Evidence RAG Pilot v0.1.0
Evidence RAG Pilot v0.1.0
A public technical prototype and portfolio demonstration of page-native, traceable RAG for engineering datasheets. This release focuses on the design and implementation; the project has not undergone a formal quantitative evaluation and makes no benchmark or SOTA claim.
Live demonstrations
The public sites preserve the original demo structure and read versioned static corpus snapshots without a GPU or API key. Dense retrieval, RRF, ingestion, and optional multimodal answering remain local-pipeline capabilities.
Corpus snapshot
- 182 indexed documents
- 338 source PDF pages
- 1,422 page-native evidence chunks
These are corpus inventory counts, not evaluation results.
Data assets
This release contains four archives: TXC source PDFs, TXC derived processing outputs, TKD source PDFs, and TKD derived processing outputs. Verify downloads with SHA256SUMS.txt. Model weights are not duplicated; the repository documents their official sources.
Code is Apache-2.0. Dataset rights and third-party materials are governed separately as described in DATA_NOTICE.md and THIRD_PARTY_NOTICES.md.