The LinkBERT architecture offloads model execution to a dedicated local microservice running on port 8006, keeping heavy ML dependencies like PyTorch and Hugging Face Transformers isolated from the main FastAPI server process.

The web application communicates with this local service over HTTP using asynchronous requests via httpx, sending flattened page text up to 20,000 characters and receiving word-level link probability scores alongside exact character spans in a single round-trip. This complete set of precalculated token probabilities enables the user interface to instantly regroup candidate anchor phrases across dynamic confidence thresholds without requiring additional model inference calls.