remdb 0.3.0__py3-none-any.whl → 0.3.127__py3-none-any.whl

rem/services/postgres/repository.py

@@ -335,3 +335,135 @@ class Repository(Generic[T]):
             row = await conn.fetchrow(sql, *params)
 
         return row[0] if row else 0
+
+    async def find_paginated(
+        self,
+        filters: dict[str, Any],
+        page: int = 1,
+        page_size: int = 50,
+        order_by: str = "created_at DESC",
+        partition_by: str | None = None,
+    ) -> dict[str, Any]:
+        """
+        Find records with page-based pagination using CTE with ROW_NUMBER().
+
+        Uses a CTE with ROW_NUMBER() OVER (PARTITION BY ... ORDER BY ...) for
+        efficient pagination with total count in a single query.
+
+        Args:
+            filters: Dict of field -> value filters (AND-ed together)
+            page: Page number (1-indexed)
+            page_size: Number of records per page
+            order_by: ORDER BY clause for row numbering (default: "created_at DESC")
+            partition_by: Optional field to partition by (e.g., "user_id").
+                         If None, uses global row numbering.
+
+        Returns:
+            Dict containing:
+            - data: List of model instances for the page
+            - total: Total count of records matching filters
+            - page: Current page number
+            - page_size: Records per page
+            - total_pages: Total number of pages
+            - has_next: Whether there are more pages
+            - has_previous: Whether there are previous pages
+
+        Example:
+            result = await repo.find_paginated(
+                {"tenant_id": "acme", "user_id": "alice"},
+                page=2,
+                page_size=20,
+                order_by="created_at DESC",
+                partition_by="user_id"
+            )
+            # result = {
+            #     "data": [...],
+            #     "total": 150,
+            #     "page": 2,
+            #     "page_size": 20,
+            #     "total_pages": 8,
+            #     "has_next": True,
+            #     "has_previous": True
+            # }
+        """
+        if not settings.postgres.enabled or not self.db:
+            logger.debug(f"Postgres disabled, returning empty {self.model_class.__name__} pagination")
+            return {
+                "data": [],
+                "total": 0,
+                "page": page,
+                "page_size": page_size,
+                "total_pages": 0,
+                "has_next": False,
+                "has_previous": False,
+            }
+
+        # Ensure connection
+        if not self.db.pool:
+            await self.db.connect()
+
+        # Type guard: ensure pool is not None after connect
+        if not self.db.pool:
+            raise RuntimeError("Failed to establish database connection")
+
+        # Build WHERE clause from filters
+        where_conditions = ["deleted_at IS NULL"]
+        params: list[Any] = []
+        param_idx = 1
+
+        for field, value in filters.items():
+            where_conditions.append(f"{field} = ${param_idx}")
+            params.append(value)
+            param_idx += 1
+
+        where_clause = " AND ".join(where_conditions)
+
+        # Build PARTITION BY clause
+        partition_clause = f"PARTITION BY {partition_by}" if partition_by else ""
+
+        # Build the CTE query with ROW_NUMBER() and COUNT() window functions
+        # This gives us pagination + total count in a single query
+        sql = f"""
+        WITH numbered AS (
+            SELECT *,
+                   ROW_NUMBER() OVER ({partition_clause} ORDER BY {order_by}) as _row_num,
+                   COUNT(*) OVER ({partition_clause}) as _total_count
+            FROM {self.table_name}
+            WHERE {where_clause}
+        )
+        SELECT * FROM numbered
+        WHERE _row_num > ${param_idx} AND _row_num <= ${param_idx + 1}
+        ORDER BY _row_num
+        """
+
+        # Calculate row range for the page
+        start_row = (page - 1) * page_size
+        end_row = page * page_size
+        params.extend([start_row, end_row])
+
+        async with self.db.pool.acquire() as conn:
+            rows = await conn.fetch(sql, *params)
+
+        # Extract total from first row (all rows have the same _total_count)
+        total = rows[0]["_total_count"] if rows else 0
+
+        # Remove internal columns and convert to models
+        data = []
+        for row in rows:
+            row_dict = dict(row)
+            row_dict.pop("_row_num", None)
+            row_dict.pop("_total_count", None)
+            data.append(self.model_class.model_validate(row_dict))
+
+        # Calculate pagination metadata
+        total_pages = (total + page_size - 1) // page_size if total > 0 else 0
+
+        return {
+            "data": data,
+            "total": total,
+            "page": page,
+            "page_size": page_size,
+            "total_pages": total_pages,
+            "has_next": page < total_pages,
+            "has_previous": page > 1,
+        }

rem/services/postgres/schema_generator.py

@@ -1,18 +1,28 @@
 """
 Schema generation utility from Pydantic models.
 
-Scans a directory of Pydantic models and generates complete database schemas including:
+Generates complete database schemas from:
+1. REM's core models (Resource, Moment, User, etc.)
+2. Models registered via rem.register_model() or rem.register_models()
+3. Models discovered from a directory scan
+
+Output includes:
 - Primary tables
 - Embeddings tables
 - KV_STORE triggers
 - Indexes (foreground and background)
 - Migrations
+- Schema table entries (for agent-like table access)
 
 Usage:
     from rem.services.postgres.schema_generator import SchemaGenerator
 
+    # Generate from registry (includes core + registered models)
     generator = SchemaGenerator()
-    schema = generator.generate_from_directory("src/rem/models/entities")
+    schema = await generator.generate_from_registry()
+
+    # Or generate from directory (legacy)
+    schema = await generator.generate_from_directory("src/rem/models/entities")
 
     # Write to file
     with open("src/rem/sql/schema.sql", "w") as f:
@@ -21,14 +31,192 @@ Usage:
 
 import importlib.util
 import inspect
+import json
+import uuid
 from pathlib import Path
-from typing import Type
+from typing import Any, Type
 
 from loguru import logger
 from pydantic import BaseModel
 
 from ...settings import settings
-from .register_type import register_type
+from ...utils.sql_paths import get_package_sql_dir
+from .register_type import register_type, should_embed_field
+
+# Namespace UUID for generating deterministic UUIDs from model names
+# Using UUID5 with this namespace ensures same model always gets same UUID
+REM_SCHEMA_NAMESPACE = uuid.UUID("6ba7b810-9dad-11d1-80b4-00c04fd430c8")  # DNS namespace
+
+
+def generate_model_uuid(fully_qualified_name: str) -> uuid.UUID:
+    """
+    Generate deterministic UUID from fully qualified model name.
+
+    Uses UUID5 (SHA-1 hash) with REM namespace for reproducibility.
+    Same fully qualified name always produces same UUID.
+
+    Args:
+        fully_qualified_name: Full module path, e.g., "rem.models.entities.Resource"
+
+    Returns:
+        Deterministic UUID for this model
+    """
+    return uuid.uuid5(REM_SCHEMA_NAMESPACE, fully_qualified_name)
+
+
+def extract_model_schema_metadata(
+    model: Type[BaseModel],
+    table_name: str,
+    entity_key_field: str,
+    include_search_tool: bool = True,
+) -> dict[str, Any]:
+    """
+    Extract schema metadata from a Pydantic model for schemas table.
+
+    Args:
+        model: Pydantic model class
+        table_name: Database table name
+        entity_key_field: Field used as entity key in kv_store
+        include_search_tool: If True, add search_rem tool for querying this table
+
+    Returns:
+        Dict with schema metadata ready for schemas table insert
+    """
+    # Get fully qualified name
+    fqn = f"{model.__module__}.{model.__name__}"
+
+    # Generate deterministic UUID
+    schema_id = generate_model_uuid(fqn)
+
+    # Get JSON schema from Pydantic
+    json_schema = model.model_json_schema()
+
+    # Find embedding fields
+    embedding_fields = []
+    for field_name, field_info in model.model_fields.items():
+        if should_embed_field(field_name, field_info):
+            embedding_fields.append(field_name)
+
+    # Build description with search capability note
+    base_description = model.__doc__ or f"Schema for {model.__name__}"
+    search_note = (
+        f"\n\nThis agent can search the `{table_name}` table using the `search_rem` tool. "
+        f"Use REM query syntax: LOOKUP for exact match, FUZZY for typo-tolerant search, "
+        f"SEARCH for semantic similarity, or SQL for complex queries."
+    ) if include_search_tool else ""
+
+    # Build spec with table metadata and tools
+    # Note: default_search_table is used by create_agent to append a description
+    # suffix to the search_rem tool when loading it dynamically
+    has_embeddings = bool(embedding_fields)
+
+    spec = {
+        "type": "object",
+        "description": base_description + search_note,
+        "properties": json_schema.get("properties", {}),
+        "required": json_schema.get("required", []),
+        "json_schema_extra": {
+            "table_name": table_name,
+            "entity_key_field": entity_key_field,
+            "embedding_fields": embedding_fields,
+            "fully_qualified_name": fqn,
+            "tools": ["search_rem"] if include_search_tool else [],
+            "default_search_table": table_name,
+            "has_embeddings": has_embeddings,
+        },
+    }
+
+    # Build content (documentation)
+    content = f"""# {model.__name__}
+
+{base_description}
+
+## Overview
+
+The `{model.__name__}` entity is stored in the `{table_name}` table. Each record is uniquely
+identified by its `{entity_key_field}` field for lookups and graph traversal.
+
+## Search Capabilities
+
+This schema includes the `search_rem` tool which supports:
+- **LOOKUP**: O(1) exact match by {entity_key_field} (e.g., `LOOKUP "entity-name"`)
+- **FUZZY**: Typo-tolerant search (e.g., `FUZZY "partial" THRESHOLD 0.3`)
+- **SEARCH**: Semantic vector search on {', '.join(embedding_fields) if embedding_fields else 'content'} (e.g., `SEARCH "concept" FROM {table_name} LIMIT 10`)
+- **SQL**: Complex queries (e.g., `SELECT * FROM {table_name} WHERE ...`)
+
+## Table Info
+
+| Property | Value |
+|----------|-------|
+| Table | `{table_name}` |
+| Entity Key | `{entity_key_field}` |
+| Embedding Fields | {', '.join(f'`{f}`' for f in embedding_fields) if embedding_fields else 'None'} |
+| Tools | {', '.join(['`search_rem`'] if include_search_tool else ['None'])} |
+
+## Fields
+
+"""
+    for field_name, field_info in model.model_fields.items():
+        field_type = str(field_info.annotation) if field_info.annotation else "Any"
+        field_desc = field_info.description or ""
+        required = "Required" if field_info.is_required() else "Optional"
+        content += f"### `{field_name}`\n"
+        content += f"- **Type**: `{field_type}`\n"
+        content += f"- **{required}**\n"
+        if field_desc:
+            content += f"- {field_desc}\n"
+        content += "\n"
+
+    return {
+        "id": str(schema_id),
+        "name": model.__name__,
+        "table_name": table_name,
+        "entity_key_field": entity_key_field,
+        "embedding_fields": embedding_fields,
+        "fqn": fqn,
+        "spec": spec,
+        "content": content,
+        "category": "entity",
+    }
+
+
+def generate_schema_upsert_sql(schema_metadata: dict[str, Any]) -> str:
+    """
+    Generate SQL UPSERT statement for schemas table.
+
+    Uses ON CONFLICT DO UPDATE for idempotency.
+
+    Args:
+        schema_metadata: Dict from extract_model_schema_metadata()
+
+    Returns:
+        SQL INSERT ... ON CONFLICT statement
+    """
+    # Escape single quotes in content and spec
+    content_escaped = schema_metadata["content"].replace("'", "''")
+    spec_json = json.dumps(schema_metadata["spec"]).replace("'", "''")
+
+    sql = f"""
+-- Schema entry for {schema_metadata['name']} ({schema_metadata['table_name']})
+INSERT INTO schemas (id, tenant_id, name, content, spec, category, metadata)
+VALUES (
+    '{schema_metadata['id']}'::uuid,
+    'system',
+    '{schema_metadata['name']}',
+    '{content_escaped}',
+    '{spec_json}'::jsonb,
+    'entity',
+    '{{"table_name": "{schema_metadata['table_name']}", "entity_key_field": "{schema_metadata['entity_key_field']}", "embedding_fields": {json.dumps(schema_metadata['embedding_fields'])}, "fqn": "{schema_metadata['fqn']}"}}'::jsonb
+)
+ON CONFLICT (id) DO UPDATE SET
+    name = EXCLUDED.name,
+    content = EXCLUDED.content,
+    spec = EXCLUDED.spec,
+    category = EXCLUDED.category,
+    metadata = EXCLUDED.metadata,
+    updated_at = CURRENT_TIMESTAMP;
+"""
+    return sql.strip()
 
 
 class SchemaGenerator:
@@ -47,9 +235,9 @@ class SchemaGenerator:
         Initialize schema generator.
 
         Args:
-            output_dir: Optional directory for output files (defaults to settings.sql_dir)
+            output_dir: Optional directory for output files (defaults to package sql dir)
         """
-        self.output_dir = output_dir or Path(settings.sql_dir)
+        self.output_dir = output_dir or get_package_sql_dir()
         self.schemas: dict[str, dict] = {}
 
     def discover_models(self, directory: str | Path) -> dict[str, Type[BaseModel]]:
@@ -225,15 +413,76 @@ class SchemaGenerator:
             create_kv_trigger=True,
         )
 
+        # Extract schema metadata for schemas table entry
+        schema_metadata = extract_model_schema_metadata(
+            model=model,
+            table_name=table_name,
+            entity_key_field=entity_key_field,
+        )
+        schema["schema_metadata"] = schema_metadata
+
         self.schemas[table_name] = schema
         return schema
 
+    async def generate_from_registry(
+        self, output_file: str | None = None, include_core: bool = True
+    ) -> str:
+        """
+        Generate complete schema from the model registry.
+
+        Includes:
+        1. REM's core models (if include_core=True)
+        2. Models registered via rem.register_model() or rem.register_models()
+
+        Args:
+            output_file: Optional output file path (relative to output_dir)
+            include_core: If True, include REM's core models (default: True)
+
+        Returns:
+            Complete SQL schema as string
+
+        Example:
+            import rem
+            from rem.models.core import CoreModel
+
+            # Register custom model
+            @rem.register_model
+            class CustomEntity(CoreModel):
+                name: str
+
+            # Generate schema (includes core + custom)
+            generator = SchemaGenerator()
+            schema = await generator.generate_from_registry()
+        """
+        from ...registry import get_model_registry
+
+        registry = get_model_registry()
+        models = registry.get_models(include_core=include_core)
+
+        logger.info(f"Generating schema from registry: {len(models)} models")
+
+        # Generate schemas for each model
+        for model_name, ext in models.items():
+            await self.generate_schema_for_model(
+                ext.model,
+                table_name=ext.table_name,
+                entity_key_field=ext.entity_key_field,
+            )
+
+        return self._generate_sql_output(
+            source="model registry",
+            output_file=output_file,
+        )
+
     async def generate_from_directory(
         self, directory: str | Path, output_file: str | None = None
     ) -> str:
         """
         Generate complete schema from all models in a directory.
 
+        Note: For most use cases, prefer generate_from_registry() which uses
+        the model registry pattern.
+
         Args:
             directory: Path to directory with Pydantic models
             output_file: Optional output file path (relative to output_dir)
@@ -248,12 +497,31 @@ class SchemaGenerator:
         for model_name, model in models.items():
             await self.generate_schema_for_model(model)
 
-        # Combine into single SQL file
+        return self._generate_sql_output(
+            source=f"directory: {directory}",
+            output_file=output_file,
+        )
+
+    def _generate_sql_output(
+        self, source: str, output_file: str | None = None
+    ) -> str:
+        """
+        Generate SQL output from accumulated schemas.
+
+        Args:
+            source: Description of schema source (for header comment)
+            output_file: Optional output file path (relative to output_dir)
+
+        Returns:
+            Complete SQL schema as string
+        """
+        import datetime
+
         sql_parts = [
             "-- REM Model Schema (install_models.sql)",
             "-- Generated from Pydantic models",
-            f"-- Source directory: {directory}",
-            "-- Generated at: " + __import__("datetime").datetime.now().isoformat(),
+            f"-- Source: {source}",
+            f"-- Generated at: {datetime.datetime.now().isoformat()}",
             "--",
             "-- DO NOT EDIT MANUALLY - Regenerate with: rem db schema generate",
             "--",
@@ -262,6 +530,7 @@ class SchemaGenerator:
             "-- 2. Embeddings tables (embeddings_<table>)",
             "-- 3. KV_STORE triggers for cache maintenance",
             "-- 4. Indexes (foreground only, background indexes separate)",
+            "-- 5. Schema table entries (for agent-like table access)",
             "",
             "-- ============================================================================",
             "-- PREREQUISITES CHECK",
@@ -307,6 +576,19 @@ class SchemaGenerator:
                 sql_parts.append(schema["sql"]["kv_trigger"])
                 sql_parts.append("")
 
+        # Add schema table entries (every entity table is also an "agent")
+        sql_parts.append("-- ============================================================================")
+        sql_parts.append("-- SCHEMA TABLE ENTRIES")
+        sql_parts.append("-- Every entity table gets a schemas entry for agent-like access")
+        sql_parts.append("-- ============================================================================")
+        sql_parts.append("")
+
+        for table_name, schema in self.schemas.items():
+            if "schema_metadata" in schema:
+                schema_upsert = generate_schema_upsert_sql(schema["schema_metadata"])
+                sql_parts.append(schema_upsert)
+                sql_parts.append("")
+
         # Add migration record
         sql_parts.append("-- ============================================================================")
         sql_parts.append("-- RECORD MIGRATION")

rem/services/postgres/service.py

@@ -190,19 +190,19 @@ class PostgresService:
 
     async def connect(self) -> None:
         """Establish database connection pool."""
-        logger.info(f"Connecting to PostgreSQL with pool size {self.pool_size}")
+        logger.debug(f"Connecting to PostgreSQL with pool size {self.pool_size}")
         self.pool = await asyncpg.create_pool(
             self.connection_string,
             min_size=1,
             max_size=self.pool_size,
             init=self._init_connection,  # Configure JSONB codec on each connection
         )
-        logger.info("PostgreSQL connection pool established")
+        logger.debug("PostgreSQL connection pool established")
 
         # Start embedding worker if available
         if self.embedding_worker and hasattr(self.embedding_worker, "start"):
             await self.embedding_worker.start()
-            logger.info("Embedding worker started")
+            logger.debug("Embedding worker started")
 
     async def disconnect(self) -> None:
         """Close database connection pool."""
@@ -211,10 +211,10 @@ class PostgresService:
         # The worker will be stopped explicitly when the application shuts down
 
         if self.pool:
-            logger.info("Closing PostgreSQL connection pool")
+            logger.debug("Closing PostgreSQL connection pool")
             await self.pool.close()
             self.pool = None
-            logger.info("PostgreSQL connection pool closed")
+            logger.debug("PostgreSQL connection pool closed")
 
     async def execute(
         self,
@@ -631,7 +631,7 @@ class PostgresService:
         table_name: str,
         embedding: list[float],
         limit: int = 10,
-        min_similarity: float = 0.7,
+        min_similarity: float = 0.3,
         tenant_id: Optional[str] = None,
     ) -> list[dict[str, Any]]:
         """

rem/services/rate_limit.py

@@ -0,0 +1,113 @@
+"""
+Rate Limit Service - Postgres-backed rate limiting.
+
+Implements tenant-aware, tiered rate limiting using PostgreSQL UNLOGGED tables
+for high performance. Supports monthly quotas and short-term burst limits.
+"""
+
+import random
+from datetime import datetime, timezone
+from enum import Enum
+from typing import Optional
+
+from loguru import logger
+
+from ..models.entities.user import UserTier
+from .postgres.service import PostgresService
+
+
+class RateLimitService:
+    """
+    Service for tracking and enforcing API rate limits.
+    
+    Uses an UNLOGGED table `rate_limits` for performance.
+    Note: Counts in UNLOGGED tables may be lost on database crash/restart.
+    """
+
+    def __init__(self, db: PostgresService):
+        self.db = db
+        
+        # Rate limits configuration
+        # Format: (limit, period_seconds)
+        # This is a simple implementation. In production, move to settings.
+        self.TIER_CONFIG = {
+            UserTier.ANONYMOUS: {"limit": 1000, "period": 3600},  # 1000/hour (for testing)
+            UserTier.FREE: {"limit": 50, "period": 2592000},      # 50/month (~30 days)
+            UserTier.BASIC: {"limit": 10000, "period": 2592000},  # 10k/month
+            UserTier.PRO: {"limit": 100000, "period": 2592000},   # 100k/month
+        }
+
+    async def check_rate_limit(
+        self, 
+        tenant_id: str, 
+        identifier: str, 
+        tier: UserTier
+    ) -> tuple[bool, int, int]:
+        """
+        Check if request is allowed under the rate limit.
+
+        Args:
+            tenant_id: Tenant identifier
+            identifier: User ID or Anonymous ID
+            tier: User subscription tier
+
+        Returns:
+            Tuple (is_allowed, current_count, limit)
+        """
+        config = self.TIER_CONFIG.get(tier, self.TIER_CONFIG[UserTier.FREE])
+        limit = config["limit"]
+        period = config["period"]
+        
+        # Construct time-window key
+        now = datetime.now(timezone.utc)
+        
+        if period >= 2592000: # Monthly
+            time_key = now.strftime("%Y-%m") 
+        elif period >= 86400: # Daily
+            time_key = now.strftime("%Y-%m-%d")
+        elif period >= 3600: # Hourly
+            time_key = now.strftime("%Y-%m-%d-%H")
+        else: # Minute/Second (fallback)
+            time_key = int(now.timestamp() / period)
+
+        key = f"{tenant_id}:{identifier}:{tier.value}:{time_key}"
+        
+        # Calculate expiry (for cleanup)
+        expires_at = now.timestamp() + period
+
+        # Atomic UPSERT to increment counter
+        # Returns the new count
+        query = """
+            INSERT INTO rate_limits (key, count, expires_at)
+            VALUES ($1, 1, to_timestamp($2))
+            ON CONFLICT (key) DO UPDATE
+            SET count = rate_limits.count + 1
+            RETURNING count;
+        """
+        
+        try:
+            count = await self.db.fetchval(query, key, expires_at)
+        except Exception as e:
+            logger.error(f"Rate limit check failed: {e}")
+            # Fail open to avoid blocking users on DB error
+            return True, 0, limit
+
+        is_allowed = count <= limit
+        
+        # Probabilistic cleanup (1% chance)
+        if random.random() < 0.01:
+            await self.cleanup_expired()
+            
+        return is_allowed, count, limit
+
+    async def cleanup_expired(self):
+        """Remove expired rate limit keys."""
+        try:
+            # Use a small limit to avoid locking/long queries
+            query = """
+                DELETE FROM rate_limits 
+                WHERE expires_at < NOW()
+            """
+            await self.db.execute(query)
+        except Exception as e:
+            logger.warning(f"Rate limit cleanup failed: {e}")

rem/services/rem/README.md

@@ -302,3 +302,17 @@ See `tests/integration/test_rem_query_evolution.py` for stage-based validation a
 * **Unified View**: The underlying SQL function `rem_traverse` uses a view `all_graph_edges` that unions `graph_edges` from all entity tables (`resources`, `moments`, `users`, etc.). This enables polymorphic traversal without complex joins in the application layer.
 * **KV Store**: Edge destinations (`dst`) are resolved to entity IDs using the `kv_store`. This requires that all traversable entities have an entry in the `kv_store` (handled automatically by database triggers).
 * **Iterated Retrieval**: REM is architected for multi-turn retrieval where LLMs conduct conversational database exploration. Each query informs the next, enabling emergent information discovery without requiring upfront schema knowledge.
+
+## Scaling & Architectural Decisions
+
+### 1. Hybrid Adjacency List
+REM implements a **Hybrid Adjacency List** pattern to balance strict relational guarantees with graph flexibility:
+*   **Primary Storage (Source of Truth):** Standard PostgreSQL tables (`resources`, `moments`, etc.) enforce schema validation, constraints, and type safety.
+*   **Graph Overlay:** Relationships are stored as "inline edges" within a JSONB column (`graph_edges`) on each entity.
+*   **Performance Layer:** A denormalized `UNLOGGED` table (`kv_store`) acts as a high-speed cache, mapping human-readable keys to internal UUIDs and edges. This avoids the traditional "join bomb" of traversing normalized SQL tables while avoiding the operational complexity of a separate graph database (e.g., Neo4j).
+
+### 2. The Pareto Principle in Graph Algorithms
+We explicitly choose **Simplicity over Full-Scale Graph Analytics**.
+*   **Hypothesis:** For LLM Agent workloads, 80% of the value is derived from **local context retrieval** (1-3 hops via `LOOKUP` and `TRAVERSE`).
+*   **Diminishing Returns:** Global graph algorithms (PageRank, Community Detection) offer diminishing returns for real-time agentic retrieval tasks. Agents typically need to answer specific questions ("Who worked on file X?"), which is a local neighborhood problem, not a global cluster analysis problem.
+*   **Future Scaling:** If deeper analysis is needed, we prefer **Graph + Vector (RAG)** approaches (using semantic similarity to find implicit links) over complex explicit graph algorithms.