I follow the assignment rule “Cite anything you borrowed. Borrowing is fine; not knowing what you borrowed is not.” and record every external resource that powers the Hiver system.
Kaggle Customer Support Dataset – source: Kaggle, CC BY‑NC‑SA 4.0 – provides authentic inbound/outbound Apple Support tweet pairs for building the historic vector index and evaluation set.
Sentence‑Transformers all‑MiniLM‑L6‑v2 (Reimers & Gurevych, 2019) – yields dense 384‑dim semantic embeddings for zero‑shot centroid intent classification and vector search, running entirely on CPU in < 35 ms.
ChromaDB (Chroma Core Team, 2023, Apache 2.0) – provides an in‑process local SQLite vector index storing @AppleSupport historical resolution pairs for intent‑filtered retrieval‑augmented drafting.
DuckDB (Raasveldt & Mühleisen, 2019) – high‑throughput out‑of‑core SQL streaming engine used in src/data/ingest_kaggle.py to isolate conversation initiators without loading the full 3‑million‑tweet CSV into memory.
Scikit‑Learn (Pedregosa et al., 2011) – supplies TF‑IDF vectorizer, Logistic Regression baseline (Baseline 2), and cohen_kappa_score for reproducible baseline modeling and inter‑rater agreement calibration.
Pydantic v2 (Colvin et al., 2023) – enforces runtime strict validation schemas ensuring fail‑closed safety and type integrity across API boundaries.
FastAPI & Uvicorn (Ramírez et al., 2018) – powers the async ASGI web server for the interactive dashboard and REST verification endpoints.
Landis & Koch (1977) rubric – benchmark scale for interpreting Cohen’s Kappa (κ = 0.7200), establishing Substantial Agreement for annotator consistency.