app.py

import os
import json
import base64
import uuid
import logging
from typing import Generator, List, Dict, Tuple, Optional

from sentence_transformers import SentenceTransformer
from openai import OpenAI
import gradio as gr
from dotenv import load_dotenv

from utils.utils import (
    get_keys_chunks,
    get_docs,
    get_top_chunk_keys,
    get_messages,
    load_knowledge_base,
)
from utils.chatLogger import ChatUploader


# --------------- Logging ---------------
def _setup_logging() -> logging.Logger:
    logging.basicConfig(
        level=logging.INFO,
        format="%(levelname)s:%(name)s:%(message)s",
    )
    return logging.getLogger(__name__)


# --------------- Initialization ---------------
def _require_env(var: str) -> str:
    val = os.getenv(var)
    if not val:
        raise RuntimeError(f"Missing required environment variable: {var}")
    return val


def initialize() -> Tuple[
    SentenceTransformer,
    List[Tuple[str, "numpy.ndarray"]],
    Dict,
    OpenAI,
    logging.Logger,
    Optional[ChatUploader],
]:
    logger = _setup_logging()
    logger.info("Initializing application...")

    load_dotenv(override=False)
    logger.info(".env loaded (override=False)")

    embedding_model_path = "ibm-granite/granite-embedding-125m-english"
    embedding_model = SentenceTransformer(embedding_model_path)
    logger.info("Embedding model loaded: %s", embedding_model_path)

    knowledge_base = load_knowledge_base()
    logger.info("Knowledge base loaded")

    pairs = list(get_keys_chunks(knowledge_base))
    if not pairs:
        raise RuntimeError("Knowledge base is empty – no chunks to encode.")
    keys, chunks = zip(*pairs)
    logger.info("KB chunks extracted: %d", len(chunks))

    chunks_encoded = embedding_model.encode(
        list(chunks),
        batch_size=64,
        convert_to_numpy=True,
        show_progress_bar=False,
    )
    keys_chunks_encoded = list(zip(keys, chunks_encoded))
    logger.info("KB chunks encoded: %d", len(keys_chunks_encoded))

    inference_api_key = _require_env("INFERENCE_API_KEY")
    openai_client = OpenAI(
        base_url="https://api.inference.net/v1",
        api_key=inference_api_key,
    )
    logger.info("OpenAI client initialized (base_url=api.inference.net)")

    chat_uploader: Optional[ChatUploader] = None
    drive_creds_b64 = os.getenv("GOOGLE_DRIVE_SERVICE_ACCOUNT_CREDENTIALS_BASE64")
    if drive_creds_b64:
        try:
            service_account_json = json.loads(
                base64.b64decode(drive_creds_b64).decode()
            )
            chat_uploader = ChatUploader(service_account_json)
            logger.info("Google Drive uploader configured")
        except Exception as e:
            logger.warning(
                "Google Drive uploader not configured (error parsing creds): %s", e
            )
            chat_uploader = None
    else:
        logger.info("Google Drive uploader not configured (no creds env var)")

    logger.info("Initialization complete")
    return (
        embedding_model,
        keys_chunks_encoded,
        knowledge_base,
        openai_client,
        logger,
        chat_uploader,
    )


(
    embedding_model,
    keys_chunksEncoded,
    knowledge_base,
    openAI_client,
    logger,
    chat_uploader,
) = initialize()


# --------------- Helpers ---------------
def _strip_think_tags(text: str) -> str:
    return text.replace("<think>", "").replace("</think>", "")


def _to_minimal(history: List[Dict[str, str]]) -> List[Dict[str, str]]:
    """Keep only role/content keys to avoid metadata/options noise in uploads."""
    minimal: List[Dict[str, str]] = []
    for m in history:
        role = m.get("role")
        content = m.get("content", "")
        if role is None:
            # ignore malformed entries
            continue
        minimal.append({"role": role, "content": content})
    return minimal


# --------------- RAG Chatbot ---------------
def rag_chatbot(
    user_message: str,
    chat_history: List[Dict[str, str]],
    browser_id: str,
) -> Generator[List[Dict[str, str]], None, None]:
    """
    Stream assistant output as a single growing message dict.
    Do NOT mutate chat_history; Gradio manages it for type="messages".
    """
    # RAG retrieval
    try:
        logger.info("RAG: encoding query & retrieving docs")
        user_query_encoded = embedding_model.encode(
            [user_message], convert_to_numpy=True
        )[0]
        top_chunk_keys = get_top_chunk_keys(
            user_query_encoded, keys_chunksEncoded, top_n=5
        )
        docs = get_docs(top_chunk_keys, knowledge_base)
        logger.info("RAG: docs retrieved=%d (top_n=5)", len(docs))
    except Exception as e:
        logger.error("RAG: retrieval failed: %s", e)
        yield [
            {
                "role": "assistant",
                "content": "⚠️ An error occurred during document retrieval. Please try again later.",
            }
        ]
        return

    # LLM stream
    try:
        logger.info(
            "LLM: opening streaming completion (model=mistralai/mistral-nemo-12b-instruct/fp-8)"
        )
        messages = get_messages(docs, user_message, chat_history)
        chat_stream = openAI_client.chat.completions.create(
            model="mistralai/mistral-nemo-12b-instruct/fp-8",
            messages=messages,
            stream=True,
        )
        logger.info("LLM: stream opened")
    except Exception as e:
        logger.error("LLM: API call failed: %s", e)
        yield [
            {
                "role": "assistant",
                "content": "⚠️ An error occurred during client API call. Please try again later.",
            }
        ]
        return

    # Stream parse → yield a single growing assistant message
    assistant_msg = {"role": "assistant", "content": ""}
    try:
        logger.info("LLM: streaming started")

        buffer = ""
        chunks_seen = 0
        content_events = 0
        chars_emitted = 0

        for chunk in chat_stream:
            chunks_seen += 1
            choices = getattr(chunk, "choices", None)
            if not choices:
                continue
            delta = getattr(choices[0], "delta", None)
            if not delta:
                continue
            piece = getattr(delta, "content", None)
            if piece is None:
                continue

            piece = _strip_think_tags(piece)
            if not piece:
                continue

            content_events += 1
            buffer += piece

            if len(buffer) >= 24 or "\n" in buffer:
                assistant_msg["content"] += buffer
                chars_emitted += len(buffer)
                yield [assistant_msg]  # append/update single assistant bubble
                buffer = ""

        if buffer:
            assistant_msg["content"] += buffer
            chars_emitted += len(buffer)
            yield [assistant_msg]

        logger.info(
            "LLM: streaming finished (chunks=%d, content_events=%d, chars=%d)",
            chunks_seen,
            content_events,
            chars_emitted,
        )

    except Exception as e:
        logger.error("LLM: streaming failed: %s", e)
        if assistant_msg["content"]:
            assistant_msg[
                "content"
            ] += "\n\n⚠️ An error occurred during LLM response streaming. Please try again later."
            yield [assistant_msg]
        else:
            yield [
                {
                    "role": "assistant",
                    "content": "⚠️ An error occurred during LLM response streaming. Please try again later.",
                }
            ]
        return

    # --- Upload transcript (optional) — reconstruct current turn explicitly
    try:
        if chat_uploader is not None:
            # Gradio passes prior turns in `chat_history`. Build latest full transcript.
            prior = _to_minimal(chat_history)
            current_user = {"role": "user", "content": user_message}
            final_history = prior + [
                current_user,
                {"role": "assistant", "content": assistant_msg["content"]},
            ]

            # Ensure we have a usable browser_id for the filename
            if not browser_id:
                browser_id = str(uuid.uuid4())
            drive_filename = f"chat__{browser_id}.json"

            logger.info(
                "Upload: writing Drive file '%s' (messages=%d, mode=overwrite)",
                drive_filename,
                len(final_history),
            )
            chat_uploader.upload_chat_history(
                final_history,
                browser_id,
                filename=drive_filename,
                mode="overwrite",  # <-- overwrite-by-name semantics
            )
            logger.info("Upload: completed")
        else:
            logger.info("Upload: skipped (uploader not configured)")
    except Exception as e:
        logger.warning("Upload: failed (non-fatal): %s", e)


# --------------- Gradio app ---------------
with gr.Blocks() as demo:
    browser_id_state = gr.BrowserState(default_value=None)

    @demo.load(inputs=browser_id_state, outputs=browser_id_state)
    def load_browser_id(current_id):
        if current_id is None or current_id == "":
            new_id = str(uuid.uuid4())
            logger.info("Browser ID created: %s", new_id)
            return new_id
        logger.info("Browser ID reused: %s", current_id)
        return current_id

    gr.ChatInterface(
        fn=rag_chatbot,
        title="Matthew Schulz's RAG Chatbot 💬🤖",
        additional_inputs=browser_id_state,
        type="messages",
        examples=[
            ["What is Matthew's educational background?", None],
            [
                "What is Matthew's current role at Visa and what problems is he solving?",
                None,
            ],
            [
                "What machine learning projects has Matthew worked on and what were the outcomes?",
                None,
            ],
            ["What research did Matthew conduct at the USC AutoDrive Lab?", None],
            [
                "What did Matthew accomplish as Lead Software Engineer at SchedGo (now EduRoute)?",
                None,
            ],
            [
                "What were the key results of Matthew’s internship at NASA’s Deep Space Network (Peraton)?",
                None,
            ],
            [
                "Give me a general background on Matthew's education and work experience.",
                None,
            ],
            [
                "Which tools, technologies, and coding practices does Matthew prefer and why?",
                None,
            ],
            [
                "What are Matthew’s strengths and weaknesses, and how is he addressing growth areas?",
                None,
            ],
            [
                "What are Matthew’s hobbies and personal interests?",
                None,
            ],
            ["Why did Matthew choose to pursue a degree in computer science?", None],
            ["Does Matthew have any leadership experience?", None],
        ],
        save_history=True,
        run_examples_on_click=False,
        cache_examples=False,
    )

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