o

rag.py

@@ -1,4 +1,4 @@
-# rag.py  –  full unbiased, zero-disk, HF-Space-safe edition
+# rag.py  –  single-index, zero-disk, HF-Space-safe edition
 from __future__ import annotations
 import os, re
 from functools import lru_cache
@@ -15,7 +15,7 @@ from supabase import create_client
 # ------------------------------------------------------------------
 # CONFIG
 # ------------------------------------------------------------------
-HF_DS        = "NimrodDev/LD_Events2"
+HF_DS        = "NimrodDev/LD_Events2"          # parquet branch auto-converted
 EMBED_MODEL  = "sentence-transformers/all-MiniLM-L6-v2"
 LLM_MODEL    = "microsoft/DialoGPT-medium"
 SUPABASE_URL = os.getenv("SUPABASE_URL")
@@ -78,7 +78,7 @@ def _company_from_text(text: str) -> str:
         return "LD Events"
     if any(k in t for k in ("lamaki", "construction", "build", "site", "bungalow", "architect")):
         return "Lamaki Designs"
-    return "LD Events"  # default when nothing matches
+    return "LD Events"
 
 def _detect_intent(text: str) -> str:
     if GREETING_RE.search(text): return "greeting"
@@ -92,14 +92,14 @@ def _fallback_answer(company: str, intent: str) -> str:
     return FALLBACKS[company].get(intent, FALLBACKS[company]["default"])
 
 # ------------------------------------------------------------------
-# RAM-ONLY DOCUMENT LOADER
+# RAM-ONLY DOCUMENT LOADER – PARQUET BRANCH
 # ------------------------------------------------------------------
 def load_texts() -> List[str]:
-    ds = load_dataset(HF_DS, split="train")
+    ds = load_dataset(HF_DS, revision="refs/convert/parquet", split="train")
     return [row["text"] for row in ds if row.get("text")]
 
 # ------------------------------------------------------------------
-# RAM-ONLY VECTOR STORE
+# SINGLE-BUILD VECTOR STORE (cached for life of worker)
 # ------------------------------------------------------------------
 @lru_cache(maxsize=1)
 def get_vectorstore() -> FAISS:
@@ -107,7 +107,7 @@ def get_vectorstore() -> FAISS:
     splitter = RecursiveCharacterTextSplitter(chunk_size=600, chunk_overlap=50)
     docs = splitter.create_documents(texts, metadatas=[{"source": HF_DS}] * len(texts))
     embeddings = HuggingFaceEmbeddings(model_name=EMBED_MODEL)
-    return FAISS.from_documents(docs, embeddings)
+    return FAISS.from_documents(docs, embeddings)   # <- built ONCE
 
 # ------------------------------------------------------------------
 # LLM
@@ -146,7 +146,7 @@ def ask_question(phone: str, question: str) -> Tuple[str, List]:
         _save_chat(phone, question, answer)
         return answer, []
 
-    # RAG path
+    # RAG path – re-uses the *same* index every call
     vs   = get_vectorstore()
     docs = vs.similarity_search(question, k=3)
     if not docs: