Hierarchical retrieval: expand a precise hit without losing its lineage

Index a child passage with a stable pointer to its parent document region and immutable source version. Retrieval ranks children, but the expansion policy decides whether to include a paragraph, section, neighboring window, or another bounded unit. Several child hits can point to the same parent. Deduplicate using the actual parent/version identity before fetching or packing it repeatedly. Retain the child match and score as selection evidence rather than pretending the parent received the same similarity score through an independent retrieval.

Expansion can reveal text beyond the matched child, so authorization must cover every served parent span under the current policy. A child-level permission is not a blanket grant for the whole document. If a parent crosses permission boundaries, fetch a permitted bounded region or reject that expansion. Parent identifiers must not be silently reassigned after an edit. Resolve them against the same corpus snapshot as the child index or explicitly detect a version mismatch and refresh rather than combining old matches with unrelated new text.

RAPTOR recursively clusters and summarizes text to build retrieval representations at multiple abstraction levels. That is different from simply returning the original parent section for a child match. A generated summary may combine distant passages or omit qualifications, so preserve its contributing source identities and generation configuration. If it is used to select evidence, follow its lineage back to original source spans for claim validation. A fluent hierarchy node is not automatically a faithful quotation or an authoritative replacement for its inputs.

A hierarchy can be a tree or a more general dependency structure depending on grouping and reuse. Updates to a leaf can invalidate summaries above it, embeddings of those summaries, and cached retrieval results. Mixed-access inputs need a derived-content permission policy that prevents a summary from leaking a restricted contributor. The simplest conservative design restricts use to principals authorized for all contributing material, though a carefully reviewed redacted derivation can define another policy. Treat that policy as an implemented data rule rather than a prompt instruction.

Compare child-only, bounded-parent, and hierarchical-summary variants on the same authorized corpus snapshot. Measure whether the required source facts are present after expansion and after budget trimming. Broad parents may improve interpretation while crowding out a second necessary source. Fusion-in-Decoder encodes retrieved passages separately before decoder fusion; its architecture is not the same as placing all parent text in one decoder-only prompt. Apply the appropriate architecture's input and compute budgets instead of assuming every retrieval-enhanced generator shares one packing rule.

The cache lab isolates repeated identity reuse. Reset, set C=3 and s=4, predict zero hits among the first sixteen requests, then Play or Step. Separately Reset, set C=4 and s=4, predict twelve initial hits, then Play or Step. This modulo-sixteen stream cycles through four addresses; three slots cannot retain the whole cycle. It contains no parent-child map, hierarchy, or document permissions. Use it to motivate measuring the number of distinct fetched identities, then validate the real expansion and source lineage independently.