[Paper]

Agentic Information Retrieval (AIR) is an advanced technique of IR that utilizes multiple Reasoning-Action (ReAct) steps combined with tool-calling to reach a user's desired information state by actively interacting with the environment.

Traditional IR (TIR) vs Agentic IR (AIR)

image.png

1. Task Scope:

TIR focuses on primarily filtering and presenting relevant information based on user queries, while AIR expects an information state from the user and actively takes action to reach that state.

2. Architecture:

TIR utilizes a fixed, domain-specific architecture that operates in a single interaction step to get to the output. Conversely, AIR employs a unified architecture in which an AI agent interacts with the environment through a recursive process of observation, reasoning, and action across multiple steps. This allows for dynamic adjustments based on user needs and context.

3. Methodologies:

TIR relies on established methods such as indexing, retrieval algorithms, scoring function, ranking, and pseudo-relevance feedback. WHEREAS AIR incorporates advanced techniques like prompting, RAG, fine-tuning with supervised and reinforcement learning, and multi-agent systems.

How does it work?

  1. Prompt describing the desired information state.
  2. Information State Initialization based on user’s input.
  3. Agent employs a policy to determine the next action based on the current information state.
  4. Agent executes the action and interacts with the environment to update the information state.
  5. Steps 3 and 4 are iterated over until the desired information state is achieved.

Key Methods

  1. Prompt engineering
  2. RAG
  3. Reflection: The agent may use the intermediate results by interacting with the environment to refine its future actions and reach to the desired information state.