Postgres

Postgres Database

R2R uses Postgres as the sole provider for relational and vector search queries. This means that Postgres is involved in handling authentication, document management, and search across R2R. For robust search capabilities, R2R leverages the pgvector extension and ts_rank to implement customizable hybrid search.

Postgres Configuration

To customize the database settings, you can modify the database section in your r2r.toml file and set corresponding environment variables or provide the settings directly in the configuration file.

1. Edit the database section in your r2r.toml file:

1
[database]
2
provider = "postgres" # currently only `postgres` is supported
3
4
# optional parameters which are typically set in the environment instead:
5
user = "your_postgres_user"
6
password = "your_postgres_password"
7
host = "your_postgres_host"
8
port = "your_postgres_port"
9
db_name = "your_database_name"
10
your_project_name = "your_project_name"

2. Alternatively, you can set the following environment variables:

$
export R2R_POSTGRES_USER=your_postgres_user
>
export R2R_POSTGRES_PASSWORD=your_postgres_password
>
export R2R_POSTGRES_HOST=your_postgres_host
>
export R2R_POSTGRES_PORT=your_postgres_port
>
export R2R_POSTGRES_DBNAME=your_database_name
>
export R2R_PROJECT_NAME=your_project_name

Advanced Postgres Features in R2R

R2R leverages several advanced Postgres features to provide powerful search and retrieval capabilities:

pgvector Extension

R2R uses the pgvector extension to enable efficient vector similarity search. This is crucial for semantic search operations. The collection.py file defines a custom Vector type that interfaces with pgvector:

1
class Vector(UserDefinedType):
2
    # ... (implementation details)
3
4
    class comparator_factory(UserDefinedType.Comparator):
5
        def l2_distance(self, other):
6
            return self.op("<->", return_type=Float)(other)
7
8
        def max_inner_product(self, other):
9
            return self.op("<#>", return_type=Float)(other)
10
11
        def cosine_distance(self, other):
12
            return self.op("<=>", return_type=Float)(other)

This allows R2R to perform efficient vector similarity searches using different distance measures.

Hybrid Search

R2R implements a sophisticated hybrid search which combines full-text search and vector similarity search. This approach provides more accurate and contextually relevant results. Key components of the hybrid search include:

1. Full-Text Search: Utilizes Postgres’s built-in full-text search capabilities with ts_rank and websearch_to_tsquery.
2. Semantic Search: Performs vector similarity search using pgvector.
3. Reciprocal Rank Fusion (RRF): Merges results from full-text and semantic searches.

In addition, R2R offers robust logical filters on metadata which include operations like eq, neq, gt, gte, lt, lte, like, ilike, in, and nin. Refer to the retrieval API documentation for all available inputs.

Indexing

Vector similarity search

R2R supports two primary indexing methods for vector similarity search through pgvector: HNSW (Hierarchical Navigable Small World) and IVF-Flat (Inverted File with Flat Storage).

HNSW offers faster search times and better recall but requires more memory and slower build times, making it ideal for production environments where query speed is critical. IVF-Flat provides a balanced approach with faster index construction and lower memory usage, suitable for scenarios requiring a trade-off between build speed and query performance.

Both methods support cosine, L2, and inner product distance measures. See the index API Reference for detailed configuration options and management endpoints.

Full-text search

R2R uses GIN (Generalized Inverted Index) indexing to optimize full-text searches:

1
Index(f"idx_{name}_fts", "fts", postgresql_using="gin"),

This indexing strategy allows for efficient full-text search.

JSON Support

R2R leverages Postgres’s JSONB type for flexible metadata storage:

1
Column(
2
    "metadata",
3
    postgresql.JSONB,
4
    server_default=text("'{}'::jsonb"),
5
    nullable=False,
6
)

This allows for efficient storage and querying of structured metadata alongside vector embeddings.

Performance Considerations

When setting up Postgres for R2R, consider the following performance optimizations:

1. Indexing: Ensure proper indexing for both full-text and vector searches. R2R automatically creates necessary indexes, but you may need to optimize them based on your specific usage patterns.

2. Hardware: For large-scale deployments, consider using dedicated Postgres instances with sufficient CPU and RAM to handle vector operations efficiently.

3. Vacuuming: Regular vacuuming helps maintain database performance, especially for tables with frequent updates or deletions.

4. Partitioning: For very large datasets, consider table partitioning to improve query performance.

By leveraging these advanced Postgres features and optimizations, R2R provides a powerful and flexible foundation for building sophisticated retrieval and search systems.