User memory preferences

.gitignore

@@ -196,4 +196,7 @@ dev-tools/mcp-mock-server/.certs/
 requirements.*.backup
 
 # Local run files
-local-run.yaml
+local-run.yaml
+
+datadir/
+INSTRUCTIONS.md

pyproject.toml

@@ -69,6 +69,7 @@ dependencies = [
     "pyasn1>=0.6.3", # LCORE-1490
     # Used for system prompt template variable rendering
     "jinja2>=3.1.0",
+    "anthropic>=0.86.0",
 ]
 
 

src/app/endpoints/query.py

@@ -65,6 +65,7 @@
     ShieldModerationResult,
     TurnSummary,
 )
+from utils.user_memory import build_instructions_with_preferences, user_memory
 from utils.vector_search import build_rag_context
 
 logger = get_logger(__name__)
@@ -183,6 +184,19 @@ async def query_endpoint_handler(
         inline_rag_context=inline_rag_context.context_text,
     )
 
+    # Extract user preferences from conversation history
+    user_preferences = await user_memory(
+        user_id,
+        client,
+        responses_params.model,
+        _skip_userid_check,
+        new_conversation=not query_request.conversation_id,
+    )
+    if user_preferences:
+        responses_params.instructions = build_instructions_with_preferences(
+            responses_params.instructions, user_preferences
+        )
+
     # Handle Azure token refresh if needed
     if (
         responses_params.model.startswith("azure")
@@ -293,6 +307,7 @@ async def retrieve_response(
             id=moderation_result.moderation_id, llm_response=moderation_result.message
         )
     try:
+        logger.info(responses_params.model_dump(exclude_none=True))
         response = await client.responses.create(
             **responses_params.model_dump(exclude_none=True)
         )

src/app/endpoints/streaming_query.py

@@ -1,4 +1,4 @@
-"""Streaming query handler using Responses API."""
+"""Streaming query handler using Responses API."""  # pylint: disable=too-many-lines
 
 import asyncio
 import datetime
@@ -95,6 +95,7 @@
 from utils.suid import get_suid, normalize_conversation_id
 from utils.token_counter import TokenCounter
 from utils.types import ReferencedDocument, ResponsesApiParams, TurnSummary
+from utils.user_memory import build_instructions_with_preferences, user_memory
 from utils.vector_search import build_rag_context
 
 logger = get_logger(__name__)
@@ -219,6 +220,19 @@ async def streaming_query_endpoint_handler(  # pylint: disable=too-many-locals
         inline_rag_context=inline_rag_context.context_text,
     )
 
+    # Extract user preferences from conversation history
+    user_preferences = await user_memory(
+        user_id,
+        client,
+        responses_params.model,
+        _skip_userid_check,
+        new_conversation=not query_request.conversation_id,
+    )
+    if user_preferences:
+        responses_params.instructions = build_instructions_with_preferences(
+            responses_params.instructions, user_preferences
+        )
+
     # Handle Azure token refresh if needed
     if (
         responses_params.model.startswith("azure")

src/models/database/conversations.py

@@ -39,6 +39,26 @@ class UserConversation(Base):  # pylint: disable=too-few-public-methods
     topic_summary: Mapped[str] = mapped_column(default="")
 
 
+class UserMemory(Base):  # pylint: disable=too-few-public-methods
+    """Model for storing cached user preference extractions."""
+
+    __tablename__ = "user_memory"
+
+    # One row per user
+    user_id: Mapped[str] = mapped_column(primary_key=True)
+
+    # The extracted preferences string (empty string if none found)
+    preferences: Mapped[str] = mapped_column(default="")
+
+    # Conversation count at the time of extraction (used for cache invalidation)
+    conversation_count: Mapped[int] = mapped_column(default=0)
+
+    updated_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True),
+        server_default=func.now(),  # pylint: disable=not-callable
+    )
+
+
 class UserTurn(Base):  # pylint: disable=too-few-public-methods
     """Model for storing turn-level metadata."""