Spaces:
Running
Zip Based Portability, New RAG Alternative Pattern for Better Contextual Awareness of Memory File Content
A current theory is we require something like RAG to pose contextual content but how do you determine which content to use in context?
This space solves the problem by statistically analyzing a zip of .md AI output files and consider each with commonality with the prompt. This allows you to prioritize even annotate the specific documents, pages, paragraphs, and sentences found in originals that make up the best return answer context for a AI operation to choose appropriate context from past memories.
Using this paradigm we can also allow 'dreaming' and memory by applying introspective attention over time.
Added double and triple sequences which if matched are higher scores. This comes from LCS Rouge research where metrics of good summaries show common LCS - longest continuous sequences which are common words in same order yet minimally show just exact matches which in itself is very useful to identify thematic matches:
