Update app.py

app.py

@@ -4,16 +4,15 @@ from langchain.chains import ConversationalRetrievalChain
 from langchain.document_loaders import PyPDFLoader, Docx2txtLoader, TextLoader
 from langchain.text_splitter import CharacterTextSplitter
 from langchain.vectorstores import Chroma
-import gradio as gr
-from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
 from sentence_transformers import SentenceTransformer
-import torch
+from transformers import pipeline
+import gradio as gr
 
-# sqlite workaround for HuggingFace Spaces
+# Workaround for sqlite in HuggingFace Spaces
 __import__('pysqlite3')
 sys.modules['sqlite3'] = sys.modules.pop('pysqlite3')
 
-# Load documents
+# 📄 Load documents
 docs = []
 for f in os.listdir("multiple_docs"):
     if f.endswith(".pdf"):
@@ -26,43 +25,37 @@ for f in os.listdir("multiple_docs"):
         loader = TextLoader(os.path.join("multiple_docs", f))
         docs.extend(loader.load())
 
-# Split docs
+# 🔗 Split into chunks
 splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=10)
 docs = splitter.split_documents(docs)
 
-# Embeddings
+# 🧠 Compute embeddings
 embedding_model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
 texts = [doc.page_content for doc in docs]
 metadatas = [{"id": i} for i in range(len(texts))]
 embeddings = embedding_model.encode(texts)
 
-# Vectorstore
+# 🗃️ Save in Chroma vectorstore
 vectorstore = Chroma(persist_directory="./db")
 vectorstore.add_texts(texts=texts, metadatas=metadatas, embeddings=embeddings)
 vectorstore.persist()
 
+# 🤖 Load free LLM with pipeline
+model_name = "google/flan-t5-large"  # small enough for CPU
+generator = pipeline("text2text-generation", model=model_name, device=-1)  # -1 → CPU
 
-# model_name = "deepseek-ai/deepseek-llm-7b-instruct" 
-# tokenizer = AutoTokenizer.from_pretrained(model_name)
-# model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16, device_map="auto")
-
-model_name = "google/flan-t5-large"
-model = AutoModelForCausalLM.from_pretrained(model_name)
-tokenizer = AutoTokenizer.from_pretrained(model_name)
-
-
-def generate(prompt):
-    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
-    outputs = model.generate(**inputs, max_new_tokens=512)
-    return tokenizer.decode(outputs[0], skip_special_tokens=True)
-
+# 🔗 Wrap the pipeline for langchain
 class HuggingFaceLLMWrapper:
+    def __init__(self, generator):
+        self.generator = generator
+
     def __call__(self, prompt, **kwargs):
-        return generate(prompt)
+        result = self.generator(prompt, max_new_tokens=512, num_return_sequences=1)
+        return result[0]['generated_text']
 
-llm = HuggingFaceLLMWrapper()
+llm = HuggingFaceLLMWrapper(generator)
 
-# QA chain
+# 🔗 Create the conversational chain
 chain = ConversationalRetrievalChain.from_llm(
     llm,
     retriever=vectorstore.as_retriever(search_kwargs={'k': 6}),
@@ -70,16 +63,21 @@ chain = ConversationalRetrievalChain.from_llm(
     verbose=False
 )
 
+# 💬 Gradio UI
 chat_history = []
 
 with gr.Blocks() as demo:
-    chatbot = gr.Chatbot([("", "Hello, I'm Thierry Decae's chatbot. Ask me about my experience, skills, eligibility, etc.")],
-                         avatar_images=["./multiple_docs/Guest.jpg", "./multiple_docs/Thierry Picture.jpg"])
+    chatbot = gr.Chatbot(
+        [("", "Hello, I'm Thierry Decae's chatbot. Ask me about my experience, skills, eligibility, etc.")],
+        avatar_images=["./multiple_docs/Guest.jpg", "./multiple_docs/Thierry Picture.jpg"]
+    )
     msg = gr.Textbox()
     clear = gr.Button("Clear")
 
     def user(query, chat_history):
+        # convert chat history to tuples
         chat_history_tuples = [(m[0], m[1]) for m in chat_history]
+        # get answer
         result = chain({"question": query, "chat_history": chat_history_tuples})
         chat_history.append((query, result["answer"]))
         return gr.update(value=""), chat_history