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

rem/services/phoenix/client.py

@@ -53,7 +53,7 @@ from datetime import datetime
 from pathlib import Path
 from typing import Any, Callable, TYPE_CHECKING, cast
 
-import pandas as pd
+import polars as pl
 from loguru import logger
 
 from .config import PhoenixConfig
@@ -64,6 +64,95 @@ if TYPE_CHECKING:
     from phoenix.client.resources.experiments.types import RanExperiment
 
 
+def dataframe_to_phoenix_dataset(
+    client: "PhoenixClient",
+    df: pl.DataFrame,
+    dataset_name: str,
+    input_keys: list[str] | None = None,
+    output_keys: list[str] | None = None,
+    metadata_keys: list[str] | None = None,
+    description: str | None = None,
+) -> "Dataset":
+    """Convert a Polars DataFrame to a Phoenix Dataset.
+
+    This function transforms a Polars DataFrame into a Phoenix Dataset by:
+    1. Extracting input columns (what agents receive)
+    2. Extracting output columns (ground truth/expected output)
+    3. Extracting metadata columns (optional labels, difficulty, etc.)
+
+    If column keys are not specified, uses smart defaults:
+    - input_keys: columns containing 'input', 'query', 'question', or 'prompt'
+    - output_keys: columns containing 'output', 'expected', 'answer', or 'response'
+    - metadata_keys: remaining columns
+
+    Args:
+        client: PhoenixClient instance
+        df: Polars DataFrame with experiment data
+        dataset_name: Name for the created Phoenix dataset
+        input_keys: Optional list of column names for inputs
+        output_keys: Optional list of column names for outputs (ground truth)
+        metadata_keys: Optional list of column names for metadata
+        description: Optional dataset description
+
+    Returns:
+        Phoenix Dataset instance
+
+    Example:
+        >>> df = pl.read_csv("golden_set.csv")
+        >>> dataset = dataframe_to_phoenix_dataset(
+        ...     client=phoenix_client,
+        ...     df=df,
+        ...     dataset_name="my-golden-set",
+        ...     input_keys=["query"],
+        ...     output_keys=["expected_output"],
+        ...     metadata_keys=["difficulty"]
+        ... )
+    """
+    columns = df.columns
+
+    # Smart defaults for column detection
+    if input_keys is None:
+        input_keys = [c for c in columns if any(
+            k in c.lower() for k in ["input", "query", "question", "prompt"]
+        )]
+        if not input_keys:
+            # Fallback: first column
+            input_keys = [columns[0]] if columns else []
+
+    if output_keys is None:
+        output_keys = [c for c in columns if any(
+            k in c.lower() for k in ["output", "expected", "answer", "response", "reference"]
+        )]
+        if not output_keys:
+            # Fallback: second column
+            output_keys = [columns[1]] if len(columns) > 1 else []
+
+    if metadata_keys is None:
+        used_keys = set(input_keys) | set(output_keys)
+        metadata_keys = [c for c in columns if c not in used_keys]
+
+    logger.debug(
+        f"DataFrame to Phoenix Dataset: inputs={input_keys}, "
+        f"outputs={output_keys}, metadata={metadata_keys}"
+    )
+
+    # Convert to list of dicts
+    records = df.to_dicts()
+
+    inputs = [{k: row.get(k) for k in input_keys} for row in records]
+    outputs = [{k: row.get(k) for k in output_keys} for row in records]
+    metadata = [{k: row.get(k) for k in metadata_keys} for row in records] if metadata_keys else None
+
+    # Create Phoenix dataset
+    return client.create_dataset_from_data(
+        name=dataset_name,
+        inputs=inputs,
+        outputs=outputs,
+        metadata=metadata,
+        description=description,
+    )
+
+
 class PhoenixClient:
     """High-level Phoenix client for REM evaluation workflows.
 
@@ -260,19 +349,22 @@ class PhoenixClient:
             "SEARCH semantic AI engineer",sarah-chen,person,medium,SEARCH
         """
         try:
-            # Load CSV
-            df = pd.read_csv(csv_file_path)
+            # Load CSV with Polars
+            df = pl.read_csv(csv_file_path)
+
+            # Convert to list of dicts
+            records = df.to_dicts()
 
             # Extract inputs
-            inputs = cast(list[dict[str, Any]], df[input_keys].to_dict("records"))
+            inputs = [{k: row.get(k) for k in input_keys} for row in records]
 
             # Extract outputs
-            outputs = cast(list[dict[str, Any]], df[output_keys].to_dict("records"))
+            outputs = [{k: row.get(k) for k in output_keys} for row in records]
 
             # Extract metadata if specified
             metadata = None
             if metadata_keys:
-                metadata = cast(list[dict[str, Any]], df[metadata_keys].to_dict("records"))
+                metadata = [{k: row.get(k) for k in metadata_keys} for row in records]
 
             return self.create_dataset_from_data(
                 name=name,
@@ -331,13 +423,16 @@ class PhoenixClient:
 
     def run_experiment(
         self,
-        dataset: "Dataset" | str,
+        dataset: "Dataset" | str | pl.DataFrame,
         task: Callable[[Any], Any] | None = None,
         evaluators: list[Callable[[Any], Any]] | None = None,
         experiment_name: str | None = None,
         experiment_description: str | None = None,
         experiment_metadata: dict[str, Any] | None = None,
         experiment_config: Any | None = None,
+        input_keys: list[str] | None = None,
+        output_keys: list[str] | None = None,
+        metadata_keys: list[str] | None = None,
     ) -> "RanExperiment":
         """Run an evaluation experiment.
 
@@ -346,14 +441,22 @@ class PhoenixClient:
         2. Agent run: Provide task function to execute agents on dataset
         3. Evaluator run: Provide evaluators to score existing outputs
 
+        Dataset can be:
+        - Phoenix Dataset instance
+        - Dataset name (string) - will be loaded from Phoenix
+        - Polars DataFrame - will be converted to Phoenix Dataset
+
         Args:
-            dataset: Dataset instance or name (required unless experiment_config provided)
+            dataset: Dataset instance, name, or Polars DataFrame
             task: Optional task function to run on each example (agent execution)
             evaluators: Optional list of evaluator functions
             experiment_name: Optional experiment name
             experiment_description: Optional description
             experiment_metadata: Optional metadata dict
             experiment_config: Optional ExperimentConfig instance (overrides other params)
+            input_keys: Column names for inputs (required if dataset is DataFrame)
+            output_keys: Column names for outputs (required if dataset is DataFrame)
+            metadata_keys: Optional column names for metadata
 
         Returns:
             RanExperiment with results
@@ -369,6 +472,16 @@ class PhoenixClient:
             ...     experiment_name="rem-v1-baseline"
             ... )
 
+        Example - With Polars DataFrame:
+            >>> df = pl.read_csv("golden_set.csv")
+            >>> experiment = client.run_experiment(
+            ...     dataset=df,
+            ...     task=run_agent,
+            ...     experiment_name="rem-v1-baseline",
+            ...     input_keys=["query"],
+            ...     output_keys=["expected_output"]
+            ... )
+
         Example - Evaluator Run (Phase 2b):
             >>> experiment = client.run_experiment(
             ...     dataset=agent_results,
@@ -407,6 +520,21 @@ class PhoenixClient:
                     else:
                         dataset = dataset_ref.path
 
+            # Convert Polars DataFrame to Phoenix Dataset
+            if isinstance(dataset, pl.DataFrame):
+                dataset_name_for_phoenix = f"{experiment_name or 'experiment'}-dataset-{datetime.now().strftime('%Y%m%d-%H%M%S')}"
+                logger.info(f"Converting Polars DataFrame to Phoenix Dataset: {dataset_name_for_phoenix}")
+                dataset = dataframe_to_phoenix_dataset(
+                    client=self,
+                    df=dataset,
+                    dataset_name=dataset_name_for_phoenix,
+                    input_keys=input_keys,
+                    output_keys=output_keys,
+                    metadata_keys=metadata_keys,
+                    description=f"Auto-created from DataFrame for experiment: {experiment_name}",
+                )
+                logger.info(f"✓ Created Phoenix Dataset: {dataset_name_for_phoenix}")
+
             # Load dataset if name provided
             if isinstance(dataset, str):
                 dataset = self.get_dataset(dataset)
@@ -454,7 +582,7 @@ class PhoenixClient:
         root_spans_only: bool = True,
         trace_id: str | None = None,
         span_id: str | None = None,
-    ) -> pd.DataFrame:
+    ) -> pl.DataFrame:
         """Query traces from Phoenix.
 
         Args:
@@ -467,7 +595,7 @@ class PhoenixClient:
             span_id: Filter by specific span ID
 
         Returns:
-            DataFrame with trace data
+            Polars DataFrame with trace data
 
         Example:
             >>> traces = client.get_traces(
@@ -492,8 +620,11 @@ class PhoenixClient:
             if span_id:
                 query_params["span_id"] = span_id
 
-            # Query traces
-            traces_df = self._client.query_spans(limit=limit, **query_params)  # type: ignore[attr-defined]
+            # Query traces (Phoenix returns pandas DataFrame)
+            pandas_df = self._client.query_spans(limit=limit, **query_params)  # type: ignore[attr-defined]
+
+            # Convert pandas to Polars
+            traces_df = pl.from_pandas(pandas_df)
 
             logger.debug(f"Retrieved {len(traces_df)} traces")
             return traces_df
@@ -535,7 +666,7 @@ class PhoenixClient:
             ... )
         """
         try:
-            # Query traces
+            # Query traces (returns Polars DataFrame)
             traces_df = self.get_traces(
                 project_name=project_name,
                 start_time=start_time,
@@ -547,12 +678,15 @@ class PhoenixClient:
             if len(traces_df) == 0:
                 raise ValueError("No traces found matching criteria")
 
+            # Convert to list of dicts for iteration
+            records = traces_df.to_dicts()
+
             # Extract inputs and outputs from traces
             inputs = []
             outputs = []
             metadata = []
 
-            for _, row in traces_df.iterrows():
+            for row in records:
                 # Extract input
                 span_input = row.get("attributes.input")
                 if span_input:
@@ -658,29 +792,169 @@ class PhoenixClient:
         label: str | None = None,
         score: float | None = None,
         explanation: str | None = None,
-    ) -> None:
-        """Add feedback annotation to a span.
+        metadata: dict[str, Any] | None = None,
+        trace_id: str | None = None,
+    ) -> str | None:
+        """Add feedback annotation to a span via Phoenix REST API.
+
+        Uses direct HTTP POST to /v1/span_annotations for reliability
+        (Phoenix Python client API changes frequently).
 
         Args:
-            span_id: Span ID to annotate
-            annotation_name: Name of the annotation (e.g., "correctness")
+            span_id: Span ID to annotate (hex string)
+            annotation_name: Name of the annotation (e.g., "correctness", "user_feedback")
             annotator_kind: Type of annotator ("HUMAN", "LLM", "CODE")
-            label: Optional label (e.g., "correct", "incorrect")
+            label: Optional label (e.g., "correct", "incorrect", "helpful")
             score: Optional numeric score (0.0-1.0)
             explanation: Optional explanation text
+            metadata: Optional additional metadata dict
+            trace_id: Optional trace ID (used if span lookup needed)
+
+        Returns:
+            Annotation ID if successful, None otherwise
         """
+        import httpx
+
         try:
-            self._client.add_span_annotation(  # type: ignore[attr-defined]
-                span_id=span_id,
-                name=annotation_name,
-                annotator_kind=annotator_kind,
-                label=label,
-                score=score,
-                explanation=explanation,
-            )
+            # Build annotation payload for Phoenix REST API
+            annotation_data = {
+                "span_id": span_id,
+                "name": annotation_name,
+                "annotator_kind": annotator_kind,
+                "result": {
+                    "label": label,
+                    "score": score,
+                    "explanation": explanation,
+                },
+                "metadata": metadata or {},
+            }
+
+            # Add trace_id if provided
+            if trace_id:
+                annotation_data["trace_id"] = trace_id
+
+            # POST to Phoenix REST API
+            annotations_endpoint = f"{self.config.base_url}/v1/span_annotations"
+            headers = {}
+            if self.config.api_key:
+                headers["Authorization"] = f"Bearer {self.config.api_key}"
+
+            with httpx.Client(timeout=5.0) as client:
+                response = client.post(
+                    annotations_endpoint,
+                    json={"data": [annotation_data]},
+                    headers=headers,
+                )
+                response.raise_for_status()
 
             logger.info(f"Added {annotator_kind} feedback to span {span_id}")
+            return span_id  # Return span_id as annotation reference
 
+        except httpx.HTTPStatusError as e:
+            logger.error(
+                f"Failed to add span feedback (HTTP {e.response.status_code}): "
+                f"{e.response.text if hasattr(e, 'response') else 'N/A'}"
+            )
+            return None
         except Exception as e:
             logger.error(f"Failed to add span feedback: {e}")
-            raise
+            return None
+
+    def sync_user_feedback(
+        self,
+        span_id: str,
+        rating: int | None = None,
+        categories: list[str] | None = None,
+        comment: str | None = None,
+        feedback_id: str | None = None,
+        trace_id: str | None = None,
+    ) -> str | None:
+        """Sync user feedback to Phoenix as a span annotation.
+
+        Convenience method for syncing Feedback entities to Phoenix.
+        Converts REM feedback format to Phoenix annotation format.
+
+        Args:
+            span_id: OTEL span ID to annotate
+            rating: User rating (-1, 1-5 scale)
+            categories: List of feedback categories
+            comment: Free-text comment
+            feedback_id: Optional REM feedback ID for reference
+            trace_id: Optional trace ID for the span
+
+        Returns:
+            Phoenix annotation ID if successful
+
+        Example:
+            >>> client.sync_user_feedback(
+            ...     span_id="abc123",
+            ...     rating=4,
+            ...     categories=["helpful", "accurate"],
+            ...     comment="Great response!"
+            ... )
+        """
+        # Convert rating to 0-1 score
+        # Rating scheme:
+        #   -1 = thumbs down → score 0.0
+        #    1 = thumbs up   → score 1.0
+        #  2-5 = star rating → normalized to 0-1 range
+        score = None
+        if rating is not None:
+            if rating == -1:
+                score = 0.0
+            elif rating == 1:
+                score = 1.0  # Thumbs up
+            elif 2 <= rating <= 5:
+                score = (rating - 1) / 4.0  # 2→0.25, 3→0.5, 4→0.75, 5→1.0
+
+        # Use primary category as label
+        label = categories[0] if categories else None
+
+        # Build explanation from comment and additional categories
+        explanation = comment
+        if categories and len(categories) > 1:
+            cats_str = ", ".join(categories[1:])
+            if explanation:
+                explanation = f"{explanation} [Categories: {cats_str}]"
+            else:
+                explanation = f"Categories: {cats_str}"
+
+        # Build metadata
+        metadata: dict[str, Any] = {
+            "rating": rating,
+            "categories": categories or [],
+        }
+        if feedback_id:
+            metadata["rem_feedback_id"] = feedback_id
+
+        return self.add_span_feedback(
+            span_id=span_id,
+            annotation_name="user_feedback",
+            annotator_kind="HUMAN",
+            label=label,
+            score=score,
+            explanation=explanation,
+            metadata=metadata,
+            trace_id=trace_id,
+        )
+
+    def get_span_annotations(
+        self,
+        span_id: str,
+        annotation_name: str | None = None,
+    ) -> list[dict[str, Any]]:
+        """Get annotations for a span.
+
+        Args:
+            span_id: Span ID to query
+            annotation_name: Optional filter by annotation name
+
+        Returns:
+            List of annotation dicts
+
+        TODO: Implement once Phoenix client exposes this method
+        """
+        # TODO: Phoenix client doesn't expose annotation query yet
+        # This is a stub for future implementation
+        logger.warning("get_span_annotations not yet implemented in Phoenix client")
+        return []

rem/services/postgres/README.md

@@ -348,8 +348,27 @@ results = await service.vector_search(
 
 ### Initialize Service
 
+There are two ways to initialize the PostgresService:
+
+**Option 1: Factory function (recommended for apps using remdb as a library)**
+
+```python
+from rem.services.postgres import get_postgres_service
+
+# Uses POSTGRES__CONNECTION_STRING from environment
+pg = get_postgres_service()
+if pg is None:
+    raise RuntimeError("Database not configured - set POSTGRES__CONNECTION_STRING")
+
+await pg.connect()
+# ... use pg ...
+await pg.disconnect()
+```
+
+**Option 2: Direct instantiation**
+
 ```python
-from rem.services.postgres import PostgresService, Repository
+from rem.services.postgres import PostgresService
 
 service = PostgresService(
     connection_string="postgresql://user:pass@localhost/remdb",
@@ -359,6 +378,9 @@ service = PostgresService(
 await service.connect()
 ```
 
+> **Note**: `get_postgres_service()` returns the service directly. It does NOT support
+> `async with` context manager syntax. Always call `connect()` and `disconnect()` explicitly.
+
 ### Using Repository Pattern
 
 **Generic Repository** for simple CRUD operations:
@@ -514,34 +536,156 @@ results = await service.vector_search(
 - HNSW parameters: `m=16, ef_construction=64` (tunable)
 - Monitor shared_buffers and work_mem
 
-## Migrations
+## Schema Management
 
-Run migrations in order:
+REM uses a **code-as-source-of-truth** approach. Pydantic models define the schema, and the database is kept in sync via diff-based migrations.
 
-```bash
-psql -d remdb -f sql/migrations/001_setup_extensions.sql
-psql -d remdb -f sql/migrations/002_kv_store_cache.sql
-psql -d remdb -f sql/generated_schema.sql
+### File Structure
+
+```
+src/rem/sql/
+├── migrations/
+│   ├── 001_install.sql          # Core infrastructure (manual)
+│   └── 002_install_models.sql   # Entity tables (auto-generated)
+└── background_indexes.sql       # HNSW vector indexes (optional)
 ```
 
-Background indexes (after data load):
+**Key principle**: Only two migration files. No incremental `003_`, `004_` files.
+
+### CLI Commands
 
 ```bash
-psql -d remdb -f sql/background_indexes.sql
+# Apply migrations (installs extensions, core tables, entity tables)
+rem db migrate
+
+# Check migration status
+rem db status
+
+# Generate schema SQL from models (for remdb development)
+rem db schema generate --models src/rem/models/entities
+
+# Validate models for schema generation
+rem db schema validate --models src/rem/models/entities
 ```
 
-## CLI Usage
+### Model Registry
 
-Generate schema from models:
+Models are discovered via the registry:
 
-```bash
-rem schema generate --models src/rem/models/entities --output sql/schema.sql
+```python
+import rem
+from rem.models.core import CoreModel
+
+@rem.register_model
+class MyEntity(CoreModel):
+    name: str
+    description: str  # Auto-embeds
+```
+
+## Using REM as a Library (Downstream Apps)
+
+When building an application that **depends on remdb as a package** (e.g., `pip install remdb`),
+there are important differences from developing remdb itself.
+
+### What Works Out of the Box
+
+1. **All core entity tables** - Resources, Messages, Users, Sessions, etc.
+2. **PostgresService** - Full database access via `get_postgres_service()`
+3. **Repository pattern** - CRUD operations for core entities
+4. **Migrations** - `rem db migrate` applies the bundled SQL files
+
+```python
+# In your downstream app (e.g., myapp/main.py)
+from rem.services.postgres import get_postgres_service
+from rem.models.entities import Message, Resource
+
+pg = get_postgres_service()
+await pg.connect()
+
+# Use core entities - tables already exist
+messages = await pg.query(Message, {"session_id": "abc"})
 ```
 
-Validate models:
+### Custom Models in Downstream Apps
+
+The `@rem.register_model` decorator registers models in the **runtime registry**, which is useful for:
+- Schema introspection at runtime
+- Future tooling that reads the registry
+
+However, **`rem db migrate` only applies SQL files bundled in the remdb package**.
+Custom models from downstream apps do NOT automatically get tables created.
+
+**Options for custom model tables:**
+
+**Option A: Use core entities with metadata**
+
+Store custom data in the `metadata` JSONB field of existing entities:
+
+```python
+resource = Resource(
+    name="my-custom-thing",
+    content="...",
+    metadata={"custom_field": "value", "another": 123}
+)
+```
+
+**Option B: Create tables manually**
+
+Write and apply your own SQL:
+
+```sql
+-- myapp/sql/custom_tables.sql
+CREATE TABLE IF NOT EXISTS conversation_summaries (
+    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    session_ref TEXT NOT NULL,
+    summary TEXT NOT NULL,
+    -- ... include CoreModel fields for compatibility
+    user_id VARCHAR(256),
+    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+    updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
+);
+```
 
 ```bash
-rem schema validate --models src/rem/models/entities
+psql $DATABASE_URL -f myapp/sql/custom_tables.sql
+```
+
+**Option C: Contribute upstream**
+
+If your model is generally useful, contribute it to remdb so it's included in
+the next release and `rem db migrate` creates it automatically.
+
+### Example: Downstream App Structure
+
+```
+myapp/
+├── main.py              # Import models, start API
+├── models/
+│   └── __init__.py      # @rem.register_model decorators
+├── sql/
+│   └── custom.sql       # Manual migrations for custom tables
+├── .env                 # POSTGRES__CONNECTION_STRING, LLM keys
+└── pyproject.toml       # dependencies = ["remdb>=0.3.110"]
+```
+
+```python
+# myapp/models/__init__.py
+import rem
+from rem.models.core import CoreModel
+
+@rem.register_model
+class ConversationSummary(CoreModel):
+    """Registered for introspection, but table created via sql/custom.sql"""
+    session_ref: str
+    summary: str
+```
+
+```python
+# myapp/main.py
+import models  # Registers custom models
+
+from rem.api.main import app  # Use REM's FastAPI app
+# Or build your own app using rem.services
 ```
 
 ## Configuration

rem/services/postgres/__init__.py

@@ -2,6 +2,7 @@
 PostgreSQL service for CloudNativePG database operations.
 """
 
+from .diff_service import DiffService, SchemaDiff
 from .repository import Repository
 from .service import PostgresService
 
@@ -20,4 +21,4 @@ def get_postgres_service() -> PostgresService | None:
     return PostgresService()
 
 
-__all__ = ["PostgresService", "get_postgres_service", "Repository"]
+__all__ = ["PostgresService", "get_postgres_service", "Repository", "DiffService", "SchemaDiff"]