From 93f0bb89272245cd30d7b472b049285b9b043dab Mon Sep 17 00:00:00 2001 From: Ross Date: Mon, 20 Oct 2025 10:43:12 +0100 Subject: [PATCH] Add LLM question import functionality and enhance question model with new fields --- sbas/forms.py | 21 ++ ...ion_condition_question_finding_and_more.py | 39 +++ sbas/migrations/0019_question_title.py | 18 ++ sbas/models.py | 9 + sbas/templates/sbas/base.html | 7 + sbas/templates/sbas/import_llm_questions.html | 11 + sbas/templates/sbas/llm_prompt_view.html | 133 ++++++++++ sbas/templates/sbas/question_detail.html | 70 ++++- sbas/urls.py | 12 + sbas/views.py | 248 ++++++++++++++++++ 10 files changed, 565 insertions(+), 3 deletions(-) create mode 100644 sbas/migrations/0018_question_condition_question_finding_and_more.py create mode 100644 sbas/migrations/0019_question_title.py create mode 100644 sbas/templates/sbas/import_llm_questions.html create mode 100644 sbas/templates/sbas/llm_prompt_view.html diff --git a/sbas/forms.py b/sbas/forms.py index fd4e5752..1101a18a 100755 --- a/sbas/forms.py +++ b/sbas/forms.py @@ -3,6 +3,7 @@ from django.forms import ( ModelForm, ModelMultipleChoiceField, ModelChoiceField, + CheckboxSelectMultiple, ChoiceField, CharField, ) @@ -22,6 +23,8 @@ from django.forms.widgets import RadioSelect, TextInput, Textarea from tinymce.widgets import TinyMCE +from dal import autocomplete + class UserAnswerForm(ModelForm): class Meta: @@ -140,6 +143,11 @@ class QuestionForm(ModelForm): "e_answer", "e_feedback", "best_answer", + "finding", + "structure", + "condition", + "presentation", + "subspecialty", ] widgets = { @@ -158,6 +166,19 @@ class QuestionForm(ModelForm): "c_feedback" : TinyMCE(attrs={'cols': 80, 'rows': 4}, mce_attrs={'height': 140}), "d_feedback" : TinyMCE(attrs={'cols': 80, 'rows': 4}, mce_attrs={'height': 140}), "e_feedback" : TinyMCE(attrs={'cols': 80, 'rows': 4}, mce_attrs={'height': 140}), + "structure": autocomplete.ModelSelect2Multiple( + url="atlas:structure-autocomplete" + ), + "finding": autocomplete.ModelSelect2Multiple( + url="atlas:finding-autocomplete" + ), + "condition": autocomplete.ModelSelect2Multiple( + url="atlas:condition-autocomplete" + ), + "presentation": autocomplete.ModelSelect2Multiple( + url="atlas:presentation-autocomplete" + ), + "subspecialty": CheckboxSelectMultiple(), } #widgets = { diff --git a/sbas/migrations/0018_question_condition_question_finding_and_more.py b/sbas/migrations/0018_question_condition_question_finding_and_more.py new file mode 100644 index 00000000..044b9bae --- /dev/null +++ b/sbas/migrations/0018_question_condition_question_finding_and_more.py @@ -0,0 +1,39 @@ +# Generated by Django 5.1.4 on 2025-10-20 08:47 + +from django.db import migrations, models + + +class Migration(migrations.Migration): + + dependencies = [ + ('atlas', '0079_casecollection_prerequisites'), + ('sbas', '0017_exam_results_supervisor_visible'), + ] + + operations = [ + migrations.AddField( + model_name='question', + name='condition', + field=models.ManyToManyField(blank=True, related_name='sbas_questions', to='atlas.condition'), + ), + migrations.AddField( + model_name='question', + name='finding', + field=models.ManyToManyField(blank=True, related_name='sbas_questions', to='atlas.finding'), + ), + migrations.AddField( + model_name='question', + name='presentation', + field=models.ManyToManyField(blank=True, related_name='sbas_questions', to='atlas.presentation'), + ), + migrations.AddField( + model_name='question', + name='structure', + field=models.ManyToManyField(blank=True, related_name='sbas_questions', to='atlas.structure'), + ), + migrations.AddField( + model_name='question', + name='subspecialty', + field=models.ManyToManyField(blank=True, related_name='sbas_questions', to='atlas.subspecialty'), + ), + ] diff --git a/sbas/migrations/0019_question_title.py b/sbas/migrations/0019_question_title.py new file mode 100644 index 00000000..2108c61f --- /dev/null +++ b/sbas/migrations/0019_question_title.py @@ -0,0 +1,18 @@ +# Generated by Django 5.1.4 on 2025-10-20 09:30 + +from django.db import migrations, models + + +class Migration(migrations.Migration): + + dependencies = [ + ('sbas', '0018_question_condition_question_finding_and_more'), + ] + + operations = [ + migrations.AddField( + model_name='question', + name='title', + field=models.CharField(blank=True, help_text='Short title for question', max_length=200, null=True), + ), + ] diff --git a/sbas/models.py b/sbas/models.py index d60638b2..430bb372 100644 --- a/sbas/models.py +++ b/sbas/models.py @@ -15,6 +15,8 @@ import reversion from django.contrib.contenttypes.fields import GenericRelation +from atlas.models import Finding, Structure, Condition, Subspecialty, Presentation + class Category(models.Model): category = models.CharField(max_length=200) @@ -24,6 +26,7 @@ class Category(models.Model): class Question(QuestionBase): + title = models.CharField(max_length=200, help_text="Short title for question", blank=True, null=True) stem = models.TextField( blank=False, help_text="Stem of the question", @@ -91,6 +94,12 @@ class Question(QuestionBase): Category, on_delete=models.SET_NULL, null=True, blank=True ) + finding = models.ManyToManyField(Finding, blank=True, related_name="sbas_questions") + structure = models.ManyToManyField(Structure, blank=True, related_name="sbas_questions") + condition = models.ManyToManyField(Condition, blank=True, related_name="sbas_questions") + presentation = models.ManyToManyField(Presentation, blank=True, related_name="sbas_questions") + subspecialty = models.ManyToManyField(Subspecialty, blank=True, related_name="sbas_questions") + #notes = GenericRelation(QuestionNote) def __str__(self): diff --git a/sbas/templates/sbas/base.html b/sbas/templates/sbas/base.html index 1b601c53..2148a89d 100644 --- a/sbas/templates/sbas/base.html +++ b/sbas/templates/sbas/base.html @@ -27,6 +27,13 @@
  • Question
    • +
  • + + +
    • {% if request.user.is_superuser %}
  • diff --git a/sbas/templates/sbas/import_llm_questions.html b/sbas/templates/sbas/import_llm_questions.html new file mode 100644 index 00000000..37dfab34 --- /dev/null +++ b/sbas/templates/sbas/import_llm_questions.html @@ -0,0 +1,11 @@ +{% extends 'sbas/base.html' %} + +{% block content %} +

    Import LLM Questions

    +
    + {% csrf_token %} + {{ form.as_p }} + +
    +

    Upload a single JSON array, a single object, or JSONL (one JSON object per line) matching the agreed schema.

    +{% endblock %} diff --git a/sbas/templates/sbas/llm_prompt_view.html b/sbas/templates/sbas/llm_prompt_view.html new file mode 100644 index 00000000..a1386b59 --- /dev/null +++ b/sbas/templates/sbas/llm_prompt_view.html @@ -0,0 +1,133 @@ +{% extends 'sbas/base.html' %} + +{% block content %} +

    LLM Prompts

    +

    + Below are the prompts used for interacting with large language models (LLMs) within the SBA system. +

    + +

    Question Generation Prompt

    + 
    +You are an expert radiologist tasked with generating high-quality multiple-choice questions for medical imaging education. Each question should consist of a clinical vignette, an image description, and five answer choices (A through E), with one correct answer and four plausible distractors.
+
+Please adhere to the following guidelines when creating each question:
+1. Clinical Vignette: Provide a brief clinical scenario that sets the context for the question including relevant patient history and clinical findings.
+2. Make sure the question focuses on key radiological concepts, findings, or diagnoses.
+3. Answer Choices: Create five answer options. Ensure that one is the correct answer and the others are plausible distractors.
+4. Explanation: Provide a concise explanation for why the correct answer is right and why the distractors are incorrect.
+5. Difficulty Level: Aim for a moderate difficulty level suitable for training radiologists.
+6. Clarity and Precision: Use clear and precise language, avoiding ambiguity.
+7. Relevance: Ensure that the question is relevant to current radiological practices and guidelines.
+8. Questions should feature metadata tags for finding, structure, condition, presentation, and subspecialty.
+9. Each question output should be a json object with the following schema:
+10. For article sources pleese reference the article doi within the explanation field in the format [doi:10.xxxx/xxxxx].
+11. For statdx sources please reference the article in the format [statdx:article_id].
+12. References should be included inline within the explanation field.
+13. Questions need to have a radiology focus, please avoid questions that are purely clinical or biochemical without imaging relevance.
+
+JSON Schema (draft-07 compatible)
+{
+"$schema": "http://json-schema.org/draft-07/schema#",
+"title": "SBA Question",
+"description": "Schema for importing LLM-generated questions into the sbas.Question model.",
+"type": "object",
+"additionalProperties": false,
+"required": [
+"title",
+"stem",
+"a_answer",
+"b_answer",
+"c_answer",
+"d_answer",
+"e_answer",
+"best_answer",
+"category"
+],
+"properties": {
+"title": {
+"type": "string",
+"description": "Short title for the question."
+},
+"stem": {
+"type": "string",
+"description": "HTML or plain text question stem. Trim leading/trailing whitespace before saving."
+},
+"a_answer": { "type": "string" },
+"a_feedback": { "type": "string", "default": "" },
+"b_answer": { "type": "string" },
+"b_feedback": { "type": "string", "default": "" },
+"c_answer": { "type": "string" },
+"c_feedback": { "type": "string", "default": "" },
+"d_answer": { "type": "string" },
+"d_feedback": { "type": "string", "default": "" },
+"e_answer": { "type": "string" },
+"e_feedback": { "type": "string", "default": "" },
+"feedback": {
+"type": "string",
+"description": "General question feedback (may include HTML)."
+},
+"best_answer": {
+"type": "string",
+"enum": ["a", "b", "c", "d", "e"],
+"description": "One of 'a','b','c','d','e'."
+},
+"category": {
+"oneOf": [
+{ "type": "string", "description": "Category name ('Central Nervous and Head & Neck', 'Paediatric', 'Genito-urinary, Adrenal, Obstetrics & Gynaecology and Breast', 'Gastro-intestinal', 'Musculoskeletal and Trauma', 'Cardiothoracic and Vascular')" }
+]
+},
+"finding": {
+"type": "array",
+"description": "References to Finding objects. Prefer numeric IDs; names allowed if importer resolves them.",
+"items": {
+"oneOf": [
+{ "type": "integer" },
+{ "type": "string" }
+]
+},
+"uniqueItems": true
+},
+"structure": {
+"type": "array",
+"items": {
+"oneOf": [
+{ "type": "integer" },
+{ "type": "string" }
+]
+},
+"uniqueItems": true
+},
+"condition": {
+"type": "array",
+"items": {
+"oneOf": [
+{ "type": "integer" },
+{ "type": "string" }
+]
+},
+"uniqueItems": true
+},
+"presentation": {
+"type": "array",
+"items": {
+"oneOf": [
+{ "type": "integer" },
+{ "type": "string" }
+]
+},
+"uniqueItems": true
+},
+"subspecialty": {
+"type": "array",
+"items": {
+"oneOf": [
+{ "type": "integer" },
+{ "type": "string", "description": "Subspecialty name (e.g. 'Haematology & Oncology', 'Vascular', 'Uroradiology', 'Thoracic', 'Paediatric', 'Obstetric and Gynaecological', 'Neuroradiology', 'Musculoskeletal', 'Head and Neck', 'Gastrointestinal and hepatobiliary', 'Cardiac', 'Breast')" }
+]
+},
+"uniqueItems": true
+},
+}
+}
+
+
    \ No newline at end of file diff --git a/sbas/templates/sbas/question_detail.html b/sbas/templates/sbas/question_detail.html index a687d5da..c0d2bf1e 100644 --- a/sbas/templates/sbas/question_detail.html +++ b/sbas/templates/sbas/question_detail.html @@ -29,14 +29,78 @@ {{ exam }} {% endfor %} +
    + Findings: {% if question.finding.exists %} +
      + {% for f in question.finding.all %} +
    • {{ f.get_link|safe }}
      • + {% endfor %} +
    + {% else %} + None + {% endif %} +
    +
    + Structures: {% if question.structure.exists %} +
      + {% for s in question.structure.all %} +
    • {{ s.get_link|safe }}
      • + {% endfor %} +
    + {% else %} + None + {% endif %} +
    +
    + Conditions: {% if question.condition.exists %} +
      + {% for c in question.condition.all %} +
    • {{ c.get_link|safe }}
      • + {% endfor %} +
    + {% else %} + None + {% endif %} +
    +
    + Presentations: {% if question.presentation.exists %} +
      + {% for p in question.presentation.all %} +
    • {{ p.get_link|safe }}
      • + {% endfor %} +
    + {% else %} + None + {% endif %} +
    +
    + Subspecialties: {% if question.subspecialty.exists %} +
      + {% for ss in question.subspecialty.all %} +
    • {{ ss.get_link|safe }}
      • + {% endfor %} +
    + {% else %} + None + {% endif %} +
    Category: {{ question.category }}
    - Author: {% for user in question.author.all %} - {{ author }}, - {% endfor %} + Author(s): + {% if question.author.exists %} + {% for u in question.author.all %} + {{ u.get_full_name|default:u.username }}{% if not forloop.last %}, {% endif %} + {% endfor %} + {% else %} + Unknown + {% endif %} +
    + +
    + Feedback: {{ question.feedback|linebreaks }}
    {% include 'question_notes.html' %} diff --git a/sbas/urls.py b/sbas/urls.py index 14c04c5c..6fb3ee3d 100644 --- a/sbas/urls.py +++ b/sbas/urls.py @@ -82,6 +82,18 @@ urlpatterns = [ views.UserAnswerDelete.as_view(), name="user_answer_delete", ), + path( + "llm_prompts/", + views.llm_prompt_view, + name="llm_prompt_view", + + ) + , + path( + "import_llm_questions/", + views.import_llm_questions, + name="import_llm_questions", + ), #path( # "exam//scores///", # views.exam_scores_cid_user, diff --git a/sbas/views.py b/sbas/views.py index 52379f92..10b381b8 100644 --- a/sbas/views.py +++ b/sbas/views.py @@ -29,6 +29,250 @@ from django.http import Http404, HttpResponseBadRequest, JsonResponse from django.http import HttpResponseRedirect, HttpResponse from .models import Question, Category, Exam, UserAnswer +from django.views.decorators.http import require_http_methods +from django.db import transaction +import re +import logging + +logger = logging.getLogger(__name__) + +from loguru import logger + +try: + import jsonschema +except Exception: + jsonschema = None + + +class LLMImportForm(forms.Form): + file = forms.FileField(required=False, help_text="Upload a JSON or JSONL file matching the SBA import schema") + raw = forms.CharField( + required=False, + widget=forms.Textarea(attrs={"rows": 10, "cols": 80}), + help_text="Or paste JSON, JSONL (newline-delimited), or multiple JSON documents separated by blank lines", + ) + + +def _resolve_m2m(model_class, value): + """Resolve an item which may be an int (pk) or a string (name/slug). + Returns a queryset or empty list of matching objects. + """ + if value is None: + return [] + if isinstance(value, int): + try: + return [model_class.objects.get(pk=value)] + except model_class.DoesNotExist: + return [] + if isinstance(value, str): + # try exact name field, then case-insensitive contains + qs = model_class.objects.filter(name__iexact=value) + if not qs.exists(): + qs = model_class.objects.filter(name__icontains=value) + return list(qs[:5]) + return [] + + +@login_required +@user_passes_test(lambda u: u.is_superuser) +@require_http_methods(["GET", "POST"]) +def import_llm_questions(request): + """Upload a JSON/JSONL file of LLM-produced questions and import them. + + Only superusers may use this view. The view validates each object against + the agreed JSON schema (draft-07) and attempts to resolve M2M references + by numeric id or by name. Returns a JSON report of created items and errors. + """ + schema = { + "$schema": "http://json-schema.org/draft-07/schema#", + "title": "SBA Question", + "type": "object", + "additionalProperties": False, + "required": [ + "title", + "stem", + "a_answer", + "b_answer", + "c_answer", + "d_answer", + "e_answer", + "best_answer", + "category", + ], + "properties": { + "title": {"type": "string"}, + "stem": {"type": "string"}, + "a_answer": {"type": "string"}, + "a_feedback": {"type": "string"}, + "b_answer": {"type": "string"}, + "b_feedback": {"type": "string"}, + "c_answer": {"type": "string"}, + "c_feedback": {"type": "string"}, + "d_answer": {"type": "string"}, + "d_feedback": {"type": "string"}, + "e_answer": {"type": "string"}, + "e_feedback": {"type": "string"}, + "feedback": {"type": "string"}, + "best_answer": {"type": "string", "enum": ["a", "b", "c", "d", "e"]}, + "category": {"oneOf": [{"type": "string"}]}, + "finding": {"type": "array", "items": {"oneOf": [{"type": "integer"}, {"type": "string"}]}, "uniqueItems": True}, + "structure": {"type": "array", "items": {"oneOf": [{"type": "integer"}, {"type": "string"}]}, "uniqueItems": True}, + "condition": {"type": "array", "items": {"oneOf": [{"type": "integer"}, {"type": "string"}]}, "uniqueItems": True}, + "presentation": {"type": "array", "items": {"oneOf": [{"type": "integer"}, {"type": "string"}]}, "uniqueItems": True}, + "subspecialty": {"type": "array", "items": {"oneOf": [{"type": "integer"}, {"type": "string"}]}, "uniqueItems": True}, + }, + } + + if request.method == "GET": + form = LLMImportForm() + return render(request, "sbas/import_llm_questions.html", {"form": form}) + + # POST + form = LLMImportForm(request.POST, request.FILES) + if not form.is_valid(): + return JsonResponse({"ok": False, "errors": ["No file uploaded or invalid form"]}, status=400) + + raw_text = None + if form.cleaned_data.get("raw"): + raw_text = form.cleaned_data.get("raw") + elif form.cleaned_data.get("file"): + f = form.cleaned_data["file"] + raw_text = f.read().decode("utf-8") + else: + return JsonResponse({"ok": False, "errors": ["No file uploaded or text provided"]}, status=400) + + # support: single JSON object, JSON array, JSONL (one JSON object per line), + # or multiple JSON documents separated by one or more blank lines. + candidates = [] + # sanitize raw_text by escaping backslashes that are not part of a valid + # JSON escape sequence. This handles inputs containing LaTeX-style + # sequences such as "\ge" which are invalid JSON (Invalid \escape). + def _sanitize_backslashes(s: str): + # valid escapes after backslash in JSON: \" \\ \/ \b \f \n \r \t and \uXXXX + # This regex finds backslashes not followed by ", \\ , /, b, f, n, r, t, or u + pattern = re.compile(r"\\(?!(?:[\"\\/bfnrtu]))") + new_s, n = pattern.subn(r"\\\\", s) + return new_s, n + + raw_text_stripped = raw_text.strip() + raw_text, sanitized_count = _sanitize_backslashes(raw_text_stripped) + logger.debug("import_llm_questions: sanitized input length %d (escaped %d backslashes)", len(raw_text), sanitized_count) + # Try whole-text JSON first (object or array) + parse_error = None + try: + parsed = json.loads(raw_text) + if isinstance(parsed, list): + candidates = parsed + elif isinstance(parsed, dict): + candidates = [parsed] + except Exception as e_outer: + parse_error = e_outer + # Try JSONL: each non-empty line is a JSON object + for i, line in enumerate(raw_text.splitlines()): + line = line.strip() + if not line: + continue + try: + candidates.append(json.loads(line)) + except Exception: + # not JSONL or some lines are multi-line JSON; fallthrough + candidates = [] + break + + # If JSONL didn't produce candidates, try splitting by blank lines + if not candidates: + blocks = [b.strip() for b in re.split(r"\n\s*\n", raw_text) if b.strip()] + if len(blocks) > 1: + for i, block in enumerate(blocks): + try: + candidates.append(json.loads(block)) + except Exception as e_block: + return JsonResponse({"ok": False, "errors": [f"Invalid JSON in block {i+1}: {e_block}"]}, status=400) + + if not candidates: + # If still empty, return the original parse error if present + msg = str(parse_error) if parse_error is not None else "No JSON objects found" + logger.debug(f"Failed to parse input: {msg}") + return JsonResponse({"ok": False, "errors": [f"Failed to parse input: {msg}"]}, status=400) + + report = {"total": len(candidates), "created": 0, "errors": [], "sanitized_backslashes": sanitized_count} + + if jsonschema is None: + return JsonResponse({"ok": False, "errors": ["jsonschema library not available in environment"]}, status=500) + + validator = jsonschema.Draft7Validator(schema) + + from atlas.models import Finding, Structure, Condition, Presentation, Subspecialty + + for idx, payload in enumerate(candidates): + errors = [] + for err in validator.iter_errors(payload): + errors.append(err.message) + if errors: + logger.debug(f"Validation errors for item {idx}: {errors}") + report["errors"].append({"index": idx, "errors": errors}) + continue + + # create the Question inside a transaction; partial failures should not leave half-written objects + try: + with transaction.atomic(): + q = Question() + q.stem = payload.get("stem", "").strip() + # map title -> not present in model; use as stem prefix or store in feedback + title = payload.get("title") + if title: + q.stem = f"{title}\n" + q.stem + + q.a_answer = payload.get("a_answer", "").strip() + q.a_feedback = payload.get("a_feedback", "") + q.b_answer = payload.get("b_answer", "").strip() + q.b_feedback = payload.get("b_feedback", "") + q.c_answer = payload.get("c_answer", "").strip() + q.c_feedback = payload.get("c_feedback", "") + q.d_answer = payload.get("d_answer", "").strip() + q.d_feedback = payload.get("d_feedback", "") + q.e_answer = payload.get("e_answer", "").strip() + q.e_feedback = payload.get("e_feedback", "") + q.feedback = payload.get("feedback", "") + q.best_answer = payload.get("best_answer") + # category resolve by name + cat_val = payload.get("category") + if isinstance(cat_val, str): + cat_obj, created = Category.objects.get_or_create(category=cat_val) + q.category = cat_obj + q.save() + + # Resolve M2Ms + for key, model_cls in ( + ("finding", Finding), + ("structure", Structure), + ("condition", Condition), + ("presentation", Presentation), + ("subspecialty", Subspecialty), + ): + vals = payload.get(key) or [] + resolved = [] + for v in vals: + if isinstance(v, int): + try: + resolved.append(model_cls.objects.get(pk=v)) + except model_cls.DoesNotExist: + # skip unknown + continue + elif isinstance(v, str): + qs = model_cls.objects.filter(name__iexact=v) + if not qs.exists(): + qs = model_cls.objects.filter(name__icontains=v) + if qs.exists(): + resolved.extend(list(qs[:5])) + if resolved: + getattr(q, key).add(*[o.pk for o in resolved]) + + report["created"] += 1 + except Exception as e: + report["errors"].append({"index": idx, "errors": [str(e)]}) + + return JsonResponse({"ok": True, "report": report}) from django_tables2 import SingleTableView, SingleTableMixin from django_filters.views import FilterView @@ -436,3 +680,7 @@ def exam_clone2(request, exam_id): new_exam = exam.clone_model() return redirect("sbas:exam_update", pk=new_exam.id) + + +def llm_prompt_view(request): + return render(request, "sbas/llm_prompt_view.html", {}) \ No newline at end of file