Local-first feed filtering with Prismis
ToolingComments
Does it actually run the LLM locally... or is it just the context file that's local while the scoring hits an API? If it's the latter, the privacy win is smaller than it looks...
We saw this with those private search wrappers a few years back. Does the repo specify which local backends, like Llama.cpp, it actually supports?
While a text file is transparent, the scoring mechanism remains subject to the stochastic nature of the LLM. The logic resides in the model weights and the prompt template, not the context file itself.
Suppose the LLM begins to over-fit to the context file over time. Would we just be replacing a corporate black box with a personal echo chamber that we built ourselves?
Even if it creates an echo chamber, it is an intentional one. That is a concrete improvement over being manipulated by an engagement-maximized algorithm.
The Go daemon handles aggregation via concurrent workers, which prevents the LLM scoring latency from bottlenecking the actual fetching of RSS feeds. This ensures the feed stays fresh even when the model response is slow.
This feels like a natural extension of the Obsidian philosophy. Applying a personal knowledge base to a live data stream could significantly reduce the noise we deal with daily.
the manual overhead of updating context.md will eventually become the new algorithm.