The short version: splitting only chops the document; four-level enrichment makes each chunk retrievably understandable. The enrichment items are breadcrumb (section path), keywords, summary, and HyDE hypothetical question, each supplying a different semantic signal.
1. Why enrich after splitting
A bare chunk is just a passage; retrieval can only measure similarity between it and the query. But the passage seldom states which chapter it belongs to, what it covers, or which question it answers. A chunk describing parameter setup may never contain the word "configuration"; a penalty clause buried in an appendix says nothing about whom it applies to.
The result is that relevant chunks rank low. The user's question and the chunk's text speak two different languages, distance in vector space grows, and Top-K slots get taken by chunks that merely look like the question. Enrichment supplies the missing signals and writes them explicitly into the chunk, beyond the raw text.
2. What the four levels add
The four levels divide the work, adding signals from four angles: location, terminology, gist, and anticipated questions.
| Enrichment | Signal added |
|---|---|
| Breadcrumb | Section path, "where this sits" |
| Keywords | Core terms, boost term recall |
| Summary | One-line gist, aids coarse rank |
| HyDE question | Hypothetical question, closes query gap |
The breadcrumb comes from the chapter tree and answers where this chunk sits. Keywords strengthen term recall—model numbers, codes, and metric names live or die by exact matching. The summary compresses the chunk into one line, so coarse ranking can decide before reading deeply. The HyDE question stands on the user's side in advance, answering how this chunk might be asked about.
3. How HyDE works
HyDE (Hypothetical Document Embeddings, Gao et al., 2022) generates "if a user asked about this, how might they phrase it" hypothetical questions for each chunk, then embeds those questions for retrieval. Because a query and a question are closer in form than a query and a passage, similarity ranking becomes more accurate. The original paper is at arXiv:2212.10496.
Two scenarios benefit most. In long documents, chunks are dense and can be approached from many angles; hypothetical questions pre-pave several roads to each one. In colloquial queries, the user says "how much pressure can it take" while the document says "rated load limit"—the hypothetical question translates between them. The cost is one extra generation per chunk at ingest, so teams usually apply it to key chunks or pair it with caching to keep the budget in check.
4. Enrichment and parent-child
Enrichment attaches to parent-child chunks, with each level taking what it needs: the child carries keywords and hypothetical questions for precise retrieval; the parent carries breadcrumb and summary for context presentation. Together, retrieval is accurate and context-aware rather than just close in vector space.
Enrichment is not a one-time investment—when rules change, chunks must be re-enriched. Whether the change helped cannot be judged by eye; compare retrieval metrics such as HitRate@K and MRR before and after, and let the numbers decide.
Enrichment comes after splitting and before ingestion. Splitting in Chunking Strategies, parent-child in Parent-Child Chunks; measuring the effect in RAG Evaluation Metrics.
Engineering Checklist: Where Each Enhancement Level Pays Off
- Breadcrumbs (section paths): enable unconditionally — near-zero cost, yet lets path context like a chapter-section reference participate in matching; biggest gains on structured documents.
- Keywords: enable for terminology-dense documents (technical manuals, contracts), compensating for the low sensitivity of embeddings to model numbers and identifiers.
- Summaries: prioritize long chunks (the legal 1500 preset); short chunks gain little from self-summaries.
- HyDE hypothetical questions: pre-generate likely user questions per chunk — biggest gains for conversational query scenarios. The method is from HyDE (Gao et al., 2022, arXiv:2212.10496).
The cost of enhancement is higher ingestion and maintenance: hypothetical questions must be rebuilt as chunks change, so enhancement levels should pair with lifecycle governance (SHA-256 idempotency, incremental diff) to rebuild only changed chunks — see Knowledge Lifecycle Governance. HyDE only pays off when ranking can exploit the rewrite signal; the fusion mechanism is covered in Hybrid Retrieval.
FAQ
Q1: What are the four enrichment levels?
Breadcrumb (section path), keywords, summary, and HyDE hypothetical question. They add semantic signals from four angles—location, terminology, gist, and anticipated questions—turning a bare chunk that relies only on raw-text similarity into something retrieval and ranking can actually understand.
Q2: What is HyDE and why does it work?
HyDE first generates hypothetical questions a user might ask about the chunk, then embeds them for retrieval. Because a query and a question are closer in form than a query and a passage, similarity ranking improves, especially for long documents and colloquial phrasing.
Q3: Which level of chunk does enrichment attach to?
On parent-child chunks: the child carries keywords and hypothetical questions for precise retrieval, while the parent carries breadcrumb and summary for context presentation. The two levels work together, so retrieval is accurate and context-aware rather than merely close in vector space.
Q4: What if you ingest without enrichment?
The bare chunk relies only on raw-text similarity and often ranks below chunks that merely look like the query. Relevant content gets buried outside Top-K, and precise term matching—model numbers, codes, metric names—suffers the most from the missing signals.
Related Articles
Need Production-Ready RAG Data Pipelines? Meet UGLYPEAR AI
A privately deployed RAG data engineering platform: parsing, cleaning, redaction, compression, chunking, governance, and evaluation in one on-premises pipeline — fully local inference, your data never leaves your domain.