app.py

#!/usr/bin/env python
# coding: utf-8

import os
import pickle
import argparse
import faiss
import numpy as np
import torch
import gradio as gr

from datasets import load_dataset
from sentence_transformers import SentenceTransformer, CrossEncoder
from transformers import (
    AutoTokenizer,
    AutoModelForSeq2SeqLM,
    pipeline as hf_pipeline,
)
import evaluate

# ── 1. Configuration ──
DATA_DIR       = os.path.join(os.getcwd(), "data")
INDEX_PATH     = os.path.join(DATA_DIR, "faiss_index.faiss")
EMB_PATH       = os.path.join(DATA_DIR, "embeddings.npy")
PCTX_PATH      = os.path.join(DATA_DIR, "passages.pkl")

MODEL_NAME     = os.getenv("MODEL_NAME", "google/flan-t5-small")
EMBEDDER_MODEL = os.getenv("EMBEDDER_MODEL", "sentence-transformers/all-MiniLM-L6-v2")
DIST_THRESHOLD = float(os.getenv("DIST_THRESHOLD", 1.0))
MAX_CTX_WORDS  = int(os.getenv("MAX_CTX_WORDS", 200))

DEVICE = 0 if torch.cuda.is_available() else -1
os.makedirs(DATA_DIR, exist_ok=True)


# ── 2. Helpers ──
def make_context_snippets(contexts, max_words=MAX_CTX_WORDS):
    snippets = []
    for c in contexts:
        words = c.split()
        if len(words) > max_words:
            c = " ".join(words[:max_words]) + " ... [truncated]"
        snippets.append(c)
    return snippets

def chunk_text(text, max_tokens, stride=None):
    words = text.split()
    if stride is None:
        stride = max_tokens // 4
    chunks, start = [], 0
    while start < len(words):
        end = start + max_tokens
        chunks.append(" ".join(words[start:end]))
        start += stride
    return chunks


# ── 3. Load & preprocess passages ──
def load_passages():
    wiki_ds   = load_dataset("rag-datasets/rag-mini-wikipedia", "text-corpus", split="passages")
    squad_ds  = load_dataset("rajpurkar/squad_v2", split="train[:100]")
    trivia_ds = load_dataset("mandarjoshi/trivia_qa", "rc", split="validation[:100]")

    wiki_passages  = wiki_ds["passage"]
    squad_passages = [ex["context"] for ex in squad_ds]
    trivia_passages = []
    for ex in trivia_ds:
        for fld in ("wiki_context", "search_context"):
            txt = ex.get(fld) or ""
            if txt:
                trivia_passages.append(txt)

    all_passages = list(dict.fromkeys(wiki_passages + squad_passages + trivia_passages))
    tokenizer   = AutoTokenizer.from_pretrained(MODEL_NAME)
    max_tokens  = tokenizer.model_max_length

    chunks = []
    for p in all_passages:
        toks = tokenizer.tokenize(p)
        if len(toks) > max_tokens:
            chunks.extend(chunk_text(p, max_tokens))
        else:
            chunks.append(p)

    print(f"[load_passages] total chunks: {len(chunks)}")
    with open(PCTX_PATH, "wb") as f:
        pickle.dump(chunks, f)
    return chunks


# ── 4. Build or load FAISS ──
def load_faiss_index(passages):
    embedder = SentenceTransformer(EMBEDDER_MODEL)
    reranker = CrossEncoder("cross-encoder/ms-marco-MiniLM-L-6-v2")

    if os.path.exists(INDEX_PATH) and os.path.exists(EMB_PATH):
        print("Loading FAISS index & embeddings…")
        index      = faiss.read_index(INDEX_PATH)
        embeddings = np.load(EMB_PATH)
    else:
        print("Encoding passages & building FAISS index…")
        embeddings = embedder.encode(
            passages,
            show_progress_bar=True,
            convert_to_numpy=True,
            batch_size=32
        )
        embeddings = embeddings / np.linalg.norm(embeddings, axis=1, keepdims=True)

        dim   = embeddings.shape[1]
        index = faiss.IndexFlatIP(dim)
        index.add(embeddings)

        faiss.write_index(index, INDEX_PATH)
        np.save(EMB_PATH, embeddings)

    return embedder, reranker, index


# ── 5. Initialize RAG components ──
def setup_rag():
    if os.path.exists(PCTX_PATH):
        with open(PCTX_PATH, "rb") as f:
            passages = pickle.load(f)
    else:
        passages = load_passages()

    embedder, reranker, index = load_faiss_index(passages)

    tok   = AutoTokenizer.from_pretrained(MODEL_NAME)
    model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_NAME)
    qa_pipe = hf_pipeline(
        "text2text-generation",
        model=model,
        tokenizer=tok,
        device=DEVICE,
        truncation=True,
        max_length=512,
        num_beams=4,
        early_stopping=True
    )

    return passages, embedder, reranker, index, qa_pipe


# ── 6. Retrieval & generation ──
def retrieve(question, passages, embedder, reranker, index, k=20, rerank_k=5):
    q_emb      = embedder.encode([question], convert_to_numpy=True)
    distances, idxs = index.search(q_emb, k)

    cands  = [passages[i] for i in idxs[0]]
    scores = reranker.predict([[question, c] for c in cands])
    top    = np.argsort(scores)[-rerank_k:][::-1]

    return [cands[i] for i in top], [distances[0][i] for i in top]

def generate(question, contexts, qa_pipe):
    lines = [
        f"Context {i+1}: {s}"
        for i, s in enumerate(make_context_snippets(contexts))
    ]
    prompt = (
        "You are a helpful assistant. Use ONLY the following contexts to answer. "
        "If the answer is not contained, say 'Sorry, I don't know.'\n\n"
        + "\n".join(lines)
        + f"\n\nQuestion: {question}\nAnswer:"
    )
    return qa_pipe(prompt)[0]["generated_text"].strip()

def retrieve_and_answer(question, passages, embedder, reranker, index, qa_pipe):
    contexts, dists = retrieve(question, passages, embedder, reranker, index)
    if not contexts or dists[0] > DIST_THRESHOLD:
        return "Sorry, I don't know.", []
    return generate(question, contexts, qa_pipe), contexts

def answer_and_contexts(question, passages, embedder, reranker, index, qa_pipe):
    ans, ctxs = retrieve_and_answer(question, passages, embedder, reranker, index, qa_pipe)
    if not ctxs:
        return ans, ""
    snippets = [
        f"Context {i+1}: {s}"
        for i, s in enumerate(make_context_snippets(ctxs))
    ]
    return ans, "\n\n---\n\n".join(snippets)


# ── 7. Evaluation routines ──
def retrieval_recall(dataset, passages, embedder, reranker, index, k=20, rerank_k=None, num_samples=100):
    hits = 0
    for ex in dataset.select(range(num_samples)):
        question     = ex["question"]
        gold_answers = ex["answers"]["text"]

        if rerank_k:
            ctxs, _ = retrieve(question, passages, embedder, reranker, index, k=k, rerank_k=rerank_k)
        else:
            q_emb      = embedder.encode([question], convert_to_numpy=True)
            distances, idxs = index.search(q_emb, k)
            ctxs = [passages[i] for i in idxs[0]]

        if any(any(ans in ctx for ctx in ctxs) for ans in gold_answers):
            hits += 1

    recall = hits / num_samples
    print(f"Retrieval Recall@{k} (rerank_k={rerank_k}): {recall:.3f} ({hits}/{num_samples})")
    return recall

def retrieval_recall_answerable(dataset, passages, embedder, reranker, index, k=20, rerank_k=None, num_samples=100):
    hits, total = 0, 0
    for ex in dataset.select(range(num_samples)):
        gold = ex["answers"]["text"]
        if not gold:
            continue
        total += 1
        question = ex["question"]

        if rerank_k:
            ctxs, _ = retrieve(question, passages, embedder, reranker, index, k=k, rerank_k=rerank_k)
        else:
            q_emb      = embedder.encode([question], convert_to_numpy=True)
            distances, idxs = index.search(q_emb, k)
            ctxs = [passages[i] for i in idxs[0]]

        if any(any(ans in ctx for ctx in ctxs) for ans in gold):
            hits += 1

    recall = hits / total if total > 0 else 0.0
    print(f"Retrieval Recall@{k} on answerable only (rerank_k={rerank_k}): {recall:.3f} ({hits}/{total})")
    return recall

def qa_eval_answerable(dataset, passages, embedder, reranker, index, qa_pipe, k=20, num_samples=100):
    squad_metric = evaluate.load("squad")
    preds, refs  = [], []
    for ex in dataset.select(range(num_samples)):
        gold = ex["answers"]["text"]
        if not gold:
            continue
        qid = ex["id"]
        answer, _ = retrieve_and_answer(ex["question"], passages, embedder, reranker, index, qa_pipe)
        preds.append({"id": qid, "prediction_text": answer})
        refs.append({"id": qid, "answers": ex["answers"]})

    results = squad_metric.compute(predictions=preds, references=refs)
    print(f"Answerable-only QA EM: {results['exact_match']:.2f}, F1: {results['f1']:.2f}")
    return results


# ── 8. Main entry ──
def main():
    passages, embedder, reranker, index, qa_pipe = setup_rag()

    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--eval", action="store_true",
        help="Run retrieval/QA evaluations on SQuAD instead of launching the demo"
    )
    args = parser.parse_args()

    if args.eval:
        squad = load_dataset("rajpurkar/squad_v2", split="validation")
        retrieval_recall(squad, passages, embedder, reranker, index, k=20, rerank_k=5, num_samples=100)
        retrieval_recall_answerable(squad, passages, embedder, reranker, index, k=20, rerank_k=5, num_samples=100)
        qa_eval_answerable(squad, passages, embedder, reranker, index, qa_pipe, k=20, num_samples=100)
    else:
        demo = gr.Interface(
            fn=lambda q: answer_and_contexts(q, passages, embedder, reranker, index, qa_pipe),
            inputs=gr.Textbox(lines=1, placeholder="Ask me anything…", label="Question"),
            outputs=[gr.Textbox(label="Answer"), gr.Textbox(label="Contexts")],
            title="🔍 RAG QA Demo",
            description="Retrieval-Augmented QA with threshold and context preview",
            examples=[
                "When was Abraham Lincoln inaugurated?",
                "What is the capital of France?",
                "Who wrote '1984'?"
            ],
            allow_flagging="never",
        )
        demo.launch()

if __name__ == "__main__":
    main()