Title: P1: Microsoft GRAG, Storm, MetaAI RAG [2024-08-02 Fri]
I am going to use multistage answering, keeping graph in a
one list of tuples with keywords, use graph partitioning
and Community summaries, use outlined textualized graphs.
NCNN from Tencent for ONNX models require converter that
distribute only in binary. :-(
😶 #dailyreport #rag #grag #airoles #retriving #tencent #nccn
Title: P1: Indexing techniques and RAG best practices [2024-08-01 Thu]
I compiled from sources PyTorch.
ONNX quantized models have boost up to 3X in speed and
memory and can be run on cheepest CPU, that is what I am
going to achieve - create approachable and presonalized
AI.
I found open source runtime for ONNX without hell of
dependencies: https://github.com/Tencent/ncnn/
I faced lack of examples and hell of dependencies for #dailyreport #rag #grag #indexing #retriving #onnx #ncnn
Title: P1: P0: Indexing techniques and RAG best practices [2024-08-01 Thu]
I have been reading original papars at arxiv and articles
about indexing techniques and RAG best practices such as
Graph-based RAG, SubGraph GRAG.
Modern history of neural retrival: #dailyreport #rag #grag #indexing #retriving #onnx #ncnn
anoncheg