Add pgAdmin migration and Colab embedding loaders

app/jobs/sync_place_embeddings.py

@@ -73,6 +73,10 @@ async def main() -> None:
         counters.upserted,
         counters.errors,
     )
+    if counters.errors:
+        raise RuntimeError(
+            f"Place embedding sync finished with {counters.errors} failed records"
+        )
 
 
 async def _flush_batch(

app/jobs/sync_post_embeddings.py

@@ -73,6 +73,10 @@ async def main() -> None:
         counters.upserted,
         counters.errors,
     )
+    if counters.errors:
+        raise RuntimeError(
+            f"Post embedding sync finished with {counters.errors} failed records"
+        )
 
 
 async def _flush_batch(

docs/fasttext_deployment.md

@@ -68,6 +68,10 @@ psql "host=<host> port=5432 dbname=nlp_vectors user=<admin> sslmode=require" `
 La migracion trunca `place_embeddings` y `post_embeddings` porque son indices
 derivados incompatibles. No toca la base transaccional de la API principal.
 
+Para hacerlo desde pgAdmin y ejecutar la carga pesada desde Google Colab consulta
+`docs/pgadmin_colab_fasttext.md`. Incluye un SQL unico para Query Tool y dos scripts
+independientes de carga.
+
 ## 4. Repoblar PGVector
 
 Con las credenciales writer en `.env`:

docs/pgadmin_colab_fasttext.md

@@ -0,0 +1,96 @@
+# Migracion FastText Con pgAdmin Y Google Colab
+
+## 1. Antes De Modificar RDS
+
+1. Publica esta version en la rama `hf-deploy` de GitHub.
+2. Crea un snapshot de RDS.
+3. Pausa el Space de Hugging Face y cualquier job de sincronizacion.
+4. No ejecutes todavia los cargadores de Colab.
+
+## 2. Migrar Desde pgAdmin
+
+En pgAdmin selecciona la base `nlp_vectors`, abre **Tools > Query Tool**, carga el
+archivo `sql/pgadmin_migrate_fasttext_300.sql` y ejecutalo completo.
+
+El script hace en una sola transaccion:
+
+- valida que ambas columnas sigan siendo `VECTOR(16)`;
+- elimina las funciones e indices incompatibles;
+- trunca solamente `place_embeddings` y `post_embeddings`;
+- cambia ambas columnas a `VECTOR(300)`;
+- reconstruye HNSW, funciones y permisos;
+- confirma al final la dimension y que ambas tablas quedaron vacias.
+
+No lo ejecutes una segunda vez: esta protegido y abortara si ya encuentra
+`VECTOR(300)`.
+
+## 3. Preparar Secrets En Colab
+
+En el panel **Secrets** de Colab agrega y habilita **Notebook access** para:
+
+| Nombre | Requerido | Uso |
+|---|---:|---|
+| `PGVECTOR_WRITER_PASSWORD` | Si | Password del rol `nlp_writer`. |
+| `MAIN_API_INTERNAL_TOKEN` | Solo si aplica | Token para leer lugares/posts de la API principal. |
+| `HF_TOKEN` | No | Puede ayudar con la descarga, pero el modelo es publico. |
+
+Los scripts incluyen los valores no secretos actuales del host, base, usuario y URL
+principal. Tambien puedes sobrescribirlos con Secrets llamados
+`PGVECTOR_HOST`, `PGVECTOR_DATABASE`, `PGVECTOR_WRITER_USER` y
+`MAIN_API_BASE_URL`.
+
+## 4. Permitir Temporalmente La IP De Colab
+
+RDS debe ser accesible desde el runtime. Obten la IP publica actual en una celda:
+
+```python
+!curl -s https://api.ipify.org
+```
+
+En el Security Group de RDS agrega temporalmente una regla de entrada TCP 5432 con
+origen `<IP_OBTENIDA>/32`. No uses `0.0.0.0/0`. Elimina la regla cuando terminen
+ambas cargas.
+
+Si RDS es privado y no tiene una ruta publica, Colab no podra conectarse directamente;
+en ese caso ejecuta los jobs desde una instancia dentro de la VPC.
+
+## 5. Ejecutar Las Cargas
+
+En un runtime de Colab con memoria suficiente:
+
+```python
+!git clone --branch hf-deploy https://github.com/AlleksDev/Frimeet-API-NLP.git
+%cd Frimeet-API-NLP
+```
+
+Primero lugares:
+
+```python
+!python scripts/colab_initial_load_places.py
+```
+
+Despues publicaciones, reutilizando dependencias y el modelo ya descargado:
+
+```python
+!python scripts/colab_initial_load_posts.py --skip-install --skip-download
+```
+
+Para una prueba sin escrituras se pueden agregar `--dry-run --max-pages 1`. Para una
+carga pequena real usa solamente `--max-pages 1`; despues puedes ejecutar otra vez
+sin esa opcion y el job completara los registros faltantes. La ejecucion completa no
+debe usar `--dry-run` ni `--max-pages`. El resultado correcto termina con `errors=0`
+y el mensaje de carga terminada.
+
+## 6. Verificar Desde pgAdmin
+
+Despues de ambas cargas abre nuevamente Query Tool y ejecuta el contenido de
+`sql/verify_fasttext_embeddings.sql`. Debe mostrar:
+
+- filas mayores que cero;
+- dimensiones minima y maxima iguales a `300`;
+- modelo `facebook/fasttext-es-vectors`;
+- version `common-crawl-300-v1`;
+- normas cercanas a `1`.
+
+Finalmente elimina la regla temporal del Security Group, actualiza las variables del
+Space y despliega la nueva imagen.

scripts/colab_initial_load_places.py

@@ -0,0 +1,152 @@
+"""Carga inicial de embeddings FastText de lugares desde Google Colab.
+
+Ejecutar desde la raiz de un clon de este repositorio:
+
+    !python scripts/colab_initial_load_places.py
+
+El script lee credenciales desde los Secrets de Colab usando los mismos nombres
+de las variables de entorno. Nunca imprime los valores secretos.
+"""
+
+from __future__ import annotations
+
+import argparse
+import os
+from pathlib import Path
+import subprocess
+import sys
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+DEFAULT_MODEL_PATH = "/content/fasttext-es/model.bin"
+
+
+def main() -> None:
+    args = _parse_args()
+    os.chdir(REPO_ROOT)
+    _configure_environment()
+
+    if not args.skip_install:
+        _run(
+            sys.executable,
+            "-m",
+            "pip",
+            "install",
+            "--quiet",
+            "-r",
+            str(REPO_ROOT / "requirements.txt"),
+        )
+
+    if not args.skip_download:
+        _run(
+            sys.executable,
+            "-m",
+            "app.shared.nlp.embeddings.download_fasttext_model",
+            "--repo-id",
+            os.environ["FASTTEXT_MODEL_REPO_ID"],
+            "--filename",
+            os.environ["FASTTEXT_MODEL_FILENAME"],
+            "--destination",
+            os.environ["FASTTEXT_MODEL_PATH"],
+        )
+
+    command = [
+        sys.executable,
+        "-m",
+        "app.jobs.initial_load_place_embeddings",
+        "--batch-size",
+        str(args.batch_size),
+        "--page-limit",
+        str(args.page_limit),
+    ]
+    if args.max_pages is not None:
+        command.extend(["--max-pages", str(args.max_pages)])
+    if args.dry_run:
+        command.append("--dry-run")
+
+    _run(*command)
+    print("Carga de lugares terminada correctamente.")
+
+
+def _configure_environment() -> None:
+    defaults = {
+        "ENV": "colab",
+        "MAIN_API_BASE_URL": "http://52.86.8.11",
+        "MAIN_API_PLACES_SEARCH_PATH": "/api/v1/places/search",
+        "MAIN_API_TIMEOUT_SECONDS": "60",
+        "MAIN_API_PLACES_PAGE_LIMIT": "50",
+        "MAIN_API_PLACES_PAGINATION_MODE": "cursor",
+        "VECTOR_STORE_PROVIDER": "aws_pgvector",
+        "PGVECTOR_HOST": "nlp-vector-db.c2jwncm87zsa.us-east-1.rds.amazonaws.com",
+        "PGVECTOR_PORT": "5432",
+        "PGVECTOR_DATABASE": "nlp_vectors",
+        "PGVECTOR_WRITER_USER": "nlp_writer",
+        "PGVECTOR_SSL_MODE": "require",
+        "EMBEDDING_PROVIDER": "fasttext",
+        "EMBEDDING_DIMENSION": "300",
+        "EMBEDDING_MODEL": "facebook/fasttext-es-vectors",
+        "EMBEDDING_VERSION": "common-crawl-300-v1",
+        "FASTTEXT_MODEL_PATH": DEFAULT_MODEL_PATH,
+        "FASTTEXT_MODEL_REPO_ID": "facebook/fasttext-es-vectors",
+        "FASTTEXT_MODEL_FILENAME": "model.bin",
+        "FASTTEXT_AUTO_DOWNLOAD": "false",
+        "LOG_LEVEL": "INFO",
+    }
+    for name, default in defaults.items():
+        os.environ[name] = _read_setting(name, default=default)
+
+    os.environ["PGVECTOR_WRITER_PASSWORD"] = _read_setting(
+        "PGVECTOR_WRITER_PASSWORD",
+        required=True,
+    )
+
+    for optional_name in (
+        "MAIN_API_INTERNAL_TOKEN",
+        "MAIN_API_AUTH_TOKEN",
+        "HF_TOKEN",
+    ):
+        value = _read_setting(optional_name)
+        if value:
+            os.environ[optional_name] = value
+
+
+def _read_setting(
+    name: str,
+    default: str | None = None,
+    required: bool = False,
+) -> str:
+    value = os.getenv(name) or _read_colab_secret(name) or default
+    if required and not value:
+        raise RuntimeError(
+            f"Falta el Secret {name!r} en Colab. Activa tambien Notebook access."
+        )
+    return value or ""
+
+
+def _read_colab_secret(name: str) -> str | None:
+    try:
+        from google.colab import userdata
+
+        value = userdata.get(name)
+        return str(value).strip() if value else None
+    except Exception:
+        return None
+
+
+def _run(*command: str) -> None:
+    subprocess.run(list(command), cwd=REPO_ROOT, check=True)
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--batch-size", type=int, default=25)
+    parser.add_argument("--page-limit", type=int, default=50)
+    parser.add_argument("--max-pages", type=int, default=None)
+    parser.add_argument("--dry-run", action="store_true")
+    parser.add_argument("--skip-install", action="store_true")
+    parser.add_argument("--skip-download", action="store_true")
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    main()

scripts/colab_initial_load_posts.py

@@ -0,0 +1,152 @@
+"""Carga inicial de embeddings FastText de publicaciones desde Google Colab.
+
+Ejecutar desde la raiz de un clon de este repositorio:
+
+    !python scripts/colab_initial_load_posts.py
+
+El script lee credenciales desde los Secrets de Colab usando los mismos nombres
+de las variables de entorno. Nunca imprime los valores secretos.
+"""
+
+from __future__ import annotations
+
+import argparse
+import os
+from pathlib import Path
+import subprocess
+import sys
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+DEFAULT_MODEL_PATH = "/content/fasttext-es/model.bin"
+
+
+def main() -> None:
+    args = _parse_args()
+    os.chdir(REPO_ROOT)
+    _configure_environment()
+
+    if not args.skip_install:
+        _run(
+            sys.executable,
+            "-m",
+            "pip",
+            "install",
+            "--quiet",
+            "-r",
+            str(REPO_ROOT / "requirements.txt"),
+        )
+
+    if not args.skip_download:
+        _run(
+            sys.executable,
+            "-m",
+            "app.shared.nlp.embeddings.download_fasttext_model",
+            "--repo-id",
+            os.environ["FASTTEXT_MODEL_REPO_ID"],
+            "--filename",
+            os.environ["FASTTEXT_MODEL_FILENAME"],
+            "--destination",
+            os.environ["FASTTEXT_MODEL_PATH"],
+        )
+
+    command = [
+        sys.executable,
+        "-m",
+        "app.jobs.initial_load_post_embeddings",
+        "--batch-size",
+        str(args.batch_size),
+        "--page-limit",
+        str(args.page_limit),
+    ]
+    if args.max_pages is not None:
+        command.extend(["--max-pages", str(args.max_pages)])
+    if args.dry_run:
+        command.append("--dry-run")
+
+    _run(*command)
+    print("Carga de publicaciones terminada correctamente.")
+
+
+def _configure_environment() -> None:
+    defaults = {
+        "ENV": "colab",
+        "MAIN_API_BASE_URL": "http://52.86.8.11",
+        "MAIN_API_POSTS_SEARCH_PATH": "/api/v1/posts/search",
+        "MAIN_API_TIMEOUT_SECONDS": "60",
+        "MAIN_API_POSTS_PAGE_LIMIT": "50",
+        "MAIN_API_POSTS_PAGINATION_MODE": "cursor",
+        "VECTOR_STORE_PROVIDER": "aws_pgvector",
+        "PGVECTOR_HOST": "nlp-vector-db.c2jwncm87zsa.us-east-1.rds.amazonaws.com",
+        "PGVECTOR_PORT": "5432",
+        "PGVECTOR_DATABASE": "nlp_vectors",
+        "PGVECTOR_WRITER_USER": "nlp_writer",
+        "PGVECTOR_SSL_MODE": "require",
+        "EMBEDDING_PROVIDER": "fasttext",
+        "EMBEDDING_DIMENSION": "300",
+        "EMBEDDING_MODEL": "facebook/fasttext-es-vectors",
+        "EMBEDDING_VERSION": "common-crawl-300-v1",
+        "FASTTEXT_MODEL_PATH": DEFAULT_MODEL_PATH,
+        "FASTTEXT_MODEL_REPO_ID": "facebook/fasttext-es-vectors",
+        "FASTTEXT_MODEL_FILENAME": "model.bin",
+        "FASTTEXT_AUTO_DOWNLOAD": "false",
+        "LOG_LEVEL": "INFO",
+    }
+    for name, default in defaults.items():
+        os.environ[name] = _read_setting(name, default=default)
+
+    os.environ["PGVECTOR_WRITER_PASSWORD"] = _read_setting(
+        "PGVECTOR_WRITER_PASSWORD",
+        required=True,
+    )
+
+    for optional_name in (
+        "MAIN_API_INTERNAL_TOKEN",
+        "MAIN_API_AUTH_TOKEN",
+        "HF_TOKEN",
+    ):
+        value = _read_setting(optional_name)
+        if value:
+            os.environ[optional_name] = value
+
+
+def _read_setting(
+    name: str,
+    default: str | None = None,
+    required: bool = False,
+) -> str:
+    value = os.getenv(name) or _read_colab_secret(name) or default
+    if required and not value:
+        raise RuntimeError(
+            f"Falta el Secret {name!r} en Colab. Activa tambien Notebook access."
+        )
+    return value or ""
+
+
+def _read_colab_secret(name: str) -> str | None:
+    try:
+        from google.colab import userdata
+
+        value = userdata.get(name)
+        return str(value).strip() if value else None
+    except Exception:
+        return None
+
+
+def _run(*command: str) -> None:
+    subprocess.run(list(command), cwd=REPO_ROOT, check=True)
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--batch-size", type=int, default=25)
+    parser.add_argument("--page-limit", type=int, default=50)
+    parser.add_argument("--max-pages", type=int, default=None)
+    parser.add_argument("--dry-run", action="store_true")
+    parser.add_argument("--skip-install", action="store_true")
+    parser.add_argument("--skip-download", action="store_true")
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    main()

sql/pgadmin_migrate_fasttext_300.sql

@@ -0,0 +1,322 @@
+-- Ejecutar completo una sola vez desde pgAdmin > Query Tool.
+-- Requisitos:
+--   1. Tener un snapshot reciente de RDS.
+--   2. Detener temporalmente el Space y cualquier job de embeddings.
+--   3. Conectarse a la base nlp_vectors como propietario de las tablas o administrador.
+--
+-- Este script elimina solamente los indices derivados de busqueda. No toca la
+-- base transaccional de la API principal. Si las columnas ya no son VECTOR(16),
+-- aborta para evitar truncar accidentalmente una migracion ya terminada.
+
+CREATE EXTENSION IF NOT EXISTS vector;
+
+BEGIN;
+
+LOCK TABLE public.place_embeddings IN ACCESS EXCLUSIVE MODE;
+LOCK TABLE public.post_embeddings IN ACCESS EXCLUSIVE MODE;
+
+DO $$
+DECLARE
+    place_type TEXT;
+    post_type TEXT;
+BEGIN
+    SELECT format_type(attribute.atttypid, attribute.atttypmod)
+    INTO place_type
+    FROM pg_attribute attribute
+    WHERE attribute.attrelid = 'public.place_embeddings'::regclass
+      AND attribute.attname = 'embedding'
+      AND NOT attribute.attisdropped;
+
+    SELECT format_type(attribute.atttypid, attribute.atttypmod)
+    INTO post_type
+    FROM pg_attribute attribute
+    WHERE attribute.attrelid = 'public.post_embeddings'::regclass
+      AND attribute.attname = 'embedding'
+      AND NOT attribute.attisdropped;
+
+    IF place_type <> 'vector(16)' OR post_type <> 'vector(16)' THEN
+        RAISE EXCEPTION
+            'Se esperaba VECTOR(16). Tipos encontrados: places=%, posts=%',
+            place_type,
+            post_type;
+    END IF;
+END $$;
+
+-- Las funciones antiguas se eliminan antes del cambio de dimension y se
+-- reconstruyen dentro de la misma transaccion.
+DROP FUNCTION IF EXISTS public.match_places(vector, integer, jsonb);
+DROP FUNCTION IF EXISTS public.match_posts(vector, integer, jsonb);
+DROP FUNCTION IF EXISTS public.upsert_place_embedding(
+    text, text, jsonb, vector, text, text, text, boolean
+);
+DROP FUNCTION IF EXISTS public.upsert_post_embedding(
+    text, text, jsonb, vector, text, text, text, boolean
+);
+
+DROP INDEX IF EXISTS public.place_embeddings_embedding_hnsw_idx;
+DROP INDEX IF EXISTS public.post_embeddings_embedding_hnsw_idx;
+
+-- No existe una conversion valida de los vectores mock de 16 dimensiones a
+-- FastText. Ambas tablas son indices derivados y se reconstruiran desde cero.
+TRUNCATE TABLE public.place_embeddings, public.post_embeddings;
+
+ALTER TABLE public.place_embeddings
+    ALTER COLUMN embedding TYPE vector(300);
+
+ALTER TABLE public.post_embeddings
+    ALTER COLUMN embedding TYPE vector(300);
+
+CREATE INDEX place_embeddings_embedding_hnsw_idx
+ON public.place_embeddings USING hnsw (embedding vector_cosine_ops);
+
+CREATE INDEX post_embeddings_embedding_hnsw_idx
+ON public.post_embeddings USING hnsw (embedding vector_cosine_ops);
+
+CREATE INDEX IF NOT EXISTS place_embeddings_metadata_gin_idx
+ON public.place_embeddings USING gin (metadata);
+
+CREATE INDEX IF NOT EXISTS post_embeddings_metadata_gin_idx
+ON public.post_embeddings USING gin (metadata);
+
+CREATE OR REPLACE FUNCTION public.match_places(
+    query_embedding vector(300),
+    match_count integer,
+    filters jsonb DEFAULT '{}'::jsonb
+)
+RETURNS TABLE (
+    external_id text,
+    document text,
+    metadata jsonb,
+    score double precision
+)
+LANGUAGE sql
+STABLE
+SECURITY DEFINER
+SET search_path = public
+AS $$
+    SELECT
+        place.external_id,
+        place.document,
+        place.metadata,
+        1 - (place.embedding <=> query_embedding) AS score
+    FROM public.place_embeddings place
+    WHERE place.is_active = true
+      AND COALESCE((filters->>'is_active')::boolean, true) = true
+      AND ((filters ? 'city') IS FALSE OR lower(place.metadata->>'city') = lower(filters->>'city'))
+      AND ((filters ? 'state') IS FALSE OR lower(place.metadata->>'state') = lower(filters->>'state'))
+      AND ((filters ? 'category') IS FALSE OR lower(place.metadata->>'category') = lower(filters->>'category'))
+      AND ((filters ? 'price_range') IS FALSE OR place.metadata->>'price_range' = filters->>'price_range')
+      AND ((filters ? 'occasion') IS FALSE OR place.metadata->>'occasion' ILIKE ('%' || (filters->>'occasion') || '%'))
+      AND (
+          (filters ? 'place_ids') IS FALSE
+          OR place.external_id IN (
+              SELECT jsonb_array_elements_text(filters->'place_ids')
+          )
+      )
+    ORDER BY place.embedding <=> query_embedding
+    LIMIT match_count;
+$$;
+
+CREATE OR REPLACE FUNCTION public.match_posts(
+    query_embedding vector(300),
+    match_count integer,
+    filters jsonb DEFAULT '{}'::jsonb
+)
+RETURNS TABLE (
+    external_id text,
+    document text,
+    metadata jsonb,
+    score double precision
+)
+LANGUAGE sql
+STABLE
+SECURITY DEFINER
+SET search_path = public
+AS $$
+    SELECT
+        post.external_id,
+        post.document,
+        post.metadata,
+        1 - (post.embedding <=> query_embedding) AS score
+    FROM public.post_embeddings post
+    WHERE post.is_active = true
+      AND COALESCE((filters->>'is_active')::boolean, true) = true
+      AND ((filters ? 'city') IS FALSE OR lower(post.metadata->>'city') = lower(filters->>'city'))
+    ORDER BY post.embedding <=> query_embedding
+    LIMIT match_count;
+$$;
+
+CREATE OR REPLACE FUNCTION public.upsert_place_embedding(
+    p_external_id text,
+    p_document text,
+    p_metadata jsonb,
+    p_embedding vector(300),
+    p_content_hash text,
+    p_embedding_model text,
+    p_embedding_version text,
+    p_is_active boolean
+)
+RETURNS void
+LANGUAGE sql
+SECURITY DEFINER
+SET search_path = public
+AS $$
+    INSERT INTO public.place_embeddings (
+        external_id,
+        document,
+        metadata,
+        embedding,
+        content_hash,
+        embedding_model,
+        embedding_version,
+        is_active,
+        updated_at
+    )
+    VALUES (
+        p_external_id,
+        p_document,
+        p_metadata,
+        p_embedding,
+        p_content_hash,
+        p_embedding_model,
+        p_embedding_version,
+        p_is_active,
+        now()
+    )
+    ON CONFLICT (external_id) DO UPDATE SET
+        document = EXCLUDED.document,
+        metadata = EXCLUDED.metadata,
+        embedding = EXCLUDED.embedding,
+        content_hash = EXCLUDED.content_hash,
+        embedding_model = EXCLUDED.embedding_model,
+        embedding_version = EXCLUDED.embedding_version,
+        is_active = EXCLUDED.is_active,
+        updated_at = now();
+$$;
+
+CREATE OR REPLACE FUNCTION public.upsert_post_embedding(
+    p_external_id text,
+    p_document text,
+    p_metadata jsonb,
+    p_embedding vector(300),
+    p_content_hash text,
+    p_embedding_model text,
+    p_embedding_version text,
+    p_is_active boolean
+)
+RETURNS void
+LANGUAGE sql
+SECURITY DEFINER
+SET search_path = public
+AS $$
+    INSERT INTO public.post_embeddings (
+        external_id,
+        document,
+        metadata,
+        embedding,
+        content_hash,
+        embedding_model,
+        embedding_version,
+        is_active,
+        updated_at
+    )
+    VALUES (
+        p_external_id,
+        p_document,
+        p_metadata,
+        p_embedding,
+        p_content_hash,
+        p_embedding_model,
+        p_embedding_version,
+        p_is_active,
+        now()
+    )
+    ON CONFLICT (external_id) DO UPDATE SET
+        document = EXCLUDED.document,
+        metadata = EXCLUDED.metadata,
+        embedding = EXCLUDED.embedding,
+        content_hash = EXCLUDED.content_hash,
+        embedding_model = EXCLUDED.embedding_model,
+        embedding_version = EXCLUDED.embedding_version,
+        is_active = EXCLUDED.is_active,
+        updated_at = now();
+$$;
+
+CREATE OR REPLACE FUNCTION public.get_place_content_hashes(
+    p_external_ids text[]
+)
+RETURNS TABLE (
+    external_id text,
+    content_hash text
+)
+LANGUAGE sql
+STABLE
+SECURITY DEFINER
+SET search_path = public
+AS $$
+    SELECT place.external_id, place.content_hash
+    FROM public.place_embeddings place
+    WHERE place.external_id = ANY(p_external_ids);
+$$;
+
+CREATE OR REPLACE FUNCTION public.get_post_content_hashes(
+    p_external_ids text[]
+)
+RETURNS TABLE (
+    external_id text,
+    content_hash text
+)
+LANGUAGE sql
+STABLE
+SECURITY DEFINER
+SET search_path = public
+AS $$
+    SELECT post.external_id, post.content_hash
+    FROM public.post_embeddings post
+    WHERE post.external_id = ANY(p_external_ids);
+$$;
+
+-- Reaplica permisos solo cuando los roles ya existen.
+DO $$
+BEGIN
+    IF EXISTS (SELECT 1 FROM pg_roles WHERE rolname = 'nlp_reader') THEN
+        EXECUTE 'GRANT USAGE ON SCHEMA public TO nlp_reader';
+        EXECUTE 'GRANT EXECUTE ON FUNCTION public.match_places(vector, integer, jsonb) TO nlp_reader';
+        EXECUTE 'GRANT EXECUTE ON FUNCTION public.match_posts(vector, integer, jsonb) TO nlp_reader';
+    END IF;
+
+    IF EXISTS (SELECT 1 FROM pg_roles WHERE rolname = 'nlp_writer') THEN
+        EXECUTE 'GRANT USAGE ON SCHEMA public TO nlp_writer';
+        EXECUTE 'GRANT EXECUTE ON FUNCTION public.get_place_content_hashes(text[]) TO nlp_writer';
+        EXECUTE 'GRANT EXECUTE ON FUNCTION public.get_post_content_hashes(text[]) TO nlp_writer';
+        EXECUTE 'GRANT EXECUTE ON FUNCTION public.upsert_place_embedding(text, text, jsonb, vector, text, text, text, boolean) TO nlp_writer';
+        EXECUTE 'GRANT EXECUTE ON FUNCTION public.upsert_post_embedding(text, text, jsonb, vector, text, text, text, boolean) TO nlp_writer';
+    END IF;
+END $$;
+
+COMMIT;
+
+-- Resultado esperado inmediatamente despues de migrar: dimension 300 y cero
+-- filas. Las filas apareceran despues de ejecutar los scripts de Colab.
+SELECT
+    table_name,
+    column_name,
+    udt_name,
+    format_type(attribute.atttypid, attribute.atttypmod) AS formatted_type
+FROM information_schema.columns column_info
+JOIN pg_attribute attribute
+  ON attribute.attrelid = (
+      quote_ident(column_info.table_schema) || '.' || quote_ident(column_info.table_name)
+  )::regclass
+ AND attribute.attname = column_info.column_name
+WHERE column_info.table_schema = 'public'
+  AND column_info.table_name IN ('place_embeddings', 'post_embeddings')
+  AND column_info.column_name = 'embedding'
+ORDER BY table_name;
+
+SELECT 'place_embeddings' AS table_name, count(*) AS rows
+FROM public.place_embeddings
+UNION ALL
+SELECT 'post_embeddings' AS table_name, count(*) AS rows
+FROM public.post_embeddings;
+