You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Start with the most coding-capable SLM that we know.
Investigate which models might be best for this purpose to be able to read, evaluate, and execute code in the XML stack universe.
Pick a good one to start with. (Is it still Qwen via ollama? Or is there something better now?)
Set this up to run in Docker for containment / privacy this time. Docker should hold everything we need to pull in the local SLM, and allow us to mount a directory for analyzing and working on projects.
No more drag and drop XML into a Claude interface! This should be more professional.
The MCP portion of it (Model Context Protocol) is what makes the SLM act as an agent instead of just a chatbot. Let's see if our mcp_server.py can be given to another SLM that isn't Claude desktop.
Currently our MCP scripts allow the agentic AI to do the following:
* Write XPath in the latest current 3.1 syntax. (When this isn't working properly, most generative AI models pull on ancient 1.0 XPath syntax. But we require the latest version and we put the tools it needs for that in its path.)
* Write XQuery (which is entirely for "pull-processing": like a database query language that lets you drill through collections of files and get results.)
* Write XSLT to run identity-transformations that "patch" and fix XML without changing the original beyond what you require.
* (We can easily develop this): Write XSLT or XQuery to output HTML + SVG visualizations that help survey something useful about the code.
* Schema validation checks: against Relax NG and Schematron.
We should be able to ask the MCP to do anything we imagine would take us a long time to do on our own and help us to understand what's needed immediately, and also show us what it will do.
We must carefully check and validate the MCP's outputs to be sure of their quality
Start with the most coding-capable SLM that we know.
Set this up to run in Docker for containment / privacy this time. Docker should hold everything we need to pull in the local SLM, and allow us to mount a directory for analyzing and working on projects.
The MCP portion of it (Model Context Protocol) is what makes the SLM act as an agent instead of just a chatbot. Let's see if our mcp_server.py can be given to another SLM that isn't Claude desktop.
Currently our MCP scripts allow the agentic AI to do the following:
* Write XPath in the latest current 3.1 syntax. (When this isn't working properly, most generative AI models pull on ancient 1.0 XPath syntax. But we require the latest version and we put the tools it needs for that in its path.)
* Write XQuery (which is entirely for "pull-processing": like a database query language that lets you drill through collections of files and get results.)
* Write XSLT to run identity-transformations that "patch" and fix XML without changing the original beyond what you require.
* (We can easily develop this): Write XSLT or XQuery to output HTML + SVG visualizations that help survey something useful about the code.
* Schema validation checks: against Relax NG and Schematron.
We should be able to ask the MCP to do anything we imagine would take us a long time to do on our own and help us to understand what's needed immediately, and also show us what it will do.
We must carefully check and validate the MCP's outputs to be sure of their quality