app.py

import json
import gradio as gr
from chunker import extract_text_from_file, chunk_text
from reranker import CrossEncoderRanker
from inference import answer_topk_longformer
import os

DEFAULT_RANKER = os.environ.get("RERANKER_MODEL", "cross-encoder/ms-marco-MiniLM-L-6-v2")

with open("questions_clauses.json", "r", encoding="utf-8") as f:
    clauses_data = json.load(f)

CLAUSES_MAP = {item["question_id_text"]: item["question"] for item in clauses_data}
CLAUSES_LABELS = list(CLAUSES_MAP.keys())

ranker = CrossEncoderRanker(model_name=DEFAULT_RANKER)

# Global state to store chunks between steps
_chunks_cache = []


def chunk_document(file, progress=gr.Progress(track_tqdm=True)):
    global _chunks_cache

    if file is None:
        return "⚠️ Please upload a contract file.", ""

    try:
        progress(0.2, desc="📄 Extracting text...")
        text = extract_text_from_file(file.name)

        if text.startswith("ERROR:"):
            return f"❌ {text}", ""

        progress(0.6, desc="✂️ Chunking document...")
        _chunks_cache = chunk_text(text)
        total_chunks = len(_chunks_cache)

        progress(1.0, desc="✅ Done!")
        # preview = "\n\n---\n\n".join(
        #     [f"**Chunk {i+1}:**\n{c}" for i, c in enumerate(_chunks_cache[:5])]
        # )
        # summary = f"✅ Document chunked into **{total_chunks} chunks**. Showing first 5 below.\n\n{preview}"
        summary = f"✅ Document chunked into **{total_chunks} chunks**. Showing first 5 below.\n\n"
        return f"✅ Ready — {total_chunks} chunks created. Now select a question and click Analyze.", summary

    except Exception as e:
        import traceback
        return f"❌ Error:\n```\n{traceback.format_exc()}\n```", ""


def run_pipeline(question_label, search_mode, top_k_chunks, top_k_answers, progress=gr.Progress(track_tqdm=True)):
    global _chunks_cache

    if not _chunks_cache:
        yield "⚠️ Please upload and chunk a document first.", "", ""
        return

    question = CLAUSES_MAP[question_label]

    try:
        total_chunks = len(_chunks_cache)

        if search_mode == "Top-K chunks (reranked)":
            yield f"🔍 Reranking {total_chunks} chunks...", "", ""
            progress(0.3, desc="🔍 Reranking chunks...")
            ranked = ranker.rank(question, _chunks_cache, top_k=int(top_k_chunks))
            selected_chunks = [chunk for chunk, score in ranked]
        else:
            yield f"📃 Using all {total_chunks} chunks...", "", ""
            progress(0.3, desc="📃 Using all chunks...")
            selected_chunks = _chunks_cache

        chunks_display = f"📄 **Total chunks:** {total_chunks} | **Using:** {len(selected_chunks)}\n\n"
        chunks_display += "\n\n---\n\n".join(
            [f"**Chunk {i+1}:**\n{c}" for i, c in enumerate(selected_chunks)]
        )

        yield "🤖 Model is analyzing the document... (this may take 30–60s on first run)", chunks_display, ""
        progress(0.65, desc="🤖 Running QA model...")
        answers = answer_topk_longformer(question, selected_chunks, top_k=int(top_k_answers))

        progress(0.95, desc="✅ Formatting results...")
        if not answers:
            yield "⚠️ No answers found. Try different chunks or question.", chunks_display, ""
            return

        answers_display = f"### ❓ Question\n> {question}\n\n---\n\n### 📋 Top Answers\n\n"
        for i, (ans, score) in enumerate(answers, 1):
            answers_display += f"**Answer {i}** — Score: `{score:.4f}`\n\n{ans}\n\n{'—'*50}\n\n"

        progress(1.0, desc="Done!")
        yield "✅ Done!", chunks_display, answers_display

    except Exception as e:
        import traceback
        yield f"❌ Error:\n```\n{traceback.format_exc()}\n```", "", ""


with gr.Blocks(title="iContract QA") as demo:
    gr.Markdown("# 📑 iContract — Legal Contract QA")
    gr.Markdown("⚠️ **First analysis may take ~30–60 seconds** as the QA model (~600MB) loads for the first time.")

    # ── Step 1 ──────────────────────────────────────────
    gr.Markdown("## Step 1 — Upload & Chunk Document")
    with gr.Row():
        file_input = gr.File(
            label="Upload Contract (PDF or TXT)",
            file_types=[".pdf", ".PDF", ".txt"]
        )
    chunk_btn = gr.Button("✂️ Chunk Document", variant="secondary")
    chunk_status = gr.Textbox(label="⏳ Status", interactive=False, value="Ready — upload a contract and chunk it.")
    chunk_preview = gr.Markdown()

    # ── Step 2 ──────────────────────────────────────────
    gr.Markdown("## Step 2 — Analyze")
    with gr.Row():
        question_select = gr.Dropdown(
            choices=CLAUSES_LABELS,
            value=CLAUSES_LABELS[0],
            label="Select a clause to analyze"
        )
    with gr.Row():
        search_mode = gr.Radio(
            choices=["Top-K chunks (reranked)", "All chunks"],
            value="Top-K chunks (reranked)",
            label="Search Mode"
        )
    with gr.Row():
        top_k_chunks = gr.Slider(1, 20, value=5, step=1, label="Top-K Chunks (reranker)")
        top_k_answers = gr.Slider(1, 10, value=3, step=1, label="Top-K Answers")

    analyze_btn = gr.Button("🔍 Analyze", variant="primary")
    analyze_status = gr.Textbox(label="⏳ Status", interactive=False, value="")

    with gr.Tabs():
        with gr.Tab("📄 Selected Chunks"):
            chunks_out = gr.Markdown()
        with gr.Tab("✅ Answers"):
            answers_out = gr.Markdown()

    chunk_btn.click(
        fn=chunk_document,
        inputs=[file_input],
        outputs=[chunk_status, chunk_preview],
    )

    analyze_btn.click(
        fn=run_pipeline,
        inputs=[question_select, search_mode, top_k_chunks, top_k_answers],
        outputs=[analyze_status, chunks_out, answers_out],
    )

demo.launch(server_name="0.0.0.0")