Contextual Retrieval
Retrieval gets stronger when chunks keep enough local context to remain meaningful after a document is split apart.
That sounds obvious, but many retrieval systems lose this early.
The problem with naive chunking
A chunk can be semantically related to a query and still be hard to retrieve correctly if it has lost the context that made it meaningful.
A sentence like:
Revenue grew by 3% over the previous quarter.
is not very useful on its own unless the system also knows:
- which company it refers to
- which time period it refers to
- what document it came from
- what surrounding topic it belongs to
What contextual retrieval changes
Instead of indexing isolated fragments, the system preserves or adds enough explanatory framing so the chunk can still be found and interpreted correctly.
In practice, retrieval quality often improves when you combine:
- lexical matching
- semantic similarity
- chunk or document context
- reranking or rank fusion
Why this matters for LLM wiki workflows
Even a wiki-first system still benefits from better retrieval.
It matters when:
- reading raw sources during ingest
- locating evidence across many source pages
- scaling beyond direct
index.mdnavigation
The larger lesson
Retrieval quality is not only about embeddings. It is also about whether the indexed units still carry enough meaning to be retrievable in the first place.
That is why context preservation matters.