app.py

import streamlit as st
import json
import numpy as np
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from sentence_transformers import SentenceTransformer

# Load the Mistral model and tokenizer
@st.cache(allow_output_mutation=True)
def load_model():
    tokenizer = AutoTokenizer.from_pretrained("mistral7b")
    model = AutoModelForSeq2SeqLM.from_pretrained("mistral7b")
    return tokenizer, model

# Load Sentence Transformer for embeddings
@st.cache(allow_output_mutation=True)
def load_sentence_transformer():
    return SentenceTransformer('all-MiniLM-L6-v2')

tokenizer, model = load_model()
sentence_transformer = load_sentence_transformer()

# Load vector store
@st.cache(allow_output_mutation=True)
def load_vectorstore():
    with open('vectorstore.json', 'r') as f:
        vectorstore = json.load(f)
    return vectorstore

vectorstore = load_vectorstore()

# Function to calculate cosine similarity
def cosine_similarity(vec1, vec2):
    return np.dot(vec1, vec2) / (np.linalg.norm(vec1) * np.linalg.norm(vec2))

# Streamlit UI
st.title("Simple RAG App with Mistral 7B")

query = st.text_input("Enter your question:")

if st.button("Get Answer"):
    if query:
        # Embed the query
        query_embedding = sentence_transformer.encode(query)

        # Find the most similar context in the vector store
        best_match = max(vectorstore, key=lambda x: cosine_similarity(query_embedding, x['embedding']))

        # Generate answer using the Mistral model
        inputs = tokenizer.encode(query + " " + best_match['text'], return_tensors='pt')
        outputs = model.generate(inputs, max_length=50, num_return_sequences=1)

        # Decode and display the answer
        answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
        st.write("**Answer:**", answer)
    else:
        st.write("Please enter a question.")