app.py

import gradio as gr
import copy
from llama_cpp import Llama
from huggingface_hub import hf_hub_download
import chromadb
from sentence_transformers import SentenceTransformer
import logging

# Initialize logging
logging.basicConfig(level=logging.INFO)

# Initialize the Llama model
llm = Llama(
    model_path=hf_hub_download(
        repo_id="microsoft/Phi-3-mini-4k-instruct-gguf",
        filename="Phi-3-mini-4k-instruct-q4.gguf",
    ),
    n_ctx=2048,
    n_gpu_layers=50,  # Adjust based on your VRAM
)

# Initialize ChromaDB Vector Store
class VectorStore:
    def __init__(self, collection_name):
        self.embedding_model = SentenceTransformer('sentence-transformers/multi-qa-MiniLM-L6-cos-v1')
        self.chroma_client = chromadb.Client()
        self.collection = self.chroma_client.create_collection(name=collection_name)

    def populate_vectors(self, texts, ids):
        embeddings = self.embedding_model.encode(texts, batch_size=32).tolist()
        for text, embedding, doc_id in zip(texts, embeddings, ids):
            self.collection.add(embeddings=[embedding], documents=[text], ids=[doc_id])

    def search_context(self, query, n_results=1):
        query_embedding = self.embedding_model.encode([query]).tolist()
        results = self.collection.query(query_embeddings=query_embedding, n_results=n_results)
        return results['documents']

# Example initialization (assuming you've already populated the vector store)
vector_store = VectorStore("embedding_vector")

def generate_text(
    message,
    history: list[tuple[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
):
    # Retrieve context from vector store
    context_results = vector_store.search_context(message, n_results=1)
    context = context_results[0] if context_results else ""

    input_prompt = f"[INST] <<SYS>>\n{system_message}\n<</SYS>>\n\n {context}\n"
    for interaction in history:
        input_prompt += f"{interaction[0]} [/INST] {interaction[1]} </s><s> [INST] "
    input_prompt += f"{message} [/INST] "

    logging.info("Input prompt:\n%s", input_prompt)  # Debugging output

    temp = ""
    output = llm(
        input_prompt,
        temperature=temperature,
        top_p=top_p,
        top_k=40,
        repeat_penalty=1.1,
        max_tokens=max_tokens,
        stop=["", " \n", "ASSISTANT:", "USER:", "SYSTEM:"],
        stream=True,
    )
    for out in output:
        temp += out["choices"][0]["text"]
        logging.info("Model output:\n%s", temp)  # Log model output
        yield temp

# Define the Gradio interface
demo = gr.Interface(
    fn=generate_text,
    title="LLM Chatbot with ChromaDB Integration",
    description="Generate responses based on context and user queries.",
    examples=[
        ["I have leftover rice, what can I make out of it?"],
        ["Can I make lunch for two people with this?"],
    ],
    inputs=[
        gr.Textbox(label="Message"),
        gr.Textbox(label="System message", default="You are a friendly Chatbot."),
        gr.Textbox(label="History", default="[('USER', 'Hi there!')]"),
        gr.Slider(minimum=1, maximum=2048, step=1, default=512, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, step=0.1, default=0.7, label="Temperature"),
        gr.Slider(minimum=0.1, maximum=1.0, step=0.05, default=0.95, label="Top-p (nucleus sampling)"),
    ],
    outputs=gr.Textbox(label="Response"),
    live=True,
)

if __name__ == "__main__":
    demo.launch()