Agentic Search
Agentic Search
A retrieval paradigm where the language model has agency over the retrieval process, deciding when to search, what queries to issue, and how to incorporate results—in contrast to static RAG pipelines with fixed retrieve-then-read structure.
Comparison with Static RAG
| Aspect | Static RAG | Agentic Search |
|---|---|---|
| Control flow | Fixed pipeline | Model-determined |
| Retrieval timing | Before generation | During reasoning |
| Query source | User query only | Model-generated subqueries |
| Number of retrievals | Fixed | Adaptive (0 to many) |
| Multi-hop reasoning | Limited | Natural |
| Training | SFT on QA pairs | Often RL with outcome reward |
Key Capabilities
- Adaptive retrieval: Search only when needed
- Query reformulation: Generate targeted subqueries
- Iterative refinement: Multiple search rounds
- Result synthesis: Combine information across retrievals
- Self-evaluation: Assess if more search is needed
System 2 Analogy
Agentic search corresponds to System 2 thinking in cognitive science:
- Deliberate, slow processing
- Explicit reasoning about information needs
- Iterative refinement
Static RAG is more like System 1:
- Fast, automatic
- Fixed pattern execution
- No adaptation
Example Systems
| System | Approach |
|---|---|
| SEARCH-R1 | RL-trained interleaved reasoning + search |
| Self-RAG | Learned retrieval decisions via special tokens |
| ReAct | Reasoning-action interleaving |
| Adaptive-RAG | Query routing by complexity |
When to Use
| Scenario | Recommendation |
|---|---|
| Simple factual QA | Static RAG |
| Multi-hop reasoning | Agentic Search |
| Latency-critical | Static RAG |
| Research queries | Agentic Search |
Connections
- Extension of Retrieval-Augmented Generation
- Enabled by Reinforcement Learning training
- Related to tool-use in LLM agents
- Implemented in SEARCH-R1
Appears In
- IR-L13 - RL for Reasoning and Search
- IR-L09 - RAG (mentioned as future direction)