aitek230telu's picture
Upload 40 files
17131ee verified
Raw History Blame Contribute Delete
9.66 kB
import json
import re
from config import Config
from modules.domain_knowledge.vector_store import VectorStore
from sentence_transformers import SentenceTransformer
class RREngine:
def __init__(self):
print("[RRE] Initializing Reasoning Engine (Multi-Location Support)...")
self.vdb = VectorStore(Config.CHROMA_DB_DIR, Config.COLLECTION_NAME)
self.model = SentenceTransformer(Config.EMBED_MODEL_NAME, device=Config.DEVICE)
print("[RRE] Ready.")
def run_pipeline_chatrec(self, state: dict) -> dict:
query_text = state.get("user_text", "")
# constraints = state.get("collected_constraints", {})
if not query_text:
state["candidates"] = []
state["candidate_count"] = 0
return state
# --- STEP 1: RETRIEVAL ---
formatted_query = f"query: {query_text}"
query_embedding = self.model.encode([formatted_query], convert_to_numpy=True, normalize_embeddings=True)[0]
# Ambil 150 kandidat
raw_docs, raw_metas, raw_dists = self.vdb.search(query_embedding, n_results=150)
candidates = []
for meta in raw_metas:
full_data = {}
if "original_json" in meta:
try: full_data = json.loads(meta["original_json"])
except: full_data = {}
if not full_data.get("title"):
full_data["title"] = meta.get("title", "Unknown Product")
full_data["price"] = meta.get("price", "0")
full_data["product_id"] = meta.get("product_id")
full_data["supplier_detail"] = {
"nama_toko": meta.get("supplier") or meta.get("seller", "Unknown Store"),
"kota": meta.get("location", ""),
"provinsi": ""
}
full_data["category_detail"] = {
"category": meta.get("category", ""),
"sub_category": ""
}
candidates.append(full_data)
state["candidates"] = candidates
state["candidate_count"] = len(candidates)
# print(candidates)
return state
def run_pipeline(self, state: dict) -> dict:
query_text = state.get("relaxed_query", "")
constraints = state.get("collected_constraints", {})
if not query_text:
state["candidates"] = []
state["candidate_count"] = 0
return state
# --- STEP 1: RETRIEVAL ---
formatted_query = f"{query_text}"
print(formatted_query)
query_embedding = self.model.encode([formatted_query], convert_to_numpy=True, normalize_embeddings=True)[0]
# Ambil 150 kandidat
raw_docs, raw_metas, raw_dists = self.vdb.search(query_embedding, n_results=150)
candidates = []
for meta in raw_metas:
full_data = {}
if "original_json" in meta:
try: full_data = json.loads(meta["original_json"])
except: full_data = {}
if not full_data.get("title"):
full_data["title"] = meta.get("title", "Unknown Product")
full_data["price"] = meta.get("price", "0")
full_data["product_id"] = meta.get("product_id")
full_data["supplier_detail"] = {
"nama_toko": meta.get("supplier") or meta.get("seller", "Unknown Store"),
"kota": meta.get("location", ""),
"provinsi": ""
}
full_data["category_detail"] = {
"category": meta.get("category", ""),
"sub_category": ""
}
candidates.append(full_data)
# --- STEP 2: HARD FILTERING ---
filtered_candidates = self._apply_hard_filters(candidates, constraints)
# --- STEP 3: ONTOLOGY VALIDATION ---
validated_candidates = self._validate_ontology(filtered_candidates)
state["candidates"] = validated_candidates
state["candidate_count"] = len(validated_candidates)
# print(validated_candidates)
return state
def _apply_hard_filters(self, items, constraints):
results = []
# --- 1. HANDLING MULTI-LOCATION (List or String) ---
raw_loc = constraints.get("location")
target_locs = [] # Kita standarisasi jadi List of Strings
if isinstance(raw_loc, list):
target_locs = [str(l).lower().strip() for l in raw_loc if l]
elif isinstance(raw_loc, str) and raw_loc:
target_locs = [raw_loc.lower().strip()]
raw_supplier = constraints.get("supplier", "")
if isinstance(raw_supplier, list):
raw_supplier = raw_supplier[0] if raw_supplier else ""
target_supplier = raw_supplier.lower().strip() if raw_supplier else ""
raw_specs = constraints.get("specs", "")
if isinstance(raw_specs, list):
# Menggabungkan ['ram 8gb', 'ssd 256'] menjadi "ram 8gb ssd 256"
target_specs = " ".join(str(s) for s in raw_specs).lower().strip()
else:
target_specs = str(raw_specs).lower().strip() if raw_specs else ""
# --- 2. HANDLING PRICE ---
price_constraint = constraints.get("price")
target_price_val = 0
price_operator = None
if isinstance(price_constraint, dict):
raw_val = price_constraint.get("value")
target_price_val = int(raw_val) if raw_val is not None else 0
price_operator = price_constraint.get("operator", "=")
for item in items:
keep = True
title = item.get("title", "").lower()
sup_detail = item.get("supplier_detail", {})
s_kota = str(sup_detail.get('kota', '')).lower()
s_prov = str(sup_detail.get('provinsi', '')).lower()
s_toko = str(sup_detail.get('nama_toko', '')).lower()
s_seller = str(item.get('seller', '')).lower()
item_loc = f"{s_kota} {s_prov}"
item_supplier_name = f"{s_seller} {s_toko}"
# --- A. FILTER LOCATION (MULTI-MATCH SUPPORT) ---
if target_locs:
# Logika: Jika user minta ['banten', 'buleleng'],
# maka produk dianggap lolos jika lokasinya mengandung 'banten' ATAU 'buleleng'.
is_loc_match = False
for t_loc in target_locs:
clean_target = t_loc.replace("kab.", "").replace("kota", "").strip()
target_tokens = clean_target.split()
# Cek 1: Exact substring match (relaxed)
if clean_target in item_loc:
is_loc_match = True
break
# Cek 2: Token match (untuk 'jakarta timur' vs 'jakarta')
token_match = False
for token in target_tokens:
if len(token) > 2 and token in item_loc:
token_match = True
break
if token_match:
is_loc_match = True
break
if not is_loc_match:
keep = False
# --- B. FILTER SUPPLIER ---
if keep and target_supplier and target_supplier not in item_supplier_name:
keep = False
# --- C. FILTER PRICE ---
if keep and target_price_val > 0:
item_price_raw = item.get("price", 0)
item_price_val = self._parse_price(item_price_raw)
if item_price_val > 0:
if price_operator == "<":
if item_price_val >= target_price_val: keep = False
elif price_operator == ">":
if item_price_val <= target_price_val: keep = False
# --- D. FILTER SPECS ---
if keep and target_specs:
spec_keywords = target_specs.split()
title_norm = re.sub(r"[^a-z0-9]", "", title)
for kw in spec_keywords:
kw_clean = re.sub(r"[^a-z0-9]", "", kw)
if not kw_clean: continue
if any(char.isdigit() for char in kw_clean):
if kw_clean not in title_norm:
keep = False
break
if keep:
results.append(item)
return results
def _parse_price(self, price_input):
try:
if price_input is None: return 0
if isinstance(price_input, (int, float)):
return int(price_input)
str_val = str(price_input).strip()
str_val = re.sub(r"(Rp|IDR|\s)", "", str_val, flags=re.IGNORECASE)
if "," in str_val: str_val = str_val.split(",")[0]
str_val = str_val.replace(".", "")
clean_str = re.sub(r"[^\d]", "", str_val)
return int(clean_str) if clean_str else 0
except:
return 0
def _validate_ontology(self, items):
valid_items = []
for item in items:
if item.get("title"):
valid_items.append(item)
return valid_items