Skip to content

BrainAPI is a knowledge graph–powered AI memory layer that transforms unstructured data into structured knowledge, enabling intelligent search, recommendations, and contextual memory for AI agents and applications.

35.5 KB
Raw Blame
"""
File: /data.py
Project: postgresql
Created Date: Sunday May 24th 2026
Author: Christian Nonis <[email protected]>
"""

from __future__ import annotations

from contextlib import contextmanager
import json
import threading
from typing import Any, Iterator, List, Literal, Optional, Tuple

import psycopg2
import psycopg2.extras

from src.adapters.interfaces.data import DataClient, SearchResult
from src.config import SEARCH_FTS_REGCONFIGS, config
from src.constants.data import (
    Brain,
    KGChanges,
    KGChangesType,
    Observation,
    StructuredData,
    TextChunk,
)

from ._naming import (
    brain_db_name,
    brain_id_from_db_name,
    is_internal_brain_db_suffix,
)
from ._provisioning import (
    borrow,
    ensure_brain_database,
    ensure_database_exists,
    get_brain_pool,
    get_system_pool,
    clear_brain_database,
    drop_brain_database,
    list_brain_database_names,
)


_SYSTEM_DDL = """
CREATE TABLE IF NOT EXISTS data_brains (
    id TEXT PRIMARY KEY,
    name_key TEXT UNIQUE NOT NULL,
    pat TEXT NOT NULL,
    document JSONB NOT NULL,
    created_at TIMESTAMPTZ NOT NULL DEFAULT NOW()
);
"""


_BRAIN_DDL = """
CREATE TABLE IF NOT EXISTS data_text_chunks (
    id TEXT PRIMARY KEY,
    text TEXT,
    metadata JSONB,
    inserted_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
    document JSONB NOT NULL
);
CREATE INDEX IF NOT EXISTS idx_data_text_chunks_inserted_at
    ON data_text_chunks (inserted_at DESC);

CREATE TABLE IF NOT EXISTS data_observations (
    id TEXT PRIMARY KEY,
    text TEXT,
    resource_id TEXT,
    metadata JSONB,
    inserted_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
    document JSONB NOT NULL
);
CREATE INDEX IF NOT EXISTS idx_data_observations_resource
    ON data_observations (resource_id);
CREATE INDEX IF NOT EXISTS idx_data_observations_labels
    ON data_observations USING gin ((metadata -> 'labels'));

CREATE TABLE IF NOT EXISTS data_structured_data (
    id TEXT PRIMARY KEY,
    data JSONB,
    types TEXT[],
    metadata JSONB,
    inserted_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
    document JSONB NOT NULL
);
CREATE INDEX IF NOT EXISTS idx_data_structured_data_types
    ON data_structured_data USING gin (types);
CREATE INDEX IF NOT EXISTS idx_data_structured_data_inserted_at
    ON data_structured_data (inserted_at DESC);

CREATE TABLE IF NOT EXISTS data_kg_changes (
    id TEXT PRIMARY KEY,
    type TEXT,
    change JSONB,
    timestamp TIMESTAMPTZ NOT NULL DEFAULT NOW(),
    document JSONB NOT NULL
);
CREATE INDEX IF NOT EXISTS idx_data_kg_changes_timestamp
    ON data_kg_changes (timestamp DESC);
"""

_SEARCH_DDL = """
ALTER TABLE data_text_chunks
    ADD COLUMN IF NOT EXISTS search_tsv tsvector
    GENERATED ALWAYS AS (to_tsvector('english', coalesce(text, ''))) STORED;
ALTER TABLE data_text_chunks
    ADD COLUMN IF NOT EXISTS search_len integer
    GENERATED ALWAYS AS (
        GREATEST(length(to_tsvector('english', coalesce(text, ''))), 0)
    ) STORED;
CREATE INDEX IF NOT EXISTS idx_data_text_chunks_search_tsv
    ON data_text_chunks USING gin (search_tsv);
"""


def _search_alt_ddl(regconfig: str) -> str:
    if regconfig not in SEARCH_FTS_REGCONFIGS:
        raise ValueError(f"Unsupported FTS regconfig {regconfig!r}")
    return f"""
ALTER TABLE data_text_chunks
    ADD COLUMN IF NOT EXISTS search_tsv_alt tsvector
    GENERATED ALWAYS AS (to_tsvector('{regconfig}', coalesce(text, ''))) STORED;
ALTER TABLE data_text_chunks
    ADD COLUMN IF NOT EXISTS search_len_alt integer
    GENERATED ALWAYS AS (
        GREATEST(length(to_tsvector('{regconfig}', coalesce(text, ''))), 0)
    ) STORED;
CREATE INDEX IF NOT EXISTS idx_data_text_chunks_search_tsv_alt
    ON data_text_chunks USING gin (search_tsv_alt);
"""


def _bm25_sql(
    regconfig: str,
    tsv_col: str,
    len_col: str,
    *,
    query_match: str | None = None,
) -> str:
    if tsv_col not in {"search_tsv", "search_tsv_alt"}:
        raise ValueError(f"Unsupported tsv column {tsv_col!r}")
    if len_col not in {"search_len", "search_len_alt"}:
        raise ValueError(f"Unsupported len column {len_col!r}")
    if regconfig not in SEARCH_FTS_REGCONFIGS and regconfig != "english":
        raise ValueError(f"Unsupported FTS regconfig {regconfig!r}")
    if query_match is None:
        query_match = "or" if tsv_col == "search_tsv_alt" else "and"
    if query_match not in {"and", "or"}:
        raise ValueError(f"Unsupported BM25 query_match {query_match!r}")
    if query_match == "and":
        match_sql = f"c.{tsv_col} @@ plainto_tsquery('{regconfig}', %s)"
    else:
        match_sql = "TRUE"
    return f"""
                    WITH query_lexemes AS (
                        SELECT unnest(tsvector_to_array(to_tsvector('{regconfig}', %s)))
                            AS lexeme
                    ),
                    coll AS (
                        SELECT
                            COUNT(*)::double precision AS n_docs,
                            COALESCE(
                                AVG(GREATEST({len_col}, 1)),
                                1
                            )::double precision AS avgdl
                        FROM data_text_chunks
                    ),
                    idf AS (
                        SELECT
                            s.word AS lexeme,
                            ln(
                                1.0 + (coll.n_docs - s.ndoc + 0.5)
                                / (s.ndoc + 0.5)
                            ) AS idf
                        FROM ts_stat(
                            'SELECT {tsv_col} FROM data_text_chunks'
                        ) AS s
                        CROSS JOIN coll
                        WHERE s.word IN (SELECT lexeme FROM query_lexemes)
                    ),
                    docs AS (
                        SELECT
                            c.id,
                            c.document,
                            c.text,
                            GREATEST(c.{len_col}, 1)::double precision AS dl,
                            lex.lexeme,
                            COALESCE(
                                array_length(lex.positions, 1),
                                1
                            )::double precision AS tf
                        FROM data_text_chunks c
                        CROSS JOIN LATERAL unnest(c.{tsv_col})
                            AS lex(lexeme, positions)
                        WHERE {match_sql}
                          AND lex.lexeme IN (SELECT lexeme FROM query_lexemes)
                    )
                    SELECT
                        d.id,
                        d.document,
                        d.text,
                        SUM(
                            COALESCE(i.idf, 0)
                            * (d.tf * (%s + 1.0))
                            / (
                                d.tf
                                + %s * (
                                    1.0 - %s
                                    + %s * d.dl
                                    / NULLIF(c.avgdl, 0)
                                )
                            )
                        ) AS bm25
                    FROM docs d
                    CROSS JOIN coll c
                    LEFT JOIN idf i ON i.lexeme = d.lexeme
                    GROUP BY d.id, d.document, d.text
                    ORDER BY bm25 DESC, d.id ASC
                    LIMIT %s
                    """


def _ilike_pattern(query_text: str) -> str:
    escaped = query_text.replace("\\", "\\\\").replace("%", "\\%").replace("_", "\\_")
    return f"%{escaped}%"


def _validate_changelog(document: dict) -> Optional[KGChanges]:
    if "change" in document and isinstance(document["change"], dict):
        change_type = document["change"].get("type")
        if isinstance(change_type, str):
            document["change"]["type"] = KGChangesType(change_type)
    if isinstance(document.get("type"), str):
        document["type"] = KGChangesType(document["type"])
    return KGChanges.model_validate(document)


class PostgreSQLDataClient(DataClient):
    """
    Data client backed by PostgreSQL with JSONB-stored documents.

    Mirrors the Mongo client's `database = brain_id` layout: a single registry
    database (``config.postgresql.system_database``) holds the ``data_brains``
    catalog, and every brain owns its own Postgres database (named
    ``brain_<sanitized_id>``) containing all of its tables.
    """

    def __init__(self):
        self._lock = threading.Lock()
        self._system_initialized = False
        self._initialized_brains: set[str] = set()
        self._search_ready_brains: set[str] = set()
        self._search_alt_brains: set[str] = set()

    @contextmanager
    def _system_connection(self) -> Iterator[psycopg2.extensions.connection]:
        self._ensure_system_schema()
        with borrow(get_system_pool()) as conn:
            yield conn

    @contextmanager
    def _brain_connection(
        self, brain_id: str
    ) -> Iterator[psycopg2.extensions.connection]:
        self._ensure_brain_schema(brain_id)
        with borrow(get_brain_pool(brain_id)) as conn:
            yield conn

    def _ensure_system_schema(self) -> None:
        if self._system_initialized:
            return
        with self._lock:
            if self._system_initialized:
                return
            ensure_database_exists(config.postgresql.system_database)
            with borrow(get_system_pool()) as conn:
                with conn.cursor() as cur:
                    cur.execute(_SYSTEM_DDL)
                conn.commit()
            self._system_initialized = True

    def _ensure_brain_schema(self, brain_id: str) -> None:
        alt_regconfig = config.search_fts_regconfig_for_brain(brain_id)
        needs_base = brain_id not in self._initialized_brains
        needs_search = (
            config.search_enabled and brain_id not in self._search_ready_brains
        )
        needs_alt = (
            config.search_enabled
            and alt_regconfig is not None
            and brain_id not in self._search_alt_brains
        )
        if not needs_base and not needs_search and not needs_alt:
            return
        with self._lock:
            alt_regconfig = config.search_fts_regconfig_for_brain(brain_id)
            needs_base = brain_id not in self._initialized_brains
            needs_search = (
                config.search_enabled and brain_id not in self._search_ready_brains
            )
            needs_alt = (
                config.search_enabled
                and alt_regconfig is not None
                and brain_id not in self._search_alt_brains
            )
            if not needs_base and not needs_search and not needs_alt:
                return
            ensure_brain_database(brain_id)
            with borrow(get_brain_pool(brain_id)) as conn:
                with conn.cursor() as cur:
                    if needs_base:
                        cur.execute(_BRAIN_DDL)
                    if needs_search:
                        cur.execute(_SEARCH_DDL)
                    if needs_alt and alt_regconfig is not None:
                        cur.execute(_search_alt_ddl(alt_regconfig))
                conn.commit()
            if needs_base:
                self._initialized_brains.add(brain_id)
            if needs_search:
                self._search_ready_brains.add(brain_id)
            if needs_alt:
                self._search_alt_brains.add(brain_id)

    def save_text_chunk(self, text_chunk: TextChunk, brain_id: str) -> TextChunk:
        document = text_chunk.model_dump(mode="json")
        with self._brain_connection(brain_id) as conn:
            with conn.cursor() as cur:
                cur.execute(
                    """
                    INSERT INTO data_text_chunks
                        (id, text, metadata, inserted_at, document)
                    VALUES (%s, %s, %s::jsonb, %s, %s::jsonb)
                    ON CONFLICT (id) DO UPDATE SET
                        text = EXCLUDED.text,
                        metadata = EXCLUDED.metadata,
                        inserted_at = EXCLUDED.inserted_at,
                        document = EXCLUDED.document
                    """,
                    (
                        text_chunk.id,
                        text_chunk.text,
                        json.dumps(text_chunk.metadata or {}, default=str),
                        text_chunk.inserted_at,
                        json.dumps(document, default=str),
                    ),
                )
            conn.commit()
        return text_chunk

    def save_observations(
        self, observations: List[Observation], brain_id: str
    ) -> List[Observation]:
        if not observations:
            return observations
        rows = [
            (
                observation.id,
                observation.text,
                observation.resource_id,
                json.dumps(observation.metadata or {}, default=str),
                observation.inserted_at,
                json.dumps(observation.model_dump(mode="json"), default=str),
            )
            for observation in observations
        ]
        with self._brain_connection(brain_id) as conn:
            with conn.cursor() as cur:
                psycopg2.extras.execute_values(
                    cur,
                    """
                    INSERT INTO data_observations
                        (id, text, resource_id, metadata, inserted_at, document)
                    VALUES %s
                    ON CONFLICT (id) DO UPDATE SET
                        text = EXCLUDED.text,
                        resource_id = EXCLUDED.resource_id,
                        metadata = EXCLUDED.metadata,
                        inserted_at = EXCLUDED.inserted_at,
                        document = EXCLUDED.document
                    """,
                    rows,
                    template="(%s, %s, %s, %s::jsonb, %s, %s::jsonb)",
                )
            conn.commit()
        return observations

    def search(
        self,
        text: str,
        brain_id: str,
        collection: str = "text_chunks",
        limit: int = 10,
    ) -> SearchResult:
        table = (
            "data_text_chunks"
            if collection in ("*", "text_chunks")
            else f"data_{collection}"
        )
        pattern = _ilike_pattern(text)
        with self._brain_connection(brain_id) as conn:
            with conn.cursor(cursor_factory=psycopg2.extras.RealDictCursor) as cur:
                cur.execute(
                    f"""
                    SELECT document
                    FROM {table}
                    WHERE text ILIKE %s
                    LIMIT %s
                    """,
                    (pattern, limit),
                )
                text_rows = cur.fetchall()
                text_chunks = [
                    TextChunk.model_validate(row["document"]) for row in text_rows
                ]
                ids = [chunk.id for chunk in text_chunks]
                observations: list[Observation] = []
                if ids:
                    cur.execute(
                        """
                        SELECT document
                        FROM data_observations
                        WHERE resource_id = ANY(%s)
                        """,
                        (ids,),
                    )
                    observations = [
                        Observation.model_validate(row["document"])
                        for row in cur.fetchall()
                    ]
        return SearchResult(text_chunks=text_chunks, observations=observations)

    def search_bm25(
        self,
        text: str,
        brain_id: str,
        collection: str = "text_chunks",
        limit: int = 10,
    ) -> List[Tuple[TextChunk, float]]:
        if not text or not text.strip() or limit <= 0:
            return []
        if collection not in ("*", "text_chunks"):
            return []
        k1 = float(config.search_bm25_k1)
        b = float(config.search_bm25_b)
        alt_regconfig = config.search_fts_regconfig_for_brain(brain_id)
        if alt_regconfig:
            sql = _bm25_sql(alt_regconfig, "search_tsv_alt", "search_len_alt")
            params = (text, k1, k1, b, b, limit)
        else:
            sql = _bm25_sql("english", "search_tsv", "search_len")
            params = (text, text, k1, k1, b, b, limit)
        with self._brain_connection(brain_id) as conn:
            with conn.cursor(cursor_factory=psycopg2.extras.RealDictCursor) as cur:
                cur.execute(
                    sql,
                    params,
                )
                rows = cur.fetchall()
        results: List[Tuple[TextChunk, float]] = []
        for row in rows:
            chunk = TextChunk.model_validate(row["document"])
            results.append((chunk, float(row["bm25"] or 0.0)))
        return results

    def get_text_chunks_by_ids(
        self, ids: List[str], with_observations: bool = False, brain_id: str = "default"
    ) -> Tuple[List[TextChunk], List[Observation]]:
        if not ids:
            return ([], [])
        with self._brain_connection(brain_id) as conn:
            with conn.cursor(cursor_factory=psycopg2.extras.RealDictCursor) as cur:
                cur.execute(
                    """
                    SELECT document FROM data_text_chunks
                    WHERE id = ANY(%s)
                    """,
                    (ids,),
                )
                text_chunks = [
                    TextChunk.model_validate(row["document"]) for row in cur.fetchall()
                ]
                observations: list[Observation] = []
                if with_observations:
                    cur.execute(
                        """
                        SELECT document FROM data_observations
                        WHERE resource_id = ANY(%s)
                        """,
                        (ids,),
                    )
                    observations = [
                        Observation.model_validate(row["document"])
                        for row in cur.fetchall()
                    ]
        return (text_chunks, observations)

    def save_structured_data(
        self, structured_data: StructuredData, brain_id: str
    ) -> StructuredData:
        document = structured_data.model_dump(mode="json")
        with self._brain_connection(brain_id) as conn:
            with conn.cursor() as cur:
                cur.execute(
                    """
                    INSERT INTO data_structured_data
                        (id, data, types, metadata, inserted_at, document)
                    VALUES (%s, %s::jsonb, %s, %s::jsonb, %s, %s::jsonb)
                    ON CONFLICT (id) DO UPDATE SET
                        data = EXCLUDED.data,
                        types = EXCLUDED.types,
                        metadata = EXCLUDED.metadata,
                        inserted_at = EXCLUDED.inserted_at,
                        document = EXCLUDED.document
                    """,
                    (
                        structured_data.id,
                        json.dumps(structured_data.data or {}, default=str),
                        list(structured_data.types or []),
                        json.dumps(structured_data.metadata or {}, default=str),
                        structured_data.inserted_at,
                        json.dumps(document, default=str),
                    ),
                )
            conn.commit()
        return structured_data

    def create_brain(self, name_key: str) -> Brain:
        brain = Brain(name_key=name_key)
        document = brain.model_dump(mode="json")
        with self._system_connection() as conn:
            with conn.cursor() as cur:
                cur.execute(
                    """
                    INSERT INTO data_brains (id, name_key, pat, document)
                    VALUES (%s, %s, %s, %s::jsonb)
                    """,
                    (
                        brain.id,
                        brain.name_key,
                        brain.pat,
                        json.dumps(document, default=str),
                    ),
                )
            conn.commit()
        self._ensure_brain_schema(name_key)
        return brain

    def get_brain(self, name_key: str) -> Optional[Brain]:
        with self._system_connection() as conn:
            with conn.cursor(cursor_factory=psycopg2.extras.RealDictCursor) as cur:
                cur.execute(
                    "SELECT document FROM data_brains WHERE name_key = %s",
                    (name_key,),
                )
                row = cur.fetchone()
        if not row:
            return None
        return Brain.model_validate(row["document"])

    def get_brain_by_pat(self, pat: str) -> Optional[Brain]:
        with self._system_connection() as conn:
            with conn.cursor(cursor_factory=psycopg2.extras.RealDictCursor) as cur:
                cur.execute(
                    "SELECT document FROM data_brains WHERE pat = %s",
                    (pat,),
                )
                row = cur.fetchone()
        if not row:
            return None
        return Brain.model_validate(row["document"])

    def get_brains_list(self) -> List[Brain]:
        with self._system_connection() as conn:
            with conn.cursor(cursor_factory=psycopg2.extras.RealDictCursor) as cur:
                cur.execute("SELECT document FROM data_brains")
                rows = cur.fetchall()
        registered = [Brain.model_validate(row["document"]) for row in rows]
        by_key = {brain.name_key: brain for brain in registered}
        known_db_names = {
            brain_db_name(brain.name_key) for brain in registered
        } | {brain_db_name(brain.id) for brain in registered}

        for db_name in list_brain_database_names():
            if db_name in known_db_names:
                continue
            suffix = brain_id_from_db_name(db_name)
            if not suffix or is_internal_brain_db_suffix(suffix):
                continue
            if suffix not in by_key:
                by_key[suffix] = Brain(name_key=suffix)

        return sorted(by_key.values(), key=lambda brain: brain.name_key)

    def clear_brain_data(self, brain_id: str) -> None:
        clear_brain_database(brain_id)

    def delete_brain_registry(self, brain_id: str) -> bool:
        with self._system_connection() as conn:
            with conn.cursor() as cur:
                cur.execute(
                    "DELETE FROM data_brains WHERE name_key = %s",
                    (brain_id,),
                )
                deleted = cur.rowcount > 0
            conn.commit()
        drop_brain_database(brain_id)
        return deleted

    def save_kg_changes(self, kg_changes: KGChanges, brain_id: str) -> KGChanges:
        document = kg_changes.model_dump(mode="json")
        type_value = (
            kg_changes.type.value
            if isinstance(kg_changes.type, KGChangesType)
            else kg_changes.type
        )
        with self._brain_connection(brain_id) as conn:
            with conn.cursor() as cur:
                cur.execute(
                    """
                    INSERT INTO data_kg_changes
                        (id, type, change, timestamp, document)
                    VALUES (%s, %s, %s::jsonb, %s, %s::jsonb)
                    ON CONFLICT (id) DO UPDATE SET
                        type = EXCLUDED.type,
                        change = EXCLUDED.change,
                        timestamp = EXCLUDED.timestamp,
                        document = EXCLUDED.document
                    """,
                    (
                        kg_changes.id,
                        type_value,
                        json.dumps(document.get("change") or {}, default=str),
                        kg_changes.timestamp,
                        json.dumps(document, default=str),
                    ),
                )
            conn.commit()
        return kg_changes

    def get_structured_data_by_id(
        self, id: str, brain_id: str
    ) -> Optional[StructuredData]:
        with self._brain_connection(brain_id) as conn:
            with conn.cursor(cursor_factory=psycopg2.extras.RealDictCursor) as cur:
                cur.execute(
                    """
                    SELECT document FROM data_structured_data
                    WHERE id = %s
                    """,
                    (id,),
                )
                row = cur.fetchone()
        if not row:
            return None
        return StructuredData.model_validate(row["document"])

    def get_structured_data_list(
        self,
        brain_id: str,
        limit: int = 10,
        skip: int = 0,
        types: list[str] = None,
        query_text: str = None,
    ) -> Tuple[list[StructuredData], int]:
        clauses: list[str] = []
        params: list[Any] = []
        if types:
            clauses.append("types && %s")
            params.append(list(types))
        if query_text:
            pattern = _ilike_pattern(query_text)
            clauses.append(
                "(data::text ILIKE %s OR metadata::text ILIKE %s "
                "OR EXISTS (SELECT 1 FROM unnest(types) t WHERE t ILIKE %s))"
            )
            params.extend([pattern, pattern, pattern])
        where_sql = f"WHERE {' AND '.join(clauses)}" if clauses else ""
        with self._brain_connection(brain_id) as conn:
            with conn.cursor(cursor_factory=psycopg2.extras.RealDictCursor) as cur:
                cur.execute(
                    f"""
                    SELECT document FROM data_structured_data
                    {where_sql}
                    ORDER BY inserted_at DESC
                    OFFSET %s LIMIT %s
                    """,
                    (*params, skip, limit),
                )
                rows = cur.fetchall()
                cur.execute(
                    f"SELECT COUNT(*) AS total FROM data_structured_data {where_sql}",
                    tuple(params),
                )
                total = cur.fetchone()["total"]
        return (
            [StructuredData.model_validate(row["document"]) for row in rows],
            int(total or 0),
        )

    def get_text_chunks(
        self,
        brain_id: str,
        limit: int = 10,
        skip: int = 0,
        query_text: str = None,
        metadata_eq: dict[str, str] | None = None,
        order: Literal["asc", "desc"] = "desc",
    ) -> Tuple[List[TextChunk], int]:
        clauses: list[str] = []
        params: list[Any] = []
        if query_text:
            pattern = _ilike_pattern(query_text)
            clauses.append("(text ILIKE %s OR metadata::text ILIKE %s)")
            params.extend([pattern, pattern])
        if metadata_eq:
            for key, value in metadata_eq.items():
                clauses.append("metadata ->> %s = %s")
                params.extend([key, value])
        where_sql = f"WHERE {' AND '.join(clauses)}" if clauses else ""
        order_sql = "ASC" if order == "asc" else "DESC"
        with self._brain_connection(brain_id) as conn:
            with conn.cursor(cursor_factory=psycopg2.extras.RealDictCursor) as cur:
                cur.execute(
                    f"""
                    SELECT document FROM data_text_chunks
                    {where_sql}
                    ORDER BY inserted_at {order_sql}
                    OFFSET %s LIMIT %s
                    """,
                    (*params, skip, limit),
                )
                rows = cur.fetchall()
                cur.execute(
                    f"SELECT COUNT(*) AS total FROM data_text_chunks {where_sql}",
                    tuple(params),
                )
                total = cur.fetchone()["total"]
        return (
            [TextChunk.model_validate(row["document"]) for row in rows],
            int(total or 0),
        )

    def get_structured_data_types(self, brain_id: str) -> list[str]:
        with self._brain_connection(brain_id) as conn:
            with conn.cursor() as cur:
                cur.execute(
                    """
                    SELECT DISTINCT unnest(types) AS t
                    FROM data_structured_data
                    WHERE types IS NOT NULL
                    ORDER BY t
                    """
                )
                return [row[0] for row in cur.fetchall() if row[0] is not None]

    def get_observation_by_id(self, id: str, brain_id: str) -> Optional[Observation]:
        with self._brain_connection(brain_id) as conn:
            with conn.cursor(cursor_factory=psycopg2.extras.RealDictCursor) as cur:
                cur.execute(
                    """
                    SELECT document FROM data_observations
                    WHERE id = %s
                    """,
                    (id,),
                )
                row = cur.fetchone()
        if not row:
            return None
        return Observation.model_validate(row["document"])

    def get_observations_list(
        self,
        brain_id: str,
        limit: int = 10,
        skip: int = 0,
        resource_id: str = None,
        labels: list[str] = None,
        query_text: str = None,
    ) -> list[Observation]:
        clauses: list[str] = []
        params: list[Any] = []
        if resource_id:
            clauses.append("resource_id = %s")
            params.append(resource_id)
        if labels:
            clauses.append("metadata -> 'labels' ?| %s")
            params.append(list(labels))
        if query_text:
            pattern = _ilike_pattern(query_text)
            clauses.append(
                "(text ILIKE %s OR resource_id ILIKE %s OR metadata::text ILIKE %s)"
            )
            params.extend([pattern, pattern, pattern])
        where_sql = f"WHERE {' AND '.join(clauses)}" if clauses else ""
        with self._brain_connection(brain_id) as conn:
            with conn.cursor(cursor_factory=psycopg2.extras.RealDictCursor) as cur:
                cur.execute(
                    f"""
                    SELECT document FROM data_observations
                    {where_sql}
                    ORDER BY inserted_at DESC
                    OFFSET %s LIMIT %s
                    """,
                    (*params, skip, limit),
                )
                rows = cur.fetchall()
        return [Observation.model_validate(row["document"]) for row in rows]

    def get_observation_labels(self, brain_id: str) -> list[str]:
        with self._brain_connection(brain_id) as conn:
            with conn.cursor() as cur:
                cur.execute(
                    """
                    SELECT DISTINCT jsonb_array_elements_text(metadata -> 'labels') AS label
                    FROM data_observations
                    WHERE metadata ? 'labels'
                      AND jsonb_typeof(metadata -> 'labels') = 'array'
                    ORDER BY label
                    """
                )
                return [row[0] for row in cur.fetchall() if row[0] is not None]

    def get_changelog_by_id(self, id: str, brain_id: str) -> Optional[KGChanges]:
        with self._brain_connection(brain_id) as conn:
            with conn.cursor(cursor_factory=psycopg2.extras.RealDictCursor) as cur:
                cur.execute(
                    """
                    SELECT document FROM data_kg_changes
                    WHERE id = %s
                    """,
                    (id,),
                )
                row = cur.fetchone()
        if not row:
            return None
        return _validate_changelog(dict(row["document"]))

    def get_changelogs_list(
        self,
        brain_id: str,
        limit: int = 10,
        skip: int = 0,
        types: list[str] = None,
        query_text: str = None,
    ) -> list[KGChanges]:
        clauses: list[str] = []
        params: list[Any] = []
        if types:
            clauses.append("type = ANY(%s)")
            params.append(list(types))
        if query_text:
            pattern = _ilike_pattern(query_text)
            clauses.append("(change::text ILIKE %s OR type ILIKE %s)")
            params.extend([pattern, pattern])
        where_sql = f"WHERE {' AND '.join(clauses)}" if clauses else ""
        with self._brain_connection(brain_id) as conn:
            with conn.cursor(cursor_factory=psycopg2.extras.RealDictCursor) as cur:
                cur.execute(
                    f"""
                    SELECT document FROM data_kg_changes
                    {where_sql}
                    ORDER BY timestamp DESC
                    OFFSET %s LIMIT %s
                    """,
                    (*params, skip, limit),
                )
                rows = cur.fetchall()
        changelogs: list[KGChanges] = []
        for row in rows:
            try:
                changelogs.append(_validate_changelog(dict(row["document"])))
            except Exception as exc:
                print(f"Error parsing changelog {row['document'].get('id')}: {exc}")
                continue
        return changelogs

    def get_changelog_types(self, brain_id: str) -> list[str]:
        with self._brain_connection(brain_id) as conn:
            with conn.cursor() as cur:
                cur.execute(
                    """
                    SELECT DISTINCT type FROM data_kg_changes
                    WHERE type IS NOT NULL
                    ORDER BY type
                    """
                )
                return [row[0] for row in cur.fetchall()]

    def update_structured_data(
        self, structured_data: StructuredData, brain_id: str
    ) -> StructuredData:
        document = structured_data.model_dump(mode="json")
        with self._brain_connection(brain_id) as conn:
            with conn.cursor() as cur:
                cur.execute(
                    """
                    UPDATE data_structured_data SET
                        data = %s::jsonb,
                        types = %s,
                        metadata = %s::jsonb,
                        inserted_at = %s,
                        document = %s::jsonb
                    WHERE id = %s
                    """,
                    (
                        json.dumps(structured_data.data or {}, default=str),
                        list(structured_data.types or []),
                        json.dumps(structured_data.metadata or {}, default=str),
                        structured_data.inserted_at,
                        json.dumps(document, default=str),
                        structured_data.id,
                    ),
                )
            conn.commit()
        return structured_data

    def get_last_text_chunks(self, brain_id: str, limit: int = 10) -> list[TextChunk]:
        with self._brain_connection(brain_id) as conn:
            with conn.cursor(cursor_factory=psycopg2.extras.RealDictCursor) as cur:
                cur.execute(
                    """
                    SELECT document FROM data_text_chunks
                    ORDER BY inserted_at DESC
                    LIMIT %s
                    """,
                    (limit,),
                )
                rows = cur.fetchall()
        return [TextChunk.model_validate(row["document"]) for row in rows]

    def get_last_structured_data(
        self, brain_id: str, limit: int = 10
    ) -> list[StructuredData]:
        with self._brain_connection(brain_id) as conn:
            with conn.cursor(cursor_factory=psycopg2.extras.RealDictCursor) as cur:
                cur.execute(
                    """
                    SELECT document FROM data_structured_data
                    ORDER BY inserted_at DESC
                    LIMIT %s
                    """,
                    (limit,),
                )
                rows = cur.fetchall()
        return [StructuredData.model_validate(row["document"]) for row in rows]


_postgresql_data_client: Optional[PostgreSQLDataClient] = None


def get_postgresql_data_client() -> PostgreSQLDataClient:
    global _postgresql_data_client
    if _postgresql_data_client is None:
        _postgresql_data_client = PostgreSQLDataClient()
    return _postgresql_data_client