RAG with Permissions

Since pgvector is built on top of Postgres, you can implement fine-grain access control on your vector database using Row Level Security (RLS). This means you can restrict which documents are returned during a vector similarity search to users that have access to them. Datafuse also supports Foreign Data Wrappers (FDW) which means you can use an external database or data source to determine these permissions if your user data doesn't exist in Datafuse.

Use this guide to learn how to restrict access to documents when performing retrieval augmented generation (RAG).

Example

In a typical RAG setup, your documents are chunked into small subsections and similarity is performed over those sections:

-- Track documents/pages/files/etc
create table documents (
  id bigint primary key generated always as identity,
  name text not null,
  owner_id uuid not null references auth.users (id) default auth.uid(),
  created_at timestamp with time zone not null default now()
);

-- Store the content and embedding vector for each section in the document
-- with a reference to original document (one-to-many)
create table document_sections (
  id bigint primary key generated always as identity,
  document_id bigint not null references documents (id),
  content text not null,
  embedding vector (384)
);

Notice how we record the owner_id on each document. Let's create an RLS policy that restricts access to document_sections based on whether or not they own the linked document:

-- enable row level security
alter table document_sections enable row level security;

-- setup RLS for select operations
create policy "Users can query their own document sections"
on document_sections for select to authenticated using (
  document_id in (
    select id
    from documents
    where (owner_id = (select auth.uid()))
  )
);

Now every select query executed on document_sections will implicitly filter the returned sections based on whether or not the current user has access to them.

For example, executing:

select * from document_sections;

as an authenticated user will only return rows that they are the owner of (as determined by the linked document). More importantly, semantic search over these sections (or any additional filtering for that matter) will continue to respect these RLS policies:

-- Perform inner product similarity based on a match_threshold
select *
from document_sections
where document_sections.embedding <#> embedding < -match_threshold
order by document_sections.embedding <#> embedding;

The above example only configures select access to users. If you wanted, you could create more RLS policies for inserts, updates, and deletes in order to apply the same permission logic for those other operations. See Row Level Security for a more in-depth guide on RLS policies.

Alternative scenarios

Every app has its own unique requirements and may differ from the above example. Here are some alternative scenarios we often see and how they are implemented in Datafuse.

Documents owned by multiple people

Instead of a one-to-many relationship between users and documents, you may require a many-to-many relationship so that multiple people can access the same document. Let's reimplement this using a join table:

create table document_owners (
  id bigint primary key generated always as identity,
  owner_id uuid not null references auth.users (id) default auth.uid(),
  document_id bigint not null references documents (id)
);

Then your RLS policy would change to:

create policy "Users can query their own document sections"
on document_sections for select to authenticated using (
  document_id in (
    select document_id
    from document_owners
    where (owner_id = (select auth.uid()))
  )
);

Instead of directly querying the documents table, we query the join table.

User and document data live outside of Datafuse

You may have an existing system that stores users, documents, and their permissions in a separate database. Let's explore the scenario where this data exists in another Postgres database. We'll use a foreign data wrapper (FDW) to connect to the external DB from within your Datafuse DB:

Let's assume your external DB contains a users and documents table like this:

create table public.users (
  id bigint primary key generated always as identity,
  email text not null,
  created_at timestamp with time zone not null default now()
);

create table public.documents (
  id bigint primary key generated always as identity,
  name text not null,
  owner_id bigint not null references public.users (id),
  created_at timestamp with time zone not null default now()
);

In your Datafuse DB, let's create foreign tables that link to the above tables:

create schema external;
create extension postgres_fdw with schema extensions;

-- Setup the foreign server
create server foreign_server
  foreign data wrapper postgres_fdw
  options (host '<db-host>', port '<db-port>', dbname '<db-name>');

-- Map local 'authenticated' role to external 'postgres' user
create user mapping for authenticated
  server foreign_server
  options (user 'postgres', password '<user-password>');

-- Import foreign 'users' and 'documents' tables into 'external' schema
import foreign schema public limit to (users, documents)
  from server foreign_server into external;

We'll store document_sections and their embeddings in Datafuse so that we can perform similarity search over them via pgvector.

create table document_sections (
  id bigint primary key generated always as identity,
  document_id bigint not null,
  content text not null,
  embedding vector (384)
);

We maintain a reference to the foreign document via document_id, but without a foreign key reference since foreign keys can only be added to local tables. Be sure to use the same ID data type that you use on your external documents table.

Since we're managing users and authentication outside of Datafuse, we have two options:

1. Make a direct Postgres connection to the Datafuse DB and set the current user every request
2. Issue a custom JWT from your system and use it to authenticate with the REST API

Direct Postgres connection

You can directly connect to your Datafuse Postgres DB using the connection info on your project's database settings page. To use RLS with this method, we use a custom session variable that contains the current user's ID:

-- enable row level security
alter table document_sections enable row level security;

-- setup RLS for select operations
create policy "Users can query their own document sections"
on document_sections for select to authenticated using (
  document_id in (
    select id
    from external.documents
    where owner_id = current_setting('app.current_user_id')::bigint
  )
);

The session variable is accessed through the current_setting() function. We name the variable app.current_user_id here, but you can modify this to any name you like. We also cast it to a bigint since that was the data type of the user.id column. Change this to whatever data type you use for your ID.

Now for every request, we set the user's ID at the beginning of the session:

set app.current_user_id = '<current-user-id>';

Then all subsequent queries will inherit the permission of that user:

-- Only document sections owned by the user are returned
select *
from document_sections
where document_sections.embedding <#> embedding < -match_threshold
order by document_sections.embedding <#> embedding;

Custom JWT with REST API

If you would like to use the auto-generated REST API to query your Datafuse database using JWTs from an external auth provider, you can get your auth provider to issue a custom JWT for Datafuse.

See the Clerk Datafuse docs for an example of how this can be done. Modify the instructions to work with your own auth provider as needed.

Now we can simply use the same RLS policy from our first example:

-- enable row level security
alter table document_sections enable row level security;

-- setup RLS for select operations
create policy "Users can query their own document sections"
on document_sections for select to authenticated using (
  document_id in (
    select id
    from documents
    where (owner_id = (select auth.uid()))
  )
);

Under the hood, auth.uid() references current_setting('request.jwt.claim.sub') which corresponds to the JWT's sub (subject) claim. This setting is automatically set at the beginning of each request to the REST API.

All subsequent queries will inherit the permission of that user:

-- Only document sections owned by the user are returned
select *
from document_sections
where document_sections.embedding <#> embedding < -match_threshold
order by document_sections.embedding <#> embedding;

Other scenarios

There are endless approaches to this problem based on the complexities of each system. Luckily Postgres comes with all the primitives needed to provide access control in the way that works best for your project.

If the examples above didn't fit your use case or you need to adjust them slightly to better fit your existing system, feel free to reach out to support and we'll be happy to assist you.