app.py

# app.py
import os
import glob
import faiss
import numpy as np

import gradio as gr
import spaces

from unstructured.partition.pdf import partition_pdf
from sentence_transformers import SentenceTransformer
from transformers import RagTokenizer, RagSequenceForGeneration

# ─── Configuration ─────────────────────────────────────────────
PDF_FOLDER = "meal_plans"
MODEL_NAME = "facebook/rag-sequence-nq"
EMBED_MODEL = "all-MiniLM-L6-v2"
TOP_K = 5

# ─── 1) LOAD + CHUNK ALL PDFs ──────────────────────────────────
rag_tokenizer = RagTokenizer.from_pretrained(MODEL_NAME)
texts, sources, pages = [], [], []

for pdf_path in glob.glob(f"{PDF_FOLDER}/*.pdf"):
    book = os.path.basename(pdf_path)
    pages_data = partition_pdf(filename=pdf_path)
    for pg_num, page in enumerate(pages_data, start=1):
        enc = rag_tokenizer(
            page.text,
            max_length=800,
            truncation=True,
            return_overflowing_tokens=True,
            stride=50,
            return_tensors="pt"
        )
        for token_ids in enc["input_ids"]:
            chunk = rag_tokenizer.decode(token_ids, skip_special_tokens=True)
            texts.append(chunk)
            sources.append(book)
            pages.append(pg_num)

# ─── 2) EMBED + BUILD FAISS INDEX ─────────────────────────────
embedder = SentenceTransformer(EMBED_MODEL)
embeddings = embedder.encode(texts, convert_to_numpy=True)
dim = embeddings.shape[1]
index = faiss.IndexFlatL2(dim)
index.add(embeddings)

# ─── 3) LOAD RAG GENERATOR ────────────────────────────────────
tokenizer = RagTokenizer.from_pretrained(MODEL_NAME)
generator = RagSequenceForGeneration.from_pretrained(MODEL_NAME)

@spaces.GPU
def respond(
    message: str,
    history: list[tuple[str,str]],
    goal: str,
    diet: list[str],
    meals: int,
    avoid: str,
    weeks: str
):
    # build prefs string
    avoid_list = [a.strip() for a in avoid.split(",") if a.strip()]
    prefs = (
        f"Goal={goal}; Diet={','.join(diet)}; "
        f"Meals={meals}/day; Avoid={','.join(avoid_list)}; Duration={weeks}"
    )
    # 1) RETRIEVE top-k chunks
    q_emb = embedder.encode([message], convert_to_numpy=True)
    D, I  = index.search(q_emb, TOP_K)
    context = "\n".join(f"[{sources[i]} p{pages[i]}] {texts[i]}" for i in I[0])

    # 2) BUILD PROMPT with guardrail
    prompt = (
        "SYSTEM: Only answer using the provided CONTEXT. "
        "If it’s not there, say \"I'm sorry, I don't know.\" \n"
        f"PREFS: {prefs}\n"
        f"CONTEXT:\n{context}\n"
        f"Q: {message}\n"
    )

    # 3) GENERATE
    inputs  = tokenizer([prompt], return_tensors="pt")
    outputs = generator.generate(**inputs, num_beams=2, max_new_tokens=200)
    answer  = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]

    # update chat history
    history = history or []
    history.append((message, answer))
    return history

# ─── 4) BUILD UI ────────────────────────────────────────────────
goal  = gr.Dropdown(["Lose weight","Bulk","Maintain"], label="Goal", value="Lose weight")
diet  = gr.CheckboxGroup(["Omnivore","Vegetarian","Vegan","Keto","Paleo","Low-Carb"], label="Diet Style")
meals = gr.Slider(1, 6, step=1, value=3, label="Meals per day")
avoid = gr.Textbox(placeholder="e.g. Gluten, Dairy, Nuts…", label="Avoidances (comma-separated)")
weeks = gr.Dropdown(["1 week","2 weeks","3 weeks","4 weeks"], label="Plan Length", value="1 week")

demo = gr.ChatInterface(
    fn=respond,
    additional_inputs=[goal, diet, meals, avoid, weeks]
)

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