Deep Research: A Systematic Survey of RAG-to-Agency Evolution

Core Insight

Why Previous Approaches Failed

Standard RAG architectures have fundamental structural limitations:

1. Fixed Retrieval Pipeline

Retrieve → Generate is a one-shot pipeline. If the first retrieval misses relevant documents, you're stuck. No mechanism for:

• Iterative refinement of queries
• Following up on partial information
• Recognizing retrieval failure and trying different approaches

2. No Query Planning

Complex questions require strategic decomposition. "What were the causes and consequences of the 2008 financial crisis?" needs multiple sub-queries, but RAG systems treat it as a single retrieval.

3. Always-Retrieve Fallacy

Standard RAG retrieves on every query, even when the model already knows the answer. Counterintuitively, irrelevant documents actively mislead LLMs. If you retrieve documents that mention similar terms but describe different concepts, the model gets confused. More retrieval ≠ better answers.

4. Context Window as Memory

Long research tasks (hours/days of work) exceed context limits. No mechanism for:

• Consolidation: Compressing findings into summaries
• Indexing: Organizing by topic for fast lookup
• Forgetting: Discarding outdated/irrelevant information

5. No Verification

Generated answers lack citations or evidence chains. Users can't audit whether claims are supported by sources or hallucinated.

The Method

The paper formalizes Deep Research as a four-component framework:

1. Query Planning

Decompose complex questions into sub-queries using one of three strategies:

• Parallel: Independent sub-questions executed simultaneously (e.g., "Who are the main characters?" and "What is the setting?")
• Sequential: Each query depends on previous answers (e.g., "Who is the CEO?" → "What did [CEO name] say about...?")
• Tree-based: Hierarchical decomposition with branching based on intermediate results

The planning strategy should match question structure—wrong decomposition causes cascading errors.

2. Information Acquisition

Three sub-components:

• Retrieval tools: Search engines, databases, APIs, code execution
• Retrieval timing: Adaptive—retrieve based on confidence, not every step
• Information filtering: Score retrieved documents for relevance, discard noise

The threshold τ is learned or tuned per domain. High-stakes domains (medical, legal) should retrieve more aggressively; general knowledge can rely on the model.

3. Memory Management

Handle long-horizon tasks that exceed context:

• Consolidation: Compress retrieved info into summaries. "These 5 documents all say X" → store "X" once.
• Indexing: Organize by topic/relevance for fast lookup. Don't linear-scan all memory on every query.
• Updating: Revise stored info when new evidence arrives. Don't keep stale facts.
• Forgetting: Actively discard outdated/irrelevant information. Memory has a cost.

4. Answer Generation

Synthesize findings into verifiable outputs:

• Explicit citations pointing to sources
• Evidence chains showing reasoning steps
• Confidence scores on claims
• Acknowledgment of gaps/uncertainties

Architecture

Key Equations

This is a survey paper—it systematizes existing work rather than proposing new equations. The key conceptual contribution is the four-component framework and capability taxonomy.

Results

This is a survey, not an empirical paper. Key findings synthesized from reviewed literature:

• Adaptive retrieval timing improves accuracy 5-15% over always-retrieve baselines
• Memory consolidation enables tasks 10× longer than context window
• Tree-based query planning best for complex multi-hop questions
• Citation generation reduces hallucination rates by 30-50%
• Forgetting mechanisms prevent memory bloat and conflicting information

Capability Progression

1. Agentic Search: Fact-finding with tools (current state for most systems)
2. Integrated Research: Multi-source synthesis with coherent reports (emerging)
3. AI Scientist: Hypothesis generation and validation (research frontier)

What Actually Matters

Key insights from surveyed papers:

Retrieval noise hurts more than no retrieval. Adding irrelevant documents degrades performance more than not retrieving at all. The model gets confused by plausible-sounding but incorrect information.

Forgetting is necessary. Without selective memory deletion, old information creates conflicts with new evidence. The model can't distinguish between "I learned X yesterday" and "X is still true."

Planning strategy must match question structure. Using parallel decomposition on sequential questions causes errors. Using sequential on parallel wastes time. Automatic strategy selection is an open problem.

Citation quality varies widely. Some systems cite sources without verifying claims actually appear there. Faithful citation requires explicit extraction and verification steps.

Assumptions & Limitations

Evaluation is immature. No standard benchmarks for full Deep Research systems—most papers evaluate components in isolation. Hard to compare end-to-end system performance.

Compute costs unclear. Long-horizon research tasks with many retrieval/planning steps have unknown scaling properties. Could be prohibitively expensive for complex queries.

Hallucination not solved. Even with citations, models can fabricate sources or misattribute claims. Verification mechanisms help but don't eliminate the problem.

Human-in-loop assumed. Current systems require human verification of final outputs. True autonomy (where you trust the output without checking) remains distant.