remdb 0.3.171__py3-none-any.whl → 0.3.230__py3-none-any.whl

rem/cli/README.md

@@ -434,6 +434,68 @@ Ensure you're using the correct model format:
 - OpenAI: `openai:gpt-4o-mini`, `openai:gpt-4o`
 - Anthropic: `anthropic:claude-sonnet-4-5-20250929`
 
+## Data Visibility: PUBLIC vs PRIVATE
+
+**IMPORTANT: All ingested data is PUBLIC by default.** This is the correct behavior
+for shared knowledge bases (ontologies, procedures, reference data).
+
+### Why PUBLIC by Default?
+
+Most data in REM should be searchable by all users:
+- Clinical ontologies (disorders, symptoms, drugs)
+- Procedures and protocols (SCID-5, PHQ-9, etc.)
+- Reference documentation
+- Shared domain knowledge
+
+The `rem_lookup()` function searches for data where `user_id IS NULL`, which means
+public data. If you set `user_id` on data, it becomes invisible to other users.
+
+### Ingesting Public Data (Default)
+
+```bash
+# Standard ingestion - data is PUBLIC
+rem process ingest ontology/procedures/ --table ontologies
+
+# From S3 - also PUBLIC
+rem process ingest s3://bucket/docs/reference.pdf
+```
+
+### Ingesting Private Data (Rare)
+
+Private data requires explicit `--make-private` flag:
+
+```bash
+# Private user data - requires --make-private and --user-id
+rem process ingest personal-notes.md --make-private --user-id user-123
+```
+
+**When to use private data:**
+- User-uploaded personal documents
+- Session-specific content
+- User notes and annotations
+
+**NEVER use private data for:**
+- Ontologies and reference material
+- Clinical procedures and protocols
+- Shared knowledge bases
+- Anything that should be searchable by agents
+
+### Common Mistake
+
+If agents can't find data via `search_rem`, the most common cause is that the data
+was ingested with a `user_id` set. Check with:
+
+```sql
+SELECT name, user_id FROM ontologies WHERE name = 'phq-9-procedure';
+-- user_id should be NULL for public data
+```
+
+Fix by setting user_id to NULL:
+```sql
+UPDATE ontologies SET user_id = NULL WHERE user_id IS NOT NULL;
+UPDATE kv_store SET user_id = NULL WHERE entity_type = 'ontologies' AND user_id IS NOT NULL;
+```
+
 ## Next Steps
 
 1. **Implement Schema Registry**

rem/cli/commands/ask.py

@@ -71,16 +71,18 @@ async def run_agent_streaming(
     max_turns: int = 10,
     context: AgentContext | None = None,
     max_iterations: int | None = None,
+    user_message: str | None = None,
 ) -> None:
     """
-    Run agent in streaming mode using agent.iter() with usage limits.
+    Run agent in streaming mode using the SAME code path as the API.
 
-    Design Pattern (from carrier):
-    - Use agent.iter() for complete execution with tool call visibility
-    - run_stream() stops after first output, missing tool calls
-    - Stream tool call markers: [Calling: tool_name]
-    - Stream text content deltas as they arrive
-    - Show final structured result
+    This uses stream_openai_response_with_save from the API to ensure:
+    1. Tool calls are saved as separate "tool" messages (not embedded in content)
+    2. Assistant response is clean text only (no [Calling: ...] markers)
+    3. CLI testing is equivalent to API testing
+
+    The CLI displays tool calls as [Calling: tool_name] for visibility,
+    but these are NOT saved to the database.
 
     Args:
         agent: Pydantic AI agent
@@ -88,88 +90,66 @@ async def run_agent_streaming(
         max_turns: Maximum turns for agent execution (not used in current API)
         context: Optional AgentContext for session persistence
         max_iterations: Maximum iterations/requests (from agent schema or settings)
+        user_message: The user's original message (for database storage)
     """
-    from pydantic_ai import UsageLimits
-    from rem.utils.date_utils import to_iso_with_z, utc_now
+    import json
+    from rem.api.routers.chat.streaming import stream_openai_response_with_save, save_user_message
 
     logger.info("Running agent in streaming mode...")
 
     try:
-        # Import event types for streaming
-        from pydantic_ai import Agent as PydanticAgent
-        from pydantic_ai.messages import PartStartEvent, PartDeltaEvent, TextPartDelta, ToolCallPart
-
-        # Accumulate assistant response for session persistence
-        assistant_response_parts = []
-
-        # Use agent.iter() to get complete execution with tool calls
-        usage_limits = UsageLimits(request_limit=max_iterations) if max_iterations else None
-        async with agent.iter(prompt, usage_limits=usage_limits) as agent_run:
-            async for node in agent_run:
-                # Check if this is a model request node (includes tool calls and text)
-                if PydanticAgent.is_model_request_node(node):
-                    # Stream events from model request
-                    request_stream: Any
-                    async with node.stream(agent_run.ctx) as request_stream:
-                        async for event in request_stream:
-                            # Tool call start event
-                            if isinstance(event, PartStartEvent) and isinstance(
-                                event.part, ToolCallPart
-                            ):
-                                tool_marker = f"\n[Calling: {event.part.tool_name}]"
-                                print(tool_marker, flush=True)
-                                assistant_response_parts.append(tool_marker)
-
-                            # Text content delta
-                            elif isinstance(event, PartDeltaEvent) and isinstance(
-                                event.delta, TextPartDelta
-                            ):
-                                print(event.delta.content_delta, end="", flush=True)
-                                assistant_response_parts.append(event.delta.content_delta)
-
-        print("\n")  # Final newline after streaming
-
-        # Get final result from agent_run
-        result = agent_run.result
-        if hasattr(result, "output"):
-            logger.info("Final structured result:")
-            output = result.output
-            from rem.agentic.serialization import serialize_agent_result
-            output_json = json.dumps(serialize_agent_result(output), indent=2)
-            print(output_json)
-            assistant_response_parts.append(f"\n{output_json}")
-
-        # Save session messages (if session_id provided and postgres enabled)
-        if context and context.session_id and settings.postgres.enabled:
-            from ...services.session.compression import SessionMessageStore
-
-            # Extract just the user query from prompt
-            # Prompt format from ContextBuilder: system + history + user message
-            # We need to extract the last user message
-            user_message_content = prompt.split("\n\n")[-1] if "\n\n" in prompt else prompt
-
-            user_message = {
-                "role": "user",
-                "content": user_message_content,
-                "timestamp": to_iso_with_z(utc_now()),
-            }
-
-            assistant_message = {
-                "role": "assistant",
-                "content": "".join(assistant_response_parts),
-                "timestamp": to_iso_with_z(utc_now()),
-            }
-
-            # Store messages with compression
-            store = SessionMessageStore(user_id=context.user_id or settings.test.effective_user_id)
-            await store.store_session_messages(
+        # Save user message BEFORE streaming (same as API, using shared utility)
+        if context and context.session_id and user_message:
+            await save_user_message(
                 session_id=context.session_id,
-                messages=[user_message, assistant_message],
                 user_id=context.user_id,
-                compress=True,
+                content=user_message,
             )
 
-            logger.debug(f"Saved conversation to session {context.session_id}")
+        # Use the API streaming code path for consistency
+        # This properly handles tool calls and message persistence
+        model_name = getattr(agent, 'model', 'unknown')
+        if hasattr(model_name, 'model_name'):
+            model_name = model_name.model_name
+        elif hasattr(model_name, 'name'):
+            model_name = model_name.name
+        else:
+            model_name = str(model_name)
+
+        async for chunk in stream_openai_response_with_save(
+            agent=agent.agent if hasattr(agent, 'agent') else agent,
+            prompt=prompt,
+            model=model_name,
+            session_id=context.session_id if context else None,
+            user_id=context.user_id if context else None,
+            agent_context=context,
+        ):
+            # Parse SSE chunks for CLI display
+            if chunk.startswith("event: tool_call"):
+                # Extract tool call info from next data line
+                continue
+            elif chunk.startswith("data: ") and not chunk.startswith("data: [DONE]"):
+                try:
+                    data_str = chunk[6:].strip()
+                    if data_str:
+                        data = json.loads(data_str)
+                        # Check for tool_call event
+                        if data.get("type") == "tool_call":
+                            tool_name = data.get("tool_name", "tool")
+                            status = data.get("status", "")
+                            if status == "started":
+                                print(f"\n[Calling: {tool_name}]", flush=True)
+                        # Check for text content (OpenAI format)
+                        elif "choices" in data and data["choices"]:
+                            delta = data["choices"][0].get("delta", {})
+                            content = delta.get("content")
+                            if content:
+                                print(content, end="", flush=True)
+                except (json.JSONDecodeError, KeyError, IndexError):
+                    pass
+
+        print("\n")  # Final newline after streaming
+        logger.info("Final structured result:")
 
     except Exception as e:
         logger.error(f"Agent execution failed: {e}")
@@ -549,7 +529,7 @@ async def _ask_async(
 
     # Run agent with session persistence
     if stream:
-        await run_agent_streaming(agent, prompt, max_turns=max_turns, context=context)
+        await run_agent_streaming(agent, prompt, max_turns=max_turns, context=context, user_message=query)
     else:
         await run_agent_non_streaming(
             agent,

rem/cli/commands/db.py

@@ -333,64 +333,120 @@ def rebuild_cache(connection: str | None):
 
 @click.command()
 @click.argument("file_path", type=click.Path(exists=True, path_type=Path))
+@click.option("--table", "-t", default=None, help="Target table name (required for non-YAML formats)")
 @click.option("--user-id", default=None, help="User ID to scope data privately (default: public/shared)")
 @click.option("--dry-run", is_flag=True, help="Show what would be loaded without loading")
-def load(file_path: Path, user_id: str | None, dry_run: bool):
+def load(file_path: Path, table: str | None, user_id: str | None, dry_run: bool):
     """
-    Load data from YAML file into database.
+    Load data from file into database.
 
-    File format:
-        - table: resources
-          key_field: name
-          rows:
-            - name: Example
-              content: Test data...
+    Supports YAML with embedded metadata, or any tabular format via Polars
+    (jsonl, parquet, csv, json, arrow, etc.). For non-YAML formats, use --table.
 
     Examples:
-        rem db load rem/tests/data/graph_seed.yaml
-        rem db load data.yaml --user-id my-user  # Private to user
-        rem db load data.yaml --dry-run
+        rem db load data.yaml                        # YAML with metadata
+        rem db load data.jsonl -t resources          # Any Polars-supported format
     """
-    asyncio.run(_load_async(file_path, user_id, dry_run))
+    asyncio.run(_load_async(file_path, table, user_id, dry_run))
 
 
-async def _load_async(file_path: Path, user_id: str | None, dry_run: bool):
+def _load_dataframe_from_file(file_path: Path) -> "pl.DataFrame":
+    """Load any Polars-supported file format into a DataFrame."""
+    import polars as pl
+
+    suffix = file_path.suffix.lower()
+
+    if suffix in {".jsonl", ".ndjson"}:
+        return pl.read_ndjson(file_path)
+    elif suffix in {".parquet", ".pq"}:
+        return pl.read_parquet(file_path)
+    elif suffix == ".csv":
+        return pl.read_csv(file_path)
+    elif suffix == ".json":
+        return pl.read_json(file_path)
+    elif suffix in {".ipc", ".arrow"}:
+        return pl.read_ipc(file_path)
+    else:
+        raise ValueError(f"Unsupported file format: {suffix}. Use any Polars-supported format.")
+
+
+async def _load_async(file_path: Path, table: str | None, user_id: str | None, dry_run: bool):
     """Async implementation of load command."""
+    import polars as pl
     import yaml
     from ...models.core.inline_edge import InlineEdge
-    from ...models.entities import Resource, Moment, User, Message, SharedSession, Schema
+    from ...models.entities import SharedSession
     from ...services.postgres import get_postgres_service
+    from ...utils.model_helpers import get_table_name
+    from ... import get_model_registry
 
     logger.info(f"Loading data from: {file_path}")
     scope_msg = f"user: {user_id}" if user_id else "public"
     logger.info(f"Scope: {scope_msg}")
 
-    # Load YAML file
-    with open(file_path) as f:
-        data = yaml.safe_load(f)
-
-    if not isinstance(data, list):
-        logger.error("YAML must be a list of table definitions")
-        raise click.Abort()
-
-    if dry_run:
-        logger.info("DRY RUN - Would load:")
-        logger.info(yaml.dump(data, default_flow_style=False))
-        return
+    suffix = file_path.suffix.lower()
+    is_yaml = suffix in {".yaml", ".yml"}
 
-    # Map table names to model classes
-    # CoreModel subclasses use Repository.upsert()
+    # Build MODEL_MAP dynamically from registry
+    registry = get_model_registry()
+    registry.register_core_models()
     MODEL_MAP = {
-        "users": User,
-        "moments": Moment,
-        "resources": Resource,
-        "messages": Message,
-        "schemas": Schema,
+        get_table_name(model): model
+        for model in registry.get_model_classes().values()
     }
 
     # Non-CoreModel tables that need direct SQL insertion
     DIRECT_INSERT_TABLES = {"shared_sessions"}
 
+    # Parse file based on format
+    if is_yaml:
+        # YAML with embedded metadata
+        with open(file_path) as f:
+            data = yaml.safe_load(f)
+
+        if not isinstance(data, list):
+            logger.error("YAML must be a list of table definitions")
+            raise click.Abort()
+
+        if dry_run:
+            logger.info("DRY RUN - Would load:")
+            logger.info(yaml.dump(data, default_flow_style=False))
+            return
+
+        table_defs = data
+    else:
+        # Polars-supported format - require --table
+        if not table:
+            logger.error(f"For {suffix} files, --table is required. Example: rem db load {file_path.name} -t resources")
+            raise click.Abort()
+
+        try:
+            df = _load_dataframe_from_file(file_path)
+        except Exception as e:
+            logger.error(f"Failed to load file: {e}")
+            raise click.Abort()
+
+        rows = df.to_dicts()
+
+        if dry_run:
+            logger.info(f"DRY RUN - Would load {len(rows)} rows to table '{table}':")
+            logger.info(f"Columns: {list(df.columns)}")
+
+            # Validate first row against model if table is known
+            if table in MODEL_MAP and rows:
+                from ...utils.model_helpers import validate_data_for_model
+                result = validate_data_for_model(MODEL_MAP[table], rows[0])
+                if result.extra_fields:
+                    logger.warning(f"Unknown fields (ignored): {result.extra_fields}")
+                if result.valid:
+                    logger.success(f"Sample row validates OK. Required: {result.required_fields or '(none)'}")
+                else:
+                    result.log_errors("Sample row")
+            return
+
+        # Wrap as single table definition
+        table_defs = [{"table": table, "rows": rows}]
+
     # Connect to database
     pg = get_postgres_service()
     if not pg:
@@ -399,23 +455,23 @@ async def _load_async(file_path: Path, user_id: str | None, dry_run: bool):
 
     await pg.connect()
 
+    # Start embedding worker for generating embeddings
+    if pg.embedding_worker:
+        await pg.embedding_worker.start()
+
     try:
         total_loaded = 0
 
-        for table_def in data:
+        for table_def in table_defs:
             table_name = table_def["table"]
-            key_field = table_def.get("key_field", "id")
             rows = table_def.get("rows", [])
 
             # Handle direct insert tables (non-CoreModel)
             if table_name in DIRECT_INSERT_TABLES:
                 for row_data in rows:
-                    # Add tenant_id if not present
-                    if "tenant_id" not in row_data:
-                        row_data["tenant_id"] = "default"
+                    # tenant_id is optional - NULL means public/shared
 
                     if table_name == "shared_sessions":
-                        # Insert shared_session directly
                         await pg.fetch(
                             """INSERT INTO shared_sessions
                                (session_id, owner_user_id, shared_with_user_id, tenant_id)
@@ -424,7 +480,7 @@ async def _load_async(file_path: Path, user_id: str | None, dry_run: bool):
                             row_data["session_id"],
                             row_data["owner_user_id"],
                             row_data["shared_with_user_id"],
-                            row_data["tenant_id"],
+                            row_data.get("tenant_id"),  # Optional - NULL means public
                         )
                         total_loaded += 1
                         logger.success(f"Loaded shared_session: {row_data['owner_user_id']} -> {row_data['shared_with_user_id']}")
@@ -434,16 +490,11 @@ async def _load_async(file_path: Path, user_id: str | None, dry_run: bool):
                 logger.warning(f"Unknown table: {table_name}, skipping")
                 continue
 
-            model_class = MODEL_MAP[table_name]  # Type is inferred from MODEL_MAP
+            model_class = MODEL_MAP[table_name]
 
-            for row_data in rows:
-                # Add user_id and tenant_id only if explicitly provided
-                # Default is public (None) - data is shared/visible to all
-                # Pass --user-id to scope data privately to a specific user
-                if "user_id" not in row_data and user_id is not None:
-                    row_data["user_id"] = user_id
-                if "tenant_id" not in row_data and user_id is not None:
-                    row_data["tenant_id"] = row_data.get("user_id", user_id)
+            for row_idx, row_data in enumerate(rows):
+                # tenant_id and user_id are optional - NULL means public/shared data
+                # Data files can explicitly set tenant_id/user_id if needed
 
                 # Convert graph_edges to InlineEdge format if present
                 if "graph_edges" in row_data:
@@ -452,30 +503,40 @@ async def _load_async(file_path: Path, user_id: str | None, dry_run: bool):
                         for edge in row_data["graph_edges"]
                     ]
 
-                # Convert any ISO timestamp strings with Z suffix to naive datetime
-                # This handles fields like starts_timestamp, ends_timestamp, etc.
+                # Convert ISO timestamp strings
                 from ...utils.date_utils import parse_iso
                 for key, value in list(row_data.items()):
                     if isinstance(value, str) and (key.endswith("_timestamp") or key.endswith("_at")):
                         try:
                             row_data[key] = parse_iso(value)
                         except (ValueError, TypeError):
-                            pass  # Not a valid datetime string, leave as-is
+                            pass
 
-                # Create model instance and upsert via repository
                 from ...services.postgres.repository import Repository
+                from ...utils.model_helpers import validate_data_for_model
 
-                instance = model_class(**row_data)
-                repo = Repository(model_class, table_name, pg)  # Type inferred from MODEL_MAP
-                await repo.upsert(instance)  # type: ignore[arg-type]
+                result = validate_data_for_model(model_class, row_data)
+                if not result.valid:
+                    result.log_errors(f"Row {row_idx + 1} ({table_name})")
+                    raise click.Abort()
+
+                repo = Repository(model_class, table_name, pg)
+                await repo.upsert(result.instance)  # type: ignore[arg-type]
                 total_loaded += 1
 
-                # Log based on model type
-                name = getattr(instance, 'name', getattr(instance, 'id', '?'))
+                name = getattr(result.instance, 'name', getattr(result.instance, 'id', '?'))
                 logger.success(f"Loaded {table_name[:-1]}: {name}")
 
         logger.success(f"Data loaded successfully! Total rows: {total_loaded}")
 
+        # Wait for embeddings to complete
+        if pg.embedding_worker and pg.embedding_worker.running:
+            queue_size = pg.embedding_worker.task_queue.qsize()
+            if queue_size > 0:
+                logger.info(f"Waiting for {queue_size} embeddings to complete...")
+            await pg.embedding_worker.stop()
+            logger.success("Embeddings generated successfully")
+
     finally:
         await pg.disconnect()
 
@@ -580,7 +641,7 @@ async def _diff_async(
 
         if not result.has_changes:
             click.secho("✓ No schema drift detected", fg="green")
-            click.echo("  Database matches Pydantic models")
+            click.echo("  Database matches source (tables, functions, triggers, views)")
             if result.filtered_count > 0:
                 click.echo()
                 click.secho(f"  ({result.filtered_count} destructive change(s) hidden by '{strategy}' strategy)", fg="yellow")
@@ -592,17 +653,34 @@ async def _diff_async(
         if result.filtered_count > 0:
             click.secho(f"   ({result.filtered_count} destructive change(s) hidden by '{strategy}' strategy)", fg="yellow")
         click.echo()
-        click.echo("Changes:")
-        for line in result.summary:
-            if line.startswith("+"):
-                click.secho(f"  {line}", fg="green")
-            elif line.startswith("-"):
-                click.secho(f"  {line}", fg="red")
-            elif line.startswith("~"):
-                click.secho(f"  {line}", fg="yellow")
-            else:
-                click.echo(f"  {line}")
-        click.echo()
+
+        # Table/column changes (Alembic)
+        if result.summary:
+            click.echo("Table Changes:")
+            for line in result.summary:
+                if line.startswith("+"):
+                    click.secho(f"  {line}", fg="green")
+                elif line.startswith("-"):
+                    click.secho(f"  {line}", fg="red")
+                elif line.startswith("~"):
+                    click.secho(f"  {line}", fg="yellow")
+                else:
+                    click.echo(f"  {line}")
+            click.echo()
+
+        # Programmable object changes (functions, triggers, views)
+        if result.programmable_summary:
+            click.echo("Programmable Objects (functions/triggers/views):")
+            for line in result.programmable_summary:
+                if line.startswith("+"):
+                    click.secho(f"  {line}", fg="green")
+                elif line.startswith("-"):
+                    click.secho(f"  {line}", fg="red")
+                elif line.startswith("~"):
+                    click.secho(f"  {line}", fg="yellow")
+                else:
+                    click.echo(f"  {line}")
+            click.echo()
 
         # Generate migration if requested
         if generate: