clean final: HF Inference API embeddings (no disk)

rag.py

@@ -6,20 +6,17 @@ from typing import List, Tuple
 
 from langchain.text_splitter import RecursiveCharacterTextSplitter
 from langchain_community.vectorstores import FAISS
-from langchain_huggingface import HuggingFaceEmbeddings, HuggingFaceEndpoint
+from langchain_huggingface import HuggingFaceEndpoint
 from langchain_core.prompts import PromptTemplate
 from langchain.chains import RetrievalQA
 from supabase import create_client
 
-# ------------------------------------------------------------------
-# CONFIG
-# ------------------------------------------------------------------
+# ------------------------------------------------------------------ CONFIG
 DATASET_API = "https://datasets-server.huggingface.co/rows"
 DATASET     = "NimrodDev/LD_Events2"
 CONFIG      = "default"
 SPLIT       = "train"
 LIMIT       = 500
-EMBED_MODEL = "sentence-transformers/all-MiniLM-L6-v2"
 LLM_MODEL   = "microsoft/DialoGPT-medium"
 SUPABASE_URL = os.getenv("SUPABASE_URL")
 SUPABASE_KEY = os.getenv("SUPABASE_KEY")
@@ -27,24 +24,20 @@ HF_TOKEN     = os.getenv("HF_TOKEN")
 
 supabase = create_client(SUPABASE_URL, SUPABASE_KEY)
 
-# ------------------------------------------------------------------
-# INTENT REGEX
-# ------------------------------------------------------------------
+# ------------------------------------------------------------------ INTENT
+import re
 GREETING_RE  = re.compile(r"\b(hi|hello|hey|good morning|good afternoon|good evening)\b", re.I)
 THANKS_RE    = re.compile(r"\b(thank|thanks|appreciate)\b", re.I)
 BYE_RE       = re.compile(r"\b(bye|goodbye|see you|later)\b", re.I)
 MONEY_RE     = re.compile(r"\b(price|cost|budget|cheap|expensive|money|usd|ksh|payment|deposit)\b", re.I)
 COMPLAIN_RE  = re.compile(r"\b(complain|bad|terrible|awful|disappointed|angry|slow|rude)\b", re.I)
 
-# ------------------------------------------------------------------
-# FALLBACKS
-# ------------------------------------------------------------------
+# ------------------------------------------------------------------ FALLBACKS
 FALLBACKS = {
     "LD Events": {
         "greeting": "Hello! 👋 I’m Amina, your assistant for *LD Events* (weddings, graduations, corporate events) "
                     "and *Lamaki Designs* (construction & architectural plans).\n\n"
-                    "Which service would you like to know about?\n\n"
-                  ,
+                    "Which service would you like to know about?\n\n",
         "money": "Our pricing depends on venue / project size. Please share a few details so we can give you a tailored quote.",
         "complain": "We’re sorry to hear this. A senior agent will contact you within 30 minutes to resolve the issue.",
         "thanks": "You’re welcome! If you need anything else, just text back.",
@@ -54,8 +47,7 @@ FALLBACKS = {
     "Lamaki Designs": {
         "greeting": "Karibu! 🏗️ I’m Amina, your assistant for *Lamaki Designs* (construction, architectural plans, project management) "
                     "and *LD Events* (weddings, graduations, corporate events).\n\n"
-                    "Which service would you like to know about?\n\n"
-                    ,
+                    "Which service would you like to know about?\n\n",
         "money": "Cost varies by project size and materials. Kindly share your plot size / plan so we can estimate for you.",
         "complain": "We apologise for the inconvenience. Our site manager will call you within 30 minutes to sort it out.",
         "thanks": "Asante! Feel free to text any time.",
@@ -64,9 +56,7 @@ FALLBACKS = {
     }
 }
 
-# ------------------------------------------------------------------
-# HELPERS
-# ------------------------------------------------------------------
+# ------------------------------------------------------------------ HELPERS
 def _company_from_text(text: str) -> str:
     t = text.lower()
     if any(k in t for k in ("ld events", "event", "wedding", "venue", "graduation")):
@@ -86,9 +76,7 @@ def _detect_intent(text: str) -> str:
 def _fallback_answer(company: str, intent: str) -> str:
     return FALLBACKS[company].get(intent, FALLBACKS[company]["default"])
 
-# ------------------------------------------------------------------
-# BULLET-PROOF ONLINE FETCH – RETURNS EMPTY LIST ON ANY ERROR
-# ------------------------------------------------------------------
+# ------------------------------------------------------------------ DATA FETCH
 @lru_cache(maxsize=1)
 def get_texts() -> List[str]:
     try:
@@ -103,38 +91,30 @@ def get_texts() -> List[str]:
         print(f"⚠ Dataset fetch failed: {e} – using empty corpus")
         return []
 
-# ------------------------------------------------------------------
-# ------------------------------------------------------------------
-# ------------------------------------------------------------------
+# ------------------------------------------------------------------ EMBEDDINGS
 @lru_cache(maxsize=1)
 def get_vectorstore() -> FAISS:
     texts = get_texts()
 
-    # --- FINAL: use HF Inference Providers router (no disk) ---------------
-       # --- FINAL: HF Inference Providers (no disk, no cache) ----------------
- # --- FINAL: HF Inference Providers (no disk, no cache) ------------------
+    # HF Inference Providers – zero disk, zero cache
     from langchain_huggingface import HuggingFaceInferenceAPIEmbeddings
     embeddings = HuggingFaceInferenceAPIEmbeddings(
-    api_key=HF_TOKEN,
-    model_name="sentence-transformers/all-MiniLM-L6-v2"
-)
-# ------------------------------------------------------------------------
-    # ------------------------------------------------------------------------
-    # ------------------------------------------------------------------------
-
-    if not texts:                       # no data → empty FAISS
+        api_key=HF_TOKEN,
+        model_name="sentence-transformers/all-MiniLM-L6-v2"
+    )
+
+    if not texts:
         return FAISS.from_texts([""], embeddings)  # dummy
 
     splitter = RecursiveCharacterTextSplitter(chunk_size=600, chunk_overlap=50)
     docs = splitter.create_documents(texts, metadatas=[{"source": DATASET}] * len(texts))
     return FAISS.from_documents(docs, embeddings)
 
-# ------------------------------------------------------------------# LLM
-# ------------------------------------------------------------------
+# ------------------------------------------------------------------ LLM
 @lru_cache(maxsize=1)
 def get_llm():
     return HuggingFaceEndpoint(
-        repo_id=LLM_MODEL,
+        repo_id="microsoft/DialoGPT-medium",
         temperature=0.1,
         max_new_tokens=150,
         huggingfacehub_api_token=HF_TOKEN
@@ -146,29 +126,19 @@ Context: {context}
 Question: {question}
 Answer:""")
 
-# ------------------------------------------------------------------
-# MAIN ENTRY – NEVER CRASHES
-# ------------------------------------------------------------------
+# ------------------------------------------------------------------ MAIN
 def ask_question(phone: str, question: str) -> Tuple[str, List]:
     intent  = _detect_intent(question)
     company = _company_from_text(question)
 
-    # unbiased interactive greeting
-    if intent == "greeting":
-        answer = _fallback_answer(company, "greeting")
-        _save_chat(phone, question, answer)
-        return answer, []
-
-    # other small-talk
-    if intent in ("thanks", "bye"):
+    if intent in ("greeting", "thanks", "bye"):
         answer = _fallback_answer(company, intent)
         _save_chat(phone, question, answer)
         return answer, []
 
-    # RAG path – same index every call (empty index → no docs → fallback)
     vs   = get_vectorstore()
     docs = vs.similarity_search(question, k=3)
-    if not docs or docs[0].page_content.strip() == "":  # empty dummy
+    if not docs or docs[0].page_content.strip() == "":
         answer = _fallback_answer(company, intent if intent in ("money", "complain") else "default")
         _save_chat(phone, question, answer)
         return answer, []