Spaces:
Sleeping
Sleeping
feat: add server-side script to rebuild FAISS index for deployment environments
Browse files- rebuild_index_server.py +59 -0
rebuild_index_server.py
ADDED
|
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
SERVER-SIDE rebuild script (Linux / Azure App Service).
|
| 3 |
+
Run this once on the deployed server to rebuild the FAISS index with all-mpnet-base-v2.
|
| 4 |
+
|
| 5 |
+
Usage:
|
| 6 |
+
cd /home/user/app
|
| 7 |
+
python rebuild_index_server.py
|
| 8 |
+
"""
|
| 9 |
+
import os
|
| 10 |
+
os.environ["OMP_NUM_THREADS"] = "1"
|
| 11 |
+
os.environ["MKL_NUM_THREADS"] = "1"
|
| 12 |
+
os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
|
| 13 |
+
|
| 14 |
+
import sys
|
| 15 |
+
sys.path.insert(0, "/home/user/app")
|
| 16 |
+
|
| 17 |
+
import numpy as np
|
| 18 |
+
import pandas as pd
|
| 19 |
+
import faiss
|
| 20 |
+
from pathlib import Path
|
| 21 |
+
from sentence_transformers import SentenceTransformer
|
| 22 |
+
|
| 23 |
+
MODEL_NAME = "all-mpnet-base-v2"
|
| 24 |
+
MODELS_DIR = Path("/home/user/app/models")
|
| 25 |
+
PARQUET_PATH = MODELS_DIR / "metadata.parquet"
|
| 26 |
+
INDEX_PATH = MODELS_DIR / "faiss_index.bin"
|
| 27 |
+
EMBED_PATH = MODELS_DIR / "project_embeddings.npy"
|
| 28 |
+
|
| 29 |
+
print(f"[SERVER REBUILD] Model: {MODEL_NAME}")
|
| 30 |
+
model = SentenceTransformer(MODEL_NAME)
|
| 31 |
+
dim = model.get_sentence_embedding_dimension()
|
| 32 |
+
print(f"[SERVER REBUILD] Embedding dim: {dim}")
|
| 33 |
+
|
| 34 |
+
df = pd.read_parquet(PARQUET_PATH)
|
| 35 |
+
print(f"[SERVER REBUILD] Loaded {len(df)} projects")
|
| 36 |
+
|
| 37 |
+
def build_text(row):
|
| 38 |
+
return f"{row.get('project_title','') or ''}. {row.get('abstract','') or ''}. {row.get('description','') or ''}".strip()
|
| 39 |
+
|
| 40 |
+
texts = df.apply(build_text, axis=1).tolist()
|
| 41 |
+
|
| 42 |
+
print("[SERVER REBUILD] Generating embeddings...")
|
| 43 |
+
embeddings = model.encode(
|
| 44 |
+
texts,
|
| 45 |
+
convert_to_numpy=True,
|
| 46 |
+
normalize_embeddings=True,
|
| 47 |
+
show_progress_bar=True,
|
| 48 |
+
batch_size=8
|
| 49 |
+
).astype("float32")
|
| 50 |
+
|
| 51 |
+
print(f"[SERVER REBUILD] Embeddings shape: {embeddings.shape}")
|
| 52 |
+
np.save(str(EMBED_PATH), embeddings)
|
| 53 |
+
|
| 54 |
+
index = faiss.IndexFlatIP(dim)
|
| 55 |
+
index.add(embeddings)
|
| 56 |
+
faiss.write_index(index, str(INDEX_PATH))
|
| 57 |
+
|
| 58 |
+
print(f"[SERVER REBUILD] Done! Index: {index.ntotal} vectors @ {dim}-dim")
|
| 59 |
+
print("[SERVER REBUILD] Restart the app process to reload the index.")
|