flyrain commented on issue #50:
URL: https://github.com/apache/polaris-tools/issues/50#issuecomment-3628522128

   On the “better context” side, I’m thinking about at least three options we 
could explore:
   
   ## MCP prompt / hints
   This is probably the easiest to start with: we keep enriching the MCP 
prompt/hints so the client sees more Polaris-specific guidance. The downside is 
that it requires updating the MCP server every time something changes, and 
we’re constrained by hint size. I also assume the hint is loaded in one shot on 
each invocation, which can burn a lot of tokens if it gets too large. Another 
downside is that URLs in hint are not allowed in MCP clients like Claude 
desktop. 
   
   ## Claude Skill
   Here we put the Polaris context into a Claude Skill, which effectively 
behaves like a dynamic system prompt. It allows for much larger and more 
instructional context than a simple MCP hint and can be updated independently 
of the server. The limitation is that this only helps when the client is 
Claude, so it doesn’t generalize to other MCP clients.
   
   ## RAG over Polaris knowledge
   A more generic, “agent-style” approach: we chunk Polaris knowledge (docs, 
examples, maybe even relevant code snippets) into a vector store and let the 
LLM retrieve what it needs on demand. This avoids pushing all context up front 
and should scale better as the knowledge grows, at the cost of having to 
maintain the RAG pipeline and index.
   
   Curious what you think about trying a lightweight version of (1) first, and 
then layering (2)/(3) as we see how far we can get.


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to