app.py

import json
import numpy as np
import gradio as gr
from sentence_transformers import SentenceTransformer
from transformers import T5Tokenizer, T5ForConditionalGeneration

# 1) Load the embedding model for retrieval
embedder = SentenceTransformer('all-MiniLM-L6-v2')

# 2) Load the T5 model and tokenizer for translation
tokenizer = T5Tokenizer.from_pretrained('t5-small')
model = T5ForConditionalGeneration.from_pretrained('t5-small')

# 3) Load JSON knowledge bases
def load_docs(path, fmt):
    with open(path, 'r') as f:
        data = json.load(f)
    return [fmt(item) for item in data]

call_docs = load_docs(
    'call_center_data.json',
    lambda d: f"Ticket {d['ticket_id']} ({d['customer']}): {d['description']}. Resolution: {d['resolution']}"
)
truck_docs = load_docs(
    'trucking_data.json',
    lambda d: f"Shipment {d['shipment_id']} from {d['origin']} to {d['destination']} is {d['status']}. Last update: {d['last_update']}"
)
docs = call_docs + truck_docs

# 4) Pre-compute embeddings for all documents
doc_embeddings = embedder.encode(
    docs,
    convert_to_numpy=True,
    normalize_embeddings=True
)

def rag_translate(query, top_k):
    # 4.1 Retrieval step
    q_emb = embedder.encode(
        [query],
        convert_to_numpy=True,
        normalize_embeddings=True
    )[0]
    sims = doc_embeddings @ q_emb
    idxs = np.argsort(sims)[::-1][:top_k]
    retrieved = [docs[i] for i in idxs]
    retrieved_text = "\n".join(retrieved)

    # 4.2 Translation step using local t5-small
    try:
        # Prepare input for translation
        spanish_input = f"translate English to Spanish: {retrieved_text}"
        french_input = f"translate English to French: {retrieved_text}"

        # Translate to Spanish
        spanish_inputs = tokenizer(spanish_input, return_tensors="pt", padding=True, truncation=True, max_length=512)
        spanish_outputs = model.generate(
            spanish_inputs["input_ids"],
            max_length=512,
            num_beams=4,
            early_stopping=True
        )
        spanish_text = tokenizer.decode(spanish_outputs[0], skip_special_tokens=True)

        # Translate to French
        french_inputs = tokenizer(french_input, return_tensors="pt", padding=True, truncation=True, max_length=512)
        french_outputs = model.generate(
            french_inputs["input_ids"],
            max_length=512,
            num_beams=4,
            early_stopping=True
        )
        french_text = tokenizer.decode(french_outputs[0], skip_special_tokens=True)

        # Combine translations
        translations = f"Spanish:\n{spanish_text}\n\nFrench:\n{french_text}"
    except Exception as e:
        translations = f"Error during translation: {e}"

    return retrieved_text, translations

# 5) Build the Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("## 🚀 Mini-RAG + Translation Demo")
    with gr.Row():
        with gr.Column(scale=3):
            query = gr.Textbox(lines=2, placeholder="Ask me anything…", label="Your question")
            top_k = gr.Slider(minimum=1, maximum=len(docs), value=1, step=1, label="Number of docs to retrieve")
            submit = gr.Button("🔍 Retrieve & Translate")
            retrieved_output = gr.Textbox(label="Retrieved Content (English)")
            translated_output = gr.Textbox(label="Translated Content (Spanish & French)")
            submit.click(
                fn=rag_translate,
                inputs=[query, top_k],
                outputs=[retrieved_output, translated_output]
            )
        with gr.Column(scale=1):
            gr.Markdown("### Example Queries")
            examples = [
                "What is the status of shipment S-2003?",
                "Where did shipment S-2002 originate?",
                "How many tickets mention payment errors?",
                "What resolution was given for T-1001?"
            ]
            for ex in examples:
                btn = gr.Button(ex)
                btn.click(lambda x=ex: x, None, query)
    gr.Markdown(
        """
        **How it works**  
        1) Retrieval: fetch top-k snippets from JSON KBs.  
        2) Translation: translate retrieved content to Spanish and French using t5-small locally.  
        """
    )

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0")