists every file with its size and SHA-256, and the parameters the app checks before using them. The app pins one commit of this repository.

Files

Path	Size	What
embedders/qwen3-embedding-0.6b-v2-q8_0.gguf	639 MB	Question embedder for retrieval (last-token pooling, 1,024 dims)
embedders/granite-embedding-107m-multilingual-q8_0.gguf	121 MB	Question embedder for SCR (CLS pooling, 384 dims)
benchmark_mobile/corpus.sqlite	31 MB	3,764 chunks with their Qwen vectors (sqlite-vec vec0 table)
benchmark_mobile/eco-vector.sqlite	17 MB	EcoVector index: centroid graph + 19 cluster HNSW graphs, one row each
benchmark_mobile/windows-int8.sqlite	34 MB	72,061 SCR window vectors (3-sentence windows, step 1), int8
mobilerag-plus.json	2 KB	Manifest

How they were built

• Embedders: converted with convert_hf_to_gguf.py of llama.cpp b10256 (the version llama.rn 0.12.9 bundles) to Q8_0. On 500 benchmark_mobile questions, MobileRAG+ with these query vectors keeps hit, evidence recall, nDCG@5 and MRR@5 within 0.4 points of the float32 sentence-transformers models.
  • Qwen: prepend the query instruction stored in the manifest, with no separator; the model appends <|endoftext|> and pools the last token.
  • granite: no instruction.
• Corpus: chunks of the benchmark_mobile source documents, embedded with UFGCEMIGONA/Qwen3-Embedding-0.6B-v2.
• EcoVector index: k-means partitions (about 200 vectors each), an HNSW graph over the centroids and one per cluster; searched with n_probe 6 and ef_search 32. Built and exported by mobile_rag (build-index, export-mobile-index).
• SCR windows: every 3-sentence window of every chunk embedded with granite-107m (float32) and quantized to int8 with one scale per window (build-scr-windows).

Sources and licenses

Source	License
UFGCEMIGONA/Qwen3-Embedding-0.6B-v2, fine-tuned from Qwen/Qwen3-Embedding-0.6B	Base model Apache 2.0; see the fine-tune's card
ibm-granite/granite-embedding-107m-multilingual	Apache 2.0
UFGCEMIGONA/benchmark_mobile source documents (corpus text)	other