Contextual Chunk Augmentation

Prepend metadata or model-generated situating text to each chunk before indexing so retrieval works when the chunk alone lacks standalone meaning.

Chunks often lack standalone context after chunking. Augment before indexing:

• Metadata — tags, keywords, titles, captions, extracted entities (preserves exact-match terms embeddings may blur).
• Questions the chunk can answer — organizing as Q&A improves match to user phrasing.
• Model-generated situating context — given the whole document and chunk, ask for a short (50–100 token) explanation of how the chunk fits the document, prepend it, then index (Anthropic-style contextual retrieval).

Augmentation raises index cost and tokens per hit but improves retrievability for RAG agents that otherwise fetch orphaned fragments under lost in the middle pressure.

---

Sources

• AI Engineering (Chip Huyen), ch. 6