Always-on grounded retrieval + role-scoped Ollama models (#62 Phases 1–3) (#65)

## Summary
- Closes Phases 1–3 of [#62](#62) (Phase 4 eval harness left for a follow-up).
- **Accuracy:** Retrieval is decided every turn (`GroundingDecider`, fails open). `FAST` no longer skips search — it only selects `OLLAMA_MODEL_FAST`. Search failures surface an explicit error instead of hallucinating from parametric memory.
- **Search:** Pluggable `services/search/` with **SearxNG primary** + DDGS failover, ranking/dedupe/rumour filtering, numbered dated source blocks, citations persisted on `Prompt.citations` and emitted as `{"v":1,"type":"citations",...}` after stream end.
- **Models:** Role-scoped `OLLAMA_MODEL_THINKING` / `_FAST` / `_UTILITY` / `OLLAMA_EMBED_MODEL=nomic-embed-text`, configurable `num_ctx`, real model name on `PromptMetric`, `reindex_embeddings` management command + loud embedding-dimension mismatch.

## SearxNG (ops)
See README **SearxNG** section. Short version: run `searxng/searxng` on the GPU host, enable `json` in `settings.yml`, set `SEARXNG_BASE_URL=http://10.0.0.128:8080` in prod/beta secrets, open `:8080` on the LAN firewall like Ollama.

## Test plan
- [x] `SKIP_RAG_INIT=1 python manage.py test chat_backend.tests` — 442 OK (6 skipped)
- [ ] Deploy beta with updated secrets (`OLLAMA_MODEL_*`, `OLLAMA_EMBED_MODEL=nomic-embed-text`, `SEARXNG_BASE_URL`)
- [ ] After embed change: `python manage.py reindex_embeddings`
- [ ] Verify `did Taylor Swift get married` in FAST and THINKING returns grounded answer with citations frame
- [ ] Kill SearxNG and confirm factual turns return search_unavailable (not Joe Alwyn hallucination); non-factual chat still worksReviewed-on: #65

llm_be/chat_backend/consumers_graph.py

@@ -11,7 +11,6 @@ from asgiref.sync import sync_to_async
 from channels.generic.websocket import AsyncWebsocketConsumer
 from channels.db import database_sync_to_async
 from langchain_core.messages import HumanMessage, AIMessage, BaseMessage
-from langchain_community.tools import DuckDuckGoSearchRun
 from langgraph.graph import StateGraph, END
 
 from .models import Conversation, Prompt, PromptMetric, DocumentWorkspace, CustomUser
@@ -30,6 +29,8 @@ from .services.title_generator import title_generator
 from .services.moderation_classifier import moderation_classifier, ModerationLabel
 from .services.prompt_classifier.prompt_classifier import PromptClassifier, PromptType
 from .services.data_analysis_service import AsyncDataAnalysisService
+from .services.grounded_chat import citations_frame, prepare_grounded_chat
+from chat_backend.ollama_config import ollama_model_for_role, resolve_chat_role
 from .utils import (
     TokenUsageCollector,
     aiter_text_chunks,
@@ -43,7 +44,6 @@ from finance.services.quotas import FeatureNotAllowed, QuotaExceeded, check_gene
 logger = logging.getLogger(__name__)
 
 CHANNEL_NAME: str = "llm_messages"
-MODEL_NAME: str = "llama3.2"
 PROMPT_CLASSIFIER = PromptClassifier()
 
 # --- Database Helpers (Reused) ---
@@ -146,7 +146,7 @@ def get_messages(conversation_id, prompt, file_string: str = None, file_type: st
     return transformed_messages, prompt_instance
 
 @database_sync_to_async
-def save_generated_message(conversation_id, message):
+def save_generated_message(conversation_id, message, citations=None):
     conversation = Conversation.objects.get(id=conversation_id)
     serializer = PromptSerializer(
         data={
@@ -159,6 +159,8 @@ def save_generated_message(conversation_id, message):
         prompt_instance = serializer.save()
         prompt_instance.conversation_id = conversation.id
         prompt_instance.save()
+        if citations is not None:
+            Prompt.objects.filter(pk=prompt_instance.pk).update(citations=citations)
     else:
         print(serializer.errors)
 
@@ -221,6 +223,8 @@ class ChatState(TypedDict):
     error: Union[str, None]
     model_name: str
     chat_user: Any
+    citations: List[Dict[str, Any]]
+    resolved_model: str
 
 
 # --- LangGraph Nodes ---
@@ -277,22 +281,7 @@ async def generation_node(state: ChatState) -> ChatState:
                 }
         return {"response_generator": {"type": "text", "content": "Image Generation is not supported at this time, but it will be soon."}}
 
-    # Feature Flag: Internet Access
-    if prompt_type == PromptType.SEARCH:
-         # Check modelName first - if FAST, we skip search regardless of settings
-         if state.get("model_name") == "FAST":
-             pass
-         elif getattr(settings, "ALLOW_INTERNET_ACCESS", False):
-             try:
-                 search = DuckDuckGoSearchRun()
-                 search_results = search.run(state["message"])
-                 messages.append(HumanMessage(content=f"Search Results: {search_results}"))
-             except Exception as e:
-                 logger.error(f"Search failed: {e}")
-                 pass
-         else:
-             pass
-
+    # Feature Flag: Internet Access / always-on grounding handled below for chat.
     if prompt_type == PromptType.RAG:
         chat_user = state.get("chat_user")
         if chat_user is not None:
@@ -318,10 +307,26 @@ async def generation_node(state: ChatState) -> ChatState:
         generator = service.generate_response(prompt_instance.message, decoded_file, file_type)
         return {"response_generator": generator}
     
-    else: # GENERAL_CHAT or others
-        service = AsyncLLMService()
-        generator = service.generate_response(messages, prompt_instance.message, conversation_id)
-        return {"response_generator": generator}
+    else:
+        # GENERAL_CHAT / SEARCH / UNKNOWN — always-on grounding (#62).
+        # FAST selects a smaller model; it no longer skips search.
+        grounded = await prepare_grounded_chat(
+            message=state["message"],
+            messages=messages,
+            model_name=state.get("model_name"),
+            conversation_id=conversation_id,
+        )
+        if grounded.error:
+            return {
+                "response_generator": grounded.error,
+                "citations": [],
+                "resolved_model": grounded.model_name or "",
+            }
+        return {
+            "response_generator": grounded.generator,
+            "citations": grounded.citations,
+            "resolved_model": grounded.model_name,
+        }
 
 
 # --- LangGraph Definition ---
@@ -475,12 +480,13 @@ class ChatConsumerGraph(AsyncWebsocketConsumer):
                 if not decoded_file:
                     decoded_file, file_type = await get_conversation_file_async(conversation_id)
 
+                resolved_model = ollama_model_for_role(resolve_chat_role(model))
                 prompt_metric = await create_prompt_metric(
                     prompt_instance.id,
                     prompt_instance.message,
                     True if file else False,
                     file_type,
-                    MODEL_NAME,
+                    resolved_model,
                     conversation_id,
                 )
 
@@ -498,6 +504,8 @@ class ChatConsumerGraph(AsyncWebsocketConsumer):
                     "error": None,
                     "model_name": model,
                     "chat_user": chat_user,
+                    "citations": [],
+                    "resolved_model": resolved_model,
                 }
                 print("Initial State: ", initial_state)
 
@@ -535,8 +543,21 @@ class ChatConsumerGraph(AsyncWebsocketConsumer):
                     tokens_in, tokens_out = usage.pair
 
                 await self.send("END_OF_THE_STREAM_ENDER_GAME_42")
-                
-                await save_generated_message(conversation_id, full_response)
+
+                citations = final_state.get("citations") or []
+                if citations:
+                    await self.send_json_message(json.dumps(citations_frame(citations)))
+
+                final_model = final_state.get("resolved_model") or resolved_model
+                if final_model and final_model != prompt_metric.model_name:
+                    prompt_metric.model_name = final_model
+                    await database_sync_to_async(prompt_metric.save)(
+                        update_fields=["model_name"]
+                    )
+
+                await save_generated_message(
+                    conversation_id, full_response, citations=citations
+                )
                 await finish_prompt_metric(
                     prompt_metric,
                     len(full_response),