kroma

API

Overview

KROMA matches concepts across two ontologies. These pages are the Python API: what you call, and what it returns.

A run loads an alignment, retrieves local context for each concept, asks a language model whether the pair matches, and refines the joined graph. The command line is the usual entry. The same steps are available as functions.

The pieces

CLIkroma run and the flags that select a config, a model, and a mode. Matching runMatchingRun prepares the task, scores pairs, and writes metrics. LoadersCSV alignments and an OWL class hierarchy. QueryParents, children, synonyms, and labels from a named source. ModelsOpenAI and Together chat models, and embedding backends. RetrievalCosine search over concatenated text and Node2Vec vectors. RefinementOnline updates while scoring, and a final offline collapse. MetricsPrecision, recall, and F1 for a binary alignment.

One run

  1. kroma run parses flags and calls MatchingRun.execute.
  2. prepare reads the method JSONL and the dataset JSON, then builds vector stores.
  3. score_pairs retrieves context, calls infer, and records the prediction.
  4. finish runs offline_refine and writes a metrics JSON file.

Install

git clone https://github.com/lamng3/kroma.git
cd kroma
./setup.sh
source .venv/bin/activate
uv pip install ".[models]"

Copy .env.example to .env and set TOGETHERAI_API_KEY or OPENAI_API_KEY. The paper is ISWC 2025.