Package sk.iway.iwcm.rag.search

Class RagChunkPostProcessor

java.lang.Object

sk.iway.iwcm.rag.search.RagChunkPostProcessor

public class RagChunkPostProcessor
extends Object

Post-processing pipeline for RAG retrieved chunks. Steps: 1. Select distinct top-K chunks by ranking score, then apply soft minimum similarity filter 2. Sort by entityId, then chunkIndex 3. Group by entity type and entity ID 4. Merge adjacent chunks within each group (removing sliding-window overlap) 5. Limit output to maxBlocks / maxCharacters Local ranking controls order; original similarity controls soft filtering, never as hard truth. The pipeline never returns zero chunks when decent top results exist.

Constructor Summary

Constructors

Constructor	Description
RagChunkPostProcessor(int topK, double minSimilarity, int maxChunkGap, int maxBlocks, int maxCharacters, int maxMergedBlockCharacters)	Configures context ranking, merging, and output limits.

Method Summary

Modifier and Type	Method	Description
List<MergedContextBlock>	process(List<VectorSearchResult> chunks)	Run the full post-processing pipeline on retrieved chunks.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Details

RagChunkPostProcessor

public RagChunkPostProcessor(int topK,
 double minSimilarity,
 int maxChunkGap,
 int maxBlocks,
 int maxCharacters,
 int maxMergedBlockCharacters)

Configures context ranking, merging, and output limits.

Parameters:

topK - number of top chunks considered for context before adaptive thresholding is applied

minSimilarity - soft similarity threshold; chunks below this are dropped only if enough top chunks remain

maxChunkGap - maximum gap between chunkIndex values to still merge (1 = adjacent only)

maxBlocks - maximum number of merged context blocks to return

maxCharacters - maximum total characters across all returned blocks

maxMergedBlockCharacters - maximum characters in a single merged block; prevents unbounded merging

Method Details

process

public List<MergedContextBlock> process(List<VectorSearchResult> chunks)

Run the full post-processing pipeline on retrieved chunks.

Parameters:

chunks - raw vector search results

Returns:

merged, filtered, and limited context blocks ready for the LLM prompt