For AI agents: The documentation index is at https://docs.digitalocean.com/llms.txt. Markdown versions of pages use the same URL with index.html.md in place of the HTML page (for example, append index.html.md to the directory path instead of opening the HTML document).

DigitalOcean Agent Data Services (ADS) processes your data once, when it arrives, and stores the results so that your agents and applications can retrieve them. Data flows extract, transcribe, enrich, redact, and embed incoming data, and a catalog keeps every result alongside its source, its attributes, and the access rules that govern it.

Because processing happens at ingest instead of at query time, a query reads stored results and does not run models again. This keeps retrieval fast and avoids paying for the same inference on every request.

Workspaces

A workspace is an isolated environment that holds your data, data flows, catalog, and query endpoint. Each workspace belongs to a team and lives in a single region. Your data is isolated from other customers’ data at rest and while it is being served, and it is never pooled with another customer’s data.

You can use multiple workspaces to separate environments, such as development and production. Access to query a workspace can be granted per workspace. For details, see Agent Data Services Access Control.

The API uses earlier names in endpoint paths, request fields, and token scopes. It uses sandbox for a workspace, such as /sandboxes, sandbox_uuids, and aidc_sandbox:read, and pipeline for a data flow, such as aidc_pipeline:create.

Workloads Supported During Private Preview

Agent Data Services supports two workloads during private preview: behavioral data and multimodal content. Both workloads use the same data flows, catalog, and query layer.

Behavioral Data

Behavioral data is the record of what users and agents saw, decided, and did, such as events, logs, click events, and agent traces. You can send your own event streams to Agent Data Services, which assembles these records into sessions, redacts sensitive fields, and stores the results as tables you can query. You can look up any session by its identifier to return the complete interaction instead of scattered raw events.

Because sessions land in the same catalog as your other data, you can join behavioral data with the content it refers to. For example, you can retrieve what a user did alongside the documents or videos they interacted with, and use that data to train recommendation models or prepare fine-tuning datasets.

Multimodal Content

Multimodal content includes documents, images, audio, and video. Agent Data Services turns each asset into AI-ready outputs, including extracted text, chunks, transcripts, scene descriptions, on-screen text, categories, moderation flags, and embeddings. Each output stays linked to the asset it came from and to the structured attributes that describe it, such as IDs, categories, and transcript segments.

For example, a query for clips where a customer talks about pricing returns ranked video segments with timestamps.

Data Sources and Ingestion

During private preview, Agent Data Services processes data that you bring into the platform. You can ingest data in the following ways:

• Connect an object storage bucket: Connect a Spaces bucket or other S3-compatible bucket. New objects that arrive in the bucket trigger your data flows automatically.
• Process an existing bucket: Point Agent Data Services at a path in an existing bucket, including an AWS S3 bucket, with read credentials. Agent Data Services fully processes and catalogs everything under that path without migrating the data.
• Upload through the API: Send data directly to your workspace using the API.

Data Flows

A data flow defines how Agent Data Services processes data in a workspace. You build a data flow by chaining built-in functions in the order you choose. Built-in functions include transcription, chunking, embedding, redaction, field extraction, session assembly, and aggregation.

Data flows follow a draft, test, and promote workflow:

1. Draft a data flow from built-in functions.
2. Test the data flow on sample data and inspect the results.
3. Promote the data flow to production. A promoted data flow runs on every new arrival without manual action.

For behavioral data, a data flow assembles events into sessions and processes each session when it closes. For multimodal content, a data flow extracts, transcribes, chunks, and enriches each asset as it arrives.

Data flows deliver records at least once, so a record can be processed more than once after a retry. To process more data in parallel, you can scale a data flow horizontally to up to 40 replicas, which divide queued work between them.

Models Used in Data Flows

Data flow steps that require a model, such as transcription, vision, and embedding, can use any model available through serverless inference on DigitalOcean Inference. You do not need to provision GPUs or set up a separate inference service.

Catalog and Lineage

Each workspace has one catalog that holds all of its content, embeddings, structured attributes, and access rules. Every data flow output lands in the catalog, and every query runs against it.

Agent Data Services stores structured and semi-structured data, such as attributes and assembled sessions, in Apache Iceberg tables, an open table format. Embeddings are stored in a managed vector database. You do not access either store directly. All queries go through the Agent Data Services query layer.

The catalog also records lineage, which traces each piece of data back to the raw arrival it came from and through each transformation applied to it. You can use lineage to find where a result came from and which data flow produced it.

Queries

Agent Data Services has a unified query layer that reads across the tables and embeddings in a workspace. During private preview, you can send queries to it in the following ways:

• API: Send queries to the Agent Data Services API.
• Control Panel: Run queries in the DigitalOcean Control Panel.
• Query credentials: Connect to the query endpoint for a workspace using its query credentials.

The query layer supports retrieval and analytical queries:

• Search: Semantic and keyword search together, filtered by structured attributes.
• Lookup: Retrieve a full assembled session or record by its ID.
• Filtered retrieval with aggregations: Retrieve records that match filters, with aggregations over a time window.
• Analytical queries: Run read-only analytical queries across the structured and semi-structured data in a workspace.

All queries are read-only. You cannot create, update, or delete data with a query. Data flows are the only way to write data to a workspace.

Governance and Data Security

Agent Data Services enforces access control, redaction, and moderation when data is processed and queried, not after results are returned.

• Access control: DigitalOcean IAM governs who can manage workspaces and data flows and who can query them. The same rules apply whether a person, an application, or an agent makes the request.
• Redaction: Data flows can redact or mask sensitive fields before anything is embedded.
• Moderation and usage rules: Moderation flags and licensing or access rules attached to an asset are applied during queries, so results that a caller is not permitted to see are not returned.
• Encryption: Data is encrypted in transit and at rest.
• Model training: Your data is never used to train DigitalOcean models. Models that process your data read it only to produce your outputs.

Agent Data Services and Knowledge Bases

Knowledge Bases and Agent Data Services both prepare your data for AI retrieval, but they cover different scopes:

• Use Knowledge Bases when you need purpose-built vector search over documents, such as files, websites, and Spaces buckets, for your agents or applications.
• Use Agent Data Services when you want to manage an entire data workload, from ingestion to queries, across structured data and unstructured content, including documents, audio, video, and images.