A tiny tutorial for starting to find similarities between building related concepts. Created for the Summer School https://linkedbuildingdata.net/ldac2024/ in Bochum, Germany
This is work in progress to pick up some of the work done in DURAARK, particularly the awesome "Interlink" system by @aothms to align different vocabularies like DBPedia, Getty AAT and bSDD.
Back then, "all we had" were some similarity searches like Levenshtein and Jaccard distances to guess the relation between nodes. Let's find out what LLMs can do fo us this time.
temporary etherpad (will be deleted in 30 days) https://yopad.eu/p/LDAC2024
from langchain_community.llms import Ollama
#question="what is the URL of the material that is created of isocyanate and polyol resin"
question="""using the SKOS vocabulary, creata a turtle triple {subject} {predicate} {object}
use a SKOS relation that describes the similarity to a subject from another vocabulary in the namespace odb: with a label 'Solid wood' / 'Glue-laminated timber board''
write a valid line usng turtle syntax.
do not write anything else"""
#question="what is the URL of the material that closest fitting to the category 'Wood' / 'Derived timber products' / '3- and 5-ply wood'? Only give the URL and a single digit that indicates how sure you are between 0.0 and 1.0"
ollama = Ollama(
base_url='http://localhost:11434',
model="mistral"
)
docs = vectorstore.similarity_search(question)
from langchain.chains import RetrievalQA
qachain=RetrievalQA.from_chain_type(ollama, retriever=vectorstore.as_retriever())
res = qachain.invoke({"query": question})
print(res['result'])which spits out
<http://dbpedia.org/resource/Beaverboard> skos:broader odb:Solid_wood .
Which is a valid line in a turtle file. Now, enjoy drawing the rest of the owl!
If you would like to collaborate, get in touch!
IANACS - I am not a computer scientist
There is a staggering stack of technology and modules involved here most of which I have only a very, very rough understanding if any at all. Take it with a grain of salt, send issuses an PRs
To do the DBPedia excersise, some background knowledge is handy. Use either
- Pieters LDAC 2024 notebooks
- LBD notebooks in the ifcopenshell-turial I made for LDAC 2022
- Mine from LDAC 2019
Based on Python, so conda/miniconda is what you need. I highly recommend Mamba. Like Conda, but on speed go fetch it here
conda env create -f environment.yml
conda activate LDAC2024
Download from Ollama Github
Download models. Let's go with some smaller ones. The 7 Billion LLama3 should do
ollama pull llama3
ollama pull nomic-embed-text
If you have a particularly weak machine, you might look into even smaller ones
The langchain-Chroma part has been adapted from here