app.py

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

# 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",
    # ),
    model_path=hf_hub_download(
        repo_id="Ankitajadhav/Phi-3-mini-4k-instruct-q4.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):
    #     embeddings = self.embedding_model.encode(texts, batch_size=32).tolist()
    #     for text, embedding in zip(texts, embeddings, ids):
    #         self.collection.add(embeddings=[embedding], documents=[text], ids=[doc_id])

        # Method to populate the vector store with embeddings from a dataset
    def populate_vectors(self, dataset):
        # Select the text columns to concatenate
        # title = dataset['train']['title_cleaned'][:1000]  # Limiting to 100 examples for the demo
        recipe = dataset['train']['recipe_new'][:1000]
        allergy = dataset['train']['allergy_type'][:1000]
        ingredients = dataset['train']['ingredients_alternatives'][:1000]

        # Concatenate the text from both columns
        texts = [f"{rep} {ingr} {alle}" for rep, ingr,alle in zip(recipe, ingredients,allergy)]
        for i, item in enumerate(texts):
            embeddings = self.embedding_model.encode(item).tolist()
            self.collection.add(embeddings=[embeddings], documents=[item], ids=[str(i)])

    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)
dataset = load_dataset('Thefoodprocessor/recipe_new_with_features_full')
vector_store = VectorStore("embedding_vector")
vector_store.populate_vectors(dataset)
def format_recipe(input_string):
    # Clean up the input
    cleaned_text = input_string.strip("[]'").replace('\\n', '\n')
    
    # Split the text into lines
    lines = cleaned_text.split('\n')
    
    # Initialize sections
    title = lines[0]
    ingredients = []
    instructions = []
    substitutions = []

    # Extract ingredients and instructions
    in_instructions = False
    for line in lines[1:]:
        if line.startswith("Instructions:"):
            in_instructions = True
            continue
        
        if in_instructions:
            if line.strip():  # Check for non-empty lines
                instructions.append(line.strip())
        else:
            if line.strip():  # Check for non-empty lines
                ingredients.append(line.strip())
    
    # Gather substitutions from the last few lines
    for line in lines:
        if ':' in line:
            substitutions.append(line.strip())

    # Format output
    formatted_recipe = f"## {title}\n\n### Ingredients:\n"
    formatted_recipe += '\n'.join(f"- {item}" for item in ingredients) + "\n\n"
    formatted_recipe += "### Instructions:\n" + '\n'.join(f"{i + 1}. {line}" for i, line in enumerate(instructions)) + "\n\n"
    
    if substitutions:
        formatted_recipe += "### Substitutions:\n" + '\n'.join(f"- **{line.split(':')[0].strip()}**: {line.split(':')[1].strip()}" for line in substitutions) + "\n"
    return formatted_recipe
    # print(formatted_recipe)
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] "

    print("Input prompt:", 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 += format_recipe(out["choices"][0]["text"])
        yield temp

# Define the Gradio interface
demo = gr.ChatInterface(
    generate_text,
    title="llama-cpp-python on GPU with ChromaDB",
    description="Running LLM with context retrieval from ChromaDB",
    examples=[
        ["I have leftover rice, what can I make out of it?"],
        ["Can I make lunch for two people with this?"],
    ],
    cache_examples=False,
    retry_btn=None,
    undo_btn="Delete Previous",
    clear_btn="Clear",
    additional_inputs=[
        gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p (nucleus sampling)"),
    ],
)

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