Deep Research — Unique AI Documentation
Deep Research
This feature is EXPERIMENTAL and under active development. It may change significantly, be discontinued, or have breaking changes without notice. Documentation may be incomplete or outdated and is NOT recommended for production use. Use at your own risk. Please refer to our Upgrade and Release Process for more information.
Functionality
The Deep Research Tool is an advanced AI-powered research assistant, or what is also referred to as a deep agent, that performs in-depth analysis and investigation across multiple sources to answer complex research questions that require longer and deeper analysis. This is usually analysis jobs running 10-15 minutes and reviewing 100+ sources.
Use Cases
The Deep Research Tool is ideal for:
- Complex Research Questions: Questions requiring multiple steps and synthesis from multiple sources
- Example: "What are the current trends in AI safety research and how do different organizations approach it?"
- Competitive Analysis: Gathering information about competitors, market trends, or industry analysis
- Example: "Compare the product strategies of the top 5 CRM vendors"
- Literature Review: Summarizing research papers, articles, or documentation on a topic
- Example: "Provide an overview of recent advances in quantum computing"
- Policy Research: Understanding regulations, compliance requirements, or internal policies
- Example: "What are the GDPR requirements for data processing and how do they apply to our use case?"
- Technical Investigation: Deep dives into technical topics requiring multiple sources
- Example: "Explain the architecture and tradeoffs of different vector database implementations"
- Market Research: Analyzing market conditions, customer sentiment, or industry trends
- Example: "What are enterprises looking for in observability platforms in 2025?"
Not recommended for:
- Simple factual queries (use standard chat)
- Real-time information updates (though web search provides current data)
- Questions answerable from a single source
Research Engines
The deep research tool requires an “Engine” that powers the research and has access to all the tools and capabilities required for the research. The “Engine” is the underlying agent that guides the agent’s decision making. Unique provides two Deep Research engines:
- Unique Engine - A custom multi-agent research engine that orchestrates multiple specialized agents
- OpenAI Engine - The OpenAI native deep research capability exposed by OpenAI via the Responses API. This engine requires access to OpenAI models hosted directly on OpenAI servers (e.g.,
litellm:openai-gpt-5).
Unique Engine (Default)
A custom multi-agent research engine that orchestrates multiple specialized agents. Users can leverage external and internal data sources to generate comprehensive answers to their research questions. The Unique Deep Research Engine provides flexible model selection allowing admins to carefully control their data security and privacy.
Available Tools:
- Web Tools: Search the web and fetch content from URLs with automatic citations (Google Engine)
- Internal Tools: Search company knowledge base and fetch internal documents including looking at uploaded documents.
- MCP (In progress): Factset, Outlook, and custom servers.
- Connectors (In progress): Quartr, SIX, and custom connectors.
- Code Execution (In progress): Run code in a secure container to perform discrete calculations, generate charts, graphs and data files.
Deep research tools can be enabled and disabled, but they are currently not configurable. The Unique engine is limited to 300 tool calls and at most 25 different research directions, allowing for thorough multi-step investigation while keeping token usage limited.
OpenAI Engine
The OpenAI Engine connects to the deep research capability exposed by OpenAI via the Responses API. This engine requires access to OpenAI models hosted directly on OpenAI servers. Users will not have access to internal data sources or control over model hosting location when choosing this option.
Available Tools:
- Web Search: Uses OpenAI's native
web_search_previewtool.
Performance and usage cost
Expected token consumption for the deep research can vary drastically depending on the model, tool configuration, and question asked.
Below is provided a reference for expected token consumption of deep research in different configurations.
| Engine | Research model | Expected consumption (tokens) | Expected runtime (minutes) | RACE score (higher is better) | LLM Approximate cost pr. run |
|---|---|---|---|---|---|
| Unique | AZURE_GPT_41_2025_0414 |
200K - 400k | 4-15 | 44.71 |
1.00$ |
| Unique | AZURE_GPT_4o_2024_1120 |
100K - 300k | 4-15 | 42.98 |
0.31$ |
| Unique | AZURE_GPT_5_2025_0807 |
200K - 500k | 4-15 | 48.81 |
1.50$ |
| Unique | litellm:anthropic-claude-sonnet-4-5 |
500K - 1M | 5-20 | 46.80 |
3.00$ |
| OpenAI | litellm:openai-gpt-5 |
100-200K | 10-15 | Please refer to OpenAI's documentation | 0.38$ |
| OpenAI | openai-o4-mini-deep-research |
100-200K | 15-30 | 46.25 |
0.30$ |
| OpenAI | openai-o3-deep-research |
200K-400K | 15-30 | 43.49 |
4.00$ |
Configuring Deep Research
Deep research is a tool available for Unique AI Chat and will, once enabled, appear in the tool section of Unique AI Space.
The tool must be set to with is exclusive checked to achieve the desired behavior of the agent. The tool has the following configurable fields:
| Field Name | Description | Type | Default Value | Engines |
|---|---|---|---|---|
engine |
The research engine to use. Options are Unique Engine and Open Ai Engine |
String (enum) | Unique Engine |
All |
engine.engine_type |
Name of the engine type. Should not be changed | String (enum) | Unique |
All |
engine.small_model |
Fast model for less demanding tasks where speed is more important than intelligence | LMI object | AZURE_GPT_4o_2024_1120 |
All |
engine.large_model |
Larger model with extended context used for synthesizing the findings and should preferably have a large context window | LMI object | AZURE_GPT_41_2025_0414 |
All |
engine.research_model |
Main research model “powering” the research | LMI object | AZURE_GPT_5_2025_0807 (Unique) / litellm:OPENAI_GPT_5 (OpenAI) |
All |
Feature flags and environment variables
Deep research requires the enablement of 2 feature flags and has additional system controls to limit the amount of runs that can be executed at any time.
| Variable | Default | Description |
|---|---|---|
FEATURE_FLAG_ENABLE_DEEP_RESEARCH_UN_12630 |
false |
Controls the visibility of the deep research tool in Unique AI Space |
DEEP_RESEARCH_MAX_EXECUTION_TIME_IN_MINUTES |
15 |
The time limit before the system automatically assumes the research has failed and sets the state to failed |
DEEP_RESEARCH_MAX_PARALLEL_EXECUTIONS_PER_COMPANY |
1 |
The number of deep research jobs that can run concurrently. Additional jobs will be queued and start when there is an available slot |
Unique deep research agent architecture
Overview
The Deep Research system employs a multi-agent architecture with nested feedback loops to conduct comprehensive research autonomously. The design mimics how human research teams operate: a lead researcher coordinates the investigation while specialized team members focus on specific topics, then findings are synthesized into a cohesive final report.
The system begins with an optional clarification phase where the agent asks follow-up questions to better understand the user's request. Once clarity is established, a comprehensive research brief is generated.
The core research process operates through two nested agent loops:
- Lead Supervisor Agent: A lead supervisor agent continuously analyzes research progress and makes strategic decisions. In each iteration, it can choose to (1) use internal reasoning to refine its strategy, (2) spawn new research sub-agents to investigate specific topics, or (3) declare research complete.
- Research Sub-Agent: Each research sub-agent operates independently in its own loop, focusing on a specific assigned topic. Sub-agents execute tools to search the web, fetch web pages, search internal knowledge bases, and retrieve documents. After each tool execution, the agent decides whether to continue investigating or conclude its research.
The critical feedback mechanism enables the supervisor to receive findings from completed sub-agents, incorporate them into its knowledge state, and spawn additional researchers if gaps remain. Multiple sub-agents can execute in parallel, enabling efficient multi-topic research.
Finally, a final report generation phase synthesizes all accumulated findings from multiple research iterations into a comprehensive, well-formatted report with validated citations.