Iterative Retrieval

The core engine of iterative retrieval. After an initial search, the system analyzes the retrieved text to identify information gaps or new entities. It then automatically rewrites the user's original query—often adding specific keywords, dates, or synonyms found in the initial results—to launch a more targeted subsequent search. This loop continues until a termination condition is met, such as reaching a maximum number of hops or achieving a high confidence score.

Unlike parallel retrieval, iterative retrieval builds a growing context buffer. Each search cycle adds new chunks of text to the working memory without discarding previous results. This allows the model to synthesize information from disparate sources that were never co-located in a single document. The challenge lies in managing the context window to avoid overflow while retaining critical bridge entities.

The system must balance two competing strategies:

• Exploitation: Deepening the search on a high-signal path by generating highly specific follow-up queries.
• Exploration: Broadening the search if the current path hits a dead end, often by relaxing query constraints or using semantic similarity to find adjacent concepts. This dynamic prevents the model from getting stuck in local maxima of irrelevant information.

An infinite loop is a critical failure mode. Robust iterative retrieval requires explicit stopping conditions:

• Information Gain Threshold: Stop when a new retrieval cycle returns no new unique tokens or entities.
• Answer Confidence: Stop when the synthesized answer passes a hallucination detector or a natural language inference (NLI) check against the accumulated context.
• Hop Limit: A hard stop after a predefined number of retrieval rounds to guarantee latency budgets.

Iterative retrieval is the search-specific implementation of the broader ReAct (Reasoning and Acting) framework. In this pattern, the 'Reasoning' step identifies what information is missing, and the 'Acting' step executes a new search query. The observation from that search is then fed back into the reasoning trace, interleaving cognitive logic with real-time data fetching.

Because steps are sequential, latency is additive. Engineering teams must budget for N * (Network Round Trip + Generation Time). Optimization strategies include:

• Streaming: Displaying intermediate results to the user to mask total processing time.
• Speculative Retrieval: Pre-fetching potential next queries in the background while the user is still formulating the initial question.
• Caching: Storing embeddings for common bridge entities to skip redundant vector lookups.