0tizm0/propertypilot-recommender

PropertyPilot Recommender

Semantic ticket similarity engine for PropertyPilot — 13,798 maintenance tickets.

Winning model: BAAI/bge-small-en-v1.5

Bake-Off (200-query stratified eval set)

Model	Precision@3	MRR	Latency
BAAI/bge-small-en-v1.5	0.337	0.453	9.2 ms
sentence-transformers/all-MiniLM-L6-v2	0.333	0.488	7.2 ms
intfloat/e5-small-v2	0.312	0.496	11.2 ms

Files

File	Description
embeddings.parquet	13,798 × 384 embeddings + ticket_id
index.faiss	FAISS IndexFlatIP (cosine similarity)
config.json	Model name, dim, query/doc prefixes
eval_results.json	Full bake-off benchmark data
bakeoff_chart.png	Visual comparison bar chart

Usage

import faiss, json
import numpy as np
from sentence_transformers import SentenceTransformer
from huggingface_hub import hf_hub_download

with open(hf_hub_download("0tizm0/propertypilot-recommender", "config.json")) as f:
    cfg = json.load(f)

index = faiss.read_index(hf_hub_download("0tizm0/propertypilot-recommender", "index.faiss"))
model = SentenceTransformer(cfg["model_name"])

query = "The pipe under my sink is leaking"
vec   = model.encode([cfg["query_prefix"] + query], normalize_embeddings=True).astype("float32")
distances, indices = index.search(vec.reshape(1, -1), 3)
print("Top-3 similar ticket indices:", indices[0])