llmproccrs / scripts /run_ingestion.py
aitek230telu's picture
Upload 3 files
fbc0c1d verified
Raw History Blame Contribute Delete
4.07 kB
import sys
import os
import json
import gc
from tqdm import tqdm
from sentence_transformers import SentenceTransformer
# Hack path agar bisa import modules/config dari root
sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
from config import Config
from modules.domain_knowledge.vector_store import VectorStore
from modules.domain_knowledge.data_processor import DataProcessor
def run():
print("="*40)
print("🚀 STARTING DATA INGESTION PROCESS")
print("="*40)
# 1. Initialize Components
processor = DataProcessor(
Config.PATHS["categories"],
Config.PATHS["supplier"],
Config.PATHS["kbli"],
Config.PATHS["supplier_kbli"]
)
vdb = VectorStore(Config.CHROMA_DB_DIR, Config.COLLECTION_NAME)
print(f"[Model] Loading Embedding Model: {Config.EMBED_MODEL_NAME}...")
model = SentenceTransformer(Config.EMBED_MODEL_NAME, device=Config.DEVICE)
print(f"[Model] Loaded on {Config.DEVICE}")
# 2. Count Total Rows (for progress bar)
products_path = Config.PATHS["products"]
if not os.path.exists(products_path):
print(f"❌ Error: File not found {products_path}")
return
total_rows = sum(1 for line in open(products_path, 'r', encoding='utf-8') if line.strip())
print(f"[Data] Total raw products to process: {total_rows}")
# 3. Processing Loop
batch_ids = []
batch_texts = []
batch_metas = []
success_count = 0
skipped_count = 0
with open(products_path, 'r', encoding='utf-8') as f:
# Gunakan TQDM untuk progress bar
for line in tqdm(f, total=total_rows, desc="Ingesting", unit="item"):
if not line.strip(): continue
# A. Parse & Enrich
try:
raw_prod = json.loads(line)
enriched_prod = processor.enrich_product(raw_prod)
if not enriched_prod:
skipped_count += 1
continue # Skip jika supplier tidak ditemukan
# B. Create Text & Metadata
text_repr = processor.create_text_representation(enriched_prod)
metadata = processor.prepare_metadata(enriched_prod)
# Gunakan product_id sebagai ID vector
pid = str(enriched_prod.get("product_id"))
# C. Add to Buffer
batch_ids.append(pid)
batch_texts.append(text_repr)
batch_metas.append(metadata)
except json.JSONDecodeError:
continue
# D. Flush Batch to DB
if len(batch_ids) >= Config.BATCH_SIZE_DB:
_flush_batch(vdb, model, batch_ids, batch_texts, batch_metas)
success_count += len(batch_ids)
# Reset Buffer & Clean Memory
batch_ids, batch_texts, batch_metas = [], [], []
gc.collect()
# E. Flush Sisa Data (Remaining Batch)
if batch_ids:
_flush_batch(vdb, model, batch_ids, batch_texts, batch_metas)
success_count += len(batch_ids)
print("\n" + "="*40)
print("✅ INGESTION COMPLETED")
print(f"Total Success Inserted : {success_count}")
print(f"Total Skipped (No Data): {skipped_count}")
print("="*40)
def _flush_batch(vdb, model, ids, texts, metas):
"""Helper function to encode and save"""
# Generate Embeddings
# Perintah 'query' (e5 model) biasanya butuh prefix 'query: ' atau 'passage: '
# Untuk document storage di DB, E5 menyarankan prefix "passage: "
processed_texts = [f"passage: {t}" for t in texts]
embeddings = model.encode(
processed_texts,
batch_size=Config.BATCH_SIZE_EMB,
convert_to_numpy=True,
show_progress_bar=False,
normalize_embeddings=True # Penting untuk cosine similarity
)
# Save to Chroma
vdb.add_documents(ids, texts, embeddings, metas)
if __name__ == "__main__":
run()