app.py

import os
import sys
from langchain.chains import ConversationalRetrievalChain
from langchain.document_loaders import PyPDFLoader, Docx2txtLoader, TextLoader
from langchain.text_splitter import CharacterTextSplitter
from langchain.vectorstores import Chroma
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.llms.base import LLM
from huggingface_hub import InferenceClient
import gradio as gr

# workaround for sqlite in HF spaces
__import__('pysqlite3')
sys.modules['sqlite3'] = sys.modules.pop('pysqlite3')

# 📄 Load documents
docs = []
for f in os.listdir("multiple_docs"):
    if f.endswith(".pdf"):
        loader = PyPDFLoader(os.path.join("multiple_docs", f))
        docs.extend(loader.load())
    elif f.endswith(".docx") or f.endswith(".doc"):
        loader = Docx2txtLoader(os.path.join("multiple_docs", f))
        docs.extend(loader.load())
    elif f.endswith(".txt"):
        loader = TextLoader(os.path.join("multiple_docs", f))
        docs.extend(loader.load())

# 🔗 Split into chunks
splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=10)
docs = splitter.split_documents(docs)

texts = [doc.page_content for doc in docs]
metadatas = [{"id": i} for i in range(len(texts))]

# 🧠 Embeddings
embedding_function = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2")

# 🗃️ Vectorstore
vectorstore = Chroma(
    persist_directory="./db",
    embedding_function=embedding_function
)
vectorstore.add_texts(texts=texts, metadatas=metadatas)
vectorstore.persist()

# 🔐 Get HF token from env variable
HF_API_KEY = os.getenv("HF_API_KEY")
if HF_API_KEY is None:
    raise ValueError("HF_API_KEY environment variable is not set.")

HF_MODEL = "deepseek-ai/deepseek-llm-7b-base"

# 🤖 Create InferenceClient bound to model
client = InferenceClient(model=HF_MODEL, token=HF_API_KEY)

# 🔷 Wrap HF client into LangChain LLM interface
class HuggingFaceInferenceLLM(LLM):
    """LLM that queries HuggingFace Inference API."""

    client: InferenceClient = client

    def _call(self, prompt, stop=None, run_manager=None, **kwargs):
        response = self.client.text_generation(
            prompt,
            max_new_tokens=512,
            temperature=0.7,
            do_sample=True,
        )
        return response

    @property
    def _llm_type(self) -> str:
        return "huggingface_inference_api"

llm = HuggingFaceInferenceLLM()

# 🔗 Conversational chain
chain = ConversationalRetrievalChain.from_llm(
    llm,
    retriever=vectorstore.as_retriever(search_kwargs={'k': 6}),
    return_source_documents=True,
    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"]
    )
    msg = gr.Textbox(placeholder="Type your question here...")
    clear = gr.Button("Clear")

    def user(query, chat_history):
        chat_history_tuples = [(m[0], m[1]) for m in chat_history]
        result = chain({"question": query, "chat_history": chat_history_tuples})
        chat_history.append((query, result["answer"]))
        return gr.update(value=""), chat_history

    msg.submit(user, [msg, chatbot], [msg, chatbot], queue=False)
    clear.click(lambda: None, None, chatbot, queue=False)

demo.launch(debug=True)  # remove share=True if running in HF Spaces