DIAL Deep Research
DIAL Deep Research prepares a research plan, searches configured sources, and produces a report with citations. This reference covers backend 0.4.0 for platform engineers familiar with DIAL Core and application types. Each application instance has its own sources, prompts, and research limits.
Research workflow
| Phase | Behavior |
|---|---|
| Preparation | Clarifies the question, prepares a plan, and waits for approval before starting research |
| Retrieval | Calls tools from the configured MCP servers and reviews gaps in the findings |
| Report | Writes a report, reviews it, and revises it within the configured version budget |
| Citations | Resolves document, dataset, and data-query citations and appends a References section |
| Progress | Streams work and review progress through DIAL stages |
The preparation tools require an approved plan. The research graph then runs retrieval, research review, report writing, and report review within the same turn.
The application retrieves evidence from its configured sources. Its prompt rules prohibit deriving figures through calculations. For Python calculations in a Quick App, see Code Interpreter.
How the service fits into DIAL
DIAL Core routes requests to Deep Research. Generic RAG's MCP interface supplies document retrieval and citation tools. Each MCP connection uses either a Core deployment ID or a direct URL/key bundle; see Connection modes.
The supported retrieval types are generic_rag for documents and statgpt for datasets. At least one server is required, with at most one server of each type. Arbitrary MCP servers are not interchangeable here: citation formats depend on the supported retrieval types. See the server types and uniqueness checks.
Dependencies and deployment
| Component | Responsibility |
|---|---|
| DIAL Core | Routes application requests, stores application properties, and provides model and file APIs. The backend's schema integration requires Core 0.41.0 or later. |
| Model deployment | Runs preparation, retrieval orchestration, and report generation. Deployment IDs can be remapped through environment variables. |
| Retrieval server | Provides the configured Generic RAG document tools, StatGPT dataset tools, or both. |
| DIAL Chat or API client | Preserves conversation state and presents the plan, progress, report, and citations. |
The application configuration model states the Core version requirement. Model defaults and overrides come from the model configuration.
The published image is epam/ai-dial-deep-research-backend:0.4.0. It runs as UID 1001 and exposes port 5000; see the Dockerfile. Deployment settings and Core registration are described in Configuration.
No DIAL platform release through 1.48 pins Deep Research. Version 0.4.0 is from 2026-10-27, and the DIAL platform chart 8.1.0 has no Deep Research dependency.
Preparation state travels in DIAL conversation messages rather than a server-side checkpointer. The backend can therefore run with multiple workers or replicas, provided the client preserves that state. See the startup implementation and API state contract.