Class DefaultRetrievalAugmentor

java.lang.Object
dev.langchain4j.rag.DefaultRetrievalAugmentor
All Implemented Interfaces:
RetrievalAugmentor
public class DefaultRetrievalAugmentor extends Object implements RetrievalAugmentor
The default implementation of RetrievalAugmentor intended to be suitable for the majority of use cases.

It's important to note that while efforts will be made to avoid breaking changes, the default behavior of this class may be updated in the future if it's found that the current behavior does not adequately serve the majority of use cases. Such changes would be made to benefit both current and future users.

This implementation is inspired by this article and this paper. It is recommended to review these resources for a better understanding of the concept.

This implementation orchestrates the flow between the following base components: 
- QueryTransformer
- QueryRouter
- ContentRetriever
- ContentAggregator
- ContentInjector
Visual representation of this flow can be found here.
For each base component listed above, we offer several ready-to-use implementations, each based on a recognized approach. We intend to introduce more such implementations over time and welcome your contributions.

The flow is as follows:
1. A Query (derived from an original UserMessage) is transformed using a QueryTransformer into one or multiple Querys.
2. Each Query is routed to the appropriate ContentRetriever using a QueryRouter. Each ContentRetriever retrieves one or multiple Contents using a Query.
3. All Contents retrieved by all ContentRetrievers using all Querys are aggregated (fused/re-ranked/filtered/etc.) into a final list of Contents using a ContentAggregator.
4. Lastly, a final list of Contents is injected into the original UserMessage using a ContentInjector.

By default, each base component (except for ContentRetriever, which needs to be provided by you) is initialized with a sensible default implementation: 
- DefaultQueryTransformer
- DefaultQueryRouter
- DefaultContentAggregator
- DefaultContentInjector
Nonetheless, you are encouraged to use one of the advanced ready-to-use implementations or create a custom one.

When there is only a single Query and a single ContentRetriever, query routing and content retrieval are performed in the same thread. Otherwise, an Executor is used to parallelize the processing. By default, a modified (keepAliveTime is 1 second instead of 60 seconds) Executors.newCachedThreadPool() is used, but you can provide a custom Executor instance.
On the asynchronous augmentAsync path, a blocking stage that is offloaded (see offloadBlocking) runs on the shared default executor (virtual threads on Java 21+, platform threads on 17-20) (non-pinning for I/O; the ExecutorProvider default), independently of the synchronous fan-out Executor above.
See Also: