Agentic API merges file search with citations and saved context Agentic API merges file search with citations and saved context

Retrieval joins the conversation

The merge joins two pieces that applications previously had to connect. The gateway can execute a file search, pass retrieved passages to the generation model and return an answer with file citations. It also saves the model-facing search exchange for a later request using previous_response_id. That matters when a follow-up refers to material found earlier. A client can keep the conversation going through the same API, instead of replaying its own reconstruction of a tool call. Public responses contain a file_search_call item, while the passage list is an optional addition. Citations can survive both completed responses and streaming output.

The storage routes arrived first

A November 7 merge exposed the upload, attachment, search and deletion routes that sit beneath the tool. It connected those routes to the gateway’s authentication configuration and request-size limits. The maintainers report testing the lifecycle with version 3.8.0 of the OpenAI Python client against a locally rebuilt gateway. That exercise covered uploading a file, attaching it to a vector store, searching with filters and deleting it. This is useful compatibility work for applications already organized around those operations. The new Responses integration adds the next layer: the gateway can conduct the search during a model turn, alongside functions executed by the client.

A local index has a deployment cost

The documented storage design separates original file bytes from searchable metadata and vectors. Uploaded bytes live in a local directory, while SQLite or PostgreSQL stores the database records. Moving an installation therefore involves both the database and its file directory. Multiple replicas need access to the same stored material. Without an embedding endpoint, retrieval uses keywords; configuring embeddings enables semantic and hybrid search. This makes deployment choices visible in application behavior. A developer cannot assume that selecting a generation model also supplies the separate embedding service. Changing a store’s embedding configuration requires creating a new store and ingesting the documents again.

Search is now a server responsibility

Agentic API is the stateful gateway in front of inference, rather than the model itself. Its repository describes moving conversation state, tool loops and continuation out of application clients. Adding document search follows that design: the server owns another multi-step operation that can otherwise be repeated across every application. The practical gain is a smaller amount of retrieval coordination in client code. The operational work moves with it. A service owner still needs storage that survives restarts, access rules for the gateway and tests that exercise a conversation after its first answer. The change is most concrete for teams running their own inference stack and wanting an existing Responses-style application to search its documents.