src/chatbot.py

from langchain_core.prompts import ChatPromptTemplate
from langchain_community.llms.huggingface_hub import HuggingFaceHub
from langchain_community.embeddings import HuggingFaceEmbeddings


from src.vectordatabase import RAG, get_vectorstore
import pandas as pd
from dotenv import load_dotenv, find_dotenv

#Load environmental variables from .env-file
#load_dotenv(find_dotenv())


embeddings = HuggingFaceEmbeddings(model_name="paraphrase-multilingual-MiniLM-L12-v2")
llm = HuggingFaceHub(
    # Try different model here
    repo_id="mistralai/Mixtral-8x7B-Instruct-v0.1",
    # repo_id="CohereForAI/c4ai-command-r-v01", # too large 69gb
    # repo_id="CohereForAI/c4ai-command-r-v01-4bit", # too large 22 gb
    # repo_id="meta-llama/Meta-Llama-3-8B", # too large 16 gb
    task="text-generation",
    model_kwargs={
        "max_new_tokens": 512,
        "top_k": 30,
        "temperature": 0.1,
        "repetition_penalty": 1.03,
        }
        #,huggingfacehub_api_token

)
# To Do: Experiment with different templates replying in german or english depending on the input language
prompt1 = ChatPromptTemplate.from_template("""<s>[INST] 
                    Instruction: Beantworte die folgende Frage auf deutsch und nur auf der Grundlage des angegebenen Kontexts:

                    Context: {context}

                    Question: {input}  
                    [/INST]"""
                    # Returns the answer in English!?
) 

prompt2 = ChatPromptTemplate.from_template("""Beantworte die folgende Frage auf deutsch und nur auf der Grundlage des angegebenen Kontexts:

        <context>
        {context}
        </context>

        Frage: {input}
        """
        # Returns the answer in German
)

 
#folder_path = 
#index_name = "speeches_1949_09_12"
#index_name = "legislature20"
#db = get



    
     

def chatbot(message, history, db_inputs, llm=llm, prompt=prompt2):
    db = get_vectorstore(inputs = db_inputs, embeddings=embeddings)
    raw_response = RAG(llm=llm, prompt=prompt, db=db, question=message)
    # Only necessary because mistral does not give beautiful outputs 
    response = raw_response['answer'].split("Antwort: ")[1]
    return response  


def keyword_search(db, query, n=10, embeddings=embeddings, method='ss', party_filter='All'):
    """
    Retrieve speech contents based on keywords using a specified method.

    Parameters:
    ----------
    db : FAISS
        The FAISS vector store containing speech embeddings.

    query : str
        The keyword(s) to search for in the speech contents.

    n : int, optional
        The number of speech contents to retrieve (default is 10).

    embeddings : Embeddings, optional
        An instance of embeddings used for embedding queries (default is embeddings).

    method : str, optional
        The method used for retrieving speech contents. Options are 'ss' (semantic search) and 'mmr' 
        (maximal marginal relevance) (default is 'ss').

    party_filter : str, optional
        A filter for retrieving speech contents by party affiliation. Specify 'All' to retrieve 
        speeches from all parties (default is 'All').

    Returns:
    -------
    pandas.DataFrame
        A DataFrame containing the speech contents, dates, and party affiliations.
    
    Notes:
    -----
    - The `db` parameter should be a FAISS vector store containing speech embeddings.
    - The `query` parameter specifies the keyword(s) to search for in the speech contents.
    - The `n` parameter determines the number of speech contents to retrieve (default is 10).
    - The `embeddings` parameter is an instance of embeddings used for embedding queries (default is embeddings).
    - The `method` parameter specifies the method used for retrieving speech contents. Options are 'ss' (semantic search) 
      and 'mmr' (maximal marginal relevance) (default is 'ss').
    - The `party_filter` parameter is a filter for retrieving speech contents by party affiliation. Specify 'All' to retrieve 
      speeches from all parties (default is 'All').
    """

    query_embedding = embeddings.embed_query(query)

    # Maximal Marginal Relevance
    if method == 'mmr':
        df_res = pd.DataFrame(columns=['Speech Content', 'Date', 'Party', 'Relevance'])
        results = db.max_marginal_relevance_search_with_score_by_vector(query_embedding, k=n)
        for doc in results:
            party = doc[0].metadata["party"]
            if party != party_filter and party_filter != 'All':
                continue
            speech_content = doc[0].page_content
            speech_date = doc[0].metadata["date"]
            score = round(doc[1], ndigits=2)
            df_res = pd.concat([df_res, pd.DataFrame({'Speech Content': [speech_content],
                                                      'Date': [speech_date],
                                                      'Party': [party],
                                                      'Relevance': [score]})], ignore_index=True)
        df_res.sort_values('Relevance', inplace=True, ascending=True)

    # Similarity Search
    else:
        df_res = pd.DataFrame(columns=['Speech Content', 'Date', 'Party'])
        results = db.similarity_search_by_vector(query_embedding, k=n)
        for doc in results:
            party = doc.metadata["party"]
            if party != party_filter and party_filter != 'All':
                continue
            speech_content = doc.page_content
            speech_date = doc.metadata["date"]
            df_res = pd.concat([df_res, pd.DataFrame({'Speech Content': [speech_content],
                                                      'Date': [speech_date],
                                                      'Party': [party]})], ignore_index=True)
    return df_res