CODEFIX: Added lazy loading and specific requirements, trying to solve problem with this not running via HuggingFace spaces

app.py

@@ -185,17 +185,26 @@ Question:
     ]
 )
 
-# 🔹 Load and process dataset once
-if "initialized" not in st.session_state:
-    dataset = load_dataset("andreska/Adrega61Docs", split="train")
-
-    def read_dataset(dataset):
-        return "\n\n".join(
-            [
-                f"Title: {item['title']}\nModule: {item['module']}\nContent: {item['content']}"
-                for item in dataset
-            ]
-        )
+# 🔹 Helper — lazy-load RAG resources ONCE on the first user interaction.
+
+def _ensure_rag_resources():
+    """Load dataset + embeddings model + FAISS index lazily (on demand).
+    HF Spaces Free tier has limited RAM, so we defer this heavy work until the first user
+    message rather than loading eagerly at container startup (which often OOMs before any user opens the chat page)."""
+
+    if "rag_initialized" in st.session_state and st.session_state["rag_initialized"]:
+        return  # resources already built for this session, skip
+
+    with st.spinner("Loading RAG resources (embeddings model + FAISS index)… this may take ~30s"):
+        dataset = load_dataset("andreska/Adrega61Docs", split="train")
+
+        def read_dataset(dataset):
+            return "\n\n".join(
+                [
+                    f"Title: {item['title']}\nModule: {item['module']}\nContent: {item['content']}"
+                    for item in dataset
+                ]
+            )
 
     raw_text = read_dataset(dataset)
     text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=100)
@@ -213,8 +222,7 @@ if "initialized" not in st.session_state:
 
     st.session_state.db = db
     st.session_state.docs = docs
-    st.session_state.initialized = True
-
+    st.session_state.rag_initialized = True
 
 # 🔹 Ollama LLM setup
 def get_llm():
@@ -242,9 +250,8 @@ def get_llm():
     )
 
 
-# 🔹 Build the LLM and retriever
+# 🔹 Build the LLM (retriever is created lazily inside the chat handler)
 llm = get_llm()
-retriever = st.session_state.db.as_retriever(search_kwargs={"k": 5})
 
 # 🔹 UI Styling
 st.markdown(
@@ -424,6 +431,8 @@ user_input = st.chat_input("Ask a question about Adrega:")
 
 # 🔹 Process chat
 if user_input:
+    _ensure_rag_resources()
+    retriever = st.session_state.db.as_retriever(search_kwargs={"k": 5})
     st.session_state.chat_history.append({"role": "user", "content": user_input})
 
     # Truncate oldest messages if history exceeds limit

requirements.txt

@@ -1,20 +1,30 @@
 datasets
 openai>=1.54.0,<2.0.0
 streamlit
-torch
+torch-cpu>=2.1  # explicit CPU build so HF Spaces Free tier doesn't try to wire CUDA
 httpx
 python-dotenv
 
-# Use compatible versions that avoid dependency conflicts
+# Core LangChain
 langchain
 langchain-community
-langchain-text-splitters
-langchain-core
+langchain-core~=0.3
+langchain-text-splitters~=0.3
 
-# integrations (use versions that support 0.3.x)
+# Integrations (compatible with langchain 0.3.x)
 langchain-huggingface
 langchain-openai
 
-sentence-transformers>=2.2,<3.0
-faiss-cpu
-accelerate
+# Embeddings + Vector store — CPU-only variant for HF Spaces (no CUDA)
+sentence-transformers>=2.6,<3.0  # pinned to avoid breaking changes between transformers 4.x versions
+faiss-cpu>=1.7,<2.0  # only load faiss CPU build; HF free tier has no GPU so faiss-gpu would never work
+
+# Model inference acceleration — pin for stable installs on free/restricted builds
+accelerate>=0.34,<1.0
+
+# HuggingFace auth helpers (used implicitly by datasets + langchain-huggingface)
+# If `HF_API_KEY` env var is set, HF Hub auto-authenticates to the corresponding account.
+huggingface_hub>=0.23,<1.0
+
+# Optional image handling for downstream RAG pipelines (PDF/image loaders — not used today but kept for future-proofing)
+Pillow>=9.5