Langhuan introduces a parent-child chunking model: child chunks handle precise recall, parent blocks provide context. Chunk boundaries can be chosen adaptively by strategy, so retrieval results are no longer forced to choose between “too large” and “too fragmented.”
Recall small fragments, return full context
Traditional fixed-length chunking runs into two problems easily: when fragments are too large, recall isn’t precise enough; when they’re too small, titles, paragraphs, and section relationships get lost. Parent-child chunking splits these two goals apart:
- Child chunks enter the vector and full-text indexes and are responsible for finding the most relevant location.
Parent blocks aggregate around those child chunks to carry the surrounding context back to the caller. The boundary strategy adapts to document structure, so headings, lists, and tables survive the chunking step instead of being cut mid-thought.
Aggregation that keeps provenance
A retrieval hit returns not only the matching fragment but also the parent block it belongs to, with the document, version, and anchor intact. You get precise recall and readable context in the same result, without trading one for the other.