plugins/ragllm.py

from global_vars import translations, t
from app import Plugin
import streamlit as st
import yaml
from litellm import completion, embedding
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity, euclidean_distances, manhattan_distances
import os
from typing import List, Dict, Any
import requests
import torch
from transformers import AutoTokenizer, AutoModel
from huggingface_hub import InferenceClient
from langchain_huggingface import HuggingFaceEmbeddings

DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
MAX_LENGTH = 512
CHUNK_SIZE = 200
CONFIG_FILE = '.llm-config.yml'

def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0]
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)


def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0]
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)

# Ajout des traductions spécifiques à ce plugin
translations["en"].update({
    "rag_plugin_loaded": "RAG LLM Plugin loaded",
    "rag_enter_text": "Enter RAG text:",
    "rag_enter_question": "Enter your question:",
    "rag_button_get_answer": "Get an answer",
    "rag_success_text_processed": "RAG text processed successfully!",
    "rag_warning_enter_text": "Please enter RAG text.",
    "rag_warning_process_text_first": "Please process the RAG text first.",
    "rag_warning_enter_question": "Please enter a question.",
    "rag_answer": "Answer:",
    "rag_citations": "Citations:",
    "rag_model_provider": "Model Provider",
    "rag_llm_model": "LLM Model",
    "rag_embedder_model": "Embedding Model",
    "rag_similarity_method": "Similarity Method",
    "rag_llm_sys_prompt": "System prompt for LLM",
    "rag_chunk_size": "Chunk size",
    "rag_top_k_chunks": "Number of chunks to use",
    "rag_default_sys_prompt": "You are an AI assistant. Your task is to analyze the provided context and answer questions based ONLY on this context. If the information is not in the context, clearly state that.",
    "rag_error_fetching_models_ollama": "Error fetching Ollama models: ",
    "rag_error_calling_llm": "Error calling LLM: ",
    "rag_processing" : "Processing...",
    "rag_hf_api_key": "HuggingFace API Token",
    "rag_config_file_missing": "Configuration file .llm-config.yml not found. This is required for Ollama and Groq providers.",
})

translations["fr"].update({
    "rag_plugin_loaded": "Plugin RAG LLM chargé",
    "rag_enter_text": "Entrez le texte RAG :",
    "rag_enter_question": "Entrez votre question :",
    "rag_button_get_answer": "Obtenir une réponse",
    "rag_success_text_processed": "Texte RAG traité avec succès!",
    "rag_warning_enter_text": "Veuillez entrer du texte RAG.",
    "rag_warning_process_text_first": "Veuillez d'abord traiter le texte RAG.",
    "rag_warning_enter_question": "Veuillez entrer une question.",
    "rag_answer": "Réponse :",
    "rag_citations": "Citations :",
    "rag_model_provider": "Fournisseur de modèle",
    "rag_llm_model": "Modèle LLM",
    "rag_embedder_model": "Modèle d'embedding",
    "rag_similarity_method": "Méthode de similarité",
    "rag_llm_sys_prompt": "Prompt système pour le LLM",
    "rag_chunk_size": "Taille des chunks",
    "rag_top_k_chunks": "Nombre de chunks à utiliser",
    "rag_default_sys_prompt": "Tu es un assistant IA. Ta tâche est d'analyser le contexte fourni et de répondre aux questions en te basant UNIQUEMENT sur ce contexte. Si l'information n'est pas dans le contexte, dis-le clairement.",
    "rag_error_fetching_models_ollama": "Erreur lors de la récupération des modèles Ollama : ",
    "rag_error_calling_llm": "Erreur lors de l'appel au LLM : ",
    "rag_processing" : "En cours de traitement...",
     "rag_hf_api_key": "Token API HuggingFace",
    "rag_config_file_missing": "Fichier de configuration .llm-config.yml non trouvé. Ce fichier est nécessaire pour les providers Ollama et Groq.",
})

class RagllmPlugin(Plugin):
    def __init__(self, name: str, plugin_manager):
        super().__init__(name, plugin_manager)
        self.embeddings = None
        self.chunks = None
        self.hf_client = None
        self.config = {}

    def load_llm_config(self) -> Dict:
        if not os.path.exists(CONFIG_FILE):
            st.warning(t("rag_config_file_missing"))
            return {}
        with open(CONFIG_FILE, 'r') as file:
            return yaml.safe_load(file)

    def get_tabs(self):
        return [{"name": "RAG", "plugin": "ragllm"}]

    def get_config_fields(self):
        fields = {
            "provider": {
                "type": "select",
                "label": t("rag_model_provider"),
                "options": [("ollama", "Ollama"), ("groq", "Groq"), ("huggingface", "HuggingFace")],
                "default": "huggingface"
            },
            "llm_model": {
                "type": "select",
                "label": t("rag_llm_model"),
                "options": [("none", "À charger...")],
                "default": "HuggingFaceH4/zephyr-7b-beta"
            },
            "embedder": {
                "type": "select",
                "label": t("rag_embedder_model"),
                "options": [
                    ("sentence-transformers/all-MiniLM-L6-v2", "all-MiniLM-L6-v2"),
                    ("nomic-ai/nomic-embed-text-v1.5", "nomic-embed-text-v1.5")
                ],
                "default": "sentence-transformers/all-MiniLM-L6-v2"
            },
            "similarity_method": {
                "type": "select",
                "label": t("rag_similarity_method"),
                "options": [
                    ("cosine", "Cosinus"),
                    ("euclidean", "Distance euclidienne"),
                    ("manhattan", "Distance de Manhattan")
                ],
                "default": "cosine"
            },
            "llm_sys_prompt": {
                "type": "textarea",
                "label": t("rag_llm_sys_prompt"),
                "default": t("rag_default_sys_prompt")
            },
            "chunk_size": {
                "type": "number",
                "label": t("rag_chunk_size"),
                "default": 200
            },
            "top_k": {
                "type": "number",
                "label": t("rag_top_k_chunks"),
                "default": 3
            }
        }
        # Add HuggingFace API key field if provider is huggingface
        if 'provider' in self.config and self.config.get('provider') == 'huggingface':
            fields["hf_api_key"] = {
                "type": "password",
                "label": t("rag_hf_api_key"),
                "default": ""
            }

        return fields

    def get_config_ui(self, config):
        updated_config = {}

        # Load config file only if provider is not huggingface
        current_provider = config.get('provider', 'ollama')
        if current_provider != 'huggingface':
            self.config = self.load_llm_config()

        for field, params in self.get_config_fields().items():
            if params['type'] == 'select':
                if field == 'llm_model':
                    provider = config.get('provider', 'huggingface')
                    models = self.get_available_models(provider)
                    try:
                        default_index = models.index(config.get(field, params['default']))
                    except ValueError:
                        default_index = 0
                    updated_config[field] = st.selectbox(
                        params['label'],
                        options=models,
                        index=default_index
                    )
                else:
                    options_list = [option[0] for option in params['options']]
                    try:
                        default_index = options_list.index(config.get(field, params['default']))
                    except ValueError:
                        default_index = 0
                    updated_config[field] = st.selectbox(
                        params['label'],
                        options=options_list,
                        format_func=lambda x: dict(params['options'])[x],
                        index=default_index
                    )
            elif params['type'] == 'textarea':
                updated_config[field] = st.text_area(
                    params['label'],
                    value=config.get(field, params['default'])
                )
            elif params['type'] == 'number':
                updated_config[field] = st.number_input(
                    params['label'],
                    value=int(config.get(field, params['default'])),
                    step=1
                )
            else:
                updated_config[field] = st.text_input(
                    params['label'],
                    value=config.get(field, params['default'])
                )

        if config.get('provider') == 'huggingface':
            updated_config['hf_api_key'] = st.text_input(
                t("rag_hf_api_key"),
                type="password",
                value=config.get('hf_api_key', '')
            )

        return updated_config

    def get_sidebar_config_ui(self, config: Dict[str, Any]) -> Dict[str, Any]:
        provider = config.get('provider', 'ollama')
        available_models = self.get_available_models(provider)
        default_model = config.get('llm_model', available_models[0] if available_models else None)

        if default_model not in available_models:
            default_model = available_models[0] if available_models else None

        selected_model = st.sidebar.selectbox(
            t("rag_llm_model"),
            options=available_models,
            index=available_models.index(default_model) if default_model in available_models else 0,
            key="ragllm_llm_model"
        )
        return {"llm_model": selected_model}

    def get_available_models(self, provider: str) -> List[str]:
        if provider == 'ollama':
            try:
                response = requests.get("http://localhost:11434/api/tags")
                models = response.json()["models"]
                return [f"ollama/{model['name']}" for model in models] + ["ollama/qwen2"]
            except Exception as e:
                st.error(f"{t('rag_error_fetching_models_ollama')}{str(e)}")
                return ["ollama/qwen2"]
        elif provider == 'groq':
            return ["groq/llama3-70b-8192", "groq/mixtral-8x7b-32768"]
        elif provider == 'huggingface':
            return ["HuggingFaceH4/zephyr-7b-beta"]
        else:
            return ["none"]

    def process_rag_text(self, rag_text: str, chunk_size: int, embedder):
        rag_text = rag_text.replace('\\n', ' ').replace('\\\'', "'")
        mots = rag_text.split()
        self.chunks = [' '.join(mots[i:i+chunk_size]) for i in range(0, len(mots), chunk_size)]
        self.embeddings = np.vstack([self.get_embedding(c, embedder) for c in self.chunks])

    def get_embedding(self, text: str, model: str) -> np.ndarray:
        if self.config.get('provider') == 'huggingface':
            if not hasattr(self, 'hf_embeddings'):
                self.hf_embeddings = HuggingFaceEmbeddings(
                    model_name=model,
                    task="feature-extraction",
                    encode_kwargs={'normalize': True}
                )
            embedding = self.hf_embeddings.embed_query(text)
            return np.array(embedding).reshape(1, -1)
        else:
            # Original embedding logic
            tokenizer = AutoTokenizer.from_pretrained(model)
            model = AutoModel.from_pretrained(model, trust_remote_code=True).to(DEVICE)
            inputs = tokenizer(text, padding=True, truncation=True, max_length=MAX_LENGTH, return_tensors="pt").to(DEVICE)
            with torch.no_grad():
                model_output = model(**inputs)
            return mean_pooling(model_output, inputs['attention_mask']).cpu().numpy()

    def calculate_similarity(self, query_embedding: np.ndarray, method: str) -> np.ndarray:
        if method == 'cosine':
            return cosine_similarity(query_embedding.reshape(1, -1), self.embeddings)[0]
        elif method == 'euclidean':
            return -euclidean_distances(query_embedding.reshape(1, -1), self.embeddings)[0]
        elif method == 'manhattan':
            return -manhattan_distances(query_embedding.reshape(1, -1), self.embeddings)[0]
        else:
            raise ValueError("Méthode de similarité non reconnue")

    def get_context(self, query: str, config: Dict[str, Any]) -> tuple:
        query_embedding = self.get_embedding(query, config['ragllm']['embedder'])
        similarities = self.calculate_similarity(query_embedding, config['ragllm']['similarity_method'])
        top_indices = np.argsort(similarities)[-config['ragllm']['top_k']:][::-1]
        context = "\n\n".join([self.chunks[i] for i in top_indices])
        return context, [self.chunks[i] for i in top_indices]

    def call_llm(self, prompt: str, sysprompt: str) -> str:
        try:
            llm_model = st.session_state.ragllm_llm_model
            if self.config.get('provider') == 'huggingface':
                if not self.hf_client:
                    self.hf_client = InferenceClient(token=self.config.get('hf_api_key'))

                messages = [
                    {"role": "system", "content": sysprompt},
                    {"role": "user", "content": prompt}
                ]

                response = self.hf_client.text_generation(
                    model=llm_model,
                    prompt=prompt,
                    max_new_tokens=512,
                    temperature=0.7,
                    stream=False
                )
                return response
            else:
                messages = [
                    {"role": "system", "content": sysprompt},
                    {"role": "user", "content": prompt}
                ]
                response = completion(model=llm_model, messages=messages)
                return response['choices'][0]['message']['content']

        except Exception as e:
            return f"{t('rag_error_calling_llm')}{str(e)}"

    def free_llm(self):
        try:
            llm_model = st.session_state.ragllm_llm_model
            if llm_model.startswith("ollama/"):
                ollama_model = llm_model.split("/")[1]
                response = requests.post(
                    "http://localhost:11434/api/generate",
                    json={
                        "model": ollama_model,
                        "prompt": "bye",
                        "keep_alive": 0
                    }
                )
                return response.json()['response']
        except Exception as e:
            return f"{t('rag_error_calling_llm')}{str(e)}"

    def process_with_llm(self, prompt: str, sysprompt: str, context: str) -> str:
        return self.call_llm(f"Contexte : {context}\n\nQuestion : {prompt}", sysprompt)

    def run(self, config):
        st.write(t("rag_plugin_loaded"))

        # Initialiser rag_text avec la valeur de session_state si elle existe, sinon utiliser une chaîne vide
        if 'rag_text' not in st.session_state:
            st.session_state.rag_text = ""
        if 'rag_question' not in st.session_state:
            st.session_state.rag_question = "Question"

        rag_text = st.text_area(t("rag_enter_text"), height=200, value=st.session_state.rag_text, key="rag_text_key")
        user_prompt = st.text_area(t("rag_enter_question"), value=st.session_state.rag_question, key="rag_prompt_key")
        st.session_state.rag_text = rag_text  # Mettre à jour la valeur dans session_state
        st.session_state.rag_question = user_prompt

        if st.button(t("rag_button_get_answer"), key="get_answer_button"):

            with st.spinner(t("rag_processing")):
                if rag_text:
                    self.process_rag_text(rag_text, config['ragllm']['chunk_size'], config['ragllm']['embedder'])
                    st.success(t("rag_success_text_processed"))
                else:
                    st.warning(t("rag_warning_enter_text"))
                if user_prompt and self.embeddings is not None:
                    context, citations = self.get_context(user_prompt, config)
                    response = self.process_with_llm(user_prompt, config['ragllm']['llm_sys_prompt'], context)

                    st.write(t("rag_answer"))
                    st.write(response)

                    st.write(t("rag_citations"))
                    for i, citation in enumerate(citations, 1):
                        st.write(f"{i}. {citation[:100]}...")
                elif self.embeddings is None:
                    st.warning(t("rag_warning_process_text_first"))
                else:
                    st.warning(t("rag_warning_enter_question"))