app.py

import gradio as gr
from huggingface_hub import InferenceClient
import evaluate
import psycopg2
from psycopg2.extras import RealDictCursor
from dotenv import load_dotenv
import os

# Load environment variables
load_dotenv()
# Load env vars
# token_hf = os.getenv('ACCESS_TOKEN')
conn_str = os.getenv('NEON_CONN_STR')

# Evaluation metrics
rouge = evaluate.load('rouge')
meteor = evaluate.load('meteor')

def respond(
    message,
    history: list[dict[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
    hf_token: gr.OAuthToken,
):
    """
    For more information on `huggingface_hub` Inference API support, please check the docs: https://huggingface.co/docs/huggingface_hub/v0.22.2/en/guides/inference
    """
    client = InferenceClient(token=hf_token.token)

    retrieved_info = get_relevant_answers(message, hf_token.token)

    messages = [{"role": "system", "content": system_message}]

    messages.extend(history)
    
    if retrieved_info:
        messages.append({
            "role": "user",
            "content":f"Answer the following question using the provided context DO NOT mention that you're referring to any context:\nContext: {retrieved_info}\nQuestion: {message}"
        })
    else:
        messages.append({
            "role": "user",
            "content": message,
        })

    response = ""

    for message in client.chat_completion(
        messages,
        model='openai/gpt-oss-20b',
        max_tokens=max_tokens,
        stream=True,
        temperature=temperature,
        top_p=top_p,
    ):
        choices = message.choices
        token = ""
        if len(choices) and choices[0].delta.content:
            token = choices[0].delta.content

        response += token
        yield response
    if retrieved_info:
        print_score(response, retrieved_info)


# Print the highest score across references
def print_score(prediction, references):
    print(f"{prediction}\n{references}\n\n=== Generation Scores ===")
    result = rouge.compute(predictions=[prediction], references=[references])
    print("ROUGELSum:", round(result["rougeLsum"], 2))
    result = meteor.compute(predictions=[prediction], references=[references])
    print("METEOR:", round(result["meteor"], 2))
    print(f"{'_'*50}\n")


# Multilangual retrieval
def get_relevant_answers(prompt, token):
    embedding_clinet = InferenceClient(
        provider="hf-inference",
        api_key=token,
    )

    with psycopg2.connect(conn_str) as conn:
        with conn.cursor(cursor_factory=RealDictCursor) as cur:
            query_embedding = embedding_clinet.feature_extraction(prompt, model="google/embeddinggemma-300m").tolist()
            cur.execute("""
                        SELECT a, q, q_embeddings_multilang <=> %s::vector AS distance
                        FROM qa
                        WHERE q_embeddings_multilang <=> %s::vector < 0.4
                        ORDER BY distance
                        LIMIT 3;
                        """,
                        (query_embedding, query_embedding))
            rows = cur.fetchall()
            for row in rows:
                print(f"\n{row['a']}\n{row['q']}\n{row['distance']}\n---")
            relevant_answers = [x['a'] for x in rows]
            print(f"Relevant Answers: {relevant_answers}", end='\n---\n')
    return relevant_answers

"""
For information on how to customize the ChatInterface, peruse the gradio docs: https://www.gradio.app/docs/chatinterface
"""
chatbot = gr.ChatInterface(
    respond,
    type="messages",
    additional_inputs=[
        gr.Textbox(
            value="""You are an assistant for question-answering about a planet named Zephyra.
- You have access to external resources which are added along with the user question.
- Use the resources provided along with the user question to answer the question.
- Think hard about whether you can answer the given question using the context that you have or not.
- If the context provided is irrelevant to the user question, answer as if no context is provided.
- If you don't know the answer, just say that you don't know.
- Answer with the same language as the question.  
- Keep the answer concise.""",
            label="System message"
        ),
        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
        gr.Slider(minimum=0, 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)",
        ),
    ],
)

with gr.Blocks() as demo:
    with gr.Sidebar():
        gr.LoginButton()
    chatbot.render()


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