Feat: Add Dummy LLM Client and Redis Lock check

docker-compose.yml

@@ -33,6 +33,8 @@ x-server: &base_server_setup
     CACHE_TEST_REDIS_URL: ${CACHE_TEST_REDIS_URL:-redis://redis:6379/11}
     CACHE_MIDDLEWARE_SECONDS: ${CACHE_MIDDLEWARE_SECONDS:-600}
     ELASTIC_SEARCH_HOST: ${ELASTIC_SEARCH_HOST:-elasticsearch://elasticsearch:9200}
+    # LLM
+    USE_DUMMY_LLM_CLIENT: ${USE_DUMMY_LLM_CLIENT:-true}
     # Appeal API
     APPEALS_USER: ${APPEALS_USER:-}
     APPEALS_PASS: ${APPEALS_PASS:-}

dref/admin.py

@@ -330,7 +330,7 @@ class ProposedActionAdmin(ReadOnlyMixin, admin.ModelAdmin):
 
 @admin.register(DrefSummary)
 class DrefSummaryAdmin(TranslationAdmin, admin.ModelAdmin):
-    list_display = ("dref", "status", "source_model_name", "source_id", "created_at", "updated_at")
+    list_display = ("dref", "status", "source", "source_id", "created_at", "updated_at")
     list_filter = ("status",)
     search_fields = ("dref__title", "dref__appeal_code")
     readonly_fields = ("source_hash", "created_at", "updated_at")
@@ -351,15 +351,10 @@ class DrefSummaryAdmin(TranslationAdmin, admin.ModelAdmin):
 
     @admin.action(description="Regenerate summary for selected DREFs")
     def regenerate_summary(self, request, queryset):
-        """Re-trigger summary generation, replaying the source the summary was last built from."""
+        """Re-trigger summary generation from the DREF's current latest approved source."""
+        queryset.filter(status=DrefSummary.SummaryStatus.PROCESSING).update(status=DrefSummary.SummaryStatus.FAILED)
         for summary in queryset:
-            generate_dref_summary.delay(
-                # Fall back to the Dref itself for rows generated before the
-                # source was tracked.
-                source_model_name=summary.source_model_name or DrefSummary.SourceModel.DREF,
-                source_id=summary.source_id or summary.dref_id,
-                overwrite=True,
-            )
+            generate_dref_summary.delay(dref_id=summary.dref_id, overwrite=True)
         self.message_user(
             request,
             f"Queued summary regeneration for {queryset.count()} DREF(s).",

dref/factories/dref.py

@@ -207,7 +207,7 @@ class Meta:
 
     dref = factory.SubFactory(DrefFactory)
     source_hash = factory.Sequence(lambda n: f"{n:064d}")
-    source_model_name = DrefSummary.SourceModel.DREF
+    source = DrefSummary.SourceModel.DREF
     source_id = factory.SelfAttribute("dref.id")
     status = DrefSummary.SummaryStatus.SUCCESS
     situational_overview = fuzzy.FuzzyText(length=100)

dref/migrations/0090_drefsummary.py

@@ -17,7 +17,7 @@ class Migration(migrations.Migration):
                 ('id', models.AutoField(auto_created=True, primary_key=True, serialize=False, verbose_name='ID')),
                 ('created_at', models.DateTimeField(auto_now_add=True)),
                 ('updated_at', models.DateTimeField(auto_now=True)),
-                ('source_model_name', models.IntegerField(choices=[(100, 'Dref'), (200, 'Dref Operational Update'), (300, 'Dref Final Report')], verbose_name='source model name')),
+                ('source', models.IntegerField(choices=[(100, 'Dref'), (200, 'Dref Operational Update'), (300, 'Dref Final Report')], help_text='The model this summary was generated from.', verbose_name='source')),
                 ('source_id', models.PositiveBigIntegerField(verbose_name='source id')),
                 ('source_hash', models.CharField(max_length=64, unique=True)),
                 ('situational_overview', models.TextField(blank=True, null=True)),

dref/models.py

@@ -1661,6 +1661,9 @@ class DrefFinalReport(models.Model):
     )
     __financial_report_id = None
 
+    # TYPING
+    id: int
+
     class Meta:
         verbose_name = _("Dref Final Report")
         verbose_name_plural = _("Dref Final Reports")
@@ -1728,11 +1731,9 @@ class SourceModel(models.IntegerChoices):
     created_at = models.DateTimeField(auto_now_add=True)
     updated_at = models.DateTimeField(auto_now=True)
 
-    # The source document the summary was last generated from. Stored so a
-    # retrigger (e.g. from the admin) can replay the same source instead of
-    # re-resolving the latest approved document.
-    source_model_name = models.IntegerField(
-        verbose_name=_("source model name"),
+    source = models.IntegerField(
+        verbose_name=_("source"),
+        help_text=_("The model this summary was generated from."),
         choices=SourceModel.choices,
     )
     source_id = models.PositiveBigIntegerField(

dref/summary.py

@@ -5,15 +5,14 @@
 import hashlib
 import json
 import logging
-from typing import Callable, Dict, List, Optional
+from typing import Callable, Dict, List, Optional, Union
 
 import tiktoken
-from django.conf import settings
-from django.utils.functional import cached_property
-from openai import AzureOpenAI
+from django.db.models import F
 
 from api.utils import get_model_name
-from dref.models import Dref, DrefFinalReport, DrefOperationalUpdate
+from dref.models import Dref, DrefFinalReport, DrefOperationalUpdate, DrefSummary
+from main.llm import get_dref_summary_llm_client
 
 logger = logging.getLogger(__name__)
 
@@ -22,6 +21,16 @@
 MAX_OUTPUT_CHARS_PER_FIELD = 1500
 MAX_INPUT_TOKENS = 10000
 
+
+# The models a DrefSummary can be generated from.
+DrefSummarySource = Union[Dref, DrefOperationalUpdate, DrefFinalReport]
+
+SOURCE_BY_MODEL: Dict[type, DrefSummary.SourceModel] = {
+    Dref: DrefSummary.SourceModel.DREF,
+    DrefOperationalUpdate: DrefSummary.SourceModel.DREF_OPERATIONAL_UPDATE,
+    DrefFinalReport: DrefSummary.SourceModel.DREF_FINAL_REPORT,
+}
+
 # DrefSummary fields — order is the iteration order for prompt assembly.
 SUMMARY_FIELDS: List[str] = [
     "situational_overview",
@@ -121,28 +130,6 @@ def _build_lessons_learned_prompt(**kwargs) -> str:
 )
 
 
-class DrefSummaryLLMClient:
-    @cached_property
-    def client(self):
-        return AzureOpenAI(
-            azure_endpoint=settings.AZURE_OPENAI_ENDPOINT,
-            api_key=settings.AZURE_OPENAI_KEY,
-            api_version="2023-05-15",
-        )
-
-    def get_response(self, messages) -> Optional[str]:
-        try:
-            response = self.client.chat.completions.create(
-                model=settings.AZURE_OPENAI_DEPLOYMENT_NAME,
-                messages=messages,
-                temperature=0.5,
-            )
-            return response.choices[0].message.content
-        except Exception as e:
-            logger.error(f"Error while generating DREF summary response: {e}", exc_info=True)
-            return None
-
-
 def count_tokens(text: str) -> int:
     try:
         return len(tiktoken.get_encoding(ENCODING_NAME).encode(text))
@@ -184,8 +171,8 @@ def _extract_fields(obj, field_names: List[str]) -> dict:
 class DrefSummaryGenerator:
     """Assembles per-section prompts from a source document and produces all summaries."""
 
-    def __init__(self, client: Optional[DrefSummaryLLMClient] = None):
-        self.client = client or DrefSummaryLLMClient()
+    def __init__(self):
+        self.client = get_dref_summary_llm_client()
 
     # Shared helpers — called by multiple extractors
 
@@ -302,8 +289,13 @@ def _extract_dref_final_kwargs(cls, final) -> Dict[str, dict]:
         return kwargs
 
     @classmethod
-    def _get_section_kwargs(cls, source_doc) -> Dict[str, dict]:
-        """Dispatch to the model-specific extractor."""
+    def get_section_kwargs(cls, source_doc: DrefSummarySource) -> Dict[str, dict]:
+        """Dispatch to the model-specific extractor.
+
+        Public so callers that need both the hash and the generated summary
+        for the same ``source_doc`` (see ``compute_source_hash``/``generate_all``)
+        can compute this once and pass it to both instead of extracting twice.
+        """
         if isinstance(source_doc, Dref):
             return cls._extract_dref_kwargs(source_doc)
         if isinstance(source_doc, DrefOperationalUpdate):
@@ -312,16 +304,47 @@ def _get_section_kwargs(cls, source_doc) -> Dict[str, dict]:
             return cls._extract_dref_final_kwargs(source_doc)
         return {}
 
+    @staticmethod
+    def get_latest_approved_source(dref: Dref) -> Optional[tuple[DrefSummary.SourceModel, DrefSummarySource]]:
+        """Most authoritative (source, source object) pair for ``dref``.
+
+        Priority: Final Report > latest Operational Update > Dref itself.
+        """
+        final_report = (
+            DrefFinalReport.objects.select_related("country", "disaster_type")
+            .filter(dref=dref, status=Dref.Status.APPROVED)
+            .order_by("-created_at")
+            .first()
+        )
+        if final_report:
+            return SOURCE_BY_MODEL[DrefFinalReport], final_report
+
+        latest_ops_update = (
+            DrefOperationalUpdate.objects.select_related("country", "disaster_type")
+            .filter(dref=dref, status=Dref.Status.APPROVED)
+            .order_by(F("operational_update_number").desc(nulls_last=True), "-created_at")
+            .first()
+        )
+        if latest_ops_update:
+            return SOURCE_BY_MODEL[DrefOperationalUpdate], latest_ops_update
+
+        if dref.status == Dref.Status.APPROVED:
+            return SOURCE_BY_MODEL[Dref], dref
+
+        return None
+
     @classmethod
-    def build_source_text(cls, source_doc) -> str:
+    def build_source_text(cls, source_doc: DrefSummarySource) -> str:
         """JSON representation of all section kwargs for inspection and debugging."""
-        return json.dumps(cls._get_section_kwargs(source_doc), indent=2, ensure_ascii=False, default=str)
+        return json.dumps(cls.get_section_kwargs(source_doc), indent=2, ensure_ascii=False, default=str)
 
     @classmethod
-    def compute_source_hash(cls, source_doc) -> str:
+    def compute_source_hash(cls, source_doc: DrefSummarySource, section_kwargs: Optional[Dict[str, dict]] = None) -> str:
         """Hash of all source content feeding the summary, for change detection."""
         model_label = get_model_name(type(source_doc))
-        payload = {"model": model_label, "id": source_doc.id, "source": cls._get_section_kwargs(source_doc)}
+        if section_kwargs is None:
+            section_kwargs = cls.get_section_kwargs(source_doc)
+        payload = {"model": model_label, "id": source_doc.id, "source": section_kwargs}
         content = json.dumps(payload, sort_keys=True, ensure_ascii=False, default=str)
         return hashlib.sha256(content.encode("utf-8")).hexdigest()
 
@@ -337,16 +360,19 @@ def _parse_response(response: str) -> Dict[str, str]:
             raise ValueError("Expected a JSON object of summary fields")
         return data
 
-    def generate_all(self, source_doc) -> Dict[str, str]:
+    def generate_all(self, source_doc: DrefSummarySource, section_kwargs: Optional[Dict[str, dict]] = None) -> Dict[str, str]:
         """Generate every summary for ``source_doc`` in a single LLM call.
 
         Always returns a value for every ``SUMMARY_FIELDS`` key (empty string
         when a section has no content), so callers overwrite stale values
-        instead of leaving them in place on regeneration.
+        instead of leaving them in place on regeneration. ``section_kwargs``
+        can be passed in by a caller that already computed it (e.g. via
+        ``compute_source_hash``) to avoid re-extracting it from ``source_doc``.
         """
         empty_results: Dict[str, str] = {field_name: "" for field_name in SUMMARY_FIELDS}
 
-        section_kwargs = self._get_section_kwargs(source_doc)
+        if section_kwargs is None:
+            section_kwargs = self.get_section_kwargs(source_doc)
         if not section_kwargs:
             logger.info(f"No source content for ({type(source_doc).__name__}) ({source_doc.id}) summary; skipping generation.")
             return empty_results