backend/core/geo_engine.py

import json
import logging
import math
import os
import tempfile
import threading
from typing import Dict, Any, Optional, List

import duckdb

from backend.core.data_catalog import get_data_catalog, describe_table

logger = logging.getLogger(__name__)


class ResultTooLargeError(Exception):
    """
    Raised when a spatial result has more geometries than can be rendered.

    Carries the counts so the caller can tell the agent exactly how far over it
    is, which is what lets it rewrite the query rather than guess.
    """

    def __init__(self, row_count: int, limit: int):
        self.row_count = row_count
        self.limit = limit
        super().__init__(
            f"Query matched {row_count:,} geometries, above the {limit:,} that can be "
            f"rendered. Aggregate (GROUP BY a region, or average per hexagon across "
            f"time) or filter (one season, one week, one country) and try again."
        )


class GeoEngine:
    """
    In-memory DuckDB Spatial engine.

    Catalog tables are exposed as lazily-created VIEWs over their source files
    (zero-copy, so startup stays fast and memory use stays low), while layers
    produced by the agent are materialized as real tables so they can be
    referenced by follow-up spatial operations.
    """

    _instance = None

    # Results are NOT truncated — a query returns every row it matched.
    #
    # There is still a physical limit: each row becomes a Python dict, then JSON,
    # then an SVG path in the browser, so a few hundred thousand geometries will
    # stall the page. Rather than silently dropping rows (which quietly makes an
    # answer wrong), a result above this size is refused with an explanation, and
    # the agent responds by aggregating or filtering — the query gets better
    # instead of the answer getting quietly worse. Non-spatial results are exempt:
    # they are cheap and never rendered as geometry.
    RENDER_REFUSAL_THRESHOLD = 400_000

    # Column names recognized as the geometry column in a result set.
    GEOMETRY_COLUMNS = ('geom', 'geometry')

    def __new__(cls):
        if cls._instance is None:
            cls._instance = super(GeoEngine, cls).__new__(cls)
            cls._instance.initialized = False
        return cls._instance

    def __init__(self):
        if self.initialized:
            return

        logger.info("Initializing GeoEngine (DuckDB)...")
        try:
            self.con = duckdb.connect(database=':memory:')
            self.con.install_extension('spatial')
            self.con.load_extension('spatial')
            logger.info("GeoEngine initialized with Spatial extension.")
        except Exception as e:
            logger.error(f"Failed to initialize GeoEngine: {e}")
            raise

        self.layers: Dict[str, str] = {}  # layer_id -> table_name
        self.catalog = get_data_catalog()

        # A single DuckDB connection is shared process-wide, but it is NOT safe to
        # use from several threads at once. The agent runs tool calls concurrently
        # and sub-agents run in parallel, so every statement goes through this
        # lock. Without it, concurrent DESCRIBE/SELECT calls deadlock and the
        # request hangs with no error. Re-entrant because several methods here
        # call one another (ensure_table_loaded -> is_table_loaded).
        self._lock = threading.RLock()

        self.initialized = True

    def fetch_all(self, sql: str) -> List[tuple]:
        """Run a read query and return all rows, holding the connection lock."""
        with self._lock:
            return self.con.execute(sql).fetchall()

    def fetch_one(self, sql: str) -> Optional[tuple]:
        """Run a read query and return the first row, holding the connection lock."""
        with self._lock:
            return self.con.execute(sql).fetchone()

    def describe_columns(self, table_name: str) -> List[tuple]:
        """Column metadata for a table, holding the connection lock."""
        with self._lock:
            return self.con.execute(f'DESCRIBE "{table_name}"').fetchall()

    def is_table_loaded(self, table_name: str) -> bool:
        """Check whether a table or view already exists in this connection."""
        with self._lock:
            try:
                self.con.execute(f'DESCRIBE "{table_name}"')
                return True
            except duckdb.Error:
                return False

    def ensure_table_loaded(self, table_name: str) -> bool:
        """
        Ensure a catalog table is available in DuckDB, creating a view over its
        source file if needed. Returns True if the table is queryable.
        """
        if self.is_table_loaded(table_name):
            return True

        file_path = self.catalog.get_file_path(table_name)
        if not file_path or not file_path.exists():
            logger.warning(f"Table {table_name} not found in catalog or file missing.")
            return False

        metadata = self.catalog.get_table_metadata(table_name)
        if not metadata:
            logger.warning(f"No metadata found for {table_name}")
            return False

        try:
            logger.info(f"Lazy loading table: {table_name}")


            # DuckDB's spatial extension reads GeoParquet geometry metadata
            # directly, so both Parquet variants use the same reader.
            if file_path.suffix == '.parquet' or metadata.get('format') in ('parquet', 'geoparquet'):
                reader = f"read_parquet('{file_path}')"
            else:
                reader = f"ST_Read('{file_path}')"

            with self._lock:
                self.con.execute(f'CREATE OR REPLACE VIEW "{table_name}" AS SELECT * FROM {reader}')
            return True
        except Exception as e:
            logger.error(f"Failed to load {table_name}: {e}")
            return False

    def get_table_schemas(self) -> str:
        """Get the schema of every currently loaded table, for LLM context."""
        result = "Currently Loaded Tables:\n\n"

        try:
            with self._lock:
                tables = self.con.execute("SHOW TABLES").fetchall()
        except duckdb.Error as e:
            logger.error(f"Error listing tables: {e}")
            return result

        for (table_name,) in tables:
            result += self._describe_table(table_name)

        return result

    def get_table_schemas_for_tables(self, table_names: List[str]) -> str:
        """
        Get the schema for specific tables, including the catalog's semantic
        description so the LLM knows what each table actually contains.
        """
        if not table_names:
            return "No tables available in the current scope.\n"

        result = "Available Tables:\n\n"
        for table_name in table_names:
            result += self._describe_table(table_name, include_description=True)
        return result

    def _describe_table(self, table_name: str, include_description: bool = False) -> str:
        """Render one table's schema as LLM context. Returns "" if unavailable."""
        try:
            with self._lock:
                columns = self.con.execute(f'DESCRIBE "{table_name}"').fetchall()
                row_count = self.con.execute(f'SELECT COUNT(*) FROM "{table_name}"').fetchone()[0]
        except duckdb.Error:
            return ""  # Not loaded yet

        out = f"### {table_name} ({row_count} rows)\n"

        if include_description:
            meta = self.catalog.get_table_metadata(table_name)
            if meta:
                desc = describe_table(meta)
                if desc and desc != "No description":
                    out += f"**Description**: {desc}\n"

        out += "Columns:\n"
        for col_name, col_type, *_ in columns:
            if col_name in self.GEOMETRY_COLUMNS:
                out += f"  - {col_name}: GEOMETRY (spatial data)\n"
            else:
                out += f"  - {col_name}: {col_type}\n"
        return out + "\n"

    def register_layer(self, layer_id: str, geojson: Dict[str, Any]) -> str:
        """
        Register a GeoJSON FeatureCollection as a queryable DuckDB table so the
        agent can reference it in later spatial operations. Returns the table name.
        """
        table_name = f"layer_{layer_id.replace('-', '_')}"

        def json_serial(obj):
            if hasattr(obj, 'isoformat'):
                return obj.isoformat()
            raise TypeError(f"Type {type(obj)} not serializable")

        tmp_path = None
        try:
            with self._lock:
                self.con.execute(f'DROP TABLE IF EXISTS "{table_name}"')

            # ST_Read needs a file, so round-trip through a temp file rather than
            # hand-unpacking nested GeoJSON into columns.
            with tempfile.NamedTemporaryFile(mode='w', suffix='.json', delete=False) as tmp:
                json.dump(geojson, tmp, default=json_serial)
                tmp_path = tmp.name

            with self._lock:
                self.con.execute(f'CREATE TABLE "{table_name}" AS SELECT * FROM ST_Read(\'{tmp_path}\')')

            self.layers[layer_id] = table_name
            logger.info(f"Registered layer {layer_id} as table {table_name}")
            return table_name

        except Exception as e:
            logger.error(f"Error registering layer {layer_id}: {e}")
            raise
        finally:
            if tmp_path and os.path.exists(tmp_path):
                os.unlink(tmp_path)

    def execute_spatial_query(self, sql: str) -> Dict[str, Any]:
        """
        Execute a SELECT and return the result as a GeoJSON FeatureCollection.

        Handles both spatial results (a geometry column is present) and plain
        aggregates (no geometry, features carry properties only). Every matching
        row is returned; nothing is truncated.
        """
        logger.info(f"Executing Spatial SQL: {sql}")

        # Held across the whole method: the result is staged in a temp table with
        # a fixed name, so two concurrent queries would overwrite each other's
        # results and return the wrong rows.
        with self._lock:
            return self._execute_spatial_query_locked(sql)

    def _execute_spatial_query_locked(self, sql: str) -> Dict[str, Any]:

        # DuckDB Spatial has no FeatureCollection writer, so the result is
        # staged in a temp table and assembled row by row below.
        self.con.execute(f"CREATE OR REPLACE TEMP TABLE query_result AS {sql}")

        columns = self.con.execute("DESCRIBE query_result").fetchall()
        col_names = [c[0] for c in columns]
        geom_col = next((c for c in col_names if c in self.GEOMETRY_COLUMNS), None)

        total_count = self.con.execute("SELECT COUNT(*) FROM query_result").fetchone()[0]
        properties: Dict[str, Any] = {"row_count": total_count}

        # Refuse before materializing: DuckDB holds the result cheaply, and the
        # expensive part is building hundreds of thousands of Python dicts. The
        # caller turns this into actionable guidance for the agent.
        if geom_col is not None and total_count > self.RENDER_REFUSAL_THRESHOLD:
            raise ResultTooLargeError(total_count, self.RENDER_REFUSAL_THRESHOLD)

        if geom_col is None:
            # Non-spatial result (e.g. a COUNT or GROUP BY aggregate).
            rows = self.con.execute("SELECT * FROM query_result").fetchall()

            features = [
                {
                    "type": "Feature",
                    "geometry": None,
                    "properties": {
                        name: self._coerce_value(value)
                        for name, value in zip(col_names, row)
                    },
                }
                for row in rows
            ]
            return {"type": "FeatureCollection", "features": features, "properties": properties}

        other_cols = [c for c in col_names if c != geom_col]
        # Quote identifiers so columns like OSM's "generator:source" survive.
        select_clause = ", ".join(
            [f'ST_AsGeoJSON("{geom_col}")'] + [f'"{c}"' for c in other_cols]
        )
        rows = self.con.execute(f"SELECT {select_clause} FROM query_result").fetchall()

        features = []
        for row in rows:
            features.append({
                "type": "Feature",
                "geometry": json.loads(row[0]) if row[0] else None,
                "properties": {
                    name: self._coerce_value(value)
                    for name, value in zip(other_cols, row[1:])
                },
            })

        return {"type": "FeatureCollection", "features": features, "properties": properties}

    @staticmethod
    def _coerce_value(value: Any) -> Any:
        """Make a DuckDB value JSON-serializable (Parquet/WKB columns yield bytes)."""
        if isinstance(value, bytes):
            try:
                return value.decode('utf-8')
            except UnicodeDecodeError:
                return str(value)
        # NaN/Infinity are legal floats but illegal JSON. They arise routinely
        # from analytics (AVG of an empty group, area of a degenerate geometry),
        # and emitting them produces a payload that the API and the browser both
        # reject with an opaque parse error. null is the honest encoding.
        if isinstance(value, float) and (math.isnan(value) or math.isinf(value)):
            return None
        return value

    def get_table_name(self, layer_id: str) -> Optional[str]:
        return self.layers.get(layer_id)

    def get_user_layers(self) -> List[str]:
        """Return the table names of all agent- and user-registered layers."""
        return list(self.layers.values())

    def drop_layer(self, layer_id: str) -> bool:
        """
        Drop a registered layer from DuckDB and internal tracking.
        Returns True if it was found and dropped.
        """
        table_name = self.layers.get(layer_id)
        if not table_name:
            logger.warning(f"Layer {layer_id} not found in GeoEngine")
            return False

        try:
            with self._lock:
                self.con.execute(f'DROP TABLE IF EXISTS "{table_name}"')
            del self.layers[layer_id]
            logger.info(f"Dropped layer {layer_id} (table: {table_name})")
            return True
        except duckdb.Error as e:
            logger.error(f"Failed to drop layer {layer_id}: {e}")
            return False


_geo_engine = None


def get_geo_engine() -> GeoEngine:
    global _geo_engine
    if _geo_engine is None:
        _geo_engine = GeoEngine()
    return _geo_engine