External Retrieval Interface (ERI)¶
About¶
The ERI acts as an interface, transforming a data source into an accessible storage for retrieval and other llm adjacent processes. It serves as a provider of helpful information to a given query. Published in 2025 by Hecking, Sommer, and Felderer, this standard enables decentralized data storage, allowing external data to be integrated into AI workflows. To use this service, the ERI-Standard defines methods for retrieving data from the source. These can be implemented in other workflows or integrated directly into LLM processes, as done by AI Studio.
Within this library, an ERI-Server is implemented, enabling users to configure and select specific retrieval algorithms, which are then accessible through an API.
The ERI-Standard¶
The ERI-Standard defines various requests for server-client communication. The full documentation and specification are available in the published paper and in the official server implementation repository on GitHub.
For authentification purpose, users can fetch available methods using the /auth/methods path. These can the be used to retrieve a session token when fulfilling authentification at /auth.
The provided ERI-Server implementation allows for either NONE or TOKEN authentification. NONE provides free access to session tokens for all users, while TOKEN compares incoming Bearer authentication with a database of valid API keys.
To fetch general information about the provided data, the /datasource endpoint provides a set name and description of the source.
Information about available retrieval algorithms can be accessed under the /retrieval/info endpoint. Here, a short description of the processes and available parameters is provided. These methods can then be used by providing a query to the /retrieval endpoint. This executes the selected algorithm internally, returning the selected data to the user.
Starting the ERI-Server¶
To start the ERI-Server, the following command can be used, executed in the raglib subdirectory: uvicorn eri:app
To configure the ERI-Server, read the configuration documentation.