app.py

# essay_assessor_gradio.py
"""
Essay Assessment Integrating Large Language Model (Gradio app)

Requirements (install with pip):
    pip install gradio pymupdf groq python-docx pandas

Environment:
    export GROQ_API_KEY="your_groq_api_key_here"

Run:
    python essay_assessor_gradio.py
"""

import os
import json
import csv
import io
from typing import List, Dict, Any, Tuple
from datetime import datetime
# extra imports (add these near the top of your file)
from docx import Document as DocxDocument    # python-docx
from pptx import Presentation as PptxPresentation  # python-pptx
from bs4 import BeautifulSoup
import re
import string


import fitz  # PyMuPDF for PDF text extraction
import gradio as gr
from groq import Groq
from docx import Document
import pandas as pd

# --- Config ---
GROQ_MODEL = "meta-llama/llama-4-maverick-17b-128e-instruct"
client = Groq(api_key = "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")  # reads GROQ_API_KEY from env var

# --------- LLM helpers ----------
def call_groq(system_prompt: str, user_prompt: str, temperature: float = 0.7, max_tokens: int = 2048) -> str:
    """
    Robust Groq caller: streams tokens and accumulates them into a final string.
    """
    completion = client.chat.completions.create(
        model=GROQ_MODEL,
        messages=[
            {"role": "system", "content": system_prompt},
            {"role": "user", "content": user_prompt},
        ],
        temperature=temperature,
        max_completion_tokens=max_tokens,
        top_p=1,
        stream=True,
        stop=None
    )
    chunks = []
    for chunk in completion:
        piece = chunk.choices[0].delta.content or ""
        chunks.append(piece)
    return "".join(chunks)

# ---------- PDF / file utils ----------
def extract_text_from_pdf_bytes(pdf_bytes: bytes, max_pages: int = 200) -> str:
    text_parts = []
    try:
        with fitz.open(stream=pdf_bytes, filetype="pdf") as doc:
            page_count = min(len(doc), max_pages)
            for i in range(page_count):
                page = doc.load_page(i)
                text_parts.append(page.get_text("text"))
    except Exception as e:
        return ""
    lines = [ln.strip() for ln in ("\n".join(text_parts)).splitlines() if ln.strip()]
    return "\n".join(lines)

def read_text_file(filepath: str) -> str:
    try:
        with open(filepath, "r", encoding="utf-8") as f:
            return f.read()
    except Exception:
        try:
            with open(filepath, "rb") as f:
                raw = f.read()
            return raw.decode("utf-8", errors="ignore")
        except Exception:
            return ""

def sanitize_text_for_preview(text: str, max_len: int = 4000) -> str:
    """
    Keep printable characters + common whitespace (newline/tab). Truncate to max_len.
    Removes null bytes and weird binary residues.
    """
    if not text:
        return ""
    # remove NULs
    text = text.replace("\x00", " ")
    # keep printable or newline/tab/carriage return
    cleaned_chars = []
    for ch in text:
        if ch.isprintable() or ch in ("\n", "\r", "\t"):
            cleaned_chars.append(ch)
        else:
            # replace control / non-printable with space to avoid merging words
            cleaned_chars.append(" ")
    cleaned = "".join(cleaned_chars)
    # collapse excessive whitespace
    cleaned = re.sub(r"[ \t]{2,}", " ", cleaned)
    # protect long strings
    return cleaned[:max_len]


def read_text_file(path: str) -> str:
    """Try reading text files robustly (txt, md)."""
    try:
        with open(path, "r", encoding="utf-8") as f:
            return f.read()
    except Exception:
        try:
            with open(path, "rb") as f:
                raw = f.read()
            return raw.decode("utf-8", errors="ignore")
        except Exception:
            return ""


def extract_text_from_docx(path: str) -> str:
    """Extract paragraph and table text from a .docx file."""
    try:
        doc = DocxDocument(path)
        parts = []
        for p in doc.paragraphs:
            if p.text and p.text.strip():
                parts.append(p.text.strip())
        # also pull text from tables
        for table in doc.tables:
            for row in table.rows:
                for cell in row.cells:
                    if cell.text and cell.text.strip():
                        parts.append(cell.text.strip())
        return "\n".join(parts)
    except Exception:
        return ""


def extract_text_from_pptx(path: str) -> str:
    """Extract text from slides in a .pptx file."""
    try:
        prs = PptxPresentation(path)
        parts = []
        for slide in prs.slides:
            for shape in slide.shapes:
                # Many shapes (pictures) don't have 'text' attr
                try:
                    text = shape.text
                except Exception:
                    text = None
                if text:
                    t = text.strip()
                    if t:
                        parts.append(t)
        return "\n".join(parts)
    except Exception:
        return ""


def extract_text_from_html(path: str) -> str:
    try:
        with open(path, "r", encoding="utf-8", errors="ignore") as f:
            html = f.read()
        soup = BeautifulSoup(html, "html.parser")
        return soup.get_text(separator="\n")
    except Exception:
        return ""


def extract_text_from_path(path: str) -> str:
    """
    Robust extractor that handles PDF, DOCX, PPTX, HTML, TXT/MD and falls back safely.
    Returns plain text (not binary).
    """
    if not path:
        return ""
    path = str(path)
    _, ext = os.path.splitext(path.lower())

    try:
        if ext == ".pdf":
            with open(path, "rb") as f:
                b = f.read()
            return extract_text_from_pdf_bytes(b)  # keep your existing function
        elif ext == ".docx":
            return extract_text_from_docx(path)
        elif ext == ".pptx":
            return extract_text_from_pptx(path)
        elif ext in (".html", ".htm"):
            return extract_text_from_html(path)
        elif ext in (".txt", ".md", ".csv"):
            return read_text_file(path)
        else:
            # Try docx/pptx first (some uploads may have wrong extension)
            txt = extract_text_from_docx(path)
            if txt:
                return txt
            txt = extract_text_from_pptx(path)
            if txt:
                return txt
            # fallback: try reading as text
            return read_text_file(path)
    except Exception:
        return ""


# ---------- Question generation ----------
SYSTEM_PROMPT_GEN = """
You are an experienced educational content creator. Given course material and a marking guide,
produce a structured list of essay and short-answer questions (no MCQs). Output ONLY valid JSON.
The output MUST be a JSON array of question objects. Each question object must include:
- id: integer
- type: "essay" or "short_answer"
- question: string (the prompt shown to students)
- rubric: string (brief marking guide / key points to look for)
- max_score: integer (e.g., 10 for essay; 3-5 for short-answer)
You MUST ensure questions are faithful to the provided course material and that rubrics reference key points.
Do NOT include any additional commentary or text outside the JSON array.
"""

def build_gen_prompt(course_text: str, marking_guide: str, n_essay: int, n_short: int) -> str:
    return f"""
Course material:
\"\"\"{course_text[:20000]}\"\"\"

Marking guide (instructor-provided). Use this to shape rubrics:
\"\"\"{marking_guide[:5000]}\"\"\"

Generate {n_essay} essay question(s) and {n_short} short-answer question(s).
Return a JSON array of question objects (see schema in system message).
"""

def parse_json_from_model_output(raw: str) -> Any:
    s = raw.strip()
    # strip fences if present
    if s.startswith("```"):
        s = s.strip("`")
        if s.startswith("json"):
            s = s[len("json"):].strip()
    start = s.find("[")
    end = s.rfind("]")
    if start != -1 and end != -1 and end > start:
        s = s[start:end+1]
    try:
        return json.loads(s)
    except Exception:
        # fallback: try to find first brace/array and parse
        try:
            return json.loads(raw)
        except Exception:
            return None

def generate_questions(course_text: str, marking_guide: str, n_essay: int, n_short: int, temperature: float = 0.6) -> Tuple[List[Dict[str, Any]], str]:
    prompt = build_gen_prompt(course_text, marking_guide, n_essay, n_short)
    raw = call_groq(SYSTEM_PROMPT_GEN, prompt, temperature=temperature, max_tokens=2048)
    parsed = parse_json_from_model_output(raw)
    questions = []
    if isinstance(parsed, list):
        # enforce id and fields
        qid = 1
        for item in parsed:
            if not isinstance(item, dict):
                continue
            qtype = item.get("type", "").lower()
            if qtype not in ("essay", "short_answer", "short-answer", "short"):
                continue
            qtype = "essay" if qtype == "essay" else "short_answer"
            question_text = item.get("question", "") or item.get("prompt", "")
            rubric = item.get("rubric", "") or item.get("marking_points", "")
            max_score = int(item.get("max_score", item.get("max", 10)))
            questions.append({
                "id": qid,
                "type": qtype,
                "question": question_text,
                "rubric": rubric,
                "max_score": max_score
            })
            qid += 1
    # if parsed is None or empty, return empty questions and raw for debug
    return questions, raw

# ---------- Grading ----------
SYSTEM_PROMPT_GRADE = """
You are a strict but fair grader. You will receive:
- Question text
- The rubric (key points, marking guide)
- Maximum score
- A student's answer

Return ONLY a JSON object with keys:
- score: integer (0 .. max_score)
- feedback: string (concise feedback and suggestions to improve; include inline markup suggestions if useful)
- annotated_answer: optional string (student's answer with short inline edits or markup for improvement)

Be specific and consistent with the rubric. Do not add any explanation outside the JSON.
"""

def build_grade_prompt(question: str, rubric: str, max_score: int, student_answer: str) -> str:
    return f"""
Question:
{question}

Rubric / Marking guide:
{rubric}

Max score: {max_score}

Student answer:
\"\"\"{student_answer}\"\"\"

Return the grade JSON as specified.
"""

def parse_json_object(raw: str) -> Dict[str, Any]:
    s = raw.strip()
    if s.startswith("```"):
        s = s.strip("`")
        if s.startswith("json"):
            s = s[len("json"):].strip()
    start = s.find("{")
    end = s.rfind("}")
    if start != -1 and end != -1 and end > start:
        s = s[start:end+1]
    try:
        return json.loads(s)
    except Exception:
        # fallback: return basic failure dict
        return {"score": 0, "feedback": "Failed to parse grader output."}

def grade_answer_with_llm(question: str, rubric: str, max_score: int, student_answer: str, temperature: float=0.2) -> Dict[str, Any]:
    prompt = build_grade_prompt(question, rubric, max_score, student_answer)
    raw = call_groq(SYSTEM_PROMPT_GRADE, prompt, temperature=temperature, max_tokens=1024)
    parsed = parse_json_object(raw)
    # clamp
    try:
        parsed["score"] = int(parsed.get("score", 0))
    except Exception:
        parsed["score"] = 0
    if parsed["score"] < 0: parsed["score"] = 0
    if parsed["score"] > max_score: parsed["score"] = max_score
    if "feedback" not in parsed:
        parsed["feedback"] = ""
    return parsed

# ---------- Export helpers ----------
def export_results_csv(questions: List[Dict[str,Any]], student_answers: List[Dict[str,Any]], results: List[Dict[str,Any]]) -> str:
    # Align by question id
    qmap = {q["id"]: q for q in questions}
    rows = []
    for res in results:
        qid = res.get("id")
        q = qmap.get(qid, {})
        rows.append({
            "id": qid,
            "type": q.get("type",""),
            "question": q.get("question",""),
            "student_answer": res.get("student_answer",""),
            "score": res.get("score", 0),
            "max_score": res.get("max", q.get("max_score", None)),
            "feedback": res.get("feedback","")
        })
    filename = f"essay_assessment_{datetime.now().strftime('%Y%m%d_%H%M%S')}.csv"
    with open(filename, "w", encoding="utf-8", newline="") as f:
        fieldnames = ["id","type","question","student_answer","score","max_score","feedback"]
        writer = csv.DictWriter(f, fieldnames=fieldnames)
        writer.writeheader()
        for r in rows:
            writer.writerow(r)
    return filename

def export_results_docx(questions: List[Dict[str,Any]], results: List[Dict[str,Any]]) -> str:
    doc = Document()
    doc.add_heading("Essay Assessment Results", level=1)
    doc.add_paragraph(f"Generated: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}")
    total_score = sum([r.get("score",0) for r in results])
    max_total = sum([r.get("max",0) for r in results])
    doc.add_paragraph(f"Total score: {total_score} / {max_total}")

    for r in results:
        qid = r.get("id")
        qtext = r.get("question","")
        student_answer = r.get("student_answer","")
        score = r.get("score",0)
        max_score = r.get("max",0)
        feedback = r.get("feedback","")
        annotated = r.get("annotated_answer","")

        doc.add_heading(f"Q{qid} ({score}/{max_score})", level=2)
        doc.add_paragraph("Question:")
        doc.add_paragraph(qtext)
        doc.add_paragraph("Student answer:")
        doc.add_paragraph(student_answer)
        if annotated:
            doc.add_paragraph("Annotated answer (suggested edits):")
            doc.add_paragraph(annotated)
        doc.add_paragraph("Feedback:")
        doc.add_paragraph(feedback)
        doc.add_paragraph("-----")

    filename = f"essay_assessment_{datetime.now().strftime('%Y%m%d_%H%M%S')}.docx"
    doc.save(filename)
    return filename

# ---------- Gradio app ----------
with gr.Blocks(title="Essay Assessment with LLM") as demo:
    gr.Markdown("# Essay Assessment Integrating Large Language Model")
    gr.Markdown(
        "Lecturers upload course material + marking guide to auto-create essay & short-answer questions. "
        "Students answer questions one-by-one. The AI grades answers against the marking guide and returns scores, feedback and annotated suggestions."
    )

    # ----- States -----
    state_course_text = gr.State("")
    state_marking_text = gr.State("")
    state_questions = gr.State([])          # list of question dicts
    state_student_answers = gr.State([])    # list of {"id", "student_answer", "timestamp"}
    state_current_idx = gr.State(0)
    state_results = gr.State([])            # list of graded results per question

    with gr.Tabs():

        # ---------------- Lecturer tab ----------------
        with gr.Tab("1) Lecturer — Upload & Generate"):
            with gr.Row():
                course_file = gr.File(label="Course Material (PDF or TXT)", type="filepath")
                marking_file = gr.File(label="Marking Guide (PDF or TXT, optional)", type="filepath")
            manual_marking = gr.Textbox(label="Or paste Marking Guide (optional)", lines=6, placeholder="Paste marking guide or rubric here (optional)")
            with gr.Row():
                n_essay = gr.Number(label="Number of essay questions", value=2, precision=0)
                n_short = gr.Number(label="Number of short-answer questions", value=2, precision=0)
                temp_slider = gr.Slider(0.0, 1.0, value=0.5, step=0.1, label="LLM creativity (temperature)")
            generate_btn = gr.Button("Generate Questions (LLM)")
            course_preview = gr.Textbox(label="Course preview (first 4000 chars)", lines=10)
            marking_preview = gr.Textbox(label="Marking guide preview (first 2000 chars)", lines=6)
            questions_json = gr.JSON(label="Generated Questions (JSON)")
            gen_debug = gr.Textbox(label="Raw model output (debug)", visible=False)

            def do_generate(course_path, marking_path, manual_marking_text, ne, ns, temp):
            # Extract texts
              course_text = ""
              marking_text = ""
              if course_path:
                course_text = extract_text_from_path(course_path) or ""
              if marking_path:
                marking_text = extract_text_from_path(marking_path) or ""
              if (not marking_text) and manual_marking_text:
                marking_text = manual_marking_text or ""

            # Basic validation: don't return empty string for JSON outputs
              if not course_text.strip():
                # Return values MUST match the outputs list exactly (see below)
                return (
                  gr.update(value=""),                      # course_preview textbox
                  gr.update(value=""),                      # marking_preview textbox
                  {},                                       # questions_json (gr.JSON) -> empty dict
                  "",                                       # state_course_text (string)
                  "",                                       # state_marking_text (string)
                  "",                                       # gen_debug (textbox)
                  []                                        # state_questions (gr.State) -> empty list
                )

    # Call LLM to generate only essays & short answers
              try:
                questions, raw = generate_questions(
                  course_text,
                  marking_text,
                  n_essay=int(ne or 0),
                  n_short=int(ns or 0),
                  temperature=float(temp or 0.6)
                )
              except Exception as e:
        # If generation fails, return safe types
                return (
                  gr.update(value=sanitize_text_for_preview(course_text, max_len=4000)),
                  gr.update(value=sanitize_text_for_preview(marking_text, max_len=2000)),
                  {},                # questions JSON safe fallback
                  course_text,
                  marking_text,
                  f"Generation failed: {e}",
                  []                 # state_questions safe fallback
                )

    # sanitize previews (never show raw binary)
              course_preview = sanitize_text_for_preview(course_text, max_len=4000)
              marking_preview = sanitize_text_for_preview(marking_text, max_len=2000)

    # questions: expected to be a Python list of dicts (or empty list)
              questions_for_state = questions if isinstance(questions, list) else []

    # return values in the exact order that the UI's generate_btn.click expects
              return (
                gr.update(value=course_preview),                    # course_preview textbox
                gr.update(value=marking_preview),                   # marking_preview textbox
                questions if questions else {},                     # questions_json (show list or {} to avoid JSON parse errors)
                course_text,                                        # state_course_text (store full text)
                marking_text,                                       # state_marking_text
                raw if raw is not None else "",                     # gen_debug raw model output (string)
                questions_for_state                                 # state_questions (actual list stored for Student tab)
              )


            generate_btn.click(
              do_generate,
              inputs=[course_file, marking_file, manual_marking, n_essay, n_short, temp_slider],
              outputs=[
                course_preview,    # course_preview (Textbox)
                marking_preview,    # (if you have marking_preview separate change appropriately)
                questions_json,        # Generated JSON shown
                state_course_text,       # store full course text
                state_marking_text,    # store source name or marking preview
                gen_debug,            # raw debug
                state_questions       # <-- NEW: store generated questions into state
            ]
          )


        # ---------------- Student tab ----------------
        with gr.Tab("2) Student — Take Test"):
            gr.Markdown("Students answer questions one at a time. Click **Start Test** to begin. Use **Submit Answer** to save each answer; when finished click **Finish & Grade** to have the AI assess everything.")
            start_btn = gr.Button("Start Test")
            progress_md = gr.Markdown("", visible=True)
            q_text = gr.Textbox(label="Question", interactive=False, lines=6)
            answer_tb = gr.Textbox(label="Your answer", lines=8, placeholder="Type your essay or short answer here...")
            submit_btn = gr.Button("Submit Answer (Save & Next)")
            finish_btn = gr.Button("Finish & Grade (AI)")
            student_status = gr.Textbox(label="Status", interactive=False, visible=False)

            # Start test: initialize index and answers list (state_student_answers)
            def start_test(questions):
                qs = questions or []
                if not qs:
                    return gr.update(value="No questions generated yet. Please ask your lecturer to create questions."), gr.update(value=""), [], 0, [], gr.update(value="No questions.", visible=True)
                # initialize
                idx0 = 0
                student_answers0 = []
                # show first question
                q0 = qs[idx0]
                prog = f"Question 1 of {len(qs)}"
                qtext_val = q0.get("question","")
                return gr.update(value=prog), gr.update(value=qtext_val), "", idx0, student_answers0, gr.update(value="Test started.", visible=True)

            start_btn.click(
                start_test,
                inputs=[state_questions],
                outputs=[progress_md, q_text, answer_tb, state_current_idx, state_student_answers, student_status]
            )

            # Submit answer & advance
            def submit_answer(answer, questions, idx, student_answers):
                qs = questions or []
                if not qs:
                    return gr.update(value="No questions."), gr.update(value=""), "", idx, student_answers, gr.update(value="No questions.", visible=True)
                # store answer for current idx
                student_answers = list(student_answers or [])
                try:
                    q = qs[idx]
                except IndexError:
                    return gr.update(value=""), gr.update(value=""), "", idx, student_answers, gr.update(value="Index out of range.", visible=True)

                student_answers.append({
                    "id": q.get("id"),
                    "type": q.get("type"),
                    "question": q.get("question",""),
                    "student_answer": answer or "",
                    "timestamp": datetime.now().isoformat()
                })
                # advance
                idx_next = idx + 1
                total = len(qs)
                if idx_next >= total:
                    # finished taking test (but not graded yet)
                    return (
                        gr.update(value=f"Completed all {total} questions. Click 'Finish & Grade' to assess."),
                        gr.update(value=""),
                        "",
                        idx_next,
                        student_answers,
                        gr.update(value="Answers saved. Ready to grade.", visible=True)
                    )
                # else show next question
                qnext = qs[idx_next]
                prog = f"Question {idx_next+1} of {total}"
                return gr.update(value=prog), gr.update(value=qnext.get("question","")), "", idx_next, student_answers, gr.update(value="Answer saved.", visible=True)

            submit_btn.click(
                submit_answer,
                inputs=[answer_tb, state_questions, state_current_idx, state_student_answers],
                outputs=[progress_md, q_text, answer_tb, state_current_idx, state_student_answers, student_status]
            )

            # Finish & grade (calls LLM graders for each answer)
            def finish_and_grade(questions, student_answers):
                qs = questions or []
                answers = student_answers or []
                if not qs:
                    return {}, "No questions to grade.", []
                if len(answers) == 0:
                    return {}, "No answers submitted.", []
                # map student answers by id for easy lookup (preserve order from qs)
                ans_map = {a["id"]: a for a in answers}
                graded = []
                total_score = 0
                max_total = 0
                for q in qs:
                    qid = q.get("id")
                    qtype = q.get("type")
                    qtext = q.get("question","")
                    rubric = q.get("rubric","")
                    max_score = int(q.get("max_score", 10))
                    student_answer_obj = ans_map.get(qid, {})
                    student_text = student_answer_obj.get("student_answer","")
                    # Grade using LLM
                    grade_obj = grade_answer_with_llm(qtext, rubric, max_score, student_text)
                    graded_item = {
                        "id": qid,
                        "type": qtype,
                        "question": qtext,
                        "student_answer": student_text,
                        "score": grade_obj.get("score", 0),
                        "max": max_score,
                        "feedback": grade_obj.get("feedback",""),
                        "annotated_answer": grade_obj.get("annotated_answer","")
                    }
                    graded.append(graded_item)
                    total_score += graded_item["score"]
                    max_total += max_score
                percentage = round(100.0 * total_score / max(1, max_total), 2)
                summary = {
                    "total_score": total_score,
                    "max_total": max_total,
                    "percentage": percentage,
                    "graded_count": len(graded)
                }
                # Save results to state_results
                return graded, json.dumps({"summary": summary, "details_count": len(graded)}), graded

            finish_btn.click(
                finish_and_grade,
                inputs=[state_questions, state_student_answers],
                outputs=[gr.JSON(label="Graded results (JSON)"), gr.Textbox(label="Summary (JSON string)"), state_results]
            )

        # ---------------- Assessment tab ----------------
        with gr.Tab("3) Assessment — Review Feedback"):
            gr.Markdown("View per-question AI assessment, feedback, annotated answer suggestions and plain text summary.")
            graded_json = gr.JSON(label="Graded results (list)")
            summary_text = gr.Textbox(label="Plaintext summary", lines=6, interactive=False)
            per_question_feedback = gr.Dataframe(headers=["id","type","score","max","feedback"], row_count=(1, "dynamic"))
            refresh_btn = gr.Button("Load latest assessment")

            def load_assessment(results):
                if not results:
                    return {}, "", pd.DataFrame([], columns=["id","type","score","max","feedback"])
                # build plain text summary
                total = sum([r.get("score",0) for r in results])
                max_total = sum([r.get("max",0) for r in results])
                pct = round(100.0 * total / max(1, max_total), 2)
                lines = [f"Total: {total} / {max_total} ({pct}%)", ""]
                rows = []
                for r in results:
                    lines.append(f"Q{r.get('id')}: {r.get('score')} / {r.get('max')} — {r.get('feedback')[:200]}")
                    rows.append([r.get("id"), r.get("type"), r.get("score"), r.get("max"), r.get("feedback")])
                return results, "\n".join(lines), pd.DataFrame(rows, columns=["id","type","score","max","feedback"])

            refresh_btn.click(load_assessment, inputs=[state_results], outputs=[graded_json, summary_text, per_question_feedback])

        # ---------------- Export tab ----------------
        with gr.Tab("4) Export"):
            gr.Markdown("Export graded assessment (CSV / DOCX).")
            export_csv_btn = gr.Button("Export results CSV")
            export_docx_btn = gr.Button("Export results DOCX")
            exported_csv = gr.File(label="Downloaded CSV")
            exported_docx = gr.File(label="Downloaded DOCX")

            def do_export_csv(questions, results):
                if not results:
                    return ""
                path = export_results_csv(questions or [], [], results or [])
                return path

            def do_export_docx(questions, results):
                if not results:
                    return ""
                path = export_results_docx(questions or [], results or [])
                return path

            export_csv_btn.click(do_export_csv, inputs=[state_questions, state_results], outputs=[exported_csv])
            export_docx_btn.click(do_export_docx, inputs=[state_questions, state_results], outputs=[exported_docx])

    # Footer/help
    with gr.Accordion("Notes & Setup", open=False):
        gr.Markdown("""
**Setup**
- Install: `pip install gradio pymupdf groq python-docx pandas`
- Set `GROQ_API_KEY` in your environment.
- Run: `python essay_assessor_gradio.py`

**How it works**
1. Lecturer uploads course material & marking guide (or pastes the guide). Click *Generate Questions*.
2. Student clicks *Start Test*, answers each question one-by-one, and clicks *Finish & Grade*.
3. AI grades answers against provided marking guide; results appear in Assessment tab.
4. Export CSV/DOCX from Export tab.

**Notes**
- This MVP expects text-based PDFs (not scanned images). For OCR, add `pytesseract` + `pdf2image`.
- The LLM is asked to return strict JSON. If generation/parsing fails, check the debug output on the Lecturer tab.
""")

if __name__ == "__main__":
    if not os.getenv("GROQ_API_KEY"):
        print("WARNING: GROQ_API_KEY not set. Set it before running the app.")
    demo.launch(debug=True)