Clojurians Log v2

It's for the Danish WordNet which is basically a huge graph of relations between the different word senses used in the Danish language. As for tips: I was already using Transit to transmit the different RDF resources as Clojure maps to my consuming CLJS web app + had an optional Turtle format download for them which used Donatello. The only real change was converting the Clojure map into JSON in way that would make it valid JSON-LD and with the keys presented in a semantic order.

I basically just go through the Clojure map and collect all prefixes, then I put the mapping between prefixes and URIs in the "@context" key of the JSON data we are building. The rest is mostly just renaming the keys in the Clojure map. One thing I had to figure out was how to present supporting data, e.g. relation labels, and I figured that the most logical way was to put those in a "@graph" at the end. I also append a short "rdfs:comment" to help the consuming LLM make more sense of the data.

Code is here: https://github.com/kuhumcst/DanNet/blob/master/src/main/dk/cst/dannet/db/export/json_ld.clj

The MCP server itself is mostly vibe-coded in Python. I researched best practices and have been prompting Claude it to analyse my Clojure code and build the MCP server based on that.

I just did a naïve EDN->JSON conversion at first, but hit a wall, so I figured that the added context of using JSON-LD will make it much easier for the LLM to understand what the data represents... which it did! And that makes a lot of sense too, as any LLM will no doubt have lots of implicit knowledge about RDF and the structure of RDF/JSON-LD, while it has to essentially resort to guessing from its immediate context whenever it's just a random JSON blob.

Super interesting, thanks for writing that up! More to follow, I’m sure, once I have digested this 😀