Reference

Hypothetical and hierarchical retrieval¶

Behavior¶

Mechanism	Corpus observation	What to infer
HyDE vector composition	QASPER evidence Recall@10 rose from `0.392` to `0.666` with a gold-answer proxy	The vector operation works. This is an oracle upper bound for generator quality
RAPTOR tree retrieval	QASPER evidence recall `0.426`. Complete recall `0.291`	Extractive summaries lose evidence. Summary quality must be evaluated
MemWalker traversal	QASPER evidence recall `0.312`. Mean nodes visited `15.7`	The measured walk budget traded fewer reads for missed leaves
RECOMP selection	Evidence recall `0.330` at `16.6%` of source tokens	Use the compression level when its evidence loss fits the task

Example structural differences

Input structure	Transformation	Observable output
Several hypothetical-answer vectors	Weighted mean, then L2 normalization	One query vector. Generated text serves as query input
Eight leaf sections	Repeated reducing partitions	Tree whose root transitively covers all eight leaves
Scored sentences over budget	Density selection, then source-order restoration	Selected sentences in original document order

The functions construct query representations and bounded navigation structures. Generation, encoding, clustering, summarization, and relevance models are injected. Mari owns shape validation, deterministic IDs, budgets, and traces.

How it works¶

hypothetical_document_embedding weights caller-encoded hypothetical answers and averages their vectors. L2 normalization produces the retrieval vector. Generated text remains a query aid. build_summary_tree validates a caller-proposed partition of every current root. Each accepted round creates stable parent nodes. Tree building ends when a round keeps the same root count. walk_summary_tree expands the highest-scoring children under explicit branch and visit budgets. It returns visited paths with the exhaustion state.

Papers

HyDE: hypothetical document embeddings RAPTOR: recursive summary trees MemWalker: bounded memory-tree navigation

Generatehypothetical answer

→

Encodenormalized query vector

→

Retrievecandidate sections

→

Organizerecursive clusters

→

Walkbranch + visit budget

Paper-derived retrieval construction¶

hyde_raptor_memwalker.py · current¶

from mari_kit.retrieval import (
    build_summary_tree, hypothetical_document_embedding, walk_summary_tree,
)

hyde_vector = hypothetical_document_embedding([
    document_encoder(text) for text in generate_hypotheses(query)
])

tree = build_summary_tree(section_text_by_id,
    cluster=lambda nodes, level: cluster_embeddings(nodes, level),
    summarize=lambda children, level: summarize(children, level))
walk = walk_summary_tree(tree, lambda node: similarity(query, node.text),
    branch_factor=2, max_visits=24)
sections = document_store.get_many(walk.leaf_ids)